window.ICML_PAPERS["54069"] = [{"id":"Lv33GhmoBC","en":"test","ko":"test","authors":"Yilun Zhao","abs":"

aaaa

","absKo":"aaaa","link":"https://openreview.net/forum?id=Lv33GhmoBC"},{"id":"gAUM8sqF5h","en":"MPINeuralODE: Multiple-Initial-Condition Physics-Informed Neural ODEs for Globally Consistent Dynamical System Learning","ko":"MPINeuralODE: 전역적으로 일관된 동역학 시스템 학습을 위한 다중 초기조건 물리 정보 기반 Neural ODE","authors":"Lake Yang, Antonio Malpica-Morales, Frank-Ioannis Papadakis-Wood, Serafim Kalliadasis","abs":"

Neural ordinary differential equations (Neural ODEs) often fit training trajectories while generalizing poorly to unseen initial conditions and long horizons. We propose MPINeuralODE, which combines a soft physics-informed residual with a Multiple-Initial-Condition (MIC) multiple-shooting curriculum whose ingredients are structurally complementary: the physics term anchors the vector-field magnitude on the support that MIC enlarges. We evaluate along three axes: out-of-sample error, long-horizon stability, and Hamiltonian drift; together exposing whether the learned dynamics have recovered the true vector field. On Lotka–Volterra, MPINeuralODE achieves the lowest out-of-sample and long-horizon MSE among data-driven methods (a 24% reduction over the baseline Neural ODE) and the lowest Hamiltonian drift, making it the strongest method across all three axes.

","absKo":"Neural ordinary differential equations (Neural ODEs)는 종종 training trajectory에는 잘 맞지만, 보지 못한 initial condition과 긴 horizon에서는 일반화가 떨어집니다. 우리는 soft physics-informed residual과 Multiple-Initial-Condition (MIC) multiple-shooting curriculum을 결합한 MPINeuralODE를 제안합니다. 이 두 요소는 구조적으로 상보적입니다. physics term은 MIC가 확장하는 support 위에서 vector-field magnitude를 고정(anchor)합니다. 우리는 out-of-sample error, long-horizon stability, Hamiltonian drift의 세 축을 따라 평가합니다. 이 세 지표를 함께 보면, 학습된 dynamics가 진짜 vector field를 얼마나 잘 복원했는지 드러납니다. Lotka–Volterra에서 MPINeuralODE는 data-driven method 중 가장 낮은 out-of-sample 및 long-horizon MSE를 달성했고(기본 Neural ODE 대비 24% 감소), Hamiltonian drift도 가장 낮아 세 축 모두에서 가장 강력한 방법임을 보였습니다.","link":"https://openreview.net/forum?id=gAUM8sqF5h"},{"id":"7rSGYMWyiG","en":"Simulation-to-Real Learning for VLBI Jet Morphology Classification","ko":"VLBI 제트 형태 분류를 위한 시뮬레이션-투-리얼 학습","authors":"Dmitrii Zagorulia, Mikhail Lisakov","abs":"

We investigate simulation-to-real learning for morphological classification of parsec-scale active galactic nucleus (AGNs) jets observed with Very Long Baseline Interferometry (VLBI) across frequencies from 1 to 90 GHz. To overcome the scarcity of labelled VLBI data, we combine real observations with physically motivated synthetic images generated from the Blandford–K\\"onigl jet model via relativistic ray tracing and realistic interferometric reconstruction. We compare supervised learning, self-supervised pretraining, and domain adaptation strategies based on a ResNet-18 backbone, and show that explicit domain adaptation substantially improves transfer to real VLBI observations, achieving an $F_1$-score of 0.86 on manually labelled Astrogeo sources. The trained model is subsequently applied to more than 130\\,000 unlabelled AGN images, enabling one of the largest parsec-scale morphology studies to date. The inferred populations reveal statistically significant differences in radio brightness between compact and extended sources, suggesting a connection between parsec-scale jet structure and observed radio emission properties.

","absKo":"우리는 1~90 GHz 주파수 대역에서 Very Long Baseline Interferometry (VLBI)로 관측된 parsec-scale active galactic nucleus (AGN) jet의 형태 분류를 위한 simulation-to-real learning을 조사한다. 라벨이 붙은 VLBI 데이터의 희소성을 극복하기 위해, 우리는 실제 관측값과 Blandford–K\\\"onigl jet model에서 생성한 물리적으로 타당한 synthetic image를 relativistic ray tracing과 현실적인 interferometric reconstruction을 통해 결합한다. 우리는 ResNet-18 backbone을 기반으로 한 supervised learning, self-supervised pretraining, domain adaptation 전략을 비교하고, 명시적 domain adaptation이 실제 VLBI 관측으로의 transfer를 상당히 개선하여 수작업으로 라벨링된 Astrogeo source에서 $F_1$-score 0.86을 달성함을 보인다. 이후 학습된 model을 130\\,000개 이상의 unlabeled AGN image에 적용하여, 지금까지 가장 큰 parsec-scale morphology 연구 중 하나를 가능하게 한다. 추론된 population은 compact source와 extended source 사이의 radio brightness에 통계적으로 유의한 차이가 있음을 보여주며, 이는 parsec-scale jet structure와 관측된 radio emission property 사이의 연관성을 시사한다.","link":"https://openreview.net/forum?id=7rSGYMWyiG"},{"id":"P0GiwDcTZ3","en":"PIEFS: Physics-Informed Eigenfunction Features with Learnable Scaling","ko":"학습 가능한 스케일링을 갖는 물리 기반 고유함수 특징","authors":"Varvara Nazarenko, Timur Lidzhiev, Alexander Tarakanov","abs":"

Spectral methods are widely used to construct representations from the geometry of data, but they often rely on a fixed kernel, graph Laplacian, or manually selected feature scaling. We propose Physics-Informed Eigenfunction Features with Learnable Scaling (PIEFS), a supervised neural representation-learning framework with a spectral inductive bias, based on a modified Dirichlet energy. In PIEFS, scalar coordinate maps are trained under empirical Gram orthogonality, a supervised linear readout, and a Dirichlet penalty in which the input gradient is transformed by a learnable metric $A(x)=\\Lambda(x)U(x)$. The diagonal factor $\\Lambda(x)$ controls anisotropic scaling, while the orthogonal factor $U(x)$ is parameterized by a structured product of Givens rotations. This construction yields task-adaptive Dirichlet-regularized coordinates rather than eigenfunctions of a fixed supervision-independent operator. Experiments on synthetic, tabular, and image-based benchmarks study the effect of identity, diagonal, and rotation-scaling metrics, and compare the resulting coordinates with classical baselines and NeuralEF. The results support PIEFS as a compact supervised spectral representation method and identify optimization stability, validation on explicit operator eigenproblems, and richer metric parameterizations as the main directions for future work.

","absKo":"Spectral method는 데이터의 geometry로부터 representation을 구성하는 데 널리 사용되지만, 고정된 kernel, graph Laplacian, 또는 수동으로 선택한 feature scaling에 의존하는 경우가 많다. 우리는 수정된 Dirichlet energy를 기반으로 하는, spectral inductive bias를 지닌 supervised neural representation-learning framework인 Physics-Informed Eigenfunction Features with Learnable Scaling(PIEFS)을 제안한다. PIEFS에서는 scalar coordinate map이 empirical Gram orthogonality, supervised linear readout, 그리고 입력 gradient가 learnable metric $A(x)=\\Lambda(x)U(x)$로 변환되는 Dirichlet penalty 하에서 학습된다. diagonal factor $\\Lambda(x)$는 anisotropic scaling을 제어하고, orthogonal factor $U(x)$는 Givens rotation의 structured product로 parameterize된다. 이 구성은 고정된 supervision-independent operator의 eigenfunction이 아니라 task-adaptive한 Dirichlet-regularized coordinate를 만들어낸다. synthetic, tabular, image-based benchmark에서 수행한 실험은 identity, diagonal, rotation-scaling metric의 효과를 조사하고, 결과 coordinate를 classical baseline 및 NeuralEF와 비교한다. 결과는 PIEFS가 compact한 supervised spectral representation method임을 뒷받침하며, optimization stability, 명시적 operator eigenproblem에서의 validation, 그리고 더 풍부한 metric parameterization이 향후 연구의 주요 방향임을 보여준다.","link":"https://openreview.net/forum?id=P0GiwDcTZ3"},{"id":"TR4M5zk68E","en":"GEMS: Molecular Structure Identification via Geodesic Navigation of the Isomer Manifold","ko":"GEMS: 이성질체 다양체의 geodesic navigation을 통한 분자 구조 식별","authors":"Utku Umur ACIKALIN, Yingheng Wang, Goncalo J. Gouveia, Tyler Schwertfeger, Frank C Schroeder, Carla P Gomes","abs":"

Identifying small-molecule structures from tandem mass spectra (MS2) is a central bottleneck in untargeted metabolomics.

Existing approaches either retrieve candidates from spectral libraries, limiting discovery to known compounds, or generate ranked final samples without maintaining validated molecular-hypothesis trajectories.

We introduce \\textbf{GEMS} (Geodesic Elucidation of Molecular Structures), a seeded isomer-navigation method that searches a chemically validated isomer manifold with valence-preserving 2-switch operations.

Rather than mapping a spectrum directly to a structure, GEMS maintains a current molecular hypothesis, compares its forward-predicted spectrum to the observation, and uses the resulting correspondence to select a local edit or HALT action.

Each retained rollout state is a valid isomer, yielding inspectable trajectories from high-quality retrieved seeds.

We train the policy with PubChem-grounded shortest-path supervision certified under the stated atom-alignment and validity model.

At inference time, multi-start search records every visited valid molecule and ranks unique candidates by forward spectral cosine and policy halt probability.

With target-excluded starts, GEMS reaches the target in at least one trajectory for 72.19\\% of held-out NIST unique spectra and obtains 63.39\\% top-10 accuracy; without additional training, it reaches the target for 32.02\\% of CASMI spectra.

GEMS achieves substantially higher NIST top-1 accuracy than recent generative baselines while using roughly two to three orders of magnitude fewer iterative model calls per query.

","absKo":"tandem mass spectrum(MS2)에서 small-molecule structure를 식별하는 것은 untargeted metabolomics의 핵심 병목이다.\n기존 접근법은 spectral library에서 candidate를 검색하여 알려진 compound로 discovery를 제한하거나, 검증된 molecular-hypothesis trajectory를 유지하지 않은 채 ranked final sample을 생성한다.\n우리는 화학적으로 검증된 isomer manifold를 valence-preserving 2-switch operation으로 탐색하는 seeded isomer-navigation method인 \\textbf{GEMS} (Geodesic Elucidation of Molecular Structures)를 소개한다.\nGEMS는 spectrum을 구조로 직접 매핑하는 대신, 현재의 molecular hypothesis를 유지하고, 그 forward-predicted spectrum을 관측값과 비교하며, 그에 따른 correspondence를 사용해 local edit 또는 HALT action을 선택한다.\n보존된 각 rollout state는 유효한 isomer이며, 고품질 retrieved seed로부터 inspect 가능한 trajectory를 제공한다.\n우리는 명시된 atom-alignment 및 validity model 하에서 인증된 PubChem-grounded shortest-path supervision으로 policy를 학습한다.\ninference 시에는 multi-start search가 방문한 모든 유효 molecule을 기록하고, forward spectral cosine과 policy halt probability로 unique candidate를 순위화한다.\ntarget-excluded start를 사용하면, GEMS는 hold-out NIST unique spectrum의 72.19\\%에 대해 최소 하나의 trajectory에서 target에 도달하고 63.39\\%의 top-10 accuracy를 달성한다. 추가 학습 없이도 CASMI spectrum의 32.02\\%에서 target에 도달한다.\nGEMS는 질의당 iterative model call을 대략 2~3자릿수 정도 더 적게 사용하면서, 최근 generative baseline보다 NIST top-1 accuracy를 훨씬 높게 달성한다.","link":"https://openreview.net/forum?id=TR4M5zk68E"},{"id":"DqmtUsPgnt","en":"Interpretable Equivariant Marks for Contrastive Cosmological Inference","ko":"대조적 우주론 추론을 위한 해석 가능한 등변 marks","authors":"Federico Semenzato, Benjamin Dan Wandelt, Liguori M., Alvise Raccanelli","abs":"

Recovering cosmological information beyond the power spectrum is a central goal for upcoming cosmological surveys, since late-time non-Gaussian signal in the matter density cannot be accessed through two-point statistics alone.

Marked statistics fold part of this information back into the two-point level by reweighting the field with non-linear functions.

We propose a neural marking scheme to generalize this process through a set of interpretable, physically motivated transformations that directly allow to interpret the gain in cosmological information at the morphological level.

We employ a contrastive learning objective to align learnable marked summaries with the underlying cosmological parameters.

At $k_{\\max}=0.2\\,h\\mathrm{Mpc}^{-1}$, our neural mark tightens the marginalized constraint on $\\sigma_8$ by $2.9\\times$ and on $\\Omega_m$ by $1.8\\times$ compared to classical marks, breaking the $\\Omega_m-\\sigma_8$ degeneracy at the Fisher information level. It further reduces the parameter MSE across our cosmological parameter prior by $1.45\\times$ over the best classical mark.

The learned latent geometry aligns with the $\\Omega_m$ and $\\sigma_8$ directions in parameter space, indicating that the contrastive objective recovers the dominant axes of cosmological information.

Our approach opens the door to more powerful, interpretable summary statistics for cosmological inference.

","absKo":"power spectrum을 넘어서는 cosmological information을 복원하는 것은 향후 cosmological survey의 핵심 목표이다. 물질 밀도의 late-time non-Gaussian signal은 two-point statistic만으로는 접근할 수 없기 때문이다.\nMarked statistic은 field를 비선형 함수로 reweighting하여 이 정보의 일부를 다시 two-point 수준으로 접어 넣는다.\n우리는 해석 가능하고 물리적으로 동기 부여된 변환 집합을 통해 이 과정을 일반화하는 neural marking scheme을 제안하며, 이를 통해 morphological level에서 cosmological information의 증가를 직접 해석할 수 있게 한다.\n우리는 learnable marked summary를 underlying cosmological parameter와 정렬하기 위해 contrastive learning objective를 사용한다.\n$k_{\\max}=0.2\\,h\\mathrm{Mpc}^{-1}$에서, 우리 neural mark는 classical mark와 비교했을 때 $\\sigma_8$에 대한 marginalized constraint를 $2.9\\times$, $\\Omega_m$에 대한 constraint를 $1.8\\times$ 더 조여 주며, Fisher information 수준에서 $\\Omega_m-\\sigma_8$ degeneracy를 깨뜨린다. 또한 우리의 cosmological parameter prior 전반에 걸쳐 parameter MSE를 최상의 classical mark 대비 $1.45\\times$ 더 줄인다.\n학습된 latent geometry는 parameter space에서 $\\Omega_m$과 $\\sigma_8$ 방향과 정렬되어 있으며, 이는 contrastive objective가 cosmological information의 지배적인 축을 복원함을 시사한다.\n우리 접근법은 cosmological inference를 위한 더 강력하고 해석 가능한 summary statistic의 길을 연다.","link":"https://openreview.net/forum?id=DqmtUsPgnt"},{"id":"r2NSBJ4ip7","en":"Geometry, Not Energy Surface, Drives the Neutral MLIP–DFT Gap in Atomistic Interaction Surrogates","ko":"원자가 상호작용 surrogate에서 중성 MLIP-DFT 간극을 좌우하는 것은 에너지 표면이 아니라 기하학이다","authors":"Rui Ding, Zixin Ding, Rodrigo P. Ferreira, Yuxin Chen, Junhong Chen","abs":"

Pretrained universal machine-learning interatomic potentials (MLIPs) are now positioned as high-fidelity surrogate simulators for electronic-structure physics, widely expected to replace explicit quantum chemistry where they cover the chemistry. We show that on non-covalent binding, this expectation is mostly right, but mostly for the wrong reason. We audit the FAIRChem UMA omol head on 5,565 probe–target dimers from 18 benchmarks across nine fidelity arms, without any task-specific fine-tuning. Three findings emerge. First, the neutral MLIP–DFT gap is geometric rather than a learned-surface error: an explicit $\\omega$B97M-V single point at the MLIP geometry leaves mean absolute error (MAE) at 1.64 kcal/mol, while reoptimizing the geometry drops it to 1.44, and MACE-omol and ORB-omol reach the same DFT-SP parity on neutral chemistry. Second, charged systems behave as a qualitatively different regime: MAE jumps to 13.83–27.21 kcal/mol with Spearman $\\rho\\approx 0$, a DFT single point does not recover it, only DFT geometry reoptimization does; and the failure persists across all three backbones. Third, the residual errors are structured: signed biases cluster by chemistry family, so a 10% per-benchmark calibration offset improves neutral MAE to 1.32 kcal/mol at zero additional cost, and a chemistry-aware replay selector (ChEMUCB-8) reaches 1.06 kcal/mol at 52% of the CREST xTB+DFT baseline cost. The audit is delivered through RAPIDS, a training-free inference-time scaffold exposed via a Model Context Protocol (MCP) interface and illustrated on a per- and polyfluoroalkyl substance (PFAS) sensing case study.

","absKo":"Pretrained universal machine-learning interatomic potentials (MLIPs)는 이제 electronic-structure physics를 위한 high-fidelity surrogate simulator로 자리매김했으며, 화학적 범위가 맞는 곳에서는 explicit quantum chemistry를 대체할 것으로 널리 기대된다. 우리는 non-covalent binding에서 이 기대가 대체로 옳지만, 대부분 잘못된 이유 때문임을 보인다. 우리는 task-specific fine-tuning 없이, nine fidelity arms에 걸친 18개 benchmark에서 5,565개의 probe-target dimer에 대해 FAIRChem UMA omol head를 감사했다. 세 가지 발견이 나온다. 첫째, neutral MLIP-DFT 차이는 learned-surface error가 아니라 geometry 차이다. MLIP geometry에서 explicit $\\omega$B97M-V single point를 계산하면 mean absolute error (MAE)는 1.64 kcal/mol로 남지만, geometry를 재최적화하면 1.44로 떨어지며, MACE-omol과 ORB-omol도 neutral chemistry에서 동일한 DFT-SP parity에 도달한다. 둘째, charged system은 질적으로 다른 regime이다. MAE가 13.83–27.21 kcal/mol로 급증하고 Spearman $\\rho\\approx 0$이 되며, DFT single point는 이를 복구하지 못하고 오직 DFT geometry reoptimization만이 가능하다. 또한 이 실패는 세 backbone 모두에서 지속된다. 셋째, residual error는 구조화되어 있다. signed bias가 chemistry family별로 군집하므로, benchmark당 10%의 calibration offset만으로도 추가 비용 없이 neutral MAE를 1.32 kcal/mol로 개선할 수 있고, chemistry-aware replay selector (ChEMUCB-8)는 CREST xTB+DFT baseline 비용의 52%에서 1.06 kcal/mol에 도달한다. 이 감사는 Model Context Protocol (MCP) interface를 통해 노출된 training-free inference-time scaffold인 RAPIDS로 제공되며, per- 및 polyfluoroalkyl substance (PFAS) sensing case study로 설명된다.","link":"https://openreview.net/forum?id=r2NSBJ4ip7"},{"id":"0TI01P3XfU","en":"Quantum Mean-Field Games: Numerical Nash Equilibria via PDE-form HJB Learning","ko":"Quantum Mean-Field Games: PDE 형식 HJB 학습을 통한 수치적 Nash 균형","authors":"Nevroz Sen","abs":"

Quantum mean-field games (QMFGs) extend classical mean-field game theory to populations of indirectly observed quantum systems governed by Belavkin stochastic master equations. Prior QMFG literature provides theoretical foundations but, to our knowledge, no computed Hamilton--Jacobi--Bellman (HJB) optimal Nash equilibria. We present two contributions. First, we identify a structural failure mode of trajectory-based discrete Bellman / deep-BSDE training on Belavkin SDEs: when the measurement diffusion direction is aligned with the value-function gradient, the Monte Carlo Bellman residual is dominated by innovation noise, and an asymptotic signal-to-noise analysis shows the Bellman-loss minimizer is limited by an alignment-dependent noise floor that does not vanish as the timestep is refined. We confirm the geometric prediction on two structurally distinct quantum benchmarks: under aligned noise the trained policy matches zero control, while in a near-orthogonal benchmark Bellman training recovers PDE-form (PINN) performance. Second, using PDE-form HJB learning inside a fictitious-play loop, we compute, to our knowledge, the first numerical Nash equilibrium for a multi-type Belavkin-filtered QMFG, exhibiting equilibria for priors $p \\in \\{0.25, 0.5, 0.75\\}$ and proving the symmetric Nash equilibrium at $p=1/2$ by reflection symmetry.

","absKo":"Quantum mean-field games(QMFGs)는 Belavkin stochastic master equation으로 지배되는 간접 관측 quantum system의 집단에 대해 classical mean-field game theory를 확장한다. 기존 QMFG 문헌은 이론적 기반을 제공하지만, 우리가 아는 한 계산된 Hamilton--Jacobi--Bellman(HJB) optimal Nash equilibrium은 없다. 우리는 두 가지 기여를 제시한다. 첫째, Belavkin SDE에 대한 trajectory-based discrete Bellman / deep-BSDE training의 구조적 failure mode를 식별한다. 측정 diffusion 방향이 value-function gradient와 정렬될 때, Monte Carlo Bellman residual은 innovation noise에 의해 지배되며, asymptotic signal-to-noise analysis는 Bellman-loss minimizer가 alignment-dependent noise floor에 의해 제한되고 timestep을 정교화해도 사라지지 않음을 보인다. 우리는 구조적으로 서로 다른 두 quantum benchmark에서 이 기하학적 예측을 확인한다. 정렬된 noise에서는 학습된 policy가 zero control과 일치하고, 거의 직교하는 benchmark에서는 Bellman training이 PDE-form(PINN) 성능을 회복한다. 둘째, fictitious-play loop 내부에서 PDE-form HJB learning을 사용하여, 우리는 multi-type Belavkin-filtered QMFG에 대한 첫 수치적 Nash equilibrium을 계산했다고 믿으며, priors $p \\in \\{0.25, 0.5, 0.75\\}$에서의 equilibrium을 보이고 reflection symmetry를 통해 $p=1/2$에서의 대칭 Nash equilibrium을 증명한다.","link":"https://openreview.net/forum?id=0TI01P3XfU"},{"id":"8bhQTEwcNL","en":"FactoryNet: A Large-Scale Dataset toward Industrial Time-Series Foundation Models","ko":"FactoryNet: 산업용 시계열 foundation model을 향한 대규모 데이터셋","authors":"Karim Othman, Jonas Petersen, Matei C. Ignuta-Ciuncanu, Gian-Alessandro Lombardi, Camilla Mazzoleni, Federico Martelli, Riccardo Maggioni, Philipp Christian Petersen","abs":"

We introduce the first universal pretraining corpus for industrial time-series data: FactoryNet. FactoryNet contains 51 million datapoints across 23,000 end-to-end task executions (13,300 real, 9,800 synthetic) on six embodiments. These are unified by a shared schema that enables robust zero-shot cross-embodiment transfer and highly parameter-efficient anomaly detection. We introduce a novel schema: Setpoint, Effort, Feedback, Context (S-E-F-C), which underlies the entire pipeline. This schema maps any actuated system into a common representational frame. The corpus spans 27 annotated anomaly types, alongside healthy baselines and counterfactual pairs across robotic manipulation and machining domains. Cross-embodiment transfer experiments yield positive results: under bias-aware metrics, our model demonstrates fair cross-embodiment transfer capabilities on the evaluated source-target pair. Additionally, 24 schema-aligned signals achieve competitive anomaly detection performance compared to high-dimensional baselines. We release FactoryNet as a growing, multi-embodiment dataset to drive progress toward industrial foundation models.

","absKo":"우리는 산업 시계열 데이터에 대한 첫 번째 universal pretraining corpus인 FactoryNet을 소개한다. FactoryNet은 6개의 embodiment에서 수행된 23,000개의 end-to-end task execution(실데이터 13,300개, synthetic 9,800개)에 걸친 5,100만 개의 datapoint를 포함한다. 이 데이터들은 강건한 zero-shot cross-embodiment transfer와 매우 parameter-efficient한 anomaly detection을 가능하게 하는 shared schema로 통합된다. 우리는 전체 pipeline의 기반이 되는 새로운 schema인 Setpoint, Effort, Feedback, Context (S-E-F-C)를 제안한다. 이 schema는 어떤 actuated system이든 공통의 representation frame으로 매핑한다. 코퍼스는 robotics manipulation과 machining 도메인 전반에 걸친 healthy baseline 및 counterfactual pair와 함께 27개의 주석된 anomaly type을 포괄한다. Cross-embodiment transfer 실험은 긍정적인 결과를 낳는다. bias-aware metric 하에서, 우리 모델은 평가된 source-target pair에서 공정한 cross-embodiment transfer 능력을 보인다. 또한 24개의 schema-aligned signal은 high-dimensional baseline과 비교해 경쟁력 있는 anomaly detection 성능을 달성한다. 우리는 industrial foundation model을 향한 진전을 이끌기 위해 성장하는 multi-embodiment dataset인 FactoryNet을 공개한다.","link":"https://openreview.net/forum?id=8bhQTEwcNL"},{"id":"DHa6Cj252v","en":"SOLENO: One Scalar Potential, Zero Divergence — Overcoming the Accuracy–Solenoidality Trade‑off","ko":"SOLENO: 하나의 스칼라 퍼텐셜, 제로 발산 - 정확도와 솔레노이드성의 상충 극복","authors":"Rahul D Ray, Charles Gretton, Matthew Hole","abs":"

Magnetic equilibrium reconstruction in tokamaks requires integrating heterogeneous

diagnostics under the fundamental constraint $\\nabla\\cdot\\mathbf{B}=0$.

Conventional learned surrogates enforce this condition through soft penalties,

inevitably trading reconstruction fidelity for physical consistency. We challenge

this paradigm by showing that a hard architectural constraint---predicting only the

poloidal flux $\\psi$ and deriving the magnetic field analytically via a physics

layer (solenoidal decoder)---is sufficient to resolve the trade-off entirely. Our

operator-learning model, \\textsc{Soleno}, guarantees $\\nabla\\cdot\\mathbf{B}=0$

to machine precision by construction, yet matches the accuracy of the best

unconstrained models on real-world MAST data, while running orders of magnitude

faster than the EFIT equilibrium solver it is trained to surrogate. A soft-penalty

variant of the identical backbone cannot replicate this result, confirming that the

physics layer, and the associated elimination of the divergence penalty

hyperparameter, is essential. Controlled experiments in both cylindrical and polar

geometries further demonstrate that the advantage is structural, not circumstantial.

The finding that a textbook analytic relation, when enforced as a hard architectural

invariant, can dominate over both model complexity and penalty-based physics informs

a strategy of \\emph{constraint-preserving operator learning} for scientific machine

learning.

","absKo":"Tokamak에서의 magnetic equilibrium reconstruction은 기본 제약 $\\nabla\\cdot\\mathbf{B}=0$ 아래에서 heterogeneous diagnostic을 통합해야 한다.\n기존의 learned surrogate는 이 조건을 soft penalty로 강제하며, 필연적으로 reconstruction fidelity와 physical consistency 사이에서 trade-off를 만든다.\n우리는 오직 poloidal flux $\\psi$만 예측하고 physics layer(solenoidal decoder)를 통해 magnetic field를 해석적으로 유도하는 hard architectural constraint가 이 trade-off를 완전히 해소하기에 충분함을 보임으로써 이 패러다임에 도전한다.\n우리의 operator-learning model, \\textsc{Soleno},는 구성상 $\\nabla\\cdot\\mathbf{B}=0$를 machine precision으로 보장하면서도, 실제 MAST data에서 최고 수준의 unconstrained model과 동일한 정확도를 달성하고, 자신이 surrogate하도록 학습된 EFIT equilibrium solver보다 수십 배에서 수백 배 빠르게 동작한다.\n동일한 backbone의 soft-penalty variant는 이 결과를 재현하지 못하며, 이는 physics layer와 divergence penalty hyperparameter의 제거가 본질적임을 확인해 준다.\n원통형 및 극좌표 기하에서 수행한 통제 실험도 이 장점이 우연적이 아니라 구조적임을 추가로 보여 준다.\n교과서적인 analytic relation을 hard architectural invariant로 강제할 때 model complexity와 penalty-based physics를 모두 능가할 수 있다는 사실은 scientific machine learning을 위한 \\emph{constraint-preserving operator learning} 전략을 시사한다.","link":"https://openreview.net/forum?id=DHa6Cj252v"},{"id":"iywEUhqGmH","en":"SciKernelBench: Evaluating LLM Kernels for Higher-Order Autodiff and Scientific Residuals","ko":"SciKernelBench: 고차 미분과 과학적 잔차를 위한 LLM Kernels 평가","authors":"Maria Ivanova, Pavel Zadorozhny, Rodion Levichev, Konstantin Sozykin, Redko Dmitry, Albert Fazlyev, Dmitry Abulkhanov, Evgeny Burnaev, Dmitrii Babaev, Egor Shvetsov","abs":"

Existing kernel generation benchmarks mainly target standard tensor operations such as matrix multiplications, convolutions, and elementwise operators. In contrast, scientific residuals require gradients, Jacobians, Laplacians, divergence, normal derivatives, and convective terms, often involving higher order autodiff paths that are not generally supported by current PyTorch compiler pipelines.

We introduce \\textbf{SciKernelBench}, a benchmark of 144 TorchPhysics derived workloads represented in the KernelBench format. The benchmark covers both isolated differential operator primitives and composed residual computations from physics informed neural networks, Deep Ritz formulations, inverse problems, time-dependent PDEs, contact mechanics, interface problems, and flow systems.

SciKernelBench evaluates generated implementations by numerical correctness, runtime performance, and implementation honesty. The last axis is essential because candidates may pass ordinary output checks by reusing PyTorch autograd, preserving the reference path, omitting derivative terms, or hard coding deterministic inputs, without producing genuine low level acceleration. Across recent frontier LLMs and test time strategies, we find that scientific residuals remain challenging. On Level 1, Claude solves all tasks under repeated sampling and refinement without execution feedback, but drops to \\(23/84\\) raw and only \\(3/84\\) honest solves under refinement with execution feedback. Gemini shows a large raw honest gap even without feedback, with \\(33/84\\) raw but only \\(11/84\\) honest solves. On Level 2, Claude reaches \\(60/60\\) raw solves under execution feedback, while Gemini and Qwen remain substantially lower. Standard correctness based success can substantially overestimate genuine kernel generation ability in several settings, especially under refinement with execution feedback. These results show that scientific residuals require evaluation protocols that test not only whether generated code runs, but whether it faithfully implements the underlying scientific computation.

","absKo":"기존의 kernel generation benchmark는 주로 matrix multiplication, convolution, elementwise operator와 같은 표준 tensor operation을 대상으로 한다. 반면 scientific residual은 gradient, Jacobian, Laplacian, divergence, normal derivative, convective term을 요구하며, 종종 현재 PyTorch compiler pipeline이 일반적으로 지원하지 않는 higher-order autodiff path를 포함한다. \n\n우리는 KernelBench 형식으로 표현된 144개의 TorchPhysics 파생 workload로 구성된 benchmark \\textbf{SciKernelBench}를 소개한다. 이 benchmark는 physics-informed neural network, Deep Ritz formulation, inverse problem, time-dependent PDE, contact mechanics, interface problem, flow system에서 나온 복합 residual 계산과 개별 differential operator primitive를 모두 포함한다.\n\nSciKernelBench는 생성된 구현을 numerical correctness, runtime performance, implementation honesty로 평가한다. 마지막 축이 필수적인 이유는, 후보가 PyTorch autograd를 재사용하거나, reference path를 유지하거나, derivative term을 생략하거나, deterministic input을 하드코딩하더라도, 일반적인 출력 검사에는 통과할 수 있지만 진정한 low-level acceleration을 제공하지는 않을 수 있기 때문이다. 최근 frontier LLM과 test-time strategy 전반에서, scientific residual은 여전히 어렵다는 점이 드러났다. Level 1에서 Claude는 실행 피드백 없이 반복 샘플링과 refinement를 통해 모든 task를 해결하지만, execution feedback을 포함한 refinement에서는 raw 기준 \\(23/84\\), honest 기준 \\(3/84\\)로 떨어진다. Gemini는 피드백이 없어도 raw와 honest 사이에 큰 격차를 보이며, \\(33/84\\) raw이지만 honest solve는 \\(11/84\\)에 그친다. Level 2에서 Claude는 execution feedback 하에 \\(60/60\\) raw solve를 달성하는 반면, Gemini와 Qwen은 훨씬 낮은 수준에 머문다. 표준 correctness 기반 성공 지표는 여러 설정에서, 특히 execution feedback이 있는 refinement 하에서, 실제 kernel generation 능력을 상당히 과대평가할 수 있다. 이러한 결과는 scientific residual이 생성된 코드가 실행되는지뿐만 아니라, 그 코드가 underlying scientific computation을 충실하게 구현하는지도 검증하는 평가 프로토콜을 필요로 한다는 점을 보여준다.","link":"https://openreview.net/forum?id=iywEUhqGmH"},{"id":"kyFh90mTN6","en":"Neural Schur Complement Operators for Coarse-to-Fine Neural Operator Transfer","ko":"거친 수준에서 세밀한 수준으로의 neural operator 전달을 위한 Neural Schur Complement Operators","authors":"Ansh Tiwari","abs":"

Neural operators are advertised as discretization-flexible, but deploying

them from a coarse grid to a finer one routinely loses accuracy that

same-grid validation cannot predict. We show that this failure is

structural rather than a training artifact: any diagnostic that depends only

on the coarse-grid restriction of a fine prediction is blind to the detail

component that lives entirely below the coarse resolution. For linear,

symmetric positive-definite finite-volume discretizations, the

energy-minimizing fine-grid state consistent with a fixed coarse prediction

is determined by an algebraic Schur-complement detail solve. We learn this

detail map with a Neural Schur Complement Operator, which preserves the

coarse prediction exactly by construction and targets the energy-optimal

fine correction, and which comes with an additive energy-norm error

decomposition and an optional projected conjugate-gradient polishing step.

On high-contrast finite-volume Darcy, the method improves over a zero-shot

lift by $1.67$ to $6.21\\times$, recovering $84.2$ to $103.1\\%$ of the

oracle Schur gain, and matches direct fine-grid training at roughly

one-tenth the fine-label budget. We also report the limits honestly: with

real coarse predictors the achievable gain is bounded by source quality,

and a model trained on one permeability morphology does not transfer to

channelized out-of-distribution media without retraining.

","absKo":"Neural operator는 discretization에 유연하다고 알려져 있지만, 거친 grid에서 더 세밀한 grid로 배포하면 same-grid validation으로는 예측할 수 없는 정확도 손실이 흔히 발생한다. 우리는 이러한 실패가 학습의 부산물이 아니라 구조적 현상임을 보인다. 즉, fine prediction의 coarse-grid restriction에만 의존하는 어떤 진단도 coarse resolution 아래에 완전히 존재하는 detail component에는 무력하다. 선형의 symmetric positive-definite finite-volume discretization에 대해, 고정된 coarse prediction과 일치하는 energy-minimizing fine-grid state는 algebraic Schur-complement detail solve에 의해 결정된다. 우리는 이 detail map을 Neural Schur Complement Operator로 학습하는데, 이는 설계상 coarse prediction을 정확히 보존하면서 energy-optimal fine correction을 목표로 하며, additive energy-norm error decomposition과 선택적 projected conjugate-gradient polishing step을 함께 제공한다. 고대비 finite-volume Darcy 문제에서 이 방법은 zero-shot lift 대비 $1.67$에서 $6.21\\times$ 향상을 보였고, oracle Schur gain의 $84.2$에서 $103.1\\%$를 회복했으며, 대략 fine-label 예산의 10분의 1로 직접 fine-grid training과 비슷한 성능을 달성했다. 또한 한계도 솔직히 보고한다. 실제 coarse predictor를 사용할 때 달성 가능한 이득은 source quality에 의해 제한되며, 하나의 permeability morphology로 학습한 모델은 재학습 없이 channelized out-of-distribution media로 전이되지 않는다.","link":"https://openreview.net/forum?id=kyFh90mTN6"},{"id":"ZMMiVPNSrN","en":"Training Neural PDE Solvers with Partial Observations","ko":"부분 관측으로 신경 PDE solver 학습하기","authors":"Harris Abdul Majid, Giannis Daras, Francesco Tudisco, Steven McDonagh","abs":"

In many scientific settings, acquiring complete observations of PDE coefficients and solutions can be expensive, hazardous, or impossible. Recent diffusion-based methods can reconstruct fields given partial observations, but require complete observations for training. We introduce Ambient Physics, a framework for learning the joint distribution of coefficient-solution pairs directly from partial observations, without requiring a single complete observation. The key idea is to randomly mask a subset of already-observed measurements and supervise on them, so the model cannot distinguish "truly unobserved" from "artificially unobserved", and must produce plausible predictions everywhere. Ambient Physics achieves state-of-the-art reconstruction performance. Compared with prior diffusion-based methods, it achieves a 62.51\\% reduction in average overall error while using 125$\\times$ fewer function evaluations. We also identify a "one-point transition'': masking a single already-observed point enables learning from partial observations across architectures and measurement patterns. Ambient Physics thus enables scientific progress in settings where complete observations are unavailable.

","absKo":"많은 scientific setting에서 PDE coefficient와 solution의 complete observation을 얻는 것은 비용이 많이 들거나, 위험하거나, 아예 불가능할 수 있다. 최근 diffusion-based method는 partial observation이 주어졌을 때 field를 재구성할 수 있지만, 학습에는 complete observation이 필요하다. 우리는 Ambient Physics를 제안하는데, 이는 단 하나의 complete observation도 요구하지 않고 partial observation만으로 coefficient-solution pair의 joint distribution을 학습하는 framework이다. 핵심 아이디어는 이미 관측된 측정값의 일부를 무작위로 마스킹하고 그것들에 대해 supervision을 제공함으로써, 모델이 \"진짜로 관측되지 않은 것\"과 \"인위적으로 관측되지 않은 것\"을 구분할 수 없게 만들고, 모든 곳에서 타당한 예측을 생성하도록 하는 것이다. Ambient Physics는 state-of-the-art reconstruction performance를 달성한다. 기존 diffusion-based method와 비교하면, function evaluation을 125$\\times$ 적게 사용하면서 평균 overall error를 62.51\\% 줄인다. 또한 우리는 \"one-point transition\"을 확인했다. 이미 관측된 점 하나를 마스킹하는 것만으로도 architecture와 measurement pattern 전반에서 partial observation으로부터 학습이 가능해진다. 따라서 Ambient Physics는 complete observation을 얻을 수 없는 setting에서도 scientific progress를 가능하게 한다.","link":"https://openreview.net/forum?id=ZMMiVPNSrN"},{"id":"3Ju0RzyUqU","en":"FlowPDE: Understanding and Justifying Flow Matching in PDE Solving","ko":"FlowPDE: PDE 풀이에서 Flow Matching을 이해하고 정당화하기","authors":"Tianxiao Li, Tianyu Huang, Hang Zhou, Yuezhou Ma, Minghao Guo, Haixu Wu, Chuang Gan, Mingsheng Long, Wojciech Matusik","abs":"

Neural PDE solvers have recently begun to incorporate generative modeling. However, applying generative models to PDE solving raises a fundamental question: the uniqueness of the solution of a well-posed PDE seems to conflict with the distributional nature of generative learning, and raises concerns about whether such models may introduce non-physical or hallucinated solutions. In this work, we provide a theoretical justification for using Flow Matching in neural PDE solving. We show that, under full problem information, the target conditional distribution collapses to a one-point distribution at the exact solution, and that classical iterative solvers can be viewed as deterministic energy-descending flow trajectories. Based on this insight, we propose **FlowPDE**, a general framework that equips existing transformer-based neural operators with a conditional flow matching model through AdaLN-style injection of PDE conditions. We provide a unified formulation for both fully observed and partially observed PDE problems. Experiments on diverse PDE benchmarks with complex geometries demonstrate that FlowPDE consistently improves over strong deterministic baselines across both full and partial observation regimes.

","absKo":"Neural PDE solver는 최근 generative modeling을 도입하기 시작했다. 그러나 generative model을 PDE solving에 적용하는 것은 근본적인 질문을 제기한다. well-posed PDE의 해의 유일성은 generative learning의 distributional nature와 충돌하는 것처럼 보이며, 이러한 모델이 비물리적이거나 환각된 해를 도입할 수 있다는 우려를 낳는다. 본 연구에서는 neural PDE solving에서 Flow Matching을 사용하는 것에 대한 이론적 정당화를 제시한다. 전체 문제 정보가 주어지면 target conditional distribution이 정확한 해에서 하나의 점으로 수축하며, 고전적 iterative solver는 결정론적 energy-descending flow trajectory로 볼 수 있음을 보인다. 이러한 통찰을 바탕으로, 우리는 기존 transformer-based neural operator에 AdaLN-style injection을 통해 PDE 조건을 주입하는 conditional flow matching model을 갖춘 일반 프레임워크 **FlowPDE**를 제안한다. 우리는 fully observed와 partially observed PDE 문제 모두를 포괄하는 통합 정식화를 제공한다. 복잡한 기하를 가진 다양한 PDE benchmark 실험에서 FlowPDE는 full observation과 partial observation regime 모두에서 강한 deterministic baseline보다 일관되게 더 나은 성능을 보였다.","link":"https://openreview.net/forum?id=3Ju0RzyUqU"},{"id":"GS0q2w3z8b","en":"Predicting Precision: Equivariant Delta-learning, Recommender Systems and Uncertainty Quantification for High-Throughput DFT","ko":"정확도 예측: 고처리량 DFT를 위한 등변성 delta-learning, 추천 시스템, 그리고 불확실성 정량화","authors":"Daniel T. Speckhard, Christian Carbogno, Matthias Scheffler, Claudia Draxl","abs":"

Density Functional Theory (DFT) is a cornerstone of materials discovery, but running highly precise calculations requires expensive convergence testing for parameters like basis-set size and k-point density. To bypass this, we use $\\Delta$-learning to predict the difference between fast, low-fidelity calculations and their converged, high-fidelity counterparts. Current machine learning models employed for similar tasks, however, are deterministic and often fail silently on out-of-distribution crystal structures. We introduce a framework that integrates Uncertainty Quantification directly into an $E(3)$-equivariant graph neural network to predict these $\\Delta$-corrections. By conditioning the network on the low-fidelity DFT settings using Feature-wise Linear Modulation (FiLM), our model adapts to the specific physical approximations made. We evaluate our approach on a newly generated dataset of over 60,000 multi-fidelity relaxations for 4,220 binary semiconductors. The framework not only acts as a highly accurate error-corrector, outperforming heavily tuned Random Forests, but also functions as a zero-cost recommender system that predicts calculation errors and recommends what settings to use before the calculations are run. Finally, we show that Deep Ensembles and latent-space distance metrics provide explicit out-of-distribution detection, allowing the model to flag its own catastrophic errors and ensure out-of-distribution reliability for high-throughput screening.

","absKo":"Density Functional Theory (DFT)는 materials discovery의 핵심이지만, basis-set size와 k-point density 같은 parameter에 대해 매우 정밀한 계산을 수행하려면 expensive convergence testing이 필요하다. 이를 우회하기 위해, 우리는 빠르고 낮은 fidelity의 계산과 그것의 converged, high-fidelity 대응물 사이의 차이를 예측하는 데 $\\Delta$-learning을 사용한다. 그러나 유사한 task에 사용되는 현재의 machine learning model은 결정론적이며, out-of-distribution crystal structure에서 종종 조용히 실패한다. 우리는 이 $\\Delta$-correction을 예측하기 위해 Uncertainty Quantification을 $E(3)$-equivariant graph neural network에 직접 통합한 프레임워크를 소개한다. Feature-wise Linear Modulation (FiLM)을 사용해 low-fidelity DFT setting을 조건으로 네트워크를 조절함으로써, 우리의 모델은 구체적인 물리적 근사에 적응한다. 우리는 4,220개의 binary semiconductor에 대해 새로 생성된 60,000개가 넘는 multi-fidelity relaxation dataset에서 이 접근법을 평가한다. 이 프레임워크는 heavily tuned Random Forest보다 우수한, 매우 정확한 error-corrector로 기능할 뿐 아니라, 계산이 실행되기 전에 계산 오차를 예측하고 어떤 setting을 사용해야 하는지 권고하는 zero-cost recommender system으로도 작동한다. 마지막으로, Deep Ensemble과 latent-space distance metric이 명시적인 out-of-distribution 탐지를 제공하여, 모델이 스스로의 치명적 오류를 표시하고 high-throughput screening을 위한 out-of-distribution reliability를 보장할 수 있음을 보인다.","link":"https://openreview.net/forum?id=GS0q2w3z8b"},{"id":"QTDa11VOKv","en":"REACT: Physically and Chemically Consistent Reconstruction of Marine Active Tracers","ko":"REACT: 해양 active tracer의 물리적·화학적으로 일관된 재구성","authors":"Wenbin Dai, Hao Zheng, Shiyu Liang, Chaofan Sun, Xueying Zhang, Hanbo Huang, Xuan Gong, Yiran Zhang, Enhui Liao","abs":"

Reconstructing global sea surface pH from sparse observations is critical for monitoring ocean acidification and understanding marine carbon cycling. Traditional assimilation and inverse models are physically grounded but costly for large-scale reconstruction. Recent black-box and physics-guided AI models improve efficiency, but are mainly designed for passive tracers, where the reconstructed variable is also the transported inventory. In contrast, pH is an active carbonate tracer: it is the prediction target, while dissolved inorganic carbon (DIC) is the conserved carbon inventory. This mismatch can produce low pH error while violating carbonate closure and source-free carbon conservation. To address this, we introduce \\textbf{REACT}, a carbon-first reconstruction framework that decouples transport, active correction, and chemical decoding. REACT transports a latent carbonate state with a conservative advection--diffusion solver, captures non-conservative carbon-cycle variations with a source module, decodes the corrected state into pH, and constrains the output through carbonate equilibrium. This design keeps pH as the target while enforcing consistency on the underlying carbon state. On simulation data, REACT reduces pH NRMSE by \\(14.7\\%\\) and chemical consistency error by \\(24.0\\%\\) over the best baseline. Cross-temporal-scale evaluations show robustness against error accumulation from coarse to fine temporal scales, and ablation studies validate the effectiveness of each component.

","absKo":"희소한 관측으로부터 전지구 해수면 pH를 재구성하는 것은 해양 산성화를 모니터링하고 해양 탄소 순환을 이해하는 데 매우 중요합니다. 전통적인 assimilation 및 inverse model은 물리적으로 타당하지만 대규모 재구성에는 비용이 큽니다. 최근의 black-box 및 physics-guided AI model은 효율성을 개선했지만, 주로 passive tracer를 대상으로 설계되었습니다. 여기서 재구성 대상 변수는 동시에 운반되는 inventory이기도 합니다. 반면 pH는 active carbonate tracer입니다. 즉, pH는 예측 target인 반면 dissolved inorganic carbon (DIC)은 보존되는 carbon inventory입니다. 이러한 불일치는 pH error는 낮지만 carbonate closure와 source-free carbon conservation을 위반하는 결과를 낳을 수 있습니다. 이를 해결하기 위해 우리는 transport, active correction, chemical decoding을 분리하는 carbon-first reconstruction framework인 \\textbf{REACT}를 제안합니다. REACT는 conservative advection--diffusion solver로 latent carbonate state를 transport하고, source module로 비보존적 carbon-cycle variation을 포착하며, 보정된 state를 pH로 decode하고, carbonate equilibrium을 통해 output을 제약합니다. 이 설계는 underlying carbon state의 일관성을 보장하면서 pH를 target으로 유지합니다. simulation data에서 REACT는 최선의 baseline 대비 pH NRMSE를 \\(14.7\\%\\), chemical consistency error를 \\(24.0\\%\\) 감소시킵니다. cross-temporal-scale 평가에서는 coarse temporal scale에서 fine temporal scale로의 error accumulation에 대한 강건성을 보였고, ablation study는 각 구성 요소의 효과를 검증합니다.","link":"https://openreview.net/forum?id=QTDa11VOKv"},{"id":"eHSiINhFoo","en":"Weight-Space Physics: Interpretable Hypernetworks for Lattice Quantum Field Theories","ko":"가중치 공간 물리학: 격자 양자장론을 위한 해석 가능한 하이퍼네트워크","authors":"Tobias Göbel, Julian R. Ebelt, Zier Mensch, Mathis Gerdes, Miranda C. N. Cheng","abs":"

Lattice field theory is the workhorse of non-perturbative physics, used to simulate phenomena from the strong nuclear force to critical phenomena in materials. Its Boltzmann distributions are parametrized analytically by coupling constants, but these bare parameters are weak predictors of observables---extracting physics typically requires extensive simulation. While normalizing flows have emerged as effective samplers at fixed couplings, it remains difficult to interpret what these networks have learned. This raises a natural question: can the flow network parameters themselves be generated for new theories, and can their physics be read off directly from the network weights? We propose lattice field theory as a testbed for neural network interpretability: because the target physics is qualitatively well-understood and smoothly varying, it provides ideal synthetic data with known ground truth. To this end we introduce JEPAWG, a Joint-Embedding Predictive Architecture--based Weight Generator that maps couplings directly to flow weights via a learned latent space. On a scalar theory at lattice sizes $6^2$, $8^2$ and $10^2$, the JEPAWG latent space recovers the correct intrinsic dimension of the underlying manifold, locates the phase transition, and encodes a finite-size shift aligned with the 2D Ising exponent $\\nu \\approx 1$, allowing us to uncover physical structure by studying the network weights alone. As a generator, JEPAWG also interpolates and extrapolates to unseen couplings effectively and remains robust to weight-space discontinuities introduced by multi-seed training data, outperforming PCA, AE, and VAE baselines.

","absKo":"Lattice field theory는 non-perturbative physics의 주력 도구로, strong nuclear force부터 material의 critical phenomenon까지 다양한 현상을 시뮬레이션하는 데 사용된다. 그 Boltzmann distribution은 coupling constant로 해석적으로 parameterized되지만, 이러한 bare parameter는 observable의 약한 예측자에 불과하며, physics를 추출하려면 보통 대규모 시뮬레이션이 필요하다. normalizing flow는 고정된 coupling에서 효과적인 sampler로 부상했지만, 이러한 network가 무엇을 학습했는지 해석하는 것은 여전히 어렵다. 이는 자연스러운 질문을 제기한다: flow network parameter 자체를 새로운 theory에 대해 생성할 수 있는가, 그리고 그 physics를 network weight에서 직접 읽어낼 수 있는가? 우리는 lattice field theory를 neural network interpretability의 testbed로 제안한다. 목표로 하는 physics가 질적으로 잘 이해되어 있고 매끄럽게 변하기 때문에, 알려진 ground truth를 가진 이상적인 synthetic data를 제공하기 때문이다. 이를 위해 우리는 JEPAWG를 소개한다. JEPAWG는 Joint-Embedding Predictive Architecture 기반 Weight Generator로, learned latent space를 통해 coupling을 flow weight로 직접 mapping한다. lattice size $6^2$, $8^2$, $10^2$에서의 scalar theory에서, JEPAWG latent space는 underlying manifold의 올바른 intrinsic dimension을 복원하고, phase transition을 위치시키며, 2D Ising exponent $\\nu \\approx 1$과 정렬되는 finite-size shift를 encode하여, network weight만을 연구함으로써 physical structure를 밝혀낼 수 있게 한다. 생성기로서 JEPAWG는 보지 못한 coupling에 대해서도 효과적으로 interpolation과 extrapolation을 수행하며, multi-seed training data로 인해 도입된 weight-space discontinuity에도 robust하고, PCA, AE, VAE baseline을 능가한다.","link":"https://openreview.net/forum?id=eHSiINhFoo"},{"id":"jonYN1gCsu","en":"GUINEAGRAD: A Differentiable Beam-Beam Interaction Simulator for Gradient-Based Collider Optimization","ko":"GUINEAGRAD: Gradient 기반 충돌기 최적화를 위한 미분 가능한 beam-beam 상호작용 시뮬레이터","authors":"Dimitrios Ntounis","abs":"

Making physical simulators differentiable unlocks gradient-based optimization. We present GuineaGrad, a JAX reimplementation of the Guinea-Pig beam-beam interaction code that computes automatic derivatives of luminosity and beamstrahlung observables via reverse-mode automatic differentiation. Validated against Guinea-Pig to within 2\\% across three collider configurations, GuineaGrad enables gradient-based multi-objective optimization in a 6-dimensional parameter space where Nelder-Mead collapses entirely, producing only degenerate or divergent solutions. Gradient-based optimization becomes wall-clock faster than derivative-free alternatives at ${\\sim}$ 6-10 parameters, establishing differentiable simulation as a practical tool for high-dimensional collider design.

","absKo":"물리 simulator를 differentiable하게 만들면 gradient-based optimization이 가능해진다. 우리는 luminosity와 beamstrahlung observable의 automatic derivative를 reverse-mode automatic differentiation으로 계산하는 Guinea-Pig beam-beam interaction code의 JAX 재구현판인 GuineaGrad를 제시한다. 세 가지 collider configuration에서 Guinea-Pig과 비교해 2\\% 이내로 검증된 GuineaGrad는 6차원 parameter space에서 gradient-based multi-objective optimization을 가능하게 한다. 이 공간에서는 Nelder-Mead가 완전히 무너져, 퇴화되거나 발산하는 해만 생성한다. Gradient-based optimization은 약 ${\\sim}$ 6-10개의 parameter부터 derivative-free alternative보다 wall-clock 기준으로 더 빨라지며, differentiable simulation이 고차원 collider design을 위한 실용적 도구임을 입증한다.","link":"https://openreview.net/forum?id=jonYN1gCsu"},{"id":"B1tPYMXQ76","en":"Advancing and Evaluating AI Agent for Frontier Physics Research","ko":"최첨단 물리학 연구를 위한 AI 에이전트의 발전과 평가","authors":"Tingjia Miao, Wenkai Jin, Jinxin Tan, Muhua Zhang, Xianghe Pang, Zexi Liu, Yuwen Du, Tian Jin, Tu Guo, Yuelin Hu, Zhengliang Zhang, Jingkun Liu, Jiejun Zhang, Yunjie Huang, Yinuo Gao, Yuhan Wang, Wenbo Li, Shuo Chen, Bingyang Zheng, Fengyang Li, Rui Ye, Yuzhi Zhang, Linfeng Zhang, Kun Chen, Wei Wang, Weinan E, Siheng Chen","abs":"

Advances of LLMs in reasoning, domain knowledge, and tool use reveal the potential of agentic science, where AI conducts autonomous research and enables AI-driven scientific discovery. Yet such paradigm remains difficult to realize in theoretical and computational physics research, where research workflows require domain knowledge, long-horizon reasoning, and reliable numerical computation. Towards agentic physics research, we present PhysMaster, a research agent system integrating adaptive multi-trajectory exploration and trusted external knowledge to conduct reliable ultra-long-horizon exploration. Further, we construct PRL-Bench, a research-oriented benchmark adapted from 100 Physical Review Letters papers. Evaluation shows that even frontier models face frequent conceptual errors, lack of advanced theoretical-physics knowledge, and unstable derivations over extended horizons, while PhysMaster substantially mitigates these limitations. Together, PhysMaster and PRL-Bench provide a pioneering paradigm for advancing and evaluating autonomous AI scientists for frontier physics research.

","absKo":"LLM의 reasoning, domain knowledge, tool use의 발전은 agentic science의 가능성을 드러낸다. 여기서 AI는 자율적으로 연구를 수행하고 AI 주도의 과학적 발견을 가능하게 한다. 그러나 이러한 패러다임은 이론물리학과 계산물리학 연구에서 구현하기가 여전히 어렵다. 이들 연구 workflow는 domain knowledge, long-horizon reasoning, 신뢰할 수 있는 수치 계산을 요구하기 때문이다. agentic physics research를 향해, 우리는 적응적 multi-trajectory exploration과 trusted external knowledge를 통합하여 신뢰할 수 있는 ultra-long-horizon exploration을 수행하는 research agent system인 PhysMaster를 제안한다. 더 나아가, 100편의 Physical Review Letters 논문에서 변형한 research-oriented benchmark인 PRL-Bench를 구축한다. 평가 결과, frontier model들조차 개념적 오류를 자주 범하고, 고급 이론물리 지식이 부족하며, 긴 horizon에서 유도 과정이 불안정한 것으로 나타났고, PhysMaster는 이러한 한계를 크게 완화한다. PhysMaster와 PRL-Bench는 함께 frontier physics research를 위한 자율 AI scientist의 발전과 평가를 위한 선구적 패러다임을 제공한다.","link":"https://openreview.net/forum?id=B1tPYMXQ76"},{"id":"xr5xIic0Ei","en":"What Do Lorentz-Equivariant Jet Taggers Learn?","ko":"Lorentz-Equivariant Jet Tagger는 무엇을 학습하는가?","authors":"Jay Agarwal, Siddharth Khare, Dhruv Kumar","abs":"

We study what Lorentz-equivariant jet taggers learn internally, using equivariance tests, linear probes and grade ablations across five models including L-GATr, L-GATr-slim and LLoCa-T. Linear probes show that equivariant models suppress frame-dependent pseudorapidity to zero while encoding jet mass and N-subjettiness strongly. Grade ablations on L-GATr reveal that bivector channels are negligible for top-quark tagging while vector-like channels are dominant but seed-variable, consistent with the network exploiting multiple representational pathways. These results characterize which physical features and algebraic grade structures carry discriminative information in equivariant taggers and may inform future development of such models.

","absKo":"우리는 equivariance test, linear probe, grade ablation을 사용해 L-GATr, L-GATr-slim, LLoCa-T를 포함한 다섯 개 model에서 Lorentz-equivariant jet tagger가 내부적으로 무엇을 학습하는지 연구한다. Linear probe 결과는 equivariant model이 frame-dependent pseudorapidity를 0으로 억제하는 동시에 jet mass와 N-subjettiness는 강하게 인코딩함을 보여준다. L-GATr에 대한 grade ablation은 top-quark tagging에서 bivector channel은 미미한 반면 vector-like channel이 지배적이며, seed-variable이라는 점을 드러내는데, 이는 network가 여러 representational pathway를 활용하고 있음을 시사한다. 이러한 결과는 equivariant tagger에서 어떤 physical feature와 algebraic grade structure가 판별 정보를 전달하는지를 특징짓고, 향후 이러한 model 개발에 도움이 될 수 있다.","link":"https://openreview.net/forum?id=xr5xIic0Ei"},{"id":"tZED9y65LW","en":"Beyond Equality: Inequality-Constrained Flow Matching for Physically Admissible Solutions","ko":"동등성을 넘어: 물리적으로 허용 가능한 해를 위한 불등식 제약 Flow Matching","authors":"Alfred Wei Lun Leong, Yatao Bian, Bryan Kian Hsiang Low","abs":"

Generative models for PDE solutions must produce \\textit{physically admissible} samples, yet state-of-the-art physics-constrained samplers enforce only equality constraints ($h(u)=0$) and ignore the inequality constraints ($g(u)\\leq 0$) that equally govern admissibility---e.g., entropy conditions and invariant-interval bounds. We introduce Inequality-aware Physics-Constrained Flow Matching (I-PCFM), a framework that \\textit{jointly handles equality and inequality constraints} at inference time without retraining. Within I-PCFM we systematically compare three general-purpose projection strategies grounded in classical constrained optimization---(A) slack-variable reformulation, (B) log-barrier augmentation, and (C) active-set projection---together with (D) Frank-Wolfe guidance, a projection-free strategy with closed-form linear-minimization oracles (LMOs) for linear constraints and a Riemannian extension for nonlinear settings. Across four PDE benchmarks, including the 1D Heat and 1D Reaction-Diffusion equations, we establish three key findings: (1) equality-only projection \\textit{degrades} inequality satisfaction by $5$--$20\\times$ relative to unconstrained generation, (2) among general-purpose strategies, active-set projection attains both the highest joint feasibility ($97\\%$ on Reaction-Diffusion) and the lowest runtime, and (3) Frank-Wolfe guidance matches active-set projection's feasibility while running up to $40\\times$ faster. The code can be found at \\url{https://anonymous.4open.science/r/ipcfm-icml-workshop/}.

","absKo":"PDE solution을 위한 generative model은 \\textit{물리적으로 허용 가능한} sample을 생성해야 하지만, 최신 physics-constrained sampler는 equality constraint($h(u)=0$)만 강제하고, admissibility를 똑같이 좌우하는 inequality constraint($g(u)\\leq 0$)는 무시한다. 예를 들어 entropy condition과 invariant-interval bound가 이에 해당한다. 우리는 Inequality-aware Physics-Constrained Flow Matching(I-PCFM)을 제안한다. 이는 재학습 없이 추론 시점에 \\textit{equality와 inequality constraint를 함께 처리}하는 framework이다. I-PCFM 내에서 우리는 고전 constrained optimization에 기반한 세 가지 일반 목적 projection 전략, 즉 (A) slack-variable reformulation, (B) log-barrier augmentation, (C) active-set projection을 체계적으로 비교하고, 여기에 (D) Frank-Wolfe guidance를 더한다. Frank-Wolfe guidance는 linear constraint에 대해 closed-form linear-minimization oracle(LMO)을 사용하고, nonlinear setting을 위한 Riemannian extension을 갖는 projection-free 전략이다. 1D Heat 및 1D Reaction-Diffusion equation을 포함한 네 개의 PDE benchmark 전반에서 우리는 세 가지 핵심 결과를 확인했다. (1) equality-only projection은 unconstrained generation에 비해 inequality 만족도를 5--20배 \\textit{악화}시킨다. (2) 일반 목적 전략 중에서는 active-set projection이 가장 높은 joint feasibility(Reaction-Diffusion에서 97\\%)와 가장 낮은 runtime을 동시에 달성한다. (3) Frank-Wolfe guidance는 active-set projection과 같은 수준의 feasibility를 유지하면서 최대 40배 더 빠르게 동작한다. 코드는 \\url{https://anonymous.4open.science/r/ipcfm-icml-workshop/}에서 확인할 수 있다.","link":"https://openreview.net/forum?id=tZED9y65LW"},{"id":"UtLX8gfw6O","en":"Physics-Informed Neural PDE Solvers via Spatio-Temporal MeanFlow","ko":"Spatio-Temporal MeanFlow를 통한 physics-informed neural PDE solver","authors":"Hanru Bai, Yuncheng Zhou, Difan Zou","abs":"

Deep learning paradigms, such as PINNs and neural operators, have significantly advanced the solving of PDEs. However, they often struggle to capture the continuous integral nature of physical systems, relying either on pointwise residuals that ignore the integral perspective or on pre-discretized temporal grids. Drawing inspiration from MeanFlow, a continuous-time integrator recently developed to efficiently solve generative ODEs, we introduce Spatio-Temporal MeanFlow, which functions as a novel PDE solver learning the finite-interval evolution of physical states. By substituting the generative velocity field with the physical PDE operator, we transform multi-step numerical integration into an efficient prediction with a freely controllable integration length. Crucially, we extend the original MeanFlow constraint from the temporal to the spatio-temporal domain, coupling time evolution with spatial consistency. This yields a unified framework naturally accommodating both time-dependent and stationary PDEs. Comprehensive experiments on benchmarks demonstrate that our approach achieves superior accuracy and inference efficiency over representative baselines. Furthermore, the proposed integral constraint enables excellent generalization to out-of-distribution initial conditions and varying spatial resolutions.

","absKo":"PINNs와 neural operator 같은 deep learning paradigm은 PDEs를 푸는 데 큰 진전을 이루었다. 그러나 이러한 방법은 물리 시스템의 연속적인 적분적 성질을 포착하는 데 종종 어려움을 겪으며, integral perspective를 무시하는 pointwise residuals에 의존하거나, 사전에 이산화된 temporal grid에 의존한다. 최근 generative ODE를 효율적으로 풀기 위해 개발된 continuous-time integrator인 MeanFlow에서 영감을 받아, 우리는 물리 state의 finite-interval evolution을 학습하는 새로운 PDE solver로 기능하는 Spatio-Temporal MeanFlow를 제안한다. 생성 velocity field를 물리 PDE operator로 대체함으로써, 우리는 multi-step numerical integration을 자유롭게 조절 가능한 integration length를 가진 효율적인 prediction으로 변환한다. 특히 우리는 원래 MeanFlow constraint를 temporal domain에서 spatio-temporal domain으로 확장하여, 시간 진화와 공간적 일관성을 결합한다. 그 결과 time-dependent PDE와 stationary PDE를 자연스럽게 모두 수용하는 통합 framework가 된다. 벤치마크에 대한 포괄적인 실험은 우리의 접근법이 대표적 baseline보다 더 우수한 정확도와 inference efficiency를 달성함을 보여준다. 더 나아가 제안한 integral constraint는 out-of-distribution initial condition과 다양한 spatial resolution에 대해 뛰어난 generalization을 가능하게 한다.","link":"https://openreview.net/forum?id=UtLX8gfw6O"},{"id":"VWWv1Npj2b","en":"One More Time: Revisiting Neural Quantum States from a Reinforcement Learning Perspective","ko":"한 번 더: 강화학습 관점에서 본 neural quantum states 재검토","authors":"Juan Agustin Duque, Sergio García-Heredia, Vinicius Hernandes, Eliska Greplova, Thomas Spriggs, Aaron Courville, Anna Dawid","abs":"

Neural quantum states (NQS) provides a flexible and scalable framework for approximating quantum many-body wavefunctions. Among NQS parameterizations, autoregressive models are especially attractive because they enable exact, independent sampling from the Born distribution, avoiding the autocorrelation and mixing issues of Markov chain methods. Yet their optimization remains comparatively underexplored: Adam is a scalable method but ignores function space geometry, while stochastic reconfiguration is principled but costly and numerically fragile in large models. To address this gap, we show that variational energy minimization can be viewed as an advantage policy-gradient problem over the Born distribution, motivating trust-region optimization for NQS training. We introduce \\emph{Proximal Wavefunction Optimization} (PWO), a trust-region algorithm that clips probability-ratio changes in the amplitude channel and wrapped phase increments in the phase channel. PWO avoids explicit matrix inversion, reuses samples across inner updates, and preserves the scalability of first-order optimization. Across Ising, Heisenberg, and frustrated $J_1-J_2$ spin chains, PWO improves stability and wall-clock convergence over Adam, minSR, and SPRING. Finally, we fine-tune a $1.5$B-parameter RWKV-7 model as a neural quantum state, demonstrating NQS optimization at a scale over three orders of magnitude beyond prior work.

","absKo":"Neural quantum states(NQS)는 quantum many-body wavefunction을 근사하기 위한 유연하고 확장 가능한 프레임워크를 제공한다. NQS parameterization 가운데 autoregressive model은 Born distribution으로부터 정확하고 독립적인 샘플링을 가능하게 하여 Markov chain 방법의 autocorrelation과 mixing 문제를 피할 수 있기 때문에 특히 매력적이다. 그러나 최적화는 상대적으로 덜 탐구되었다. Adam은 확장 가능하지만 function space geometry를 무시하고, stochastic reconfiguration은 원칙적이지만 대규모 모델에서는 비용이 많이 들고 수치적으로 취약하다. 이 간극을 메우기 위해, 우리는 variational energy minimization을 Born distribution 위의 advantage policy-gradient 문제로 볼 수 있음을 보이고, 이를 통해 NQS 학습에 trust-region 최적화를 동기 부여한다. 우리는 \\emph{Proximal Wavefunction Optimization}(PWO)을 도입한다. 이는 amplitude channel에서는 probability-ratio 변화를, phase channel에서는 wrapped phase increment를 clipping하는 trust-region 알고리즘이다. PWO는 명시적인 matrix inversion을 피하고, inner update 사이에 샘플을 재사용하며, first-order optimization의 확장성을 유지한다. Ising, Heisenberg, 그리고 frustrated $J_1-J_2$ spin chain 전반에서 PWO는 Adam, minSR, SPRING보다 안정성과 wall-clock 수렴을 개선한다. 마지막으로, 우리는 1.5B 파라미터 RWKV-7 모델을 neural quantum state로 fine-tuning하여, 이전 연구보다 세 자릿수 이상 큰 규모에서 NQS 최적화를 시연한다.","link":"https://openreview.net/forum?id=VWWv1Npj2b"},{"id":"oQKRvOLYbb","en":"IKNO: Infinite-order Kernel Neural Operators","ko":"IKNO: 무한 차수 kernel neural operator","authors":"Pengyuan Zhu, Ivor Tsang, Yueming Lyu","abs":"

Neural operators have achieved significant success in modern scientific computing due to their flexibility and strong generalization capabilities. Existing models, however, primarily rely on first-order kernel integral approximations, which severely limit their expressivity. To address this, we propose the Infinite-order Kernel Neural Operator (IKNO), which constructs neural operators via infinite-order kernel integrals and admits an elegant closed-form finite approximation. We develop two complementary infinite-order neural operator constructions: IKNO-Vanilla, which applies the full-kernel resolvent on the product grid via Kronecker eigendecomposition, and IKNO-TP, an alternative tensor-product operator that composes per-axis resolvents. Furthermore, we develop fast computation schemes for both variants of IKNO, which achieve outstanding global information aggregation while maintaining high computational efficiency. Empirically, we evaluate our IKNO on both time-dependent and time-independent benchmarks with arbitrary input shapes, including large-scale industrial datasets. Extensive experiments demonstrate that the IKNO method consistently achieves the SOTA accuracy with significant improvements on nearly all benchmark datasets while maintaining scalability to very large point clouds.

","absKo":"Neural operator는 유연성과 강한 generalization 능력 덕분에 현대 과학 계산에서 큰 성공을 거두었다. 그러나 기존 모델은 주로 first-order kernel integral approximation에 의존하며, 이는 표현력을 심각하게 제한한다. 이를 해결하기 위해 우리는 Infinite-order Kernel Neural Operator(IKNO)를 제안한다. IKNO는 infinite-order kernel integral을 통해 neural operator를 구성하며, 우아한 closed-form finite approximation을 허용한다. 우리는 상호보완적인 두 가지 infinite-order neural operator 구성을 개발한다: IKNO-Vanilla는 Kronecker eigendecomposition을 통해 product grid 위에서 full-kernel resolvent를 적용하고, IKNO-TP는 축별 resolvent를 합성하는 대안적 tensor-product operator이다. 더 나아가 우리는 IKNO의 두 변형 모두를 위한 빠른 계산 scheme을 개발했으며, 이는 높은 계산 효율을 유지하면서 뛰어난 global information aggregation을 달성한다. 실증적으로 우리는 arbitrary input shape을 갖는 time-dependent 및 time-independent benchmark, 대규모 산업 데이터셋을 포함한 다양한 과제에서 IKNO를 평가한다. 광범위한 실험은 IKNO 방법이 거의 모든 benchmark dataset에서 유의미한 향상과 함께 일관되게 SOTA 정확도를 달성하며, 매우 큰 point cloud까지 확장 가능함을 보여준다.","link":"https://openreview.net/forum?id=oQKRvOLYbb"},{"id":"UZsb3AffU4","en":"Equivariant Graph Neural Networks Improve Optical Spectra Prediction for Materials Screening","ko":"등변 graph neural network가 materials screening을 위한 optical spectra prediction을 향상시킨다","authors":"Kasper Helverskov Petersen, François R J Cornet, Martin Ovesen, Mikkel Jordahn, Kristian S. Thygesen, Mikkel N. Schmidt","abs":"

Scalable prediction of optical spectra is a critical component of high-throughput materials screening for optoelectronic applications such as solar cells. Existing surrogate models are trained on spectra computed from lower levels of theory or rely on rotation-invariant scalar features, limiting their geometric expressiveness. We explore the use of equivariant graph neural networks for optical spectra prediction, adapting GotenNet to this task and evaluating it on multiple datasets including a recently published collection of 10,533 structures with spectra computed at the level of the random phase approximation (RPA). The proposed model outperforms the current state of the art, with the largest gains in the 0-8 eV range and on predicting the static real permittivity, both of particular relevance for thin-film optics.

","absKo":"광학 스펙트럼의 scalable prediction은 solar cell과 같은 optoelectronic application을 위한 high-throughput materials screening의 핵심 구성요소이다. 기존 surrogate model은 더 낮은 theory level에서 계산된 spectrum으로 학습되거나 rotation-invariant scalar feature에 의존하기 때문에, 기하학적 표현력이 제한된다. 우리는 optical spectrum prediction에 equivariant graph neural network를 적용하는 방안을 탐구하며, 이 작업에 맞게 GotenNet을 조정하고, 최근 발표된 random phase approximation(RPA) level에서 계산된 spectrum을 가진 10,533개 구조의 데이터셋을 포함한 여러 dataset에서 평가한다. 제안한 model은 현재 state of the art를 능가하며, 특히 0-8 eV 구간과 thin-film optics에서 특히 중요한 static real permittivity 예측에서 가장 큰 향상을 보였다.","link":"https://openreview.net/forum?id=UZsb3AffU4"},{"id":"oAe1wRN76R","en":"21cmEMUv3: a Hybrid Diffusion-LSTM Emulator for Joint Inference of Cosmic Dawn and Reionisation","ko":"21cmEMUv3: 우주 새벽과 재이온화의 공동 추론을 위한 하이브리드 diffusion-LSTM 에뮬레이터","authors":"Daniela Breitman","abs":"

The cosmic dawn (CD) and epoch of reionisation (EoR) mark the first billion years of cosmic history, during which the first stars and galaxies formed and drove the final major phase transition of the Universe as their radiation ionised the neutral intergalactic medium (IGM), culminating in the present-day Universe.

Over the past decade, a surge of observations has begun to illuminate these periods, with upcoming observations of the redshifted 21-cm line of neutral hydrogen — a direct tracer of the neutral IGM — expected to revolutionise our understanding of the first luminous sources.

Robustly interpreting these observations with Bayesian inference requires repeatedly simulating the entire Universe across high-dimensional cosmological and astrophysical parameter spaces, making ML a critical tool for tractable inference, with emulation emerging as a powerful approach.

We present {\\tt 21cmEMUv3}, an emulator trained on {\\tt 21cmFASTv3} simulations modelling two galaxy populations, conditioned on eleven parameters ($\\sigma_8$ plus ten astrophysical parameters), and designed to jointly predict seven summary observables including 21-cm power spectra, IGM thermal history, and UV luminosity functions. We emulate the 2D 21-cm PS via score-based diffusion and the remaining six summaries via long short-term memory (LSTM) networks, achieving sub-percent median accuracy across all outputs.

We showcase the emulator by producing forecasts for the upcoming Square Kilometre Array (SKA), finding that SKA will enable percent-level constraints on the neutral hydrogen fraction at EoR midpoint, while further demonstrating that inference pipelines remain highly sensitive to model misspecification, an important methodological challenge for robust scientific inference with upcoming observations.

","absKo":"cosmic dawn(CD)과 epoch of reionisation(EoR)은 우주 역사 첫 10억 년을 나타내며, 이 시기 최초의 별과 은하가 형성되었고, 그 복사로 중성 intergalactic medium(IGM)이 이온화되면서 우주의 마지막 주요 phase transition이 진행되어 오늘날의 우주에 이르렀다.\n지난 10년간 관측의 급증으로 이 시기들이 점차 드러나기 시작했으며, 중성 수소의 redshifted 21-cm line에 대한 향후 관측은 - 중성 IGM의 직접 tracer인 이 신호는 - 최초의 luminous source에 대한 우리의 이해를 혁신할 것으로 기대된다.\n이러한 관측을 Bayesian inference로 견고하게 해석하려면, 고차원 cosmological 및 astrophysical parameter space 전반에 걸쳐 우주 전체를 반복적으로 시뮬레이션해야 하므로, ML은 tractable inference를 위한 핵심 도구가 되며, 그중 emulation이 강력한 접근법으로 부상하고 있다.\n우리는 {\\tt 21cmEMUv3}를 제시한다. 이는 두 개의 galaxy population을 모델링하는 {\\tt 21cmFASTv3} simulation으로 학습된 emulator로, 11개의 parameter($\\sigma_8$와 10개의 astrophysical parameter)에 조건화되며, 21-cm power spectrum, IGM thermal history, UV luminosity function을 포함한 7개의 summary observable을 jointly predict하도록 설계되었다. 우리는 2D 21-cm PS는 score-based diffusion으로, 나머지 6개 summary는 long short-term memory(LSTM) network로 emulate하며, 모든 output에서 sub-percent 수준의 median accuracy를 달성한다.\n우리는 다가오는 Square Kilometre Array(SKA)에 대한 forecast를 생성함으로써 이 emulator의 활용 가능성을 보여준다. 그 결과 SKA는 EoR midpoint에서 중성 수소 fraction에 대해 percent-level constraint를 가능하게 할 것이며, 동시에 inference pipeline이 model misspecification에 여전히 매우 민감하다는 점도 추가로 입증한다. 이는 향후 관측을 활용한 견고한 scientific inference에서 중요한 방법론적 과제이다.","link":"https://openreview.net/forum?id=oAe1wRN76R"},{"id":"IcLmHEb5HO","en":"Adaptive Interaction Graphs for Particle Simulation","ko":"입자 시뮬레이션을 위한 적응형 상호작용 그래프","authors":"Aiden Zhou","abs":"

Learned particle simulators based on graph neural networks achieve strong one-step accuracy, but errors compound over long horizons. An underexplored variable is the interaction graph: existing methods fix its topology via $k$-nearest neighbors or a static radius rule, regardless of local model confidence. We propose making this graph adaptive: a per-particle variance head, trained jointly with the acceleration head under a

heteroscedastic Gaussian NLL loss, drives a trajectory in which high-uncertainty particles receive an expanded neighborhood. This is done at little extra inference cost by using the previous step's uncertainty estimate. A key discovery is that the variance head learns a meaningful notion of uncertainty: high-variance particles concentrate near complex regions, such as splash zones or free surfaces. When this signal drives graph topology, the resulting AdaptGNS simulator achieves a strict Pareto improvement on WaterDrop and a modest gain on Sand. Given the model's stronger performance on WaterDrop, we hypothesize that adaptive graphs are most useful when complexity is concentrated in space. Our code can be found at https://github.com/aidenzhou8/AdaptGNS.

","absKo":"graph neural network 기반의 learned particle simulator는 단일 step 정확도는 높지만, 장기 horizon에서는 오차가 누적된다. 충분히 탐구되지 않은 변수는 interaction graph이다. 기존 방법은 local model confidence와 무관하게 $k$-nearest neighbors 또는 static radius rule로 topology를 고정한다. 우리는 이 graph를 adaptive하게 만들 것을 제안한다. 즉, heteroscedastic Gaussian NLL loss 아래에서 acceleration head와 함께 공동으로 학습되는 per-particle variance head가 trajectory를 유도하여, uncertainty가 높은 particle에 더 넓은 neighborhood를 부여한다. 이는 이전 step의 uncertainty estimate를 사용함으로써 추가 inference cost를 거의 들이지 않고 수행된다. 핵심 발견은 variance head가 의미 있는 uncertainty 개념을 학습한다는 점이다. 높은 variance를 보이는 particle은 splash zone이나 free surface 같은 복잡한 영역 주변에 집중된다. 이 신호가 graph topology를 구동할 때, 결과적인 AdaptGNS simulator는 WaterDrop에서 strict Pareto improvement를 달성하고 Sand에서 소폭 향상을 보인다. 모델이 WaterDrop에서 더 강한 성능을 보인다는 점을 고려할 때, 우리는 adaptive graph가 복잡성이 공간적으로 집중될 때 가장 유용하다고 가정한다. 코드는 https://github.com/aidenzhou8/AdaptGNS 에서 확인할 수 있다.","link":"https://openreview.net/forum?id=IcLmHEb5HO"},{"id":"s5picDaxfo","en":"How Much Memory Do We Need? Adaptive Memory Gate for Neural Operators","ko":"얼마나 많은 memory가 필요한가? neural operator를 위한 적응형 memory gate","authors":"Jihyeon Hur, Yongseok Kwon, Min-Gi Jo, Jeongwhan Choi, Noseong Park","abs":"

Neural operators have emerged as a powerful data-driven approach for solving time-dependent PDEs. Among recent advances, memory-augmented neural operators explicitly incorporate past states and have achieved remarkable performance under low-resolution observation settings. However, existing approaches apply a fixed memory weight regardless of observation conditions, such as resolution or physical parameters, limiting their adaptability. Our preliminary experiments reveal that optimal memory weight varies with resolution and viscosity, implying that a fixed memory weight cannot simultaneously optimize performance across diverse settings. We propose AMGFNO, which dynamically modulates memory weight through a learnable gate. On the Kuramoto-Sivashinsky and Burgers' equations, AMGFNO achieves 55--79\\% nRMSE reduction over at low resolution, with the learned gate value automatically decreasing from $\\bar{g} \\approx 0.7$ to near-zero as resolution increases.

","absKo":"Neural operator는 시간 의존 PDE를 푸는 강력한 data-driven 접근으로 떠올랐다. 최근의 발전 중 memory-augmented neural operator는 과거 상태를 명시적으로 통합하며, 저해상도 관측 설정에서 뛰어난 성능을 보였다. 그러나 기존 접근은 해상도나 물리적 파라미터 같은 관측 조건과 무관하게 고정된 memory weight를 적용하므로 적응성이 제한된다. 우리의 예비 실험은 최적의 memory weight가 해상도와 viscosity에 따라 달라진다는 점을 보여주며, 이는 고정된 memory weight로는 다양한 설정 전반에서 성능을 동시에 최적화할 수 없음을 시사한다. 우리는 learnable gate를 통해 memory weight를 동적으로 조절하는 AMGFNO를 제안한다. Kuramoto-Sivashinsky 방정식과 Burgers' 방정식에서 AMGFNO는 저해상도에서 nRMSE를 55--79\\% 줄였고, 학습된 gate 값은 해상도가 증가함에 따라 평균 $\\bar{g} \\approx 0.7$에서 거의 0에 가깝게 자동으로 감소했다.","link":"https://openreview.net/forum?id=s5picDaxfo"},{"id":"eUp3ZsDgeK","en":"Hybrid Neural-MPM for Accelerated and Controllable Fluid Simulations","ko":"가속 가능하고 제어 가능한 유체 시뮬레이션을 위한 Hybrid Neural-MPM","authors":"Hong Huang, Jingxuan Xu, Chuhang Zou, Manolis Savva, Yunchao Wei, Wuyang Chen","abs":"

We propose **Hybrid Neural-MPM**, a neural physics framework for accelerated, controllable fluid simulation.

While traditional physics-based methods offer high accuracy, they are often computationally prohibitive and prone to latency.

Recent machine-learning approaches have reduced these costs, yet most still struggle to support the demands of interactive applications.

To bridge this gap, we introduce a novel hybrid method that integrates

*numerical simulation, neural physics, and generative control.*

Our neural physics model balances low-latency performance with physical fidelity by employing a fallback safeguard to classical numerical solvers.

Furthermore, we design a diffusion-based controller, trained via a reverse modeling strategy, to generate external dynamic force fields for intuitive fluid manipulation.

Our system demonstrates improved performance across diverse 2D/3D scenarios, material types, and obstacle interactions, achieving accelerated simulations

($11\\sim 29$\\% latency reduced)

while enabling fluid control guided by user-friendly freehand sketches.

We include video results in the supplement, and commit to releasing models and data upon acceptance.

","absKo":"우리는 가속화되고 제어 가능한 fluid simulation을 위한 neural physics framework인 **Hybrid Neural-MPM**을 제안한다.\n전통적인 physics-based method는 높은 정확도를 제공하지만, 계산 비용이 매우 크고 지연(latency)에 취약한 경우가 많다.\n최근 machine-learning 기반 접근은 이러한 비용을 줄였지만, 대부분은 여전히 상호작용형 응용의 요구를 지원하는 데 어려움을 겪는다.\n이 간극을 메우기 위해 우리는\n*numerical simulation, neural physics, and generative control*\n을 통합하는 새로운 hybrid method를 도입한다.\n우리의 neural physics model은 classical numerical solver로의 fallback safeguard를 사용하여 낮은 지연 성능과 물리적 충실도 사이의 균형을 맞춘다.\n또한, reverse modeling 전략을 통해 학습된 diffusion-based controller를 설계하여, 직관적인 fluid 조작을 위한 외부 동적 힘장(dynamic force field)을 생성한다.\n우리 시스템은 다양한 2D/3D 시나리오, 물질 유형, obstacle 상호작용 전반에서 향상된 성능을 보이며, 사용자 친화적인 freehand sketch로 유도되는 fluid control을 가능하게 하면서 가속화된 simulation\n($11\\sim 29$\\% latency reduced)\n을 달성한다.\n보충 자료에 video 결과를 포함하였으며, 채택 시 모델과 데이터를 공개할 것을 약속한다.","link":"https://openreview.net/forum?id=eUp3ZsDgeK"},{"id":"ehu3SPjrpP","en":"QUIVER: Quantum-Informed Views for Enhanced Representations in Large ML Models","ko":"QUIVER: 대형 ML 모델의 표현 향상을 위한 양자-정보 기반 관점","authors":"Aritra Bal, Michael Binder, Markus Klute, Benedikt Maier, Michael Spannowsky","abs":"

Large machine learning models benefit substantially from multimodal inputs that provide a complementary view of the same example. We introduce QUIVER (QUantum Informed Views for Enhanced Representations), a paradigm that enriches classical data-driven features with a quantum Fisher view: a geometrically motivated, basis-independent summary of higher-order correlations captured by a variational quantum circuit (VQC) trained to perform the same task. Unlike classical feature augmentation, the quantum Fisher information matrix encodes the intrinsic geometry of the learned quantum state manifold.This feature map, motivated by quantum information theory, is ordinarily non-trivial to model classically. However, it can reveal statistical structure that additional classical data or model capacity finds difficult to learn, making it a complementary modality. We demonstrate that QUIVER improves standard performance metrics on two benchmark datasets from very different fields: the JetClass dataset for predicting jet flavor at the Large Hadron Collider (LHC), and the QM9 dataset for predicting molecule properties. The core contribution, however, is domain-agnostic: the quantum Fisher view can be fused into a broad class of model architectures via targeted modifications to the base architecture, to incorporate information about the quantum geometry of the problem.

","absKo":"대규모 machine learning model은 같은 example을 보완적으로 바라보는 multimodal input에서 상당한 이득을 얻는다. 우리는 QUIVER(QUantum Informed Views for Enhanced Representations)를 소개한다. 이는 classical data-driven features에 quantum Fisher view를 더하는 패러다임으로, 같은 task를 수행하도록 학습된 variational quantum circuit (VQC)이 포착한 higher-order correlation의 geometrically motivated, basis-independent summary를 제공한다. classical feature augmentation과 달리 quantum Fisher information matrix는 학습된 quantum state manifold의 intrinsic geometry를 인코딩한다. quantum information theory에서 동기를 얻은 이 feature map은 일반적으로 classical하게 모델링하기에 비자명하다. 그러나 추가적인 classical data나 model capacity로는 학습하기 어려운 statistical structure를 드러낼 수 있어, 보완적 modality가 된다. 우리는 QUIVER가 서로 매우 다른 두 benchmark dataset에서 표준 성능 지표를 개선함을 보인다. 하나는 Large Hadron Collider (LHC)에서 jet flavor를 예측하는 JetClass dataset이고, 다른 하나는 molecule property를 예측하는 QM9 dataset이다. 그러나 핵심 기여는 domain-agnostic하다. quantum Fisher view는 base architecture에 대한 targeted modification을 통해 다양한 model architecture에 fuse될 수 있으며, 문제의 quantum geometry에 대한 정보를 통합할 수 있다.","link":"https://openreview.net/forum?id=ehu3SPjrpP"},{"id":"6HCUxH4vdm","en":"SNAP-FM: Sparse Nonlinear Accelerated Projection for Physics-Constrained Generative Modeling","ko":"SNAP-FM: 물리 제약 생성 모델링을 위한 sparse nonlinear accelerated projection","authors":"Alaina Kolli, Theodoros Xenakis, Utkarsh Utkarsh, Pengfei Cai, Rafael Gomez-Bombarelli, Alan Edelman, Christopher Vincent Rackauckas","abs":"

Generative models have emerged as scalable surrogates for physical simulation, yet they offer no guarantee that their outputs respect the conservation laws, boundary conditions, and nonlinear invariants that govern the underlying physics. Constrained sampling closes this gap, enforcing such constraints exactly at inference time without retraining, but at a computational cost: projection, correction and trajectory-optimization steps are repeated during sampling, with these steps becoming expensive for nonlinear constraints.

Standard ML frameworks exacerbate this: their dense tensor algebra and limited sparse solver composability obscure the structure that physical constraints naturally induce, making efficient batched nonlinear optimization difficult to realize in practice. We address this bottleneck by exploiting the structure that sample-wise batching and local PDE couplings induce in the projection subproblems -- namely, block-sparse Jacobian and KKT systems -- exposing this structure using ExaModels.jl and solving the resulting sparse nonlinear programs with MadNLP.jl and GPU sparse factorization. Applied to Physics-Constrained Flow Matching (PCFM), on PDE benchmarks with linear, nonlinear, one-dimensional, and two-dimensional constraints, this approach accelerates nonlinear constraint projection while maintaining constraint satisfaction. These results show that sparse GPU nonlinear optimization is a practical foundation for constrained generative sampling in scientific machine learning.

","absKo":"Generative model은 물리 시뮬레이션을 위한 확장 가능한 surrogate로 부상했지만, 그 출력이 underlying physics를 지배하는 conservation law, boundary condition, nonlinear invariant를 준수한다는 보장은 없다. Constrained sampling은 이 공백을 메우며, 재학습 없이 inference time에 이러한 제약을 정확히 강제한다. 그러나 계산 비용이 따른다. projection, correction, trajectory-optimization step이 sampling 동안 반복되며, nonlinear constraint에서는 이 step이 특히 비싸진다.\n표준 ML framework는 이 문제를 악화시킨다. dense tensor algebra와 제한된 sparse solver composability는 물리 제약이 자연스럽게 유도하는 구조를 가려서, 효율적인 batched nonlinear optimization을 실제로 구현하기 어렵게 만든다. 우리는 sample-wise batching과 local PDE coupling이 projection subproblem에 유도하는 구조, 즉 block-sparse Jacobian과 KKT system을 활용하여 이 병목을 해결한다. 이 구조를 ExaModels.jl로 노출하고, MadNLP.jl과 GPU sparse factorization으로 생성된 sparse nonlinear program을 푼다. 이를 Physics-Constrained Flow Matching(PCFM)에 적용하여, linear, nonlinear, one-dimensional, two-dimensional constraint를 가진 PDE benchmark에서 이 접근법은 constraint satisfaction을 유지하면서 nonlinear constraint projection을 가속한다. 이러한 결과는 sparse GPU nonlinear optimization이 scientific machine learning에서 constrained generative sampling을 위한 실용적인 기반임을 보여준다.","link":"https://openreview.net/forum?id=6HCUxH4vdm"},{"id":"1eJBXqNBvS","en":"Deriving an Analytical Expression for the Rise of Type Ia Supernovae with Symbolic Regression and High Cadence data","ko":"기호 회귀와 고빈도 데이터로 Ia형 초신성 증가에 대한 해석식 유도","authors":"Mark Kanex, Qinan Wang","abs":"

Early excess emission in Type Ia supernovae light curves provides a direct clue to the progenitor systems and explosion mechanism of these standardizable candles. Yet the traditional baseline model, i.e., the empirical ‘power-law’ model developed from low cadence data, is increasingly limited for modern high cadence data due to its restricted time coverage and sensitivity to the chosen fitting window. Here we use symbolic regression to search for compact analytic descriptions of sub-hour-cadence SNe Ia light curves observed by TESS and Kepler, and examine whether events with early excess emission occupy a distinct region of the learned parameter space. Our approach is able to discover functions with inherent invariance and separate different light curve components. We find that a model combining quadratic and sinusoidal terms provides the best description of the observed rise. These preliminary results demonstrate the potential of symbolic regression as an interpretable, data-driven approach for discovering empirical structure in astronomical data.

","absKo":"Type Ia supernova light curve에서 초기 과잉 방출(early excess emission)은 표준화 가능한 candle의 progenitor system과 폭발 메커니즘에 대한 직접적인 단서를 제공한다. 그러나 저주기(low cadence) 데이터로부터 개발된 경험적 ‘power-law’ 모델이라는 전통적 baseline은, 시간 범위가 제한적이고 적합(fitting) 윈도우 선택에 민감하다는 이유로 현대의 고주기(high cadence) 데이터에는 점점 더 한계를 드러낸다. 본 연구에서는 symbolic regression을 사용해 TESS와 Kepler로 관측된 서브-시간 간격(sub-hour-cadence) SNe Ia light curve에 대한 간결한 해석적 설명을 탐색하고, 초기 과잉 방출이 있는 사건들이 학습된 parameter space의 뚜렷한 영역을 차지하는지 살펴본다. 우리의 접근법은 고유한 불변성(invariance)을 지닌 함수를 발견하고 서로 다른 light curve 구성 요소를 분리할 수 있다. 우리는 quadratic term과 sinusoidal term을 결합한 모델이 관측된 상승(rise)을 가장 잘 설명한다는 사실을 발견했다. 이러한 예비 결과는 astronomical data에서 경험적 구조를 발견하는 해석 가능하고 데이터 기반의 접근법으로서 symbolic regression의 잠재력을 보여준다.","link":"https://openreview.net/forum?id=1eJBXqNBvS"},{"id":"5ZIFFutCKi","en":"QMaxCal: Path-Space Regularization for Open Quantum Control via Girsanov's Theorem","ko":"QMaxCal: Girsanov 정리를 활용한 개방형 양자 제어를 위한 경로공간 정규화","authors":"Merijn Moody, Zier Mensch, Miranda C. N. Cheng, Peter G. Bolhuis, Max Welling","abs":"

Reliable quantum control in the presence of decoherence requires

policies that combat the effect of environmental noise on the

controlled dynamics. Open quantum systems under continuous monitoring

generate classical measurement records whose drift depends on the

noise experienced by the system; the records of two evolutions

sharing the same decoherence channels differ only in this drift, so

Girsanov's theorem yields a closed-form, differentiable estimator of

the KL divergence between their trajectory distributions. We

instantiate this estimator with two physically motivated reference

measures, yielding two regularizers that both drive the system

toward states where the effects of decoherence are minimal: the Wiener KL, $(KL_{\\mathrm{W}})$ which is empirically more effective under

certain conditions on the noise model, and the drift-variance

regularizer $(R_{\\mathrm{DV}})$, which works for all noise models. Both are

qualitatively distinct from existing penalties on control fluence or

smoothness: they penalize the observable consequences of control on

the decoherence channels rather than the control amplitude itself.

The regularizers outperform unregularized gradient-based and

reinforcement-learning baselines across a range of open quantum

systems---including single- and multi-qubit benchmarks and a multi-qubit chain

calibrated to a published snapshot of the IBM Kingston

processor---along several axes of evaluation: final-state fidelity,

robustness to mismatch in the assumed noise model (gains grow

from $+17$ pp at training noise to $+27$ pp under $2.5\\times$

noise mismatch), and occupation of forbidden states. The regularizers reduce infidelity

by up to 50\\%, with $\\sim$16\\% gains on the calibrated IBM Kingston

chain.

","absKo":"decoherence가 존재하는 환경에서 신뢰할 수 있는 quantum control을 수행하려면, 제어된 dynamics에 대한 환경 잡음의 영향을 상쇄하는 policy가 필요하다. 지속적으로 모니터링되는 open quantum system은 시스템이 겪는 noise에 따라 drift가 달라지는 classical measurement record를 생성한다. 동일한 decoherence channel을 공유하는 두 evolution의 record는 이 drift에서만 차이가 나므로, Girsanov's theorem을 통해 두 trajectory distribution 사이 KL divergence의 closed-form, differentiable estimator를 얻을 수 있다. 우리는 이 estimator를 물리적으로 동기가 있는 두 reference measure로 구체화하여, decoherence의 효과가 최소인 상태로 시스템을 유도하는 두 regularizer를 얻었다. 하나는 Wiener KL, $(KL_{\\mathrm{W}})$로, 특정 noise model 조건에서 경험적으로 더 효과적이며, 다른 하나는 drift-variance regularizer $(R_{\\mathrm{DV}})$로, 모든 noise model에서 작동한다. 두 regularizer는 control fluence나 smoothness에 대한 기존 penalty와 질적으로 다르다. 즉, control amplitude 자체가 아니라 decoherence channel에 대한 control의 관측 가능한 결과를 penalize한다. 이 regularizer들은 단일 및 다중 qubit benchmark와 published snapshot of the IBM Kingston processor에 맞춘 multi-qubit chain을 포함한 다양한 open quantum system에서, 미규제 gradient-based 및 reinforcement-learning baseline보다 여러 평가 축에서 우수했다. 평가 축에는 final-state fidelity, 가정한 noise model 불일치에 대한 robustness(학습 noise에서는 +17pp, 2.5배 noise mismatch에서는 +27pp로 향상), 그리고 forbidden state 점유가 포함된다. 이 regularizer들은 infidelity를 최대 50\\%까지 줄였으며, IBM Kingston chain에서는 약 16\\%의 향상을 보였다.","link":"https://openreview.net/forum?id=5ZIFFutCKi"},{"id":"7xf2daF2vi","en":"Quality-Weighted Determinantal Point Processes for Sample Efficient DFT Surrogate Models","ko":"샘플 효율적인 DFT surrogate 모델을 위한 품질 가중 DPP","authors":"Shirley X. Yu, Adji Bousso Dieng","abs":"

Density functional theory (DFT) is widely used to compute electronic structure and materials properties, but it is expensive to run. When compute is limited, practitioners must choose which candidate systems to evaluate. We study this selection problem on JARVIS-DFT, a large standardized dataset of DFT labeled crystal structures and a realistic testbed for training machine learning models to predict expensive DFT computed properties. We propose quality-weighted determinantal point process (QW-DPP), a label-free coreset selection method for training DFT surrogate models. QW-DPP selects candidates that are both diverse through an RBF kernel and representative through a density-based quality score from unlabeled composition fingerprints. On JARVIS-DFT, QW-DPP reduces test MAE compared with random sampling by up to 8.9% at a 1% labeling budget across formation energy and bulk modulus. This improvement holds across Ridge, Bayesian Ridge, and Gradient Boosting. We find that QW-DPP is most useful at very small budgets because it improves worst case coverage of feature space.

","absKo":"Density functional theory (DFT)는 electronic structure와 materials property를 계산하는 데 널리 사용되지만, 실행 비용이 크다. 계산 자원이 제한될 때, 실무자는 어떤 후보 system을 평가할지 선택해야 한다. 우리는 DFT labeled crystal structure의 대규모 표준화 dataset이자, 비용이 큰 DFT 계산 property를 예측하는 machine learning model 학습을 위한 현실적인 testbed인 JARVIS-DFT에서 이 선택 문제를 연구한다. 우리는 DFT surrogate model 학습을 위한 label-free coreset selection 방법인 quality-weighted determinantal point process (QW-DPP)를 제안한다. QW-DPP는 RBF kernel을 통해 다양성과, unlabeled composition fingerprint에서 얻은 density 기반 quality score를 통해 대표성을 모두 갖춘 candidate를 선택한다. JARVIS-DFT에서 QW-DPP는 formation energy와 bulk modulus 전반에서 1% labeling budget일 때 random sampling 대비 test MAE를 최대 8.9%까지 감소시킨다. 이 개선은 Ridge, Bayesian Ridge, Gradient Boosting 전반에서 유지된다. 우리는 QW-DPP가 매우 작은 budget에서 특히 유용하다는 것을 발견했는데, 이는 feature space의 최악 경우 coverage를 개선하기 때문이다.","link":"https://openreview.net/forum?id=7xf2daF2vi"},{"id":"VNdLLU6SkE","en":"Text-Guided Diffusion Model for Adaptive Reconstruction through Scattering Media","ko":"산란 매질을 통한 적응적 복원을 위한 텍스트 기반 Diffusion 모델","authors":"Fakhriyya Mammadova, Chanseok Lee, Mooseok Jang","abs":"

Optical imaging through scattering media presents a fundamentally ill-posed inverse problem, as inhomogeneous materials scramble incident light into seemingly random speckle patterns, destroying phase information and diffraction-limited resolution. While the optical memory effect allows for the recovery of object autocorrelation from scattered intensity, traditional phase-retrieval algorithms routinely stagnate on complex or dense objects. In this paper, we introduce a text-guided diffusion model framework to solve the inverse problem of imaging through scattering layers, leveraging the strong generative priors of diffusion models and the semantic disambiguation of text conditioning. We evaluate our method on the MNIST and Quick, Draw! datasets using experimentally obtained speckle measurements through an optical diffuser. Our results demonstrate that textual prompts effectively resolve the inherent ambiguities of phase loss, significantly outperforming traditional correlation-based imaging techniques. As a foundational proof of concept, this work isolates the phase-retrieval reconstruction step to demonstrate that explicit semantic priors can resolve topological ambiguities where classical methods fail.

","absKo":"산란 매질을 통한 광학 이미징은 본질적으로 ill-posed inverse problem을 제시한다. 비균질한 물질은 입사광을 겉보기에는 무작위한 speckle pattern으로 뒤섞어 phase information과 diffraction-limited resolution을 파괴하기 때문이다. optical memory effect를 이용하면 산란 강도에서 object autocorrelation을 복원할 수 있지만, 기존 phase-retrieval 알고리즘은 복잡하거나 밀집된 객체에서 흔히 정체된다. 본 논문에서는 text-guided diffusion model framework를 도입하여 scattering layers를 통과한 이미징의 inverse problem을 해결한다. 이는 diffusion model의 강력한 generative prior와 text conditioning의 semantic disambiguation을 활용한다. 우리는 optical diffuser를 통과한 실험적으로 획득한 speckle measurements를 사용하여 MNIST와 Quick, Draw! datasets에서 제안법을 평가하였다. 결과는 textual prompt가 phase loss가 내재하는 모호성을 효과적으로 해소하며, 전통적인 correlation-based imaging techniques를 크게 능가함을 보여준다. foundational proof of concept으로서, 본 연구는 phase-retrieval reconstruction step에 초점을 맞추어 explicit semantic prior가 classical methods가 실패하는 topological ambiguity를 어떻게 해소할 수 있는지 보여준다.","link":"https://openreview.net/forum?id=VNdLLU6SkE"},{"id":"wuJjHvzTwC","en":"Ion-Pot and Ion-Curr: Ion-Aware Benchmarks for Electrostatic Field and Observable Prediction in Complex Geometries","ko":"Ion-Pot와 Ion-Curr: 복잡한 기하에서의 정전장 및 관측값 예측을 위한 이온 인지 벤치마크","authors":"Jingqian Liu, Pinhao Gu, Aleksei Aksimentiev","abs":"

Electrostatic interactions and ionic transport are central to biomolecular function, but continuum solvers for these problems remain costly at scale. Recent learning-based surrogates have shown promise, yet most existing settings focus on fixed ionic conditions and pay limited attention to the role of mobile ions. We introduce two complementary benchmarks for ion-aware biomolecular surrogate modeling: Ion-Pot, for electrostatic potential prediction around biomolecules under varying ionic conditions, and Ion-Curr, for ionic current prediction in complex nanopore-based biomolecular systems. Together, these tasks span two related prediction regimes: high-dimensional field reconstruction and observable-level transport prediction. We further develop voxel-based convolutional models tailored to each setting and show that they provide accurate and efficient surrogates for both tasks. For Ion-Pot, explicit conditioning on ion concentration improves predictive performance over no-ion-input baselines, highlighting the importance of ion-aware learning for electrostatic field prediction. For Ion-Curr, a lightweight 39K-parameter architecture accurately predicts solver-derived current labels from complex nanopore geometries and analyte configurations, suggesting that surrogate complexity may depend strongly on the dimensionality and structure of the prediction target. These results establish Ion-Pot and Ion-Curr as a paired testbed for biomolecular electrostatics under varying ionic conditions and geometry-dependent ion-transport surrogate modeling. The implementations and instructions for downloading the Ion-Curr dataset are available at https://github.com/jingqian-liu/Ion-Pot_Ion-Curr/tree/main.

","absKo":"Electrostatic interaction과 ionic transport는 biomolecular function의 핵심이지만, 이 문제들을 위한 continuum solver는 규모가 커질수록 비용이 많이 든다. 최근 learning-based surrogate는 가능성을 보였지만, 대부분의 기존 설정은 고정된 ionic condition에 초점을 맞추고 mobile ion의 역할에는 상대적으로 덜 주목했다. 우리는 ion-aware biomolecular surrogate modeling을 위한 두 개의 상보적인 benchmark를 소개한다. Ion-Pot은 다양한 ionic condition 하에서 biomolecule 주변의 electrostatic potential prediction을 위한 것이고, Ion-Curr는 complex nanopore-based biomolecular system에서 ionic current prediction을 위한 것이다. 이 두 과제는 high-dimensional field reconstruction과 observable-level transport prediction이라는 서로 관련된 두 prediction regime를 포괄한다. 우리는 각 설정에 맞춘 voxel-based convolutional model을 추가로 개발했으며, 두 과제 모두에서 정확하고 효율적인 surrogate를 제공함을 보였다. Ion-Pot에서는 ion concentration에 대한 명시적 conditioning이 ion-input이 없는 baseline보다 예측 성능을 향상시켜, electrostatic field prediction에서 ion-aware learning의 중요성을 강조한다. Ion-Curr에서는 39K-parameter의 lightweight architecture가 복잡한 nanopore geometry와 analyte configuration으로부터 solver-derived current label을 정확하게 예측하여, surrogate의 복잡도가 prediction target의 차원성과 구조에 크게 좌우될 수 있음을 시사한다. 이러한 결과는 Ion-Pot과 Ion-Curr를 다양한 ionic condition과 geometry-dependent ion-transport surrogate modeling을 위한 paired testbed로 정립한다. 구현 및 Ion-Curr dataset 다운로드 지침은 https://github.com/jingqian-liu/Ion-Pot_Ion-Curr/tree/main 에서 제공된다.","link":"https://openreview.net/forum?id=wuJjHvzTwC"},{"id":"t8hgNJB1Dr","en":"Template Matching, Not Time Learning: A Diagnostic for Self-Supervised Light Curve Encoders","ko":"시간 학습이 아니라 템플릿 매칭: 자기지도 광도곡선 인코더를 위한 진단법","authors":"Emma Chickles","abs":"

We propose period regression on catalogued variable stars as a diagnostic for whether a self-supervised encoder has learned each source's period from its raw irregularly-sampled brightness ("absolute time"), or has merely template-matched class identity. Decomposing predicted-log-period variance into between- and within-class components separates "the encoder identifies the class and predicts the class-mean period" from "the encoder reads the period off the individual source." Applied to eight encoders — including pretrained Chronos-T5 (46M) and MOMENT (110M), a 4.4M cadence-as-channel BiGRU, and a 4.8M continuous-time SSL transformer — the diagnostic returns the same answer: 60–70% of period-regression $R^2$ is between-class, and within-class Spearman $\\rho$ stays at $0.20$–$0.26$ (95% CIs over multiple seeds and data splits). A nonlinear MLP probe does not change this: it raises overall $R^2$ ($0.66 \\to 0.77$ for the BiGRU) by sharpening class templates while $\\rho_{\\mathrm{within}}$ stays $\\le 0.31$, whereas a direct Lomb–Scargle extractor reaches $\\rho_{\\mathrm{within}} = 0.78$ on the same data — the within-source signal is recoverable, and the encoders are under-extracting it. As a control, retraining Chronos-T5 from scratch at matched scale collapses period $R^2$ from $0.685$ to $0.186$, isolating cross-domain pretraining as the dominant driver of its downstream win — but not of genuine time encoding. No encoder tested, including the published ASTROMER checkpoint, has learned absolute time on irregular ZTF photometry.

","absKo":"우리는 catalogued variable stars에서 period regression을 self-supervised encoder가 각 source의 period를 raw irregularly-sampled brightness(\"absolute time\")에서 학습했는지, 아니면 class identity만 template-matching했는지를 진단하는 지표로 제안한다. predicted-log-period variance를 between-class 성분과 within-class 성분으로 분해하면, \"encoder가 class를 식별하고 class-mean period를 예측한다\"는 경우와 \"encoder가 개별 source에서 period를 읽어낸다\"는 경우를 구분할 수 있다. pretrained Chronos-T5 (46M)와 MOMENT (110M), 4.4M cadence-as-channel BiGRU, 4.8M continuous-time SSL transformer를 포함한 8개의 encoder에 적용해도 진단 결과는 동일하다. period-regression $R^2$의 60–70%가 between-class이며, within-class Spearman $\\rho$는 여러 seed와 data split에 대한 95% CI에서 $0.20$–$0.26$에 머문다. nonlinear MLP probe도 이를 바꾸지 못한다. BiGRU에서 overall $R^2$를 $0.66 \\to 0.77$로 높이지만, class templates를 더 선명하게 만들 뿐 $\\rho_{\\mathrm{within}}$은 $\\le 0.31$에 머무른다. 반면 direct Lomb–Scargle extractor는 같은 데이터에서 $\\rho_{\\mathrm{within}} = 0.78$에 도달한다. 즉 within-source signal은 복구 가능하며, encoder들은 이를 충분히 추출하지 못하고 있다. 통제 실험으로, Chronos-T5를 같은 규모에서 scratch부터 재학습하면 period $R^2$가 $0.685$에서 $0.186$으로 붕괴하여, cross-domain pretraining이 downstream 성능 향상의 주된 요인임을 분리해 보여주지만, 진정한 time encoding의 요인은 아님을 시사한다. published ASTROMER checkpoint를 포함해 테스트한 어떤 encoder도 irregular ZTF photometry에서 absolute time을 학습하지 못했다.","link":"https://openreview.net/forum?id=t8hgNJB1Dr"},{"id":"x7tbgnfCbd","en":"Learning to Trigger: Reinforcement Learning at Large Hadron Collider","ko":"트리거를 학습하기: Large Hadron Collider에서의 Reinforcement Learning","authors":"Zixin Ding, Shaghayegh Emami, Giovanna Salvi, Cecilia Tosciri, Abhijith Gandrakota, Ngadiuba Jennifer, Nhan Tran, Christian Herwig, David Miller, Yuxin Chen","abs":"

High-throughput scientific facilities such as the Large Hadron Collider depend on real-time event filtering (\\textit{triggering}) under tight constraints on bandwidth, latency, and storage. In practice, trigger menus are largely static and hand-tuned and can become suboptimal as detector conditions, pileup, and background composition drift over time. We cast online threshold tuning as a sequential decision-making problem: a reinforcement learning agent ingests streaming summaries of recent rates and signal-sensitive features and updates trigger thresholds to maximize signal efficiency while tracking a target background rate within a tolerance band. We adapt Group-Filtered Policy Optimization (GFPO) to streaming control and introduce two variants (GFPO-F, GFPO-FR) that enforce background rate feasibility during training. On a benchmark that emulates realistic collider operation, we study two representative triggers: a total transverse energy ($H_{T}$) trigger sensitive to pileup variation, and an anomaly-detection (AD)

trigger based on reconstruction loss for rare or non-standard

signatures. On Monte Carlo streams, our agent increases the fraction of in-tolerance time intervals by 38\\% ($H_T$) and 30\\% (AD), with a cumulative gain of up to 5\\% in signal efficiency on those in-tolerance intervals. Transferring from simulation to \\emph{real} collision data (CMS Run 283408), the same agent, without fine-tuning, achieves a 40\\% ($H_T$) and 25\\% (AD) in-tolerance improvement over baselines, with an additional 1\\% signal-efficiency gain on both triggers. To our knowledge, this is the first demonstration of RL-based trigger control on real Large Hadron Collider collision data.

","absKo":"Large Hadron Collider와 같은 고처리량 과학 시설은 대역폭, 지연 시간, 저장 용량에 대한 엄격한 제약 하에서 실시간 이벤트 필터링(\\textit{triggering})에 의존한다. 실제로 trigger menu는 대체로 정적이고 수작업으로 조정되며, 시간이 지나면서 detector 조건, pileup, background 구성의 변화에 따라 비최적 상태가 될 수 있다. 우리는 online threshold tuning을 순차적 의사결정 문제로 정식화한다. 즉, reinforcement learning agent가 최근 rate와 signal-sensitive feature의 스트리밍 요약을 입력으로 받아 trigger threshold를 갱신하고, 허용 오차 범위 내에서 target background rate를 추적하면서 signal efficiency를 최대화한다. 우리는 Group-Filtered Policy Optimization (GFPO)을 streaming control에 맞게 적응시키고, 학습 중 background rate의 feasibility를 강제하는 두 변형(GFPO-F, GFPO-FR)을 도입한다. 실제 collider 운영을 모사한 benchmark에서, 우리는 두 가지 대표적 trigger를 연구한다. 하나는 pileup 변화에 민감한 total transverse energy ($H_{T}$) trigger이고, 다른 하나는 드문 또는 비표준 signature를 위한 reconstruction loss 기반 anomaly-detection (AD) \\ntrigger이다. Monte Carlo stream에서 우리 agent는 tolerance 내 시간 구간의 비율을 38\\% ($H_T$), 30\\% (AD) 증가시키며, 해당 tolerance 내 구간에서 signal efficiency의 누적 향상은 최대 5\\%에 이른다. simulation에서 \\emph{real} collision data (CMS Run 283408)로 전이했을 때, 동일한 agent는 fine-tuning 없이 baseline 대비 tolerance 내 개선을 40\\% ($H_T$), 25\\% (AD) 달성하며, 두 trigger 모두에서 추가로 1\\%의 signal-efficiency 향상을 보인다. 우리가 아는 한, 이는 실제 Large Hadron Collider collision data에서 RL 기반 trigger control을 시연한 첫 사례이다.","link":"https://openreview.net/forum?id=x7tbgnfCbd"},{"id":"5DSYaLjMMo","en":"Wavelet Variance Equipartition as a Threshold for World-Model Quality and Quantum Kernel TN-Simulability","ko":"세계 모델 품질과 quantum kernel TN 시뮬 가능성의 기준으로서의 Wavelet Variance Equipartition","authors":"Chon Fai Kam, Xavier Cadet, Miloud Bessaf, Frederic CADET","abs":"

While world models excel at learning compact representations of complex environments, they lack a principled, physics-grounded metric for assessing the structural fidelity of their latent spaces. We identify the wavelet scaling exponent $\\alpha$ as a critical diagnostic for this structural regularity, proposing that optimal representations should satisfy \\emph{variance equipartition} ($\\alpha \\approx 1/2$)---a condition mirroring the constant energy flux within Kolmogorov's inertial range. We formally establish $\\alpha = 1/2$ as a sharp transition boundary for the classical simulability of amplitude-encoded quantum kernels. Using tensor-network theory, we prove that latents with $\\alpha > 1/2$ reside in an area-law phase admitting efficient classical emulation, whereas $\\alpha < 1/2$ triggers a volume-law phase where the required Matrix Product State bond dimension $\\chi$ grows exponentially with the qubit count $n$. Empirical analysis of pre-trained VideoMAE latents reveals a fundamental dichotomy: while spatial token sequences approach the physical equipartition limit ($\\alpha \\approx 0.423$), the permutation-invariant feature channels exhibit unstructured disorder ($\\alpha \\approx -0.123$). This channel-wise complexity forces real-world latents deep into the volume-law phase, providing a data-driven necessary structural condition for tensor-network simulation hardness. Finally, to rigorously quantify the measurement overhead of high-dimensional quantum representations, we apply Weingarten calculus to derive the exact analytical variance of the scrambled transition probability under a 2-design ensemble. We prove that this variance scales strictly as $\\Var[X] = \\Theta(d^{-2})$. We confirm this exact scaling numerically with a log-log slope of $-1.881$ ($R^2 = 0.999$), explicitly identifying a formidable ``shot-noise wall'' demanding a critical measurement budget of $M = \\Omega(d^2)$ that constrains the scalability of quantum machine learning.

","absKo":"World model은 복잡한 환경의 compact representation을 학습하는 데 뛰어나지만, latent space의 구조적 fidelity를 평가하기 위한 원리 기반의 physics-grounded metric은 부족하다. 우리는 wavelet scaling exponent $\\alpha$를 이러한 구조적 regularity의 핵심 진단 지표로 식별하고, 최적의 representation은 \\emph{variance equipartition}($\\alpha \\approx 1/2$)을 만족해야 한다고 제안한다. 이는 Kolmogorov의 inertial range 내부에서 일정한 energy flux가 유지되는 조건을 반영한다. 우리는 $\\alpha = 1/2$가 amplitude-encoded quantum kernel의 고전적 시뮬레이션 가능성에 대한 sharp transition boundary임을 엄밀히 보인다. tensor-network theory를 사용해, $\\alpha > 1/2$인 latent는 효율적인 고전적 emulation이 가능한 area-law phase에 속하는 반면, $\\alpha < 1/2$는 필요한 Matrix Product State bond dimension $\\chi$가 qubit 수 $n$에 대해 지수적으로 증가하는 volume-law phase를 유발함을 증명한다. 사전 학습된 VideoMAE latent에 대한 경험적 분석은 근본적인 이분법을 드러낸다. spatial token sequence는 물리적 equipartition 한계($\\alpha \\approx 0.423$)에 접근하는 반면, permutation-invariant feature channel은 구조화되지 않은 disorder($\\alpha \\approx -0.123$)를 보인다. 이러한 channel-wise complexity는 실제 latent를 volume-law phase 깊숙이 몰아넣으며, tensor-network simulation의 난해함을 위한 데이터 기반의 필요 구조 조건을 제공한다. 마지막으로, 고차원 quantum representation의 measurement overhead를 엄밀히 정량화하기 위해 우리는 Weingarten calculus를 적용하여 2-design ensemble에서 scrambled transition probability의 정확한 analytical variance를 유도한다. 우리는 이 variance가 엄밀히 $\\Var[X] = \\Theta(d^{-2})$로 스케일링함을 증명한다. 또한 log-log slope가 $-1.881$($R^2 = 0.999$)인 수치적 검증을 통해 이 정확한 스케일링을 확인하고, $M = \\Omega(d^2)$의 critical measurement budget을 요구하는 강력한 ``shot-noise wall''을 명시적으로 식별함으로써 quantum machine learning의 확장성을 제약함을 보인다.","link":"https://openreview.net/forum?id=5DSYaLjMMo"},{"id":"gYbqOvL2Ld","en":"Bounds on Extrapolation across Phase Transitions with Generalized Regression","ko":"일반화된 회귀를 통한 phase transition 전반의 외삽 한계","authors":"Jeffrey Wei, Manolis Zampetakis, John Sous","abs":"

Phase transitions occur when a physical system undergoes a dramatic transformation as it crosses a transition. We study whether regression methods trained only in one phase can predict the physical properties that characterize the unobserved phase. Our test sets consist of the analytically tractable Heisenberg model and angle-resolved photoemission spectroscopy measurements of a venerable high transition temperature superconductor. We find that polynomial regression and kernel methods recover the unobserved phase structure, while standard neural networks fail. To bound the error, we adopt the generalized eigenvalue problem (GEVP) into an upper bound on the worst-case transfer coefficient, defined as the ratio of mean squared error (MSE) loss on the extrapolation region to the MSE loss on the training region. Empirically, we find that our GEVP bound stays approximately constant with respect to the empirical transfer coefficient across training-set sizes and distance between the training and extrapolation region for both datasets, providing a principled framework for studying extrapolation feasibility.

","absKo":"Phase transition은 물리 시스템이 transition을 통과할 때 겪는 극적인 변환이다. 우리는 한 phase에서만 학습된 regression method가 관측되지 않은 phase를 특징짓는 물리적 성질을 예측할 수 있는지 연구한다. 테스트셋은 해석적으로 다루기 쉬운 Heisenberg model과, 오랜 역사를 가진 high transition temperature superconductor의 angle-resolved photoemission spectroscopy 측정으로 구성된다. 우리는 polynomial regression과 kernel method가 관측되지 않은 phase 구조를 복원하는 반면, 표준 neural network는 실패함을 발견한다. 오차를 상한하기 위해 generalized eigenvalue problem(GEVP)을 worst-case transfer coefficient의 upper bound로 도입한다. 여기서 transfer coefficient는 extrapolation region의 mean squared error(MSE) loss를 training region의 MSE loss로 나눈 비율로 정의된다. 경험적으로, 우리는 GEVP bound가 두 데이터셋 모두에서 training-set size와 training region과 extrapolation region 사이의 거리 전반에 걸쳐 empirical transfer coefficient에 대해 거의 일정하게 유지됨을 확인했으며, 이는 extrapolation 가능성을 연구하기 위한 원리적 framework를 제공한다.","link":"https://openreview.net/forum?id=gYbqOvL2Ld"},{"id":"Ph1Jj4A049","en":"LieSolver: PDE-Constrained Learning for IBVPs via Lie Symmetries","ko":"LieSolver: Lie 대칭을 통한 IBVP용 PDE 제약 학습","authors":"René Pascal Klausen, Ivan Timofeev, Jonas Naujoks, Johannes Frank, Thomas Wiegand, Sebastian Lapuschkin, Wojciech Samek","abs":"

Initial-boundary value problems (IBVPs) provide the essential framework for modelling a wide range of phenomena in physics and engineering. We introduce a novel method for efficiently solving IBVPs using Lie symmetries to enforce the associated partial differential equation (PDE) exactly by construction. By leveraging symmetry transformations, our model embeds the underlying physical laws and learns the solution solely from initial and boundary data. Consequently, the boundary loss directly quantifies domain-wide error, enabling rigorous error estimation for well-posed IBVPs. We implement *LieSolver* and demonstrate its application to linear homogeneous PDEs, showing that it outperforms physics-informed neural networks (PINNs) in both speed and accuracy while yielding compact models. Overall, our approach significantly enhances the efficiency and reliability of predictions for PDE-constrained problems.

","absKo":"Initial-boundary value problem(IBVP)은 물리와 공학의 광범위한 현상을 모델링하는 데 필수적인 framework를 제공한다. 우리는 연관된 partial differential equation(PDE)을 구성상 정확하게 enforce하기 위해 Lie symmetry를 사용하는 IBVP의 효율적 해법을 위한 새로운 방법을 소개한다. symmetry transformation을 활용하여, 우리의 model은 underlying physical law를 내재화하고 initial 및 boundary data로부터만 solution을 학습한다. 그 결과 boundary loss는 domain 전체의 error를 직접 정량화하여 well-posed IBVP에 대한 엄밀한 error estimation을 가능하게 한다. 우리는 *LieSolver*를 구현하고 선형 homogeneous PDE에의 적용을 보이며, physics-informed neural network(PINN)보다 속도와 정확도 모두에서 우수하면서도 더 compact한 model을 산출함을 보인다. 전반적으로 우리의 접근은 PDE-constrained problem에 대한 예측의 효율성과 신뢰성을 크게 향상시킨다.","link":"https://openreview.net/forum?id=Ph1Jj4A049"},{"id":"1yRpUPopFM","en":"NIGO: Lyapunov-Stable Continuous-Time Neural Operators for Long-Horizon PDE Dynamics","ko":"NIGO: 장기 지평 PDE 동역학을 위한 Lyapunov-안정 continuous-time neural operators","authors":"Ashish Musale, ajay waghumbare, VINDUJA T, Upasna Singh","abs":"

Neural operators have become a promising framework for learning solution operators of partial differential equations, yet their long-horizon use remains limited by the stability of autoregressive rollout. In many spatiotemporal settings, learned time-advancement maps must be composed repeatedly, causing small local errors to accumulate into trajectory drift, spectral artifacts, or non-physical energy growth. We propose **NIGO**, a continuous-time neural operator that replaces unconstrained discrete rollout maps with a dissipative infinitesimal generator in latent space. Given an initial PDE state, NIGO encodes the field into a compact latent representation and evolves it analytically through a matrix exponential. The generator is parameterized as a skew-symmetric transport component plus a negative-semidefinite dissipative component, which enforces non-increasing latent energy by construction. A physics-inference network predicts the generator coefficients from the initial state and conditioning variables, while temporal refinement and Lipschitz-controlled decoding map stable latent trajectories back to physical fields. We evaluate NIGO on Burgers, cylinder flow, Kuramoto--Sivashinsky, shallow-water, and incompressible Navier--Stokes systems, covering decaying, periodic, and chaotic regimes. Across extended rollouts, NIGO avoids the blowup and freezing modes observed in standard neural-operator and physics-informed baselines. Ablations show that the dissipative generator is necessary for stable long-horizon evolution, and mechanistic analyses reveal learned structure consistent with viscous scaling, operator locality, and advection--dissipation balance. These results indicate that hard-constrained continuous-time latent dynamics provide a principled foundation for stable neural operators for PDE forecasting.

","absKo":"Neural operator는 partial differential equation의 solution operator를 학습하는 유망한 프레임워크가 되었지만, 장기 사용은 autoregressive rollout의 안정성 한계로 제약되어 왔습니다. 많은 spatiotemporal setting에서 학습된 time-advancement map을 반복해서 합성해야 하므로, 작은 local error가 누적되어 trajectory drift, spectral artifact, 또는 비물리적인 energy growth로 이어집니다. 우리는 제약 없는 discrete rollout map을 latent space의 dissipative infinitesimal generator로 대체하는 continuous-time neural operator인 **NIGO**를 제안합니다. 초기 PDE state가 주어지면 NIGO는 field를 compact latent representation으로 인코딩하고 matrix exponential을 통해 이를 해석적으로 진화시킵니다. generator는 skew-symmetric transport component와 negative-semidefinite dissipative component로 매개변수화되며, 이는 설계상 non-increasing latent energy를 보장합니다. physics-inference network는 초기 state와 conditioning variable로부터 generator coefficient를 예측하고, temporal refinement와 Lipschitz-controlled decoding은 안정적인 latent trajectory를 물리적 field로 되돌립니다. 우리는 Burgers, cylinder flow, Kuramoto--Sivashinsky, shallow-water, incompressible Navier--Stokes 시스템에서 NIGO를 평가했으며, decaying, periodic, chaotic regime를 포괄합니다. 긴 rollout 전반에 걸쳐 NIGO는 표준 neural-operator 및 physics-informed baseline에서 관찰되는 blowup과 freezing mode를 피합니다. Ablation은 안정적인 장기 진화를 위해 dissipative generator가 필수적임을 보여주며, 기계적 분석은 viscous scaling, operator locality, advection--dissipation balance와 일치하는 학습된 구조를 드러냅니다. 이러한 결과는 강한 제약을 둔 continuous-time latent dynamics가 PDE forecasting을 위한 안정적인 neural operator의 원리적 기반을 제공함을 시사합니다.","link":"https://openreview.net/forum?id=1yRpUPopFM"},{"id":"F0DB2JDpb3","en":"Generalization vs. Memorization in Partial Differential Equation Emulators: Or, Training Dynamics of Cross-Time and Cross-Class Gradient Alignment","ko":"부분미분방정식 emulator에서 일반화 대 기억: 또는 시간 간 및 클래스 간 gradient alignment의 학습 동역학","authors":"James Amarel, Nicolas Hengartner, Robyn Miller, Kamaljeet Singh, Siddharth Mansingh, Arvind Mohan, Benjamin Migliori, Emily Taylor, Alexei Skurikhin, Earl Lawrence, Gerd J. Kunde","abs":"

Neural emulators of solution operators to partial differential equations capable of reliable generalization stand to accelerate scientific discovery by extrapolating beyond training regimes and performing over varied initial condition classes. Assessing model generalization capabilities requires distinguishing solution operator-consistent behavior from memorization of feature derived correlations, which can be achieved by using influence functions to characterize how information is shared across examples with different temporal indices and/or of different initial conditions. We compute the hat-matrix metric-weighted gradient alignment by inverting the relevant Hessian restricted to the mini-batch gradient-spanned subspace, enabling exact measurement of cross-time and cross-class coherence during training. This diagnostic is architecture-agnostic and can be efficiently computed. Our results indicate that gradient alignment between training examples is governed by their feature-space similarity. We find that gradient coherence across initial condition classes is suppressed due to feature-space dissimilarity. Moreover, we quantify the temporal decay of gradient alignment induced by standard one-step autoregressive training, revealing that gradient coherence is not extended in time, which limits long-horizon extrapolation performance. Cross-time influence exhibits a negative correlation with rollout error across horizons, indicating that the measured response geometry contains information predictive of inference-time degradation.

","absKo":"신뢰할 수 있는 일반화가 가능한 partial differential equation의 solution operator용 neural emulator는 training regime를 넘어 extrapolation하고 다양한 initial condition class에서 작동함으로써 과학적 발견을 가속할 수 있다. 모델의 generalization capability를 평가하려면 solution operator와 일관된 행동과 feature에서 파생된 correlation을 memorization하는 행동을 구분해야 하며, 이는 서로 다른 temporal index와/또는 서로 다른 initial condition을 가진 example들 사이에서 정보가 어떻게 공유되는지 특성화하기 위해 influence function을 사용함으로써 달성할 수 있다. 우리는 mini-batch gradient가 span하는 subspace로 제한된 관련 Hessian을 역행하여 hat-matrix metric-weighted gradient alignment를 계산하며, 이를 통해 training 동안 cross-time 및 cross-class coherence를 정확하게 측정할 수 있다. 이 진단법은 architecture에 독립적이며 효율적으로 계산할 수 있다. 우리의 결과는 training example 간 gradient alignment가 feature-space similarity에 의해 좌우됨을 보여준다. 우리는 initial condition class 전반의 gradient coherence가 feature-space dissimilarity 때문에 억제됨을 발견한다. 또한 표준 one-step autoregressive training이 유도하는 gradient alignment의 temporal decay를 정량화하여, gradient coherence가 시간적으로 확장되지 않으며 이것이 long-horizon extrapolation 성능을 제한함을 보인다. Cross-time influence는 rollout error와 horizon 전반에 걸쳐 음의 상관관계를 보이며, 이는 측정된 response geometry가 inference-time degradation을 예측하는 정보를 담고 있음을 시사한다.","link":"https://openreview.net/forum?id=F0DB2JDpb3"},{"id":"lmj83YIyiz","en":"Fluid-Gen-Zero: Grounding Pretrained Video Generators in Physics without Training","ko":"Fluid-Gen-Zero: 학습 없이 물리 법칙을 pretrained video generator에 grounding하기","authors":"Hong Huang, Yuqiu Liu, Chenyu You, Daniel F Martin, Chuhang Zou, Wuyang Chen","abs":"

We present **Fluid-Gen-Zero**, a training-free framework for physics-aware fluid video generation that elegantly decouples physical reasoning from appearance synthesis. Our key insight is to delegate motion dynamics to a physics simulator while preserving the appearance modeling capacity of pretrained video generators. We bridge these two domains through a two-level agentic workflow: *generation-time planning*, where a rule-grounded VLM agent parses intent, observes simulation evidence, and organizes generation into physically feasible clips, and *latent-space guidance*, which injects simulation-derived signals into the denoising process via region-aware latent wrapping. This plug-and-play design is compatible with the latest video models. We further introduce a benchmark for evaluating fluid-object interactions. Across Tora (CogVideoX-based) and VACE (Wan-based) backbones, Fluid-Gen-Zero reduces object trajectory error by 30.0\\% to 48.0\\%, fluid trajectory error by up to 21.1\\%, and receives 62.86\\% to 87.14\\% overall human preference over baselines. Code and data will be released upon acceptance.

","absKo":"We present **Fluid-Gen-Zero**, physics-aware fluid video generation을 위한 training-free framework로, physical reasoning과 appearance synthesis를 우아하게 분리한다. 우리의 핵심 통찰은 motion dynamics는 physics simulator에 맡기고, appearance modeling 능력은 pretrained video generator가 유지하도록 하는 것이다. 우리는 이 두 영역을 두 단계의 agentic workflow로 연결한다. *generation-time planning*에서는 rule-grounded VLM agent가 intent를 해석하고 simulation evidence를 관찰하며 generation을 물리적으로 가능한 clip으로 조직하고, *latent-space guidance*에서는 region-aware latent wrapping을 통해 simulation에서 유도된 signal을 denoising process에 주입한다. 이 plug-and-play 설계는 최신 video model과 호환된다. 또한 fluid-object interaction을 평가하기 위한 benchmark도 새로 제안한다. Tora(CogVideoX-based)와 VACE(Wan-based) backbone 전반에서 Fluid-Gen-Zero는 object trajectory error를 30.0\\%에서 48.0\\%까지 줄이고, fluid trajectory error는 최대 21.1\\% 감소시키며, baseline 대비 전체 human preference에서 62.86\\%에서 87.14\\%를 얻는다. Code와 data는 채택 시 공개할 예정이다.","link":"https://openreview.net/forum?id=lmj83YIyiz"},{"id":"Itf2hU4EPG","en":"PINNfluence: Interpreting PINNs through Influence Functions","ko":"PINNfluence: influence function을 통한 PINN 해석","authors":"Aleksander Krasowski, Jonas Naujoks, Moritz Weckbecker, Galip Ümit Yolcu, Thomas Wiegand, Sebastian Lapuschkin, Wojciech Samek, René Pascal Klausen","abs":"

Physics-informed neural networks (PINNs) have emerged as a powerful deep learning approach for solving partial differential equations (PDEs) in the physical sciences, yet their behavior remains largely opaque and is typically understood through failure mode analyses rather than explicit interpretability.

To address this issue, we introduce PINNfluence, a training data attribution framework for interpreting PINNs based on influence functions.

By extending influence functions to composite physics-informed training objectives, we enable fine-grained attribution between predictions, loss components, and training data points.

Through benchmark experiments across various PDEs, we demonstrate that influence patterns provide granular diagnostics that distinguish structural characteristics across well-trained and poorly-trained PINNs.

PINNfluence thus opens a new avenue for understanding and improving the reliability of PINNs through the lens of their data.

","absKo":"Physics-informed neural networks (PINNs)는 물리과학에서 partial differential equations (PDEs)를 푸는 강력한 deep learning 접근법으로 부상했지만, 그 동작은 여전히 대부분 불투명하며 일반적으로 explicit interpretability보다는 failure mode 분석을 통해 이해되어 왔다.\n이 문제를 해결하기 위해 우리는 influence functions에 기반한 PINN 해석용 training data attribution framework인 PINNfluence를 제안한다.\ninfluence functions를 composite physics-informed training objective로 확장함으로써, 예측, loss component, training data point 사이의 세밀한 attribution을 가능하게 한다.\n다양한 PDE에 대한 benchmark 실험을 통해, 우리는 influence pattern이 잘 학습된 PINN과 덜 학습된 PINN을 구분하는 세밀한 진단을 제공함을 보인다.\n따라서 PINNfluence는 data의 관점에서 PINNs의 신뢰성을 이해하고 개선하기 위한 새로운 길을 연다.","link":"https://openreview.net/forum?id=Itf2hU4EPG"},{"id":"RIojKtUCrJ","en":"Learning Interface Breakup: A Geometry-Conditioned Latent Surrogate for Spray Formation","ko":"인터페이스 분리 학습: spray 형성의 geometry-conditioned latent surrogate","authors":"Julius H Ramlau, Friedrich Hastedt, Tolga Birdal, Antonio Del rio chanona, Nausheen Basha, Omar K Matar","abs":"

Designing spray nozzles requires predicting how geometry shapes transient two-phase breakup, but high-fidelity volume-of-fluid (VOF) simulations with adaptive mesh refinement (AMR) are too expensive for iterative design exploration. Standard surrogate models are also challenged by this setting because both the liquid--gas interface and the underlying adaptive discretization evolve across time and geometries. We introduce a geometry-conditioned latent surrogate trained on 797 two-phase nozzle simulations that addresses this by encoding the AMR cell-density field, rather than the full multi-channel flow state, as a compact proxy for where the solver concentrates resolution. From this representation, the model reconstructs transient density evolution and nozzle geometry, and a lightweight second stage recovers the remaining flow variables. On held-out simulations, the method accurately captures key interface dynamics while reducing inference time to 0.045 seconds per trajectory, corresponding to a speed-up of more than $6\\times10^4$ relative to Basilisk CFD. These results suggest that AMR refinement structure can serve as a compact and learnable representation for geometry-conditioned surrogate modeling of transient two-phase flows.

","absKo":"스프레이 노즐 설계는 geometry가 transient two-phase breakup을 어떻게 형성하는지 예측해야 하지만, adaptive mesh refinement (AMR)를 포함한 고충실도 volume-of-fluid (VOF) simulation은 반복적인 설계 탐색에는 너무 비용이 크다. 표준 surrogate model도 이 환경에서는 어려움을 겪는데, liquid--gas interface와 underlying adaptive discretization이 시간과 geometry에 따라 함께 진화하기 때문이다. 우리는 797개의 two-phase nozzle simulation으로 학습된 geometry-conditioned latent surrogate를 제안하며, 전체 multi-channel flow state가 아니라 AMR cell-density field를 solver가 해상도를 집중하는 위치를 나타내는 압축된 proxy로 인코딩함으로써 이 문제를 해결한다. 이 representation으로부터 모델은 transient density evolution과 nozzle geometry를 재구성하고, 경량의 두 번째 stage가 나머지 flow variable을 복원한다. 보류된 simulation에서 이 방법은 핵심 interface dynamics를 정확하게 포착하면서 추론 시간을 trajectory당 0.045초로 줄여, Basilisk CFD 대비 6\\times10^4$ 이상 빠른 속도를 달성한다. 이러한 결과는 AMR refinement structure가 transient two-phase flow의 geometry-conditioned surrogate modeling을 위한 압축 가능하고 학습 가능한 representation으로 기능할 수 있음을 시사한다.","link":"https://openreview.net/forum?id=RIojKtUCrJ"},{"id":"FYLB0utjqH","en":"Probabilistic Data-Driven Modelling of Astrophysical Transients: The Neural Process Family for Ultrafast and Class-Agnostic Light Curve Reconstruction","ko":"천체 transient의 확률적 데이터 기반 모델링: 초고속 및 class-agnostic light curve 재구성을 위한 Neural Process 계열","authors":"Siddharth Chaini, federica bianco, Ashish Mahabal","abs":"

Astronomical light curves --- the sparse, irregular, multi-wavelength time series produced by photometric surveys --- are notoriously hard to model, with Gaussian Processes (GPs) the current standard despite their poor scaling, *a priori* kernel specification, and inability to share information across objects.

We introduce the use of Attentive Neural Processes (ANPs) as a fast, data-driven, morphologically agnostic interpolator: a single meta-learned model handles 14 explosive transient classes without specifying the class at inference.

Evaluated on realistic LSST cadences, the ANP outperforms seven GP- and neural-network-based benchmarks on all regression, feature-recovery, and probabilistic-calibration metrics, while running in $\\sim 10^{-6}$ s per light curve --- four to five orders of magnitude faster than benchmark models --- fast enough for on-line use on the LSST alert stream of $\\sim 10^{7}$ events/night.

We demonstrate how the ANP avoids neural networks' overconfidence and GPs' underconfidence, delivering sharp, calibrated uncertainties suitable for downstream science.

This work establishes the neural process family as a scalable, probabilistic foundation for real-time transient science in the LSST era.

","absKo":"천문학적 light curve --- 측광 survey에서 생성되는 희소하고 불규칙하며 다중 파장의 time series --- 는 모델링하기가 매우 어렵고, Gaussian Processes (GPs)가 현재 표준이지만 확장성이 좋지 않고, *a priori* kernel specification이 필요하며, 객체 간 정보를 공유할 수 없다는 한계가 있다.\n우리는 Attentive Neural Processes (ANPs)를 빠르고 데이터 기반이며 형태에 구애받지 않는 interpolator로 사용하는 방법을 제안한다. 하나의 meta-learned model이 추론 시 class를 지정하지 않고도 14개의 explosive transient class를 처리한다.\n현실적인 LSST cadence에서 평가한 결과, ANP는 모든 regression, feature-recovery, probabilistic calibration metric에서 7개의 GP- 및 neural-network-based benchmark를 능가했으며, light curve당 약 $\\sim 10^{-6}$ s로 동작하여 benchmark model보다 4~5자릿수 더 빠르다. 이는 LSST alert stream의 약 $\\sim 10^{7}$ events/night에 대해 online 사용이 가능할 만큼 빠르다.\n우리는 ANP가 neural network의 과도한 confidence와 GP의 과소한 confidence를 어떻게 피하는지 보여주며, downstream science에 적합한 sharp하고 calibrated uncertainty를 제공함을 입증한다.\n이 연구는 LSST 시대의 real-time transient science를 위한 확장 가능하고 확률적인 foundation으로서 neural process family를 확립한다.","link":"https://openreview.net/forum?id=FYLB0utjqH"},{"id":"6VOQxRx5mJ","en":"Crystalite: A Lightweight Transformer for Efficient Crystal Modeling","ko":"Crystalite: 효율적인 결정 모델링을 위한 경량 Transformer","authors":"Tin Hadži Veljković, Joshua Rosenthal, Ivor Lončarić, Jan-Willem van de Meent","abs":"

Generative models for crystalline materials often rely on equivariant graph neural networks, which capture geometric structure well but are costly to train and slow to sample. We present Crystalite, a lightweight diffusion Transformer for crystal modeling built around two simple inductive biases. The first is Chemically Structured Atom Tokens, a compact atom representation that replaces high-dimensional one-hot encodings and is better suited to continuous diffusion. The second is the Geometry Enhancement Module (GEM), which injects periodic minimum-image pair geometry directly into attention through additive geometric biases. Together, these components preserve the simplicity and efficiency of a standard Transformer while making it better matched to the structure of crystalline materials. Crystalite achieves state-of-the-art results on crystal structure prediction benchmarks, and de novo generation, attaining the best Stable, Unique, and Novel (SUN) discovery score among the evaluated baselines while sampling substantially faster than geometry-heavy alternatives.

","absKo":"결정성 물질을 위한 generative model은 종종 equivariant graph neural network에 의존하는데, 이는 기하 구조를 잘 포착하지만 학습 비용이 높고 샘플링이 느리다. 우리는 두 가지 간단한 inductive bias를 중심으로 구성된 경량 diffusion Transformer인 Crystalite를 제시한다. 첫 번째는 Chemically Structured Atom Tokens로, 고차원 one-hot encoding을 대체하는 간결한 원자 표현이며 연속 diffusion에 더 적합하다. 두 번째는 Geometry Enhancement Module(GEM)로, 주기적 minimum-image pair geometry를 additive geometric bias를 통해 attention에 직접 주입한다. 이 두 구성요소는 표준 Transformer의 단순성과 효율성을 유지하면서도 결정성 물질의 구조에 더 잘 맞도록 만든다. Crystalite는 crystal structure prediction benchmark와 de novo generation에서 state-of-the-art 결과를 달성했으며, 평가한 baseline들 중 가장 높은 Stable, Unique, and Novel(SUN) discovery score를 기록하는 동시에 기하 구조가 무거운 대안들보다 훨씬 빠르게 샘플링한다.","link":"https://openreview.net/forum?id=6VOQxRx5mJ"},{"id":"iuK0FfmcMG","en":"Stargazer: A Scalable Model-fitting Benchmark Environment for AI Agents under Astrophysical Constraints","ko":"Stargazer: 천체물리학적 제약 하에서 AI Agents를 위한 확장 가능한 모델 피팅 벤치마크 환경","authors":"Xinge Liu, Terry Jingchen Zhang, Bernhard Schölkopf, Zhijing Jin, Kristen Menou","abs":"

The rise of autonomous AI agents suggests that dynamic benchmark environments with built-in feedback on scientifically grounded tasks are needed to evaluate the capabilities of these agents in research work. We introduce Stargazer, a scalable environment for evaluating AI agents on dynamic, iterative physics-grounded model-fitting tasks using inference on radial-velocity (RV) time series data. Stargazer comprises 120 tasks across three difficulty tiers, including 20 real archival cases, covering diverse scenarios ranging from high-SNR single-planet systems to complex multi-planetary configurations requiring involved low-SNR analysis. Our evaluation of eight frontier agents reveals a gap between numerical optimization and adherence to physical constraints: although agents often achieve a good statistical fit, they frequently fail to recover correct physical system parameters, a limitation that persists even when agents are equipped with vanilla skills. Furthermore, increasing test-time compute yields only marginal gains, with excessive token usage often reflecting recursive failure loops rather than meaningful exploration. Stargazer presents an opportunity to train, evaluate, scaffold, and scale strategies on a model-fitting problem of practical research relevance today. Our methodology to design a simulation-driven environment for AI agents presumably generalizes to many other model-fitting problems across scientific domains.

","absKo":"자율 AI agent의 부상은, 이들 agent의 연구 작업 능력을 평가하기 위해 과학적으로 근거가 있는 과제에 대한 내장 피드백을 갖춘 동적 benchmark 환경이 필요함을 시사합니다. 우리는 radial-velocity (RV) time series data에 대한 inference를 사용해 동적으로 반복되는 물리 기반 model-fitting task에서 AI agent를 평가하는 확장 가능한 환경 Stargazer를 소개합니다. Stargazer는 3개 difficulty tier에 걸친 120개 task로 구성되며, 그중 20개는 실제 archival case입니다. 이들은 high-SNR single-planet system부터 복잡한 low-SNR 분석이 필요한 다중 행성 구성까지 다양한 시나리오를 포괄합니다. 8개의 frontier agent에 대한 우리의 평가는 수치 최적화와 물리 제약 준수 사이의 격차를 드러냅니다. agent는 종종 좋은 통계적 fit을 얻지만, 올바른 물리 시스템 parameter를 회복하는 데는 자주 실패하며, 이 한계는 vanilla skill를 장착해도 계속됩니다. 또한 test-time compute를 늘려도 개선은 제한적이며, 과도한 token 사용은 의미 있는 탐색이라기보다 재귀적 실패 루프를 반영하는 경우가 많습니다. Stargazer는 오늘날 실제 연구 관련성을 지닌 model-fitting 문제에서 전략을 학습, 평가, scaffolding, 확장할 기회를 제공합니다. AI agent를 위한 simulation-driven environment를 설계하는 우리의 방법론은 과학 분야 전반의 많은 다른 model-fitting 문제에도 아마 일반화될 것입니다.","link":"https://openreview.net/forum?id=iuK0FfmcMG"},{"id":"Kbr7JLu7Wh","en":"LM-PIELM: A Lagrange-Multiplier Framework for Hard Constraint Enforcement in Physics-Informed Extreme Learning Machines for PDE-Governed Forward and Inverse Problems","ko":"LM-PIELM: PDE 지배 순방향 및 역문제에서 물리 정보 기반 Extreme Learning Machines의 하드 제약 강제를 위한 라그랑주 승수 프레임워크","authors":"Eshan Kulkarni, Rishi Mishra, Balaji Srinivasan","abs":"

Physics-Informed Extreme Learning Machines (PIELMs) solve PDEs as linear least-squares problems by fixing the feature layer and learning only the output weights, but boundary conditions are usually imposed through soft penalties with problem-dependent weights. We propose LM-PIELM, a Lagrange-multiplier formulation that imposes boundary conditions as algebraic equality constraints using the linear-in-parameters structure of PIELMs. The resulting constrained least-squares problem leads to a linear saddle-point system with no boundary penalty tuning, no iterative dual updates, and no problem-specific analytical constrained ansatz. An equivalent augmented-system implementation avoids explicit normal-equation Gram matrices and improves robustness in ill-conditioned regimes. Since the constraint block uses only boundary operator evaluations, the same formulation applies to mixed boundary conditions, irregular geometries,

and inverse problems including parameter identification and source recovery from noisy data. Across forward and inverse benchmarks, the augmented-system form preserves small boundary residuals, improves solution accuracy in ill-conditioned regimes, and keeps boundary admissibility essentially decoupled from sensor count, noise level, and reconstruction quality.

","absKo":"Physics-Informed Extreme Learning Machines (PIELMs)는 feature layer를 고정하고 output weight만 학습함으로써 PDE를 linear least-squares 문제로 풀이하지만, boundary condition은 보통 문제에 의존하는 가중치를 갖는 soft penalty로 부과된다. 우리는 PIELM의 linear-in-parameters 구조를 사용해 boundary condition을 algebraic equality constraint로 부과하는 Lagrange multiplier formulation인 LM-PIELM을 제안한다. 그 결과 constrained least-squares 문제는 boundary penalty tuning이 필요 없고, iterative dual update도 없으며, 문제별 analytical constrained ansatz도 필요 없는 linear saddle-point system으로 귀결된다. 동등한 augmented-system 구현은 explicit normal-equation Gram matrix를 피하고 ill-conditioned regime에서의 robustness를 향상시킨다. constraint block은 boundary operator evaluation만 사용하므로, 동일한 formulation을 mixed boundary condition, irregular geometry,\n그리고 noisy data로부터의 parameter identification 및 source recovery를 포함한 inverse problem에도 적용할 수 있다. forward 및 inverse benchmark 전반에서 augmented-system form은 작은 boundary residual을 유지하고, ill-conditioned regime에서 solution accuracy를 개선하며, boundary admissibility를 sensor 수, noise level, reconstruction quality와 본질적으로 분리된 상태로 유지한다.","link":"https://openreview.net/forum?id=Kbr7JLu7Wh"},{"id":"3OMwjNsx90","en":"Towards a Physics Foundation Model","ko":"물리 파운데이션 모델을 향하여","authors":"Florian Wiesner, Zoë J. Gray, Matthias Wessling, Stephen Baek","abs":"

Foundation models have revolutionized natural language processing through a "train once, deploy anywhere" paradigm, where a single pre-trained model adapts to countless downstream tasks without retraining. Access to a Physics Foundation Model (PFM) would be transformative---democratizing access to high-fidelity simulations, accelerating scientific discovery, and eliminating the need for specialized solver development. Yet current physics-aware machine learning approaches remain fundamentally limited to single, narrow domains and require retraining for each new system. We present the General Physics Transformer (GPhyT), trained on 1.8 TB of diverse simulation data, that demonstrates foundation model capabilities are achievable for physics. Our key insight is that transformers can learn to infer governing dynamics from context, enabling a single model to simulate fluid-solid interactions, shock waves, thermal convection, and multi-phase dynamics without being told the underlying equations. GPhyT achieves three critical breakthroughs: (1) superior performance across multiple physics domains, outperforming SOTA multi-physics architectures by more than 7x, (2) plausible zero-shot generalization to entirely unseen physical systems through in-context learning, and (3) more stable long-term predictions through long-horizon rollouts. By establishing that a single model can learn generalizable physical principles from data alone, this work opens the path toward a universal PFM that could transform computational science and engineering.

","absKo":"Foundation models는 하나의 사전학습된 모델이 재학습 없이 무수한 downstream task에 적응하는 \"train once, deploy anywhere\" 패러다임을 통해 natural language processing을 혁신해 왔다. Physics Foundation Model(PFM)에 대한 접근은 고충실도 시뮬레이션에 대한 접근을 민주화하고, 과학적 발견을 가속하며, 특수한 solver 개발의 필요성을 없애는 등 변혁적일 것이다. 그러나 현재의 physics-aware machine learning 접근법은 본질적으로 단일하고 좁은 도메인에만 제한되어 있으며, 새로운 시스템마다 재학습이 필요하다. 우리는 1.8 TB의 다양한 simulation data로 학습된 General Physics Transformer(GPhyT)를 제시하며, foundation model의 역량이 물리 분야에서도 달성 가능함을 보인다. 우리의 핵심 통찰은 transformer가 context로부터 governing dynamics를 추론하도록 학습할 수 있어, 기본 방정식을 명시적으로 알려주지 않아도 하나의 모델이 fluid-solid interactions, shock waves, thermal convection, multi-phase dynamics를 시뮬레이션할 수 있다는 점이다. GPhyT는 세 가지 중요한 돌파구를 달성한다: (1) 여러 physics domain 전반에서 SOTA multi-physics architecture보다 7배 이상 우수한 성능, (2) in-context learning을 통한 완전히 보지 못한 physical system으로의 그럴듯한 zero-shot generalization, (3) long-horizon rollout을 통한 더 안정적인 장기 예측. 데이터만으로 하나의 모델이 일반화 가능한 physical principle을 학습할 수 있음을 입증함으로써, 이 연구는 computational science와 engineering을 바꿀 수 있는 universal PFM으로 가는 길을 연다.","link":"https://openreview.net/forum?id=3OMwjNsx90"},{"id":"BIdT3cqx0E","en":"Conservation-Constrained Fourier Neural Operators for Physically Consistent PDE Simulation","ko":"물리적으로 일관된 PDE 시뮬레이션을 위한 보존 제약 Fourier Neural Operator","authors":"Kabir Jain, Aayam Bansal, Ishaan Gangwani, Ziming Qiu, Raghav Agarwal","abs":"

Neural operators have enabled rapid surrogate modeling of partial differential equations but systematically violate conservation laws, limiting their physical fidelity and long-horizon stability. We introduce CC-FNO, a modified Fourier Neural Operator that enforces mass conservation via a differentiable Leray projection applied to the output velocity field, projecting it onto the divergence-free subspace in Fourier space at negligible computational cost ($<2\\%$ overhead). On 2D incompressible Navier--Stokes at $Re \\in \\{100, 500, 1000, 2000\\}$, CC-FNO reduces divergence by five orders of magnitude ($\\sim 10^{-5}$ vs.\\ $\\sim 3.0$ for FNO, $\\sim 10^{-2}$ for PINO) while matching baseline accuracy across seven model comparisons including FNO, PINO, Stream-FNO, U-Net, and FNO with pushforward training. We additionally propose CC-FNO-E, which stabilizes autoregressive rollouts through a per-step energy bound that acts as a stability regularizer, preventing catastrophic amplitude amplification. This extends stable rollout horizons by $17$--$52\\%$ over FNO across all regimes, though as a stability regularizer rather than a physical conservation law, it systematically underestimates energy growth by $67$--$84\\%$ at the final rollout step in forced flow. A systematic penalty weight sweep confirms that soft constraints cannot match hard projection regardless of tuning, and an ablation of forcing-aware energy bounds reveals that simple one-sided capping outperforms physically motivated alternatives. Our results establish a clear constraint hierarchy (none $<$ soft $<$ hard $<$ hard+energy stabilizer) and demonstrate that hard conservation constraints provide ``free'' physical consistency without sacrificing accuracy.

","absKo":"Neural operator는 partial differential equations의 빠른 surrogate modeling을 가능하게 했지만, conservation law를 체계적으로 위반하여 물리적 충실도와 장기 안정성을 제한한다. 우리는 출력 velocity field에 differentiable Leray projection을 적용해 divergence-free subspace로 사상함으로써 mass conservation을 강제하는, 수정된 Fourier Neural Operator인 CC-FNO를 소개한다. 이 과정의 계산 오버헤드는 무시할 만한 수준(<2%)이다. 2D incompressible Navier--Stokes에서 $Re \\in \\{100, 500, 1000, 2000\\}$에 대해, CC-FNO는 divergence를 5자릿수 규모로 줄이면서($\\sim 10^{-5}$ 대 $\\sim 3.0$ for FNO, $\\sim 10^{-2}$ for PINO) FNO, PINO, Stream-FNO, U-Net, FNO with pushforward training을 포함한 7개 모델 비교 전반에서 baseline 정확도를 맞춘다. 또한 우리는 autoregressive rollout을 step별 energy bound로 안정화하는 CC-FNO-E를 제안하며, 이는 stability regularizer로 작동하여 재앙적인 amplitude amplification을 방지한다. 이 방법은 모든 regime에서 FNO 대비 안정적인 rollout horizon을 17--52% 연장하지만, physical conservation law가 아니라 stability regularizer이므로 forced flow의 최종 rollout step에서 energy growth를 67--84% 체계적으로 과소추정한다. 체계적인 penalty weight sweep은 soft constraint가 tuning과 무관하게 hard projection을 따라갈 수 없음을 확인하며, forcing-aware energy bound에 대한 ablation은 단순한 one-sided capping이 물리적으로 동기부여된 대안보다 우수함을 보여준다. 우리의 결과는 명확한 constraint hierarchy (none $<$ soft $<$ hard $<$ hard+energy stabilizer)를 확립하고, hard conservation constraint가 정확도를 희생하지 않으면서 ``free''한 물리적 일관성을 제공함을 보여준다.","link":"https://openreview.net/forum?id=BIdT3cqx0E"},{"id":"uX2ryvmaay","en":"Amortized Neural-Symbolic Inverse Dynamics: Robust and Invariant Parameter Inference via Latent Dynamics","ko":"Amortized Neural-Symbolic Inverse Dynamics: 잠재 동역학을 통한 강건하고 불변적인 파라미터 추론","authors":"Matteo Rufolo, Dario Piga, Marco Forgione","abs":"

We propose an amortized neural-symbolic framework for physical parameter inference in dynamical systems. Building on the CAMEL meta-learning framework, whose latent task embeddings carry provable identifiability guarantees, we train a post-hoc MLP to directly recover physical parameters that are nonlinearly related to trajectory data, without per-trajectory optimization or explicit knowledge of the dynamics. We distill this neural mapping into closed-form symbolic expressions via symbolic regression, achieving robust out-of-distribution generalization that validates the recovered physics. To handle realistic conditions, we address measurement noise through domain randomization and improve robustness to unobserved nuisance variables via a structured grouped training strategy. Experiments on a simulated 2-DOF robotic arm show competitive accuracy with classical identification baselines at $\\sim1000\\times$ faster inference, with interpretable parameter mappings and superior noise robustness.

","absKo":"우리는 동역학 시스템에서 물리 파라미터 추론을 위한 amortized neural-symbolic 프레임워크를 제안한다. provable identifiability 보장을 갖는 latent task embedding을 제공하는 CAMEL meta-learning 프레임워크를 기반으로, 우리는 trajectory 데이터와 비선형적으로 관련된 물리 파라미터를 per-trajectory optimization이나 dynamics에 대한 명시적 지식 없이 직접 복원하도록 post-hoc MLP를 학습한다. 우리는 symbolic regression을 통해 이 neural mapping을 closed-form symbolic expression으로 distill하며, 복원된 physics를 검증하는 강건한 out-of-distribution generalization을 달성한다. 현실적인 조건을 다루기 위해, measurement noise는 domain randomization으로 처리하고 structured grouped training 전략을 통해 관측되지 않은 nuisance variable에 대한 robustness를 향상시킨다. 시뮬레이션된 2-DOF robotic arm 실험에서, 이 방법은 해석 가능한 parameter mapping과 우수한 noise robustness를 유지하면서, classical identification baseline과 경쟁력 있는 정확도를 $\\sim1000\\times$ 더 빠른 inference로 달성한다.","link":"https://openreview.net/forum?id=uX2ryvmaay"},{"id":"dNHJzFC0Pd","en":"Accelerating Posterior Inference from Pulsar Light Curves via Learned Latent Representations and Local Simulator-Guided Optimization","ko":"학습된 잠재 표현과 지역 시뮬레이터 유도 최적화를 통한 pulsar 광곡선의 posterior 추론 가속","authors":"Farhana Taiyebah, Abu Bucker Siddik, Indronil Bhattacharjee, Diane Oyen, Soumi De, Greg Olmschenk, Constantinos Kalapotharakos","abs":"

Posterior inference from pulsar light curves is commonly performed using Markov chain Monte Carlo (MCMC), which is accurate but computationally expensive. We introduce a framework that accelerates inference while maintaining accuracy by combining learned latent representations with simulator-guided optimization. A masked U-Net is pretrained to reconstruct complete light curves from partial observations and to produce informative latent embeddings. For a query light curve, we retrieve similar simulations from a precomputed bank using distances in this embedding space, yielding an initial empirical approximation to the posterior over parameters. This estimate is then refined via local hill-climbing updates guided by a forward simulator, progressively concentrating on higher-likelihood regions. Experiments on the observed light curve of PSR J0030+0451, captured by NASA’s Neutron Star Interior Composition Explorer (NICER), show that our method recovers key posterior support and multimodal structure while remaining broadly consistent with MCMC, achieving a 120x speedup and reducing inference time from 24 hours to 12 minutes.

","absKo":"pulsar light curve로부터의 posterior inference는 일반적으로 Markov chain Monte Carlo (MCMC)로 수행되며, 이는 정확하지만 계산 비용이 크다. 우리는 학습된 latent representation과 simulator-guided optimization을 결합하여 정확도를 유지하면서 inference를 가속하는 framework를 제안한다. masked U-Net은 부분 관측으로부터 complete light curve를 복원하고 정보성 있는 latent embedding을 생성하도록 사전학습된다. query light curve에 대해, 우리는 이 embedding space의 거리로 사전 계산된 bank에서 유사한 simulation을 retrieval하여 parameter posterior의 초기 empirical approximation을 얻는다. 그런 다음 이 추정치는 forward simulator가 안내하는 local hill-climbing update를 통해 정제되며, 점차 더 높은 likelihood 영역에 집중한다. NASA의 Neutron Star Interior Composition Explorer (NICER)가 관측한 PSR J0030+0451의 light curve에 대한 실험은, 우리의 방법이 MCMC와 전반적으로 일관성을 유지하면서 key posterior support와 multimodal structure를 복원하고, 120배의 speedup과 inference time을 24시간에서 12분으로 단축함을 보여 준다.","link":"https://openreview.net/forum?id=dNHJzFC0Pd"},{"id":"pDQoUjLrZ0","en":"When Does Critique Improve AI-Assisted Theoretical Physics? SCALAR: Structured Critic--Actor Loop for Agentic Reasoning","ko":"비평은 언제 AI 보조 이론물리학을 개선하는가? SCALAR: agentic reasoning을 위한 구조화 critic-actor 루프","authors":"Vasilis Niarchos, Constantinos Papageorgakis, Alexander G. Stapleton, Sokratis Trifinopoulos","abs":"

As large language models (LLMs) show increasing promise on research-level physics reasoning tasks and agentic AI becomes more common, a practical question emerges: How does the interaction between researchers and agents affect the results?

We study this using SCALAR (Structured Critic--Actor Loop for Agentic Reasoning), an Actor--Critic--Judge pipeline applied to quantum field theory and string theory problems. The Actor proposes solutions, the Critic provides iterative feedback, and an independent Judge evaluates the transcript against reference solutions. We vary the Actor persona, the Critic feedback strategy, and the Actor model family and scale. Multi-turn dialogue improves over single-shot attempts throughout, but both the mechanism of improvement and the value of different prompting choices depend strongly on the Actor--Critic pairing. Increasing the scale within one model family (e.g. from the 8B-parameter DeepSeek-R1 variant to DeepSeek-R1 70B) improves some easier-problem behavior, but does not remove the hardest bottleneck we observe. Critic feedback strategy matters most clearly in the asymmetric Actor--Critic setting (e.g., a lightweight Haiku Actor guided by a stronger Sonnet Critic), where constructive feedback improves mean-score outcomes. In same-family Actor--Critic settings, strategy effects are weaker: lenient feedback is sometimes favored, while strict and adversarial feedback are not beneficial. Taken together, SCALAR provides a controlled testbed for evaluating which interaction structures help or hinder AI-driven scientific discovery.

","absKo":"대형 language model(LLM)이 연구 수준의 physics reasoning task에서 점점 더 유망한 성능을 보이고 agentic AI가 더 흔해지면서, 실용적인 질문이 떠오른다. 연구자와 agent의 상호작용은 결과에 어떤 영향을 미치는가?\n우리는 이 질문을 SCALAR(Structured Critic--Actor Loop for Agentic Reasoning)를 사용해 연구했다. SCALAR는 quantum field theory와 string theory 문제에 적용된 Actor--Critic--Judge pipeline이다. Actor는 해법을 제안하고, Critic은 반복적인 피드백을 제공하며, 독립적인 Judge는 transcript를 reference solution과 비교해 평가한다. 우리는 Actor persona, Critic feedback strategy, Actor model family와 scale을 변화시킨다. multi-turn dialogue는 전반적으로 single-shot 시도보다 향상되지만, 개선 메커니즘과 서로 다른 prompting choice의 가치는 Actor--Critic pairing에 강하게 의존한다. 하나의 model family 내부에서 scale을 키우는 것(예: 8B-parameter DeepSeek-R1 variant에서 DeepSeek-R1 70B로 이동)은 더 쉬운 문제에서 일부 행동을 개선하지만, 우리가 관찰한 가장 어려운 병목을 제거하지는 못한다. Critic feedback strategy는 특히 비대칭적인 Actor--Critic setting에서 가장 분명하게 중요하다(예: 더 강한 Sonnet Critic이 이끄는 가벼운 Haiku Actor). 이 경우 건설적인 feedback은 mean-score 결과를 개선한다. 같은 family의 Actor--Critic setting에서는 strategy 효과가 더 약하다. 때로는 관대한 feedback이 선호되지만, 엄격하거나 적대적인 feedback은 도움이 되지 않는다. 종합하면 SCALAR는 어떤 interaction structure가 AI 기반 과학적 발견을 돕거나 방해하는지 평가하기 위한 통제된 testbed를 제공한다.","link":"https://openreview.net/forum?id=pDQoUjLrZ0"},{"id":"A0k7VU6iRG","en":"Self-Monitoring Warmup: Data-Driven Confidence Rectification for PINN Optimizers","ko":"Self-Monitoring Warmup: PINN 최적화를 위한 데이터 기반 신뢰도 보정","authors":"Giangiacomo Mercatali, Jie Chen, Alexandros Kalousis","abs":"

Adaptive optimizers such as Adam, RAdam, and SOAP decide how aggressively to trust their adaptive learning rate from a fixed rule: a closed-form factor, an iteration-count schedule, or a distributional assumption on gradients. In physics-informed neural networks (PINNs), where gradients couple multi-scale IC, BC, and PDE-residual terms through differential operators, these assumptions fit poorly with the observed training dynamics: gradients are heterogeneous across parameters, and Adam- and SOAP-style updates exhibit early-training gradient distortion. We argue that the optimizer should set its rectification from the empirical statistics of its own second-moment estimates, rather than from the iteration count, and propose Adaptive Confidence Rectification (ACR): a data-driven warmup in which each parameter's rectification is set by the empirical volatility of its second-moment estimate, turned into a bounded per-parameter confidence. ACR makes no assumption on the gradient distribution and acts independently on each parameter. We show ACR keeps the variance its rectification adds to the effective learning rate bounded—vanishing as confidence rises—and strictly contracts the step under low confidence. The method is general and can be combined with any adaptive optimizer; we demonstrate it on SOAP, Sophia, and AdaHessian, and in the first-order setting against RAdam, where it consistently improves performance across four PDEs (Allen–Cahn, Wave, KdV, Burgers) and three model scales.

","absKo":"Adam, RAdam, SOAP 같은 adaptive optimizer는 고정된 규칙에 기반해 adaptive learning rate를 얼마나 공격적으로 신뢰할지 결정한다. 예를 들어 닫힌형태의 factor, iteration-count schedule, gradient에 대한 distributional assumption이 그것이다. physics-informed neural network (PINN)에서는 gradient가 differential operator를 통해 multi-scale IC, BC, PDE-residual 항들을 결합하기 때문에, 이러한 가정은 관찰되는 학습 동역학과 잘 맞지 않는다. gradient는 parameter마다 이질적이며, Adam 및 SOAP 스타일 업데이트는 학습 초기에 gradient distortion을 보인다. 우리는 optimizer가 iteration count가 아니라 자신의 second-moment estimate의 경험적 통계에 따라 rectification을 설정해야 한다고 주장하고, Adaptive Confidence Rectification (ACR)을 제안한다. 이는 각 parameter의 rectification을 second-moment estimate의 경험적 변동성에 의해 정하는 data-driven warmup으로, 이를 bounded per-parameter confidence로 변환한다. ACR은 gradient distribution에 대해 아무 가정도 하지 않으며 각 parameter에 독립적으로 작동한다. 우리는 ACR이 rectification이 effective learning rate에 더하는 분산을 bounded 상태로 유지하고, confidence가 증가할수록 그 분산이 사라지며, 낮은 confidence에서는 step을 엄격히 줄인다는 것을 보인다. 이 방법은 일반적이며 어떤 adaptive optimizer와도 결합할 수 있다. 우리는 SOAP, Sophia, AdaHessian에 적용하고, first-order setting에서는 RAdam과 비교하여 이를 입증했으며, Allen–Cahn, Wave, KdV, Burgers의 네 PDE와 세 가지 model scale 전반에서 일관되게 성능을 향상시킨다.","link":"https://openreview.net/forum?id=A0k7VU6iRG"},{"id":"YYolr0ezd7","en":"Trading off Controllability and Realism in Generative Physics Emulators","ko":"생성형 물리 에뮬레이터에서 제어 가능성과 사실성의 절충","authors":"Francesco Immorlano, Elijah Tavares, Felix Draxler, Juan Nathaniel, Violette Launeau, Padhraic Smyth, Pierre Gentine, Stephan Mandt","abs":"

Machine-learned emulators trained on numerical models' outputs inherit systematic biases relative to real-world observations. While adapting these emulators toward observations can enhance their realism, it often reduces their diversity from a carefully constructed ensemble of simulations. This creates a fundamental \\textit{controllability-realism tradeoff}, where gaining accuracy comes at the cost of losing the distinct identities of the underlying simulators. In this work, we formalize this tradeoff for conditional generative emulators and introduce \\textit{condition-to-condition guidance}. Our method is an inference-time approach that interpolates between simulator and observational conditional score functions controlled by a guidance scale $\\gamma$. We further provide a practical procedure for selecting an optimal $\\gamma ^\\star$. We evaluate our framework on a climate science use case and compare it with a classical bias correction method.

","absKo":"수치 모델의 출력으로 학습된 machine-learned emulator는 실제 관측값에 비해 체계적인 bias를 그대로 계승한다. 이러한 emulator를 관측값에 맞게 적응시키면 현실성을 높일 수 있지만, 신중하게 구성된 simulation ensemble로부터의 다양성이 종종 감소한다. 이는 정확도를 얻는 대가로 underlying simulator의 고유한 정체성을 잃게 되는 근본적인 \\textit{controllability-realism tradeoff}를 만들어낸다. 본 연구에서는 conditional generative emulator에 대해 이 tradeoff를 정식화하고, \\textit{condition-to-condition guidance}를 제안한다. 우리의 방법은 guidance scale $\\gamma$로 제어되는 simulator conditional score function과 observational conditional score function 사이를 보간하는 inference-time 접근법이다. 또한 최적의 $\\gamma ^\\star$를 선택하기 위한 실용적인 절차를 제시한다. 우리는 climate science 사례에 대해 이 framework를 평가하고 classical bias correction method와 비교한다.","link":"https://openreview.net/forum?id=YYolr0ezd7"},{"id":"8imTXiO0Ax","en":"LLM-Guided Neural Architecture Search for Robust Co-Design of Physical Neural Networks","ko":"물리 신경망의 강건한 co-design을 위한 LLM-guided neural architecture search","authors":"Tyler King, Timothee Leleu","abs":"

Deploying neural networks on unconventional hardware demands architectures that co-optimize task accuracy and platform-specific constraints such as energy cost, physical non-idealities, and numerical precision. Existing neural architecture search (NAS) methods are typically tailored to a single hardware family, limiting cross-platform comparison and generalization. We introduce Unconventional Hardware Neural Architecture Search (UH-NAS), a hardware-agnostic, LLM-guided NAS framework that integrates language models as evolutionary operators to co-optimize accuracy and inference energy. By exposing hardware as a swappable backend with per-platform energy models, physical constraints, and non-ideality simulators, UH-NAS enables fair system-level comparisons across various backends without modifying the search algorithm. Tested on optical MZI hardware, UH-NAS discovers significantly more diverse architectures than conventional baselines while outperforming existing LLM-to-NAS approaches. Additional ablations on architecture robustness under non-idealities and the role of system prompts highlight the importance of architecture–hardware co-design for emerging computing platforms.

","absKo":"비전통적 hardware에 neural network를 배치하려면 task 정확도와 에너지 비용, 물리적 비이상성, numerical precision 같은 platform-specific 제약을 함께 최적화하는 architecture가 필요하다. 기존 neural architecture search (NAS) 방법은 대개 단일 hardware family에 맞춰져 있어 platform 간 비교와 일반화가 제한된다. 우리는 Unconventional Hardware Neural Architecture Search (UH-NAS)를 제안한다. 이는 hardware-agnostic한 LLM-guided NAS framework로, language model을 evolutionary operator로 통합해 accuracy와 inference energy를 공동 최적화한다. hardware를 플랫폼별 energy model, 물리적 제약, non-ideality simulator를 갖춘 교체 가능한 backend로 노출함으로써, UH-NAS는 search algorithm을 수정하지 않고도 다양한 backend 간 공정한 system-level 비교를 가능하게 한다. optical MZI hardware에서 테스트한 결과, UH-NAS는 기존 baseline보다 훨씬 더 다양한 architecture를 발견하면서도 기존 LLM-to-NAS 접근법보다 우수한 성능을 보였다. non-ideality 하에서의 architecture robustness와 system prompt의 역할에 대한 추가 ablation은 emerging computing platform에서 architecture-hardware co-design의 중요성을 강조한다.","link":"https://openreview.net/forum?id=8imTXiO0Ax"},{"id":"ZHz2oVJBvZ","en":"Hierarchical Discovery of Adiabatic Hamiltonian Paths and RL Schedules for Quantum Linear System Solvers","ko":"양자 선형 시스템 solver를 위한 단열 Hamiltonian 경로와 RL 스케줄의 계층적 발견","authors":"Wang chenglong","abs":"

Quantum linear-system solvers can provide

asymptotic speedups under structured access as-

sumptions, but finite-size performance depends

strongly on the spectral encoding, Hamiltonian

path, traversal schedule, and implementation

model. We study a hierarchical framework for the

quantum linear systems problem (QLSP) that sep-

arates admissible Hamiltonian-path design from

adaptive traversal control. The outer loop searches

constrained deformations of analytically valid adi-

abatic backbones within implementable opera-

tor libraries, while the inner loop is restricted

to residual schedule control around a derivative-

aware local-adiabatic prior. This residual class

can be optimized by derivative-free search, dif-

ferentiable optimal control, or family-wise RL;

the present experiments instantiate derivative-

free residual search and frame RL as the nat-

ural amortized-control extension rather than as

unconstrained Hamiltonian invention. In this

sense, the framework acts as a two-stage auto-

matic algorithm-design pipeline: the outer layer

automatically adapts valid path families, while the

inner layer automatically refines residual traversal

schedules. We do not seek to improve worst-case

asymptotic query complexity beyond optimal-

scaling QLSP solvers; instead, the framework tar-

gets engineering-relevant finite-size fidelity–time

tradeoffs, family-dependent adaptation, and diag-

nostic transparency. Simulations show the clearest

gains on structured gap-amplified and precondi-

tioned families. Ancilla-assisted extensions can

improve gap structure, subspace anchoring, and

preconditioning, but do not evade known optimal-

scaling query-complexity limits.

","absKo":"Quantum linear-system solver는 구조화된 access assumption 하에서 asymptotic speedup을 제공할 수 있지만, finite-size 성능은 spectral encoding, Hamiltonian path, traversal schedule, 구현 model에 강하게 의존한다. 우리는 quantum linear systems problem (QLSP)을 위한 계층적 framework를 연구하며, 이 framework는 허용 가능한 Hamiltonian-path 설계와 적응적 traversal 제어를 분리한다. outer loop는 구현 가능한 operator library 내에서 분석적으로 valid한 adiabatic backbone의 제약된 deformation을 탐색하고, inner loop는 derivative-aware local-adiabatic prior 주변의 residual schedule control로 제한된다. 이 residual class는 derivative-free search, differentiable optimal control, family-wise RL로 최적화할 수 있다. 본 실험에서는 derivative-free residual search를 구현하며, RL은 제약 없는 Hamiltonian invention이 아니라 natural amortized-control extension으로 해석한다. 이런 의미에서 이 framework는 두 단계의 automatic algorithm-design pipeline으로 작동한다. outer layer는 valid path family를 자동으로 적응시키고, inner layer는 residual traversal schedule을 자동으로 정교화한다. 우리는 optimal-scaling QLSP solver를 넘어 최악의 경우 asymptotic query complexity를 개선하려는 것이 아니다. 대신, 이 framework는 공학적으로 중요한 finite-size fidelity-time tradeoff, family-dependent adaptation, 진단적 투명성을 목표로 한다. 시뮬레이션은 구조화된 gap-amplified family와 preconditioned family에서 가장 분명한 향상을 보여준다. Ancilla-assisted extension은 gap structure, subspace anchoring, preconditioning을 개선할 수 있지만, 알려진 optimal-scaling query-complexity 한계를 회피하지는 못한다.","link":"https://openreview.net/forum?id=ZHz2oVJBvZ"},{"id":"MEZ1otYgXS","en":"Spectrally Regularized Latent Flow Matching for Turbulence Generation","ko":"난류 생성을 위한 스펙트럼 정규화 잠재 flow matching","authors":"Khalid Rafiq, Aditya G. Nair","abs":"

Latent diffusion and flow matching have emerged as leading approaches for synthetic turbulence generation, yet they systematically

under-represent dissipation-range amplitudes. We introduce a latent flow matching framework with a spectrally regularized compression stage that directly targets this failure mode. On a $256^2$ DNS dataset at $Re_f \\approx 2250$, replacing an MSE-trained VAE with a zone-weighted log-spectral objective raises deep-dissipation retained spectral power from 25\\% to 94\\% in reconstruction and from 20\\% to 79\\% in unconditional generation. The improved latent representation also yields a substantially better sampling cost-fidelity tradeoff: the MSE-trained latent space imposes a fundamental quality ceiling near DD bias $-0.70$ that no integrator or step-count can overcome, while the spectrally regularized latent space reaches DD bias $-0.117$ at just $20$ function evaluations. Mechanistically, encoder-decoder swap experiments show that the improvement is driven primarily by encoder-induced latent reorganization rather than decoder capacity, while a support-amplitude decomposition reveals that MSE-trained models behave as conservative suppression models, minimizing pointwise error by attenuating intermittent high-wavenumber structure. Both pipelines recover the second-order structure function and the correct sign of $S_3$, indicating the correct cascade direction without explicit supervision. A small residual gap in the magnitude of $S_3$ suggests that phase-coherent triadic organization remains a complementary axis to amplitude fidelity for future generative turbulence models.

","absKo":"Latent diffusion과 flow matching은 synthetic turbulence 생성의 선도적 접근으로 부상했지만, dissipation-range amplitude를 체계적으로 과소표현한다. 우리는 이 실패 모드를 직접 겨냥하는 spectrally regularized compression stage를 갖춘 latent flow matching framework를 소개한다. $Re_f \\approx 2250$의 $256^2$ DNS dataset에서, MSE로 학습된 VAE를 zone-weighted log-spectral objective로 대체하면 reconstruction에서 deep-dissipation retained spectral power가 25\\%에서 94\\%로, unconditional generation에서는 20\\%에서 79\\%로 상승한다. 개선된 latent representation은 sampling cost-fidelity tradeoff도 크게 향상시킨다. MSE-trained latent space는 integrator나 step-count로도 극복할 수 없는 quality ceiling을 DD bias 약 -0.70 부근에 부과하는 반면, spectrally regularized latent space는 단지 $20$ function evaluation만으로 DD bias -0.117에 도달한다. 메커니즘적으로는 encoder-decoder swap experiment가 이 개선이 decoder capacity보다 encoder가 유도한 latent reorganization에 의해 주로 발생함을 보여주며, support-amplitude decomposition은 MSE-trained model이 conservative suppression model처럼 동작하여 점별 오류를 최소화하기 위해 간헐적인 high-wavenumber structure를 감쇠함을 드러낸다. 두 pipeline 모두 second-order structure function과 $S_3$의 올바른 부호를 복원하여, 명시적 supervision 없이도 cascade 방향이 올바름을 보인다. $S_3$의 크기에 남아 있는 작은 격차는, phase-coherent triadic organization이 향후 generative turbulence model에서 amplitude fidelity와 상보적인 축으로 남아 있음을 시사한다.","link":"https://openreview.net/forum?id=MEZ1otYgXS"},{"id":"ItlBE5738K","en":"Synthetic Physics as Supervision: Learning Real-World Physical Reasoning in Vision-Language Models","ko":"합성 물리학을 감독 신호로: Vision-Language Model이 현실 세계 물리 추론을 학습하기","authors":"Swastik R, Natesha B V","abs":"

Vision-language models (VLMs) remain unreliable on visually grounded physics reasoning in real-world media, and human-labelled physics supervision is expensive. We study whether *synthetic* simulator scenes alone can supply an effective training signal. We fine-tune Qwen3-VL-30B on 14,597 rigid-body and fluid scenes with free-text answers read directly from simulator state, using no human annotation and no local GPU. On PhysBench Test (n=9,786), this synthetic-only SFT lifts accuracy from 40.7% to 47.6% (+6.9pp), improves 27/39 subtasks, and yields a +20.1pp gain on the Scene domain. A follow-up GRPO stage with simulator-verifiable rewards preserves this aggregate gain (+7.0pp over baseline) while adding targeted improvements on the un-trained `general:relationships` domain, and a data-fidelity probe at both stages shows the gains track the physical quality of the synthetic signal. On a vision-essential ScienceQA-Physics probe, prior textbook-physics ability is *not* significantly degraded, so the model remains a generalist VLM. The full pipeline costs under \\$30 in hosted-API credits, providing evidence that synthetic physics is a practical and reproducible supervision signal for grounded visual reasoning in VLMs.

","absKo":"Vision-language models (VLMs)은 실제 미디어에서 시각적으로 grounded된 physics reasoning에 대해 여전히 신뢰하기 어렵고, human-labeled physics supervision은 비용이 많이 든다. 우리는 *synthetic* simulator scene만으로도 효과적인 training signal을 제공할 수 있는지를 연구한다. 우리는 human annotation도 local GPU도 사용하지 않고, simulator state에서 직접 읽은 free-text answer를 사용해 14,597개의 rigid-body 및 fluid scene으로 Qwen3-VL-30B를 fine-tune했다. PhysBench Test(n=9,786)에서 이 synthetic-only SFT는 정확도를 40.7%에서 47.6%로(+6.9pp) 끌어올리고, 27/39개의 subtask를 개선하며, Scene domain에서 +20.1pp 향상을 낸다. simulator-verifiable reward를 사용하는 후속 GRPO 단계는 이 aggregate gain(+7.0pp over baseline)을 유지하면서도 아직 학습되지 않은 `general:relationships` domain에서 목표 지향적 개선을 추가로 제공한다. 두 단계 모두에서 수행한 data-fidelity probe는 이러한 향상이 synthetic signal의 물리적 품질을 따라간다는 것을 보여준다. vision이 필수인 ScienceQA-Physics probe에서는 사전의 textbook-physics 능력이 *유의하게* 저하되지 않아, 모델은 여전히 generalist VLM으로 남아 있다. 전체 파이프라인 비용은 hosted API credit 기준 30달러 미만이며, 이는 synthetic physics가 VLM에서 grounded visual reasoning을 위한 실용적이고 재현 가능한 supervision signal임을 뒷받침한다.","link":"https://openreview.net/forum?id=ItlBE5738K"},{"id":"YCnZJEONzO","en":"Spectral Bias in Variational Quantum Machine Learning","ko":"변분 양자 머신러닝에서의 스펙트럴 바이어스","authors":"Callum Duffy, Marcin Jastrzebski, Sarah Malik","abs":"

In this work, we investigate the phenomenon of spectral bias in quantum machine learning (QML), in classical machine learning this manifests as models fitting low-frequency components of a target function before high-frequency ones. We perform the first comprehensive study characterising how this phenomena manifests in parameterised quantum circuits (PQCs) both theoretically and empirically. Leveraging the established formulation of PQCs as Fourier series, we prove that spectral bias in this setting can arise from the ``redundancy'' of the Fourier coefficients, which denotes the number of terms in the analytical form of the model contributing to the same frequency component. The choice of data encoding scheme dictates the degree of redundancy for a Fourier coefficient. We then further demonstrate this empirically with several different encoding schemes. We show this has direct consequences for quantum-assisted simulation of physical systems such as PDEs. Additionally, we demonstrate that PQCs with greater redundancy exhibit increased robustness to random perturbations in their parameters at the corresponding frequencies.

","absKo":"본 연구에서는 quantum machine learning(QML)에서의 spectral bias 현상을 조사한다. 고전적 machine learning에서 이는 모델이 대상 함수의 고주파 성분보다 저주파 성분을 먼저 적합하는 현상으로 나타난다. 우리는 parameterized quantum circuit(PQC)에서 이 현상이 어떻게 나타나는지 이론적·경험적으로 특성화한 최초의 종합 연구를 수행한다. PQC를 Fourier series로 정식화한 기존 틀을 활용하여, 이 설정에서의 spectral bias는 Fourier coefficient의 ``redundancy''에서 비롯될 수 있음을 증명한다. 여기서 redundancy는 동일한 frequency component에 기여하는 모델의 해석적 형태 내 term의 수를 의미한다. 데이터 encoding scheme의 선택은 Fourier coefficient의 redundancy 정도를 결정한다. 이어서 우리는 여러 다른 encoding scheme을 사용해 이를 경험적으로 추가 입증한다. 이는 PDE와 같은 물리 시스템의 quantum-assisted simulation에 직접적인 결과를 갖는다는 점을 보인다. 또한 redundancy가 더 큰 PQC는 해당 frequency에서 파라미터의 무작위 perturbation에 대해 더 높은 robustness를 보인다는 사실도 입증한다.","link":"https://openreview.net/forum?id=YCnZJEONzO"},{"id":"wppmweesLn","en":"Amortized Simulation-Based Inference of Colliding-Wind Binaries from Short, Noisy Image Time Series","ko":"짧고 잡음 많은 이미지 시계열로부터 충돌풍 쌍성의 amortized simulation-based inference","authors":"Niklas Knöll, Tobias Buck, Lorenzo Branca, Giuseppe Viterbo","abs":"

Colliding-wind binaries (CWBs), which are systems of two massive stars whose supersonic winds collide into bow shocks, encode rich information about stellar wind properties in their multi-frequency emission, e.g. images in the H$\\alpha$, X-ray, and radio wavelengths. Inferring physical parameters (mass-loss rates, terminal wind velocities, orbital elements) from short time-series observations is a compelling but challenging inverse problem, because the forward hydrodynamic simulator is computationally expensive and the likelihood is intractable. We adopt a factorized spatio-temporal architecture for amortized posterior inference that separates spatial encoding from temporal aggregation. This design aligns with the structure of the underlying physical process of local morphology and global dynamical evolution, induces time-translation equivariance in the learned representation, and improves identifiability in low-signal regimes. Coupled with a neural spline flow conditioned on these spatio-temporal embeddings of 10-frame H$\\alpha$ photon-count time series, we present a complete simulation-based inference pipeline for CWBs. Our method jointly infers seven physical parameters from synthetic observations under realistic detector noise, with posteriors verified as well-calibrated via TARP and SBC diagnostics. The approach naturally expands posterior width in information-poor regimes (low photon counts) and robustly recovers orbital parameters and mass-loss rates, demonstrating the feasibility of amortized likelihood-free inference for this challenging astrophysical inverse problem.

","absKo":"Colliding-wind binaries(CWBs)는 초거대질량 별 두 개로 이루어진 계로, 이들의 초음속 wind가 bow shock으로 충돌한다. 이들은 H$\\alpha$, X-ray, radio 파장의 image와 같은 다중 주파수 방출에 항성 wind 특성에 대한 풍부한 정보를 담고 있다. 짧은 시계열 관측으로부터 질량손실률, 종단 wind 속도, 궤도 요소와 같은 물리적 parameter를 추론하는 일은 매력적이지만 어려운 inverse problem이다. 왜냐하면 forward hydrodynamic simulator는 계산 비용이 크고 likelihood는 비가법적이기 때문이다. 우리는 공간 인코딩과 시간 집계를 분리하는 amortized posterior inference를 위한 factorized spatio-temporal architecture를 채택한다. 이 설계는 국소 morphology와 전역 동역학 진화라는 underlying physical process의 구조와 맞닿아 있으며, 학습된 representation에 time-translation equivariance를 유도하고, low-signal regime에서 identifiability를 향상시킨다. 10-frame H$\\alpha$ photon-count time series의 spatio-temporal embedding으로 조건화된 neural spline flow와 결합하여, 우리는 CWBs를 위한 완전한 simulation-based inference pipeline을 제시한다. 우리 방법은 현실적인 detector noise 하에서 synthetic observation으로부터 일곱 개의 물리적 parameter를 공동 추론하며, posterior는 TARP와 SBC diagnostic을 통해 잘 보정되었음이 검증된다. 이 접근은 정보가 부족한 regime(low photon count)에서 posterior width를 자연스럽게 넓히고, 궤도 parameter와 질량손실률을 강건하게 복원하여, 이 까다로운 천체물리 inverse problem에 대한 amortized likelihood-free inference의 실현 가능성을 입증한다.","link":"https://openreview.net/forum?id=wppmweesLn"},{"id":"606d4x5CjS","en":"torch-sla: Differentiable Sparse Linear Algebra with Adjoint Solvers and Sparse Tensor Parallelism for PyTorch","ko":"torch-sla: PyTorch용 adjoint solver와 sparse tensor parallelism을 갖춘 미분 가능한 sparse linear algebra","authors":"Mingyuan chi, Shizheng Wen","abs":"

Differentiable sparse linear algebra is foundational for scientific machine learning, yet PyTorch lacks a unified library for it: \\texttt{torch.sparse} provides only low-level kernels and a non-differentiable, CPU-only \\texttt{spsolve}, and \\texttt{torch.linalg} is dense-only. We present \\torchsla{}, an open-source library that fills this gap. It exposes a single autograd-aware API for direct, iterative, nonlinear, and eigenvalue solvers across five interchangeable backends---SciPy and Eigen on CPU, cuDSS, CuPy, and a PyTorch-native iterative solver on GPU---with automatic dispatch by device and problem size. The library further supports batched solves over shared or distinct sparsity patterns and distributed multi-GPU execution via domain decomposition with halo exchange. These capabilities are made scalable by an O(1)-graph adjoint differentiation framework and an autograd-compatible distributed halo-exchange layer.

","absKo":"Differentiable sparse linear algebra는 scientific machine learning의 기초이지만, PyTorch에는 이를 위한 통합 라이브러리가 없다. `torch.sparse`는 low-level kernel만 제공하고 non-differentiable한 CPU 전용 `spsolve`만 있으며, `torch.linalg`는 dense-only이다. 우리는 이 공백을 메우는 오픈소스 라이브러리 \\torchsla{}를 제시한다. 이 라이브러리는 CPU에서는 SciPy와 Eigen, GPU에서는 cuDSS, CuPy, PyTorch-native iterative solver라는 다섯 개의 상호 교체 가능한 backend 전반에서 direct, iterative, nonlinear, eigenvalue solver를 위한 단일 autograd-aware API를 제공하며, device와 problem size에 따라 자동으로 dispatch한다. 또한 이 라이브러리는 공유되거나 서로 다른 sparsity pattern에 대한 batched solve와, domain decomposition과 halo exchange를 통한 distributed multi-GPU 실행을 지원한다. 이러한 기능은 O(1)-graph adjoint differentiation framework와 autograd-compatible distributed halo-exchange layer를 통해 확장 가능하게 구현된다.","link":"https://openreview.net/forum?id=606d4x5CjS"},{"id":"Q9ZOO12Iy1","en":"Coupling-Robust Accuracy in Multiphysics PINNs via Kronecker-Preconditioned Optimization","ko":"Kronecker 전처리 최적화를 통한 multiphysics PINN의 coupling-강건 정확도","authors":"Youngjae Park, Jaemin Kim, Junghwa Hong","abs":"

Physics-informed neural networks (PINNs) for coupled multiphysics systems suffer systematic accuracy degradation as inter-equation coupling strengthens. We systematically evaluate how PINN accuracy scales with coupling strength and show that combining the Kronecker-preconditioned optimizer SOAP with inverse-gradient-norm loss balancing (SOAP+GN) yields coupling-robust accuracy. Across 234 experiments spanning three 1D systems of increasing nonlinearity and a 2D electroosmotic flow benchmark, SOAP+GN maintains final-epoch $L_2$ degradation $\\leq 1.1\\times$ (ratio of strong- to weak-coupling error) even as coupling parameters vary over one to two orders of magnitude, compared with $> 10^2\\times$ for Adam+GN. In the nonlinear Nernst--Planck--Poisson system, $L_\\infty$-norm balancing (LRA) catastrophically fails due to weight explosion to $O(10^9)$, whereas GN remains stable---revealing a structural limitation of LRA for nonlinearly coupled PDEs. SOAP+GN further scales to a 2D, 6-PDE electroosmotic flow system at EDL-resolved conditions---a regime that all prior PINN electrokinetics studies have avoided through simplified physics---where Adam+GN fails entirely ($L_2 > 0.9$).

","absKo":"결합된 multiphysics 시스템에 대한 physics-informed neural networks(PINNs)는 방정식 간 coupling이 강해질수록 체계적으로 정확도가 저하된다. 우리는 PINN 정확도가 coupling strength에 따라 어떻게 스케일하는지 체계적으로 평가하고, Kronecker-preconditioned optimizer인 SOAP과 inverse-gradient-norm loss balancing(SOAP+GN)을 결합하면 coupling에 강인한 정확도를 얻을 수 있음을 보인다. 비선형성이 점점 커지는 1D 시스템 3개와 2D electroosmotic flow benchmark를 아우르는 234개의 실험에서, SOAP+GN은 coupling parameter가 한두 자릿수(order) 범위로 변해도 최종 epoch의 $L_2$ degradation(강결합 오차 대비 약결합 오차 비율)이 $\\leq 1.1\\times$를 유지하는 반면, Adam+GN은 $> 10^2\\times$에 달한다. 비선형 Nernst--Planck--Poisson 시스템에서는 $L_\\infty$-norm balancing(LRA)이 가중치가 $O(10^9)$까지 폭증하면서 치명적으로 실패하는데, GN은 안정성을 유지한다. 이는 비선형으로 결합된 PDE에서 LRA가 구조적으로 갖는 한계를 드러낸다. 또한 SOAP+GN은 EDL-resolved 조건의 2D, 6-PDE electroosmotic flow 시스템으로도 확장되며, 이는 기존 PINN electrokinetics 연구들이 단순화된 물리로 회피해 온 영역이다. 이 regime에서는 Adam+GN이 완전히 실패한다($L_2 > 0.9$).","link":"https://openreview.net/forum?id=Q9ZOO12Iy1"},{"id":"svMJ3yH4dq","en":"The Encoding-Learning Tradeoff in Scientific Machine Learning: A Risk-Asymmetric Study of Brusselator Reaction-Diffusion Forecasting","ko":"과학적 머신러닝에서의 인코딩-학습 트레이드오프: Brusselator 반응-확산 예측의 위험 비대칭 연구","authors":"Vejandla Sampath Kumar, Prathamesh Dinesh Joshi, Raj Dandekar, Rajat Dandekar, Sreedath Panat","abs":"

Hybrid modeling blends physics and data, but the risks of this balance are rarely quantified. Data-driven models struggle to extrapolate, while physics-informed models can fail catastrophically when the encoded science is wrong.This trade-off is evaluated using 270 models on a chemical reaction–diffusion system under multiple noise and data regimes. Hard-coded equations are contrasted with flexible surrogates learned from observations. Incorrect physics inflates forecast error by up to 6,400×, whereas learning the same term from data incurs only a 1.7× penalty. These results show that misspecifying physics is far riskier than allowing the model to learn.

","absKo":"Hybrid modeling은 physics와 data를 결합하지만, 이러한 균형의 위험은 거의 정량화되지 않았다. Data-driven model은 extrapolation에 어려움을 겪는 반면, physics-informed model은 내재된 science가 잘못되었을 때 치명적으로 실패할 수 있다. 이 trade-off는 여러 noise 및 data regime 하에서 화학 reaction–diffusion system에 대한 270개 model을 사용해 평가된다. hard-coded equation은 관측으로부터 학습된 유연한 surrogate와 대비된다. 잘못된 physics는 forecast error를 최대 6,400×까지 부풀리는 반면, 같은 term을 data로부터 학습하는 것은 단지 1.7×의 penalty만 초래한다. 이러한 결과는 physics를 잘못 명세하는 것이 model이 학습하도록 허용하는 것보다 훨씬 더 위험하다는 점을 보여 준다.","link":"https://openreview.net/forum?id=svMJ3yH4dq"},{"id":"K9dZkgjzRU","en":"Signal-Preserving Cosmic Microwave Background Cleaning with Multi-Scale Learning","ko":"다중 스케일 학습을 통한 신호 보존형 우주 마이크로파 배경 정제","authors":"helen shao, Fiona McCarthy, Miles Cranmer, Blake D. Sherwin","abs":"

Large-scale Cosmic Microwave Background (CMB) B-mode polarization is a key probe of inflation, but is obscured by complex Galactic dust foregrounds. Reliable foreground removal must suppress this contamination without distorting the underlying primordial signal. This is typically done with Internal Linear Combination (ILC), which requires multi-frequency data and is limited to second-order statistics. We present \\textit{single}-frequency cleaning with CNNs that exploit inter-scale correlations in Galactic dust structure to estimate large-scale foregrounds from small-scale modes, while preserving the primordial signal by construction. Using only small-scale B-modes, our model reduces the MSE from $1.4\\times10^{-3}\\,\\mu\\mathrm{K}^2$ for single-channel ILC to $3.4\\times10^{-4}\\,\\mu\\mathrm{K}^2$; adding same-frequency temperature and E-polarization further reduces it to $2.1\\times10^{-4}\\,\\mu\\mathrm{K}^2$. Hybrid models combining multi-frequency information with inter-scale learning improve over ILC residuals by $\\mathcal{O}(10^3)$. Our results show that inter-scale and cross-component structure offers substantial foreground cleaning information beyond multi-frequency correlations, enabling ML models that are both physically reliable and complementary to frequency-based CMB foreground removal. However, incorporating realistic instrumental effects and improving generalization across simulations remain key challenges.

","absKo":"대규모 Cosmic Microwave Background (CMB) B-mode polarization은 inflation의 핵심 탐침이지만, 복잡한 Galactic dust foreground에 의해 가려진다. 신뢰할 수 있는 foreground removal은 이 오염을 억제하면서도 underlying primordial signal을 왜곡하지 않아야 한다. 이는 일반적으로 Internal Linear Combination (ILC)으로 수행되며, multi-frequency data를 필요로 하고 second-order statistics에 제한된다. 우리는 Galactic dust 구조의 inter-scale correlation을 활용해 small-scale mode로부터 large-scale foreground를 추정하는 \\textit{single}-frequency cleaning with CNNs를 제안하며, 이때 primordial signal은 설계상 보존된다. small-scale B-mode만 사용할 때, 우리 모델은 single-channel ILC의 MSE $1.4\\times10^{-3}\\,\\mu\\mathrm{K}^2$를 $3.4\\times10^{-4}\\,\\mu\\mathrm{K}^2$로 낮춘다. 여기에 same-frequency temperature와 E-polarization을 추가하면 $2.1\\times10^{-4}\\,\\mu\\mathrm{K}^2$까지 더 줄어든다. multi-frequency information과 inter-scale learning을 결합한 hybrid model은 ILC residual보다 $\\mathcal{O}(10^3)$ 수준으로 개선된다. 우리의 결과는 inter-scale 및 cross-component structure가 multi-frequency correlation을 넘어서는 상당한 foreground cleaning 정보를 제공하며, frequency-based CMB foreground removal과 상보적인 동시에 물리적으로 신뢰할 수 있는 ML model을 가능하게 함을 보여준다. 다만 현실적인 instrumental effect를 통합하고 simulation 전반에 걸친 generalization을 개선하는 것은 여전히 핵심 과제이다.","link":"https://openreview.net/forum?id=K9dZkgjzRU"},{"id":"aQyk6xF2tS","en":"From Experiments to Expertise: Scientific Knowledge Consolidation for AI-Driven Computational Physics","ko":"실험에서 전문성으로: AI 기반 computational physics를 위한 과학 지식 통합","authors":"Haonan Huang","abs":"

While large language models have transformed AI agents into proficient executors of computational materials science, performing a hundred simulations does not make a researcher. What distinguishes research from routine execution is the progressive accumulation of knowledge -- learning which approaches fail, recognizing patterns across systems, and applying understanding to new problems. However, the prevailing paradigm in AI-driven computational science treats each execution in isolation, largely discarding hard-won insights between runs. Here we present QMatSuite, an open-source platform closing this gap. Agents record findings with full provenance, retrieve knowledge before new calculations, and in dedicated reflection sessions correct erroneous findings and synthesize observations into cross-compound patterns. In benchmarks on a six-step quantum-mechanical simulation workflow, accumulated knowledge reduces reasoning overhead by 67% and improves accuracy from 47% to 3% deviation from literature -- and when transferred to an unfamiliar material, achieves 1% deviation with zero pipeline failures.

","absKo":"large language model은 AI agent를 computational materials science의 유능한 실행자로 바꾸었지만, 시뮬레이션을 백 번 수행한다고 해서 연구자가 되는 것은 아닙니다. 연구를 일상적 실행과 구분하는 것은 지식의 점진적 축적입니다. 어떤 접근이 실패하는지 배우고, 시스템 전반의 패턴을 인식하며, 새 문제에 이해를 적용하는 일입니다. 그러나 AI-driven computational science의 지배적 패러다임은 각 실행을 고립적으로 다루며, 실행 사이에 어렵게 얻은 통찰을 대부분 버립니다. 여기서는 이 간극을 메우는 open-source platform인 QMatSuite를 소개합니다. agent는 provenance가 완전한 상태로 findings를 기록하고, 새 계산에 앞서 지식을 검색하며, 전용 reflection session에서 잘못된 findings를 수정하고 관찰을 cross-compound pattern으로 종합합니다. 여섯 단계 quantum-mechanical simulation workflow benchmark에서 누적된 지식은 reasoning overhead를 67% 줄이고, accuracy를 literature 대비 47%에서 3% deviation으로 개선했습니다. 또한 낯선 material에 transfer했을 때는 pipeline failure 없이 1% deviation을 달성했습니다.","link":"https://openreview.net/forum?id=aQyk6xF2tS"},{"id":"ppQvDCH4qi","en":"LLMs with in-context learning for Algorithmic Theoretical Physics","ko":"In-context learning을 활용한 알고리즘 이론물리학용 LLMs","authors":"Anamaria Hell, Leander Thiele","abs":"

Theoretical physicists typically spend much time on algorithmic computations. These, while conceptually simple, can nevertheless be time-consuming and contain subtleties that are crucial for the broader research question. Motivated by the recent improvement of Large Language Models (LLMs), we investigate to what degree LLMs equipped with a computer algebra system (CAS) runtime and sufficiently informative context can reliably carry out these algorithmic computations.

We task the agent with solving problems from a restricted group, namely cosmological perturbations in modified theories of gravity.

We find that worked examples in the context are essential, a small number of representative examples being more effective than a longer context in which solution strategies are duplicated.

Conversely, a purely abstract description of the algorithm is insufficient context.

Through qualitative analysis we find a strong correlation between the agent's capabilities and the specific strategies available in the context. With an ideal context, we evaluate the agent's ability on par with a first-year PhD student.

","absKo":"이론물리학자들은 대개 algorithmic computation에 많은 시간을 쓴다. 이 계산들은 개념적으로는 단순하지만, 시간이 많이 걸릴 수 있고 더 넓은 연구 질문에서 매우 중요한 미묘한 점들을 포함한다. 최근 Large Language Model(LLM)의 향상에 동기를 얻어, 우리는 computer algebra system(CAS) runtime과 충분히 정보가 풍부한 context를 갖춘 LLM이 이러한 algorithmic computation을 어느 정도까지 신뢰성 있게 수행할 수 있는지 조사한다.\n 우리는 agent에게 제한된 집합의 문제를 풀게 하는데, 구체적으로는 modified theories of gravity에서의 cosmological perturbation이다.\n 우리는 context 속 worked example이 필수적이며, 전략이 중복된 더 긴 context보다 소수의 대표적인 예제가 더 효과적임을 발견했다.\n 반대로, algorithm에 대한 순수한 추상적 설명만으로는 충분한 context가 되지 않는다.\n 질적 분석을 통해 우리는 agent의 능력과 context에서 이용 가능한 특정 전략 사이에 강한 상관관계가 있음을 발견했다. 이상적인 context에서는, 우리는 이 agent의 능력을 1년 차 PhD student와 비슷한 수준으로 평가한다.","link":"https://openreview.net/forum?id=ppQvDCH4qi"},{"id":"kRlSwbfiEQ","en":"Bifurcation Preservation as a Physics Diagnostic for Neural Phase-Field Surrogates","ko":"신경 phase-field surrogate를 위한 물리 진단으로서의 bifurcation 보존","authors":"Alejandro Salinas-Medina, Anisleidy Gonzalez-Mitjans, Xue Liu","abs":"

A common approach for evaluating neural surrogates of phase-field equations is aggregate field error against a reference solver, a measure that can overlook bifurcations: abrupt shifts between qualitatively distinct outcomes, e.g., whether a phase-field droplet dissolves or persists. We propose evaluating neural phase-field surrogates in terms of their capacity for bifurcation preservation. We demonstrate the diagnostic on the Cahn-Hilliard (CH) critical droplet boundary using a droplet-aware Fourier Neural Operator, which reaches a moderate held-out rollout error, with relative $L^2 = 0.153$, however, it shifts the threshold-stable $m = 8$ dissolve/persist boundary by $16.2\\%$. This persists for different random seeds, training budgets, exact zero-mode conservation, a U-Net baseline, and an entire 2D $(m, R_0)$ transition curve including out-of-training regimes. We then compared Allen-Cahn (AC), conservative AC, and CH models using the same epoch-checkpoint protocol. AC preserved the bifurcation once training had converged but fails when undertrained into the CH rollout-error band, and adding mass conservation to AC reduces preservation from from $13/24$ checkpoints to $4/24$. Conservation is implicated as a contributing CH-specific feature, with operator order, free-energy curvature, and branch stability as remaining candidates.

","absKo":"phase-field equation의 neural surrogate를 평가하는 일반적인 접근법은 reference solver와의 aggregate field error이며, 이 척도는 bifurcation을 놓칠 수 있다. 즉, phase-field droplet이 소멸하는지 지속되는지와 같이 질적으로 다른 outcome 사이의 abrupt shift를 간과한다. 우리는 neural phase-field surrogate를 bifurcation preservation 능력의 관점에서 평가할 것을 제안한다. 우리는 droplet-aware Fourier Neural Operator를 사용해 Cahn-Hilliard (CH) critical droplet boundary에서 이 진단법을 시연하며, 이 모델은 moderate held-out rollout error를 달성하지만(relative $L^2 = 0.153$), threshold-stable한 $m = 8$ dissolve/persist boundary를 $16.2\\%$ 이동시킨다. 이 현상은 서로 다른 random seed, training budget, exact zero-mode conservation, U-Net baseline, 그리고 out-of-training regime을 포함한 전체 2D $(m, R_0)$ transition curve에서도 지속된다. 이어서 같은 epoch-checkpoint protocol로 Allen-Cahn (AC), conservative AC, CH model을 비교했다. AC는 training이 수렴한 뒤에는 bifurcation을 보존했지만, CH rollout-error band로 undertrain되면 실패했으며, AC에 mass conservation을 추가하면 보존률이 13/24 checkpoint에서 4/24 checkpoint로 감소했다. Conservation은 CH-specific feature의 기여 요인으로 시사되며, operator order, free-energy curvature, branch stability가 남은 후보로 제시된다.","link":"https://openreview.net/forum?id=kRlSwbfiEQ"},{"id":"NjE9a2uwx9","en":"Learning Design Skills as Memory Policies for Agentic Photonic Inverse Design","ko":"행위적 photonic inverse design에서 memory policy로 설계 기술을 학습하기","authors":"Shengchao Chen, Ting Shu, Sufen Ren","abs":"

Photonic crystal fiber (PCF) inverse design remains challenging because candidate geometries must satisfy coupled optical targets under expensive electromagnetic simulation. Existing pipelines improve surrogate prediction or one-shot parameter recommendation, but they do not accumulate reusable design knowledge across iterative trials. We formulate PCF inverse design as a memory-policy learning problem and propose SkillPCF, a closed-loop agent framework that combines a physics-guided memory skill bank, reinforcement-learned skill selection, and simulator-grounded skill evolution. We further construct a real-world dataset with 479 expert interaction traces (2,507 spans) and 553 memory-dependent evaluation queries covering dispersion engineering, loss optimization, and multi-objective design. Experiments across multiple LLM backbones and classical baselines show that SkillPCF achieves stronger design-quality and efficiency trade-offs under practical simulation budgets, demonstrating the effectiveness of our proposed memory-skill learning paradigm for physics-aware PCF inverse design.

","absKo":"Photonic crystal fiber(PCF) inverse design은 후보 geometry가 비용이 큰 electromagnetic simulation 하에서 결합된 optical target을 만족해야 하므로 여전히 어렵다. 기존 pipeline은 surrogate prediction이나 one-shot parameter recommendation을 개선하지만, 반복 trial 전반에 걸쳐 재사용 가능한 design knowledge를 축적하지는 못한다. 우리는 PCF inverse design을 memory-policy learning problem으로 정식화하고, physics-guided memory skill bank, reinforcement-learned skill selection, simulator-grounded skill evolution을 결합한 closed-loop agent framework인 SkillPCF를 제안한다. 또한 dispersion engineering, loss optimization, multi-objective design을 포괄하는 실제 world dataset을 479개의 expert interaction trace(2,507 span)와 553개의 memory-dependent evaluation query로 구성하여 만든다. 여러 LLM backbone과 classical baseline에 걸친 실험에서 SkillPCF는 실용적인 simulation budget 하에서 더 강한 design quality와 efficiency trade-off를 달성하여, physics-aware PCF inverse design을 위한 우리가 제안한 memory-skill learning paradigm의 효과를 입증한다.","link":"https://openreview.net/forum?id=NjE9a2uwx9"},{"id":"vcx6YRHRBi","en":"Bayesian Cosmic Void Finding with Graph Flows","ko":"그래프 flow를 이용한 베이지안 우주 void 탐색","authors":"Leander Thiele","abs":"

The most underdense regions in the universe are known as cosmic voids. They contain cosmological information and are of interest for astroparticle physics. Finding genuine matter underdensities in sparse galaxy surveys is an underconstrained problem. Traditional void finding algorithms produce deterministic void catalogs, neglecting the probabilistic nature of the problem. We present a method to sample from the stochastic mapping from galaxy catalogs to arbitrary void definitions. Our algorithm uses a deep graph neural network to evolve "test particles" according to a flow-matching objective. We demonstrate the method in a simplified example setting and outline steps to generalize it towards practically usable void finders. Trained on a deterministic teacher, the model performs well but has considerable stochasticity which we interpret as regularization. Cosmological information in the predicted void catalogs outperforms the teacher.

","absKo":"우주의 가장 희박한 영역은 cosmic void로 알려져 있다. 이들은 cosmological information을 담고 있으며 astroparticle physics에서도 관심 대상이다. 희소한 galaxy survey에서 진짜 matter underdensity를 찾는 일은 underconstrained 문제이다. 기존 void finding 알고리즘은 결정론적 void catalog를 생성하며, 문제의 확률적 특성을 간과한다. 우리는 galaxy catalog에서 임의의 void definition으로의 stochastic mapping에서 샘플링하는 방법을 제시한다. 우리의 알고리즘은 deep graph neural network를 사용해 flow-matching objective에 따라 \"test particles\"를 진화시킨다. 우리는 이 방법을 단순화된 예시 설정에서 시연하고, 실용적으로 사용할 수 있는 void finder로 일반화하기 위한 단계를 개괄한다. 결정론적 teacher로 학습된 모델은 성능이 좋지만 상당한 stochasticity를 보이며, 우리는 이를 regularization으로 해석한다. 예측된 void catalog에 포함된 cosmological information은 teacher를 능가한다.","link":"https://openreview.net/forum?id=vcx6YRHRBi"},{"id":"GPCdVV3vMj","en":"Plausible but Wrong: A Case Study on Agentic Failures in Astrophysical Workflows","ko":"그럴듯하지만 틀렸다: 천체물리 워크플로에서 agentic 실패 사례 연구","authors":"Shivam Rawat, Lucie Flek","abs":"

Agentic AI systems are increasingly being integrated into scientific

workflows, yet their behavior under realistic conditions remains

insufficiently understood. We evaluate CMBAgent across two workflow

paradigms and eighteen astrophysical tasks. In the One-Shot setting,

access to domain-specific context yields an approximately ${\\sim}6\\times$

performance improvement ($0.85$ vs.\\ $\\approx\\!0$ without context), with

the primary failure mode being silent incorrect computation---syntactically

valid code that produces plausible but inaccurate results. In the Deep

Research setting, the system frequently exhibits silent failures across

stress tests, producing physically inconsistent posteriors without

self-diagnosis. Overall, performance is strong on well-specified tasks

but degrades on problems designed to probe reasoning limits, often

without visible error signals. These findings highlight that the most

concerning failure mode in agentic scientific workflows is not overt

failure, but confident generation of incorrect results. We release our

evaluation framework to facilitate systematic reliability analysis of

scientific AI agents

","absKo":"Agentic AI system은 점점 더 scientific workflow에 통합되고 있지만, 현실적인 조건에서의 행동은 여전히 충분히 이해되지 않고 있다. 우리는 두 가지 workflow paradigm과 열여덟 개의 astrophysical task에서 CMBAgent를 평가한다. One-Shot 설정에서는 domain-specific context에 접근할 수 있을 때 성능이 약 ${\\sim}6\\times$ 향상되었고 ($0.85$ 대 \\ $\\approx\\!0$ without context), 주요 실패 양상은 조용한 잘못된 계산이었다. 즉, 문법적으로는 유효하지만 그럴듯하나 부정확한 결과를 산출하는 코드이다. Deep Research 설정에서는 system이 stress test 전반에서 자주 silent failure를 보이며, 자기 진단 없이 물리적으로 일관되지 않은 posterior를 생성했다. 전반적으로 성능은 잘 명세된 task에서는 우수하지만, reasoning 한계를 검증하도록 설계된 문제에서는 저하되며, 종종 눈에 띄는 오류 신호 없이 실패한다. 이러한 결과는 agentic scientific workflow에서 가장 우려스러운 실패 양상이 명시적인 실패가 아니라, 자신감 있게 잘못된 결과를 생성하는 것임을 보여준다. 우리는 scientific AI agent의 체계적인 신뢰성 분석을 촉진하기 위해 평가 프레임워크를 공개한다","link":"https://openreview.net/forum?id=GPCdVV3vMj"},{"id":"9IdnFv2TmQ","en":"QMP-Bench: A Research-Level Benchmark for Autonomous End-to-End Quantum Many-Body Simulations","ko":"QMP-Bench: 자율적 종단간 양자 다체 시뮬레이션을 위한 연구급 벤치마크","authors":"Ken Deng, Xiangfei Wang, Guijing Duan, Chen Mo, Junkun Huang, Runqing Zhang, Ling Qian, Zhiguo Huang, Jize Han, Di Luo","abs":"

Although LLM-driven systems exhibit remarkable potential for automated scientific discovery, the community severely lacks realistic, end-to-end benchmark datasets for actual research, particularly for scientific simulations. To address this gap, we introduce QMP-Bench, a research-level dataset comprising $100$ end-to-end simulation tasks extracted directly from $21$ high-impact articles in quantum many-body physics. Unlike human-curated problems, QMP-Bench challenges agents to autonomously reproduce physical results from scratch using domain-specific numerical libraries including ITensors, NetKet, Qiskit, and ORCA. Baseline evaluations of frontier LLMs on these tasks reveal severe struggles with domain-specific programming errors and physical hallucinations, whereas introducing a framework equipped with verification mechanisms significantly improves the performance. Consequently, QMP-Bench serves as a rigorous testbed for driving the development of reliable and physics-informed AI physicists.

","absKo":"LLM 기반 시스템이 자동 과학 발견에서 놀라운 잠재력을 보이지만, 특히 과학 시뮬레이션에 대해 실제 연구를 위한 현실적인 end-to-end benchmark dataset은 커뮤니티에 심각하게 부족하다. 이 격차를 해소하기 위해, 우리는 quantum many-body physics의 high-impact 논문 21편에서 직접 추출한 100개의 end-to-end simulation task로 구성된 연구 수준 데이터셋 QMP-Bench를 제안한다. 인간이 선별한 문제와 달리, QMP-Bench는 ITensors, NetKet, Qiskit, ORCA와 같은 도메인 특화 numerical library를 사용하여 에이전트가 물리 결과를 처음부터 자율적으로 재현하도록 요구한다. 이러한 task에 대한 frontier LLM의 baseline 평가에서는 도메인 특화 프로그래밍 오류와 물리적 hallucination에 심각하게 고전하는 것으로 나타났으며, 검증 메커니즘을 갖춘 프레임워크를 도입하면 성능이 크게 향상된다. 결과적으로 QMP-Bench는 신뢰할 수 있고 물리 지식이 반영된 AI physicist 개발을 촉진하는 엄격한 testbed 역할을 한다.","link":"https://openreview.net/forum?id=9IdnFv2TmQ"},{"id":"MeAFOZnrvM","en":"Semigroup Consistency as a Diagnostic for Learned Physics Simulators","ko":"학습된 물리 시뮬레이터의 진단을 위한 semigroup 일관성","authors":"Lennon J. Shikhman","abs":"

Learned physics simulators are often evaluated by one-step or short-horizon prediction error, but these metrics can miss failures in temporal composition and long-horizon rollout. For autonomous, state-complete systems, exact solution maps satisfy a semigroup law: direct evolution over $s+t$ should agree with evolution over $s$ followed by $t$. We propose normalized semigroup error as a post hoc, model-agnostic diagnostic comparing these direct and composed learned predictions. On one-dimensional heat and Burgers dynamics with time-conditioned ConvNet and FNO baselines, semigroup error is positively associated with rollout degradation, with trajectory-level Spearman correlation $\\rho=0.635$ and $95$% CI $[0.621,0.649]$. Semigroup regularization has mixed effects, supporting semigroup consistency primarily as an evaluation diagnostic rather than a universally beneficial training objective.

","absKo":"학습된 physics simulator는 흔히 one-step 또는 short-horizon prediction error로 평가되지만, 이러한 지표는 temporal composition과 long-horizon rollout의 실패를 놓칠 수 있다. autonomous한 state-complete system에서 exact solution map은 semigroup law를 만족한다. 즉, $s+t$에 대한 직접 진화는 $s$에 대한 진화 뒤 $t$에 대한 진화와 일치해야 한다. 우리는 이러한 direct prediction과 composed learned prediction을 비교하는 post hoc, model-agnostic 진단 지표인 normalized semigroup error를 제안한다. time-conditioned ConvNet과 FNO baseline을 사용한 one-dimensional heat 및 Burgers dynamics에서, semigroup error는 rollout degradation과 양의 상관을 보였으며, trajectory-level Spearman correlation은 $\\rho=0.635$, 95% CI는 $[0.621,0.649]$였다. Semigroup regularization은 효과가 엇갈렸으며, 이는 semigroup consistency가 보편적으로 유익한 training objective라기보다 주로 evaluation diagnostic으로서의 가치를 가진다는 점을 뒷받침한다.","link":"https://openreview.net/forum?id=MeAFOZnrvM"},{"id":"PUEt2m9lzC","en":"PRISM: Position-encoded Regressive Inverse Spectral Model for Multilayer Thin-Film Design","ko":"PRISM: 다층 박막 설계를 위한 위치 인코딩 회귀형 역 스펙트럼 모델","authors":"Runtian Wang, Renhao Xue, Baige Chen, Hao Wu","abs":"

The inverse problem of multilayer thin-film optical coatings design represents a complex combinatorial-continuous optimization challenge. We present $\\textbf{PRISM}$ (Position-encoded Regressive Inverse Spectral Model), a unified decoder-only autoregressive transformer that streamlines this process by jointly predicting discrete material selection and continuous thickness regression within a single backbone. PRISM introduces two primary architectural innovations: (1) $\\textit{spectrum prefix conditioning}$, which utilizes standard prefix tokens for in-context target injection, and (2) $\\textit{cumulative-depth Rotary Position Embeddings}$, which encode continuous thickness directly into the positional representation to preserve the physical spatial relationships of the stack. Our benchmarks demonstrate that a PRISM-13M model reduces MAE by over 50\\% compared to other transformer baselines while utilizing only one-fifth of the parameters. Furthermore, a 44M-parameter variant achieves state-of-the-art performance (MAE = 0.010) on our in-distribution validation benchmark and operates significantly faster than simulated annealing, offering a highly efficient alternative to classical optimization methods.

","absKo":"다층 thin-film optical coating 설계의 inverse problem은 복잡한 조합-연속 최적화 과제를 이룬다. 우리는 $\\textbf{PRISM}$(Position-encoded Regressive Inverse Spectral Model)을 제안한다. 이는 단일 backbone 안에서 discrete material selection과 continuous thickness regression을 함께 예측함으로써 이 과정을 단순화하는 통합 decoder-only autoregressive transformer이다. PRISM은 두 가지 주요 architecture 혁신을 도입한다: (1) in-context target injection을 위한 standard prefix token을 활용하는 $\\textit{spectrum prefix conditioning}$, (2) stack의 물리적 공간 관계를 보존하기 위해 continuous thickness를 positional representation에 직접 인코딩하는 $\\textit{cumulative-depth Rotary Position Embeddings}$이다. 우리의 benchmark는 PRISM-13M model이 다른 transformer baseline 대비 parameter의 5분의 1만 사용하면서 MAE를 50\\% 이상 줄임을 보여준다. 더 나아가 44M-parameter variant는 우리 in-distribution validation benchmark에서 state-of-the-art performance(MAE = 0.010)를 달성하며 simulated annealing보다 훨씬 빠르게 동작하여, classical optimization method에 대한 매우 효율적인 대안을 제공한다.","link":"https://openreview.net/forum?id=PUEt2m9lzC"},{"id":"YDTdNs4H4Z","en":"PAINT: Parallel-in-time Neural Twins for Dynamical System Reconstruction","ko":"PAINT: 동역학 시스템 재구성을 위한 병렬 시간 신경 트윈","authors":"Andreas Radler, Vincent Seyfried, Johannes Brandstetter, Thomas Lichtenegger","abs":"

Neural surrogates have shown great potential for simulating dynamical systems in real time. We envision *Neural Twins* as an evolution of neural surrogates that continuously consume measurements at test time, acting as digital replicas of real systems and enabling context-specific decision-making.

We argue that a natural requirement for Neural Twins is that they remain *on-trajectory*: staying close to the true system state rather than drifting due to compounding errors. Yet most deep learning approaches for reconstructing high-dimensional dynamical systems are *autoregressive*: they predict states from previously predicted ones, which makes them inherently susceptible to such drift.

We find that *parallel-in-time* modeling is a viable but underexplored alternative. Rather than sequential rollout, these models jointly predict a full window of states from a corresponding window of measurements. Such models are conceptually simple, scale to state spaces with thousands of dimensions, and avoid rollout degradation. We show theoretically and empirically that autoregressive models generally fail to satisfy the on-trajectory property, whereas parallel-in-time models satisfy it by construction.

We call this family of methods **PAINT** (**Pa**rallel-**i**n-time **N**eural **T**wins). On a challenging two-dimensional turbulent fluid dynamics problem, our PAINT instantiation using Flow Matching closely follows the trajectory and reconstructs states from sparse measurements with high fidelity.

These results establish parallel-in-time modeling as a viable alternative to autoregressive approaches and motivate further investigation of this direction.

","absKo":"Neural surrogate는 dynamical system을 실시간으로 시뮬레이션하는 데 큰 잠재력을 보여주었다. 우리는 *Neural Twins*를 테스트 시점에 측정을 지속적으로 소비하는 neural surrogate의 진화형으로 상정하며, 실제 시스템의 디지털 복제본처럼 동작하면서 상황별 의사결정을 가능하게 하는 것을 목표로 한다.\n우리는 Neural Twins에 대한 자연스러운 요구사항은 *on-trajectory*를 유지하는 것이라고 주장한다. 즉, 누적 오차로 인해 발산하지 않고 실제 시스템 상태에 가깝게 머물러야 한다는 것이다. 그러나 고차원 dynamical system을 재구성하기 위한 대부분의 deep learning 접근은 *autoregressive*하며, 이전에 예측한 state로부터 다음 state를 예측하기 때문에 본질적으로 이러한 drift에 취약하다.\n우리는 *parallel-in-time* 모델링이 실현 가능한 동시에 아직 충분히 탐구되지 않은 대안임을 보인다. 순차적 rollout 대신, 이들 모델은 대응하는 측정 window로부터 하나의 전체 state window를 공동으로 예측한다. 이러한 모델은 개념적으로 단순하고, 수천 차원의 state space로 확장 가능하며, rollout degradation을 피한다. 우리는 이론적 및 실증적으로 autoregressive 모델은 일반적으로 on-trajectory 속성을 만족하지 못하는 반면, parallel-in-time 모델은 구조적으로 이를 만족함을 보인다.\n우리는 이 방법군을 **PAINT** (**Pa**rallel-**i**n-time **N**eural **T**wins)라고 부른다. 도전적인 2차원 난류 유체역학 문제에서, Flow Matching을 사용한 PAINT 구현은 trajectory를 매우 가깝게 추종하며 sparse measurement로부터 state를 높은 fidelity로 재구성한다.\n이 결과는 parallel-in-time 모델링이 autoregressive 접근의 실질적인 대안임을 확립하고, 이 방향에 대한 추가 연구를 촉진한다.","link":"https://openreview.net/forum?id=YDTdNs4H4Z"},{"id":"mh1eBzETBf","en":"Interpreting the Synchronization Gap: The Hidden Mechanism Inside Diffusion Transformers","ko":"동기화 간극 해석하기: Diffusion Transformer 내부의 숨겨진 메커니즘","authors":"Emil Albrychiewicz, Andrés Franco Valiente, Li-Ching Chen, Viola Zixin Zhao","abs":"

In this paper, we investigate how the synchronization gap is mechanistically realized within pretrained Diffusion Transformers (DiTs) by constructing an explicit architectural realization of replica coupling by embedding two generative trajectories into a joint token sequence, modulated by a symmetric cross attention gate with variable coupling strength $g$. Through a linearized analysis of the attention difference, we show that the replica interaction decomposes mechanistically and we empirically validate our theoretical framework on a pretrained DiT-XL/2 model by tracking commitment and per layer internal mode energies. Our results reveal that: (1) the synchronization gap is an intrinsic architectural property of DiTs that persists even when external coupling is turned off; (2) as predicted by our spatial routing bounds, the gap completely collapses under strong coupling $g\\rightarrow1$; (3) the gap is strictly depth localized, emerging sharply only within the final layers of the Transformer; and (4) global, low frequency structures consistently commit before local, high frequency details. Ultimately, our findings provide a mechanistic interpretation of how Diffusion Transformers resolve generative ambiguity, isolating speciation transitions to the terminal layers of the network.

","absKo":"본 논문에서는 replica coupling의 명시적 architectural realization을 구성하여, 두 개의 generative trajectory를 joint token sequence로 embedding하고 variable coupling strength $g$를 갖는 대칭적 cross attention gate로 이를 변조함으로써, pretrained Diffusion Transformer (DiT) 내부에서 synchronization gap이 기계적으로 어떻게 실현되는지 조사한다. attention difference에 대한 선형화 분석을 통해 replica interaction이 기계적으로 분해됨을 보이고, commitment와 layer별 internal mode energy를 추적하여 pretrained DiT-XL/2 model에서 우리의 이론적 틀을 실증적으로 검증한다. 우리의 결과는 다음을 보여준다: (1) synchronization gap은 external coupling이 꺼져 있어도 지속되는 DiT의 본질적 architectural property이다; (2) 우리의 spatial routing bound가 예측하듯, $g\\rightarrow1$의 강한 coupling 아래에서 gap은 완전히 붕괴한다; (3) gap은 엄격히 depth-localized되어 Transformer의 마지막 layer에서만 급격히 나타난다; 그리고 (4) global한 저주파 구조가 local한 고주파 세부사항보다 일관되게 먼저 commit된다. 궁극적으로 우리의 발견은 Diffusion Transformer가 generative ambiguity를 해결하는 방식을 기계적으로 해석하며, speciation transition이 네트워크의 terminal layer에 국한됨을 밝힌다.","link":"https://openreview.net/forum?id=mh1eBzETBf"},{"id":"pzZpC69gC7","en":"Strong Gravitational Lensing Posterior Sampling in Pixel-Space Using Diffusion Models and Recurrent Inference Machines","ko":"diffusion models와 recurrent inference machine을 이용한 pixel space 강중력렌즈 posterior sampling","authors":"Guillaume Payeur, Laurence Perreault-Levasseur, Gabriel Missael Barco, Yashar Hezaveh","abs":"

Modeling galaxy-galaxy strong gravitational lenses to infer the brightness of the source galaxy and the mass distribution of the foreground galaxy is computationally challenging, particularly for high-resolution, high signal-to-noise observations. In this regime, high-dimensional representations of both the source and the foreground mass distribution are necessary to model the data down to the noise level. This inference problem has been challenging for both traditional and machine learning-based methods because of its high dimensionality and its non-linearity in the foreground mass distribution. We present a method to generate joint posterior samples of the source galaxy and foreground mass distribution as pixelated images conditioned on observations. The method combines diffusion-based generative modeling and recurrent inference machines. It can model realistic gravitational lensing simulations with background and foreground galaxies drawn from cosmological hydrodynamical simulations down to the noise level.

","absKo":"은하-은하 강중력렌즈를 모델링하여 source galaxy의 밝기와 전경 galaxy의 질량 분포를 추론하는 일은 계산적으로 매우 어렵다. 특히 고해상도, 고 signal-to-noise 관측에서는 더욱 그렇다. 이 regime에서는 noise level까지 데이터를 모델링하기 위해 source와 전경 질량 분포 모두에 대한 고차원 표현이 필요하다. 이 추론 문제는 높은 차원성과 전경 질량 분포에 대한 비선형성 때문에 전통적 방법과 machine learning 기반 방법 모두에 어려움을 주어 왔다. 우리는 관측에 조건부인 pixelated image로서 source galaxy와 전경 질량 분포의 joint posterior sample을 생성하는 방법을 제시한다. 이 방법은 diffusion-based generative modeling과 recurrent inference machine을 결합한다. 이를 통해 cosmological hydrodynamical simulation에서 추출한 background 및 foreground galaxy를 이용한 현실적인 gravitational lensing simulation을 noise level까지 모델링할 수 있다.","link":"https://openreview.net/forum?id=pzZpC69gC7"},{"id":"MmdqKqU5Fw","en":"MōLe-Λ: Learning the Coupled-Cluster Response State for Energies, Gradients, and Properties","ko":"MōLe-Λ: 에너지, 기울기, 그리고 물성을 위한 coupled-cluster 반응 상태 학습","authors":"Andreas Burger, Luca Thiede, Abdulrahman Aldossary, Jorge A. Campos-Gonzalez-Angulo, Alex Zook, Jérôme Florian Gonthier, Alan Aspuru-Guzik","abs":"

Coupled-cluster (CC) theory is often considered the gold standard of quantum chemistry, but its high computational cost limits routine access to accurate energies, forces and response properties.

While the right-hand $T$-amplitudes determine the correlated wavefunction, many practically important observables additionally require the left-hand $\\Lambda$-amplitudes.

We introduce MōLe-$\\Lambda$, an extension of Molecular Orbital Learning (MōLe) that predicts the full ground-state coupled-cluster singles and doubles (CCSD) response state by jointly learning right-hand amplitudes $(T_1,T_2)$ and left-hand amplitudes $(\\Lambda_1,\\Lambda_2)$ from localized Hartree--Fock molecular orbitals.

Architecturally, MōLe-$\\Lambda$ extends MōLe with $\\Lambda_1$ and $\\Lambda_2$ readouts that mirror the symmetry constraints of the $T_1$ and $T_2$ heads, while preserving the original equivariant orbital encoder, odd sign-equivariant decoding, locality and size-extensivity.

The resulting model yields accurate CC-quality energies and forces, while simultaneously recovering dipoles, quadrupoles, polarizabilities, the electron density, and 2-electron observables such as the pair density.

We show that MōLe-$\\Lambda$ further extends the speed advantage of MōLe over full CCSD while substantially expanding the accessible properties, providing a route to wavefunction-level surrogate models for correlated quantum chemistry.

","absKo":"Coupled-cluster (CC) theory는 종종 quantum chemistry의 gold standard로 간주되지만, 높은 계산 비용 때문에 정확한 energy, force, response property에 대한 일상적 접근이 제한됩니다. 오른쪽 $T$-amplitude가 correlated wavefunction을 결정하지만, 실제로 중요한 많은 observables는 왼쪽 $\\Lambda$-amplitude도 추가로 필요로 합니다. 우리는 Molecular Orbital Learning (MōLe)의 확장인 MōLe-$\\Lambda$를 소개합니다. 이 방법은 localized Hartree--Fock molecular orbital로부터 오른쪽 amplitude $(T_1,T_2)$와 왼쪽 amplitude $(\\Lambda_1,\\Lambda_2)$를 공동 학습하여 완전한 ground-state coupled-cluster singles and doubles (CCSD) response state를 예측합니다. 구조적으로 MōLe-$\\Lambda$는 $\\Lambda_1$과 $\\Lambda_2$ readout을 추가하여 $T_1$ 및 $T_2$ head의 대칭 제약을 반영하면서, 원래의 equivariant orbital encoder, odd sign-equivariant decoding, locality, size-extensivity는 유지합니다. 그 결과 모델은 정확한 CC-quality energy와 force를 산출하는 동시에 dipole, quadrupole, polarizability, electron density, 그리고 pair density 같은 2-electron observable도 회복합니다. 우리는 MōLe-$\\Lambda$가 full CCSD 대비 MōLe의 속도 이점을 더욱 확장하면서 접근 가능한 property를 크게 넓혀, correlated quantum chemistry를 위한 wavefunction-level surrogate model로 가는 경로를 제공함을 보입니다.","link":"https://openreview.net/forum?id=MmdqKqU5Fw"},{"id":"7zp3ZfoSX4","en":"Contrastive Image-Metadata Pre-Training for Materials Transmission Electron Microscopy","ko":"재료 TEM을 위한 contrastive 이미지-메타데이터 pre-training","authors":"Georgia Channing, Debora Keller, Marta D Rossell, Philip Torr, Stig Helveg, Henrik Eliasson","abs":"

The vast majority of transmission electron microscopy (TEM) data never gets published and ends up on a backup drive until deleted to free up space. These left-over datasets are rich in detail and variation, often paired with automatically saved metadata of instrument state and acquisition parameters. In this work, we introduce a dataset of 7,330 high-angle annular dark-field scanning-TEM (HAADF-STEM) images from a single instrument to learn a joint embedding space between image metadata and HAADF image. These embeddings link image style with acquisition parameters, which allows us to train a generative style transfer network that can convert experimental images into the style they would have had if they were recorded with different instrument parameters. We evaluate the performance of the network and explore the usefulness of the technique for physical denoising.

","absKo":"전송전자현미경(TEM) 데이터의 압도적 다수는 출판되지 못한 채, 공간을 비우기 위해 삭제될 때까지 백업 드라이브에 남아 있다. 이렇게 남겨진 데이터셋은 세부 정보와 변이가 풍부하며, 종종 장비 상태와 획득 파라미터의 자동 저장 메타데이터와 함께 제공된다. 본 연구에서는 단일 장비에서 얻은 7,330장의 high-angle annular dark-field scanning-TEM (HAADF-STEM) 이미지를 포함하는 데이터셋을 도입하여, 이미지 메타데이터와 HAADF 이미지 사이의 joint embedding space를 학습한다. 이러한 embedding은 이미지 스타일과 획득 파라미터를 연결하며, 이를 통해 서로 다른 장비 파라미터로 기록되었을 경우의 스타일로 실험 이미지를 변환할 수 있는 generative style transfer network를 학습할 수 있다. 우리는 네트워크의 성능을 평가하고, 물리적 denoising에 대한 이 기법의 유용성을 탐구한다.","link":"https://openreview.net/forum?id=7zp3ZfoSX4"},{"id":"GPbFukEIDQ","en":"Deep Ritz Neural Operator for Dendritic Growth via Energy Splitting","ko":"에너지 분할을 이용한 dendritic growth용 Deep Ritz Neural Operator","authors":"Chih-Kang Huang, Miha Založnik, Ludovick Gagnon, Benoît Appolaire","abs":"

We propose a neural operator approach bridging classical convex-concave splitting with physics-informed learning to accelerate phase-field simulations. By training via an energy-splitting variational formulation, we enforce the energy dissipation property of the underlying models. We further introduce a Reaction-Diffusion Neural Operator architecture, specifically designed to incorporate the operator-splitting of diffusion and reaction terms in the model equation. For the anisotropic dendritic growth simulation, we show that our approach provides better generalization and higher accuracy than the data-driven training approach, while achieving faster inference than traditional Fourier spectral methods.

","absKo":"우리는 phase-field simulation을 가속하기 위해 classical convex-concave splitting과 physics-informed learning을 연결하는 neural operator 접근법을 제안한다. energy-splitting variational formulation으로 학습함으로써, 우리는 underlying model의 energy dissipation property를 강제한다. 또한 diffusion term과 reaction term의 operator-splitting을 model equation에 통합하도록 특별히 설계된 Reaction-Diffusion Neural Operator architecture를 도입한다. anisotropic dendritic growth simulation에서, 우리의 접근법은 data-driven training approach보다 더 나은 generalization과 더 높은 정확도를 제공하면서, 전통적인 Fourier spectral method보다 더 빠른 inference를 달성함을 보인다.","link":"https://openreview.net/forum?id=GPbFukEIDQ"},{"id":"K9bHv3Y3Uf","en":"QuantumLean-Bench: A Unified Benchmark for Informal and Formal Quantum Reasoning","ko":"QuantumLean-Bench: 비공식 및 형식적 양자 추론을 위한 통합 벤치마크","authors":"Isha Goswami, Anushka Paulchoudhury, Robert Joseph George, Anima Anandkumar","abs":"

Existing Large Language Model benchmarks typically evaluate either informal reasoning or formal proof generation, and are often limited to just one subject area or type of task. To address this gap in robust benchmarks, we introduce QuantumLean-Bench, a unified benchmark of 932 problems spanning chemistry, computing, cryptography, information theory, physics, and quantum machine learning. Each problem pairs an informal natural-language response with a formal counterpart in Lean, enabling direct evaluation of whether models preserve mathematical structure across reasoning modalities. We evaluate several frontier and open-source models, including GPT-5.4 and Gemini 3.1, using a task-aware 0/1/2 rubric and report both coverage and full-benchmark performance. We find that while frontier models achieve near-ceiling performance on informal reasoning (GPT-5.4: 0.981 normalized average, 96.1\\% strict accuracy), performance drops substantially in the formal setting (0.621 normalized average, 36.1\\% strict accuracy). These results highlight a persistent gap in current model capabilities: strong natural-language reasoning does not reliably translate into precise formal representations.

","absKo":"기존 Large Language Model benchmark는 대개 비형식적 추론 또는 형식적 증명 생성 중 하나만을 평가하며, 한 주제 영역이나 한 종류의 task에만 제한되는 경우가 많다. 이러한 강건한 benchmark의 공백을 메우기 위해, 우리는 chemistry, computing, cryptography, information theory, physics, quantum machine learning 전반의 932개 문제를 아우르는 통합 benchmark인 QuantumLean-Bench를 소개한다. 각 문제는 비형식적 natural-language 응답과 Lean의 형식적 대응물을 한 쌍으로 구성하여, 모델이 서로 다른 추론 modality 간에 수학적 구조를 보존하는지 직접 평가할 수 있게 한다. 우리는 task-aware 0/1/2 rubric을 사용하여 GPT-5.4와 Gemini 3.1을 포함한 여러 frontier 및 open-source model을 평가하고, coverage와 전체 benchmark 성능을 모두 보고한다. frontier model은 비형식적 추론에서 거의 상한에 가까운 성능을 보였지만(GPT-5.4: 0.981 normalized average, 96.1\\% strict accuracy), 형식적 setting에서는 성능이 크게 떨어졌다(0.621 normalized average, 36.1\\% strict accuracy). 이러한 결과는 현재 모델 능력에 지속적인 간극이 있음을 보여준다. 강한 natural-language reasoning이 정밀한 formal representation으로 안정적으로 이어지지는 않는다는 점이다.","link":"https://openreview.net/forum?id=K9bHv3Y3Uf"},{"id":"Le0DXiZKlk","en":"Harmonic Sampling: Zero-Parameter Flows on Compact Manifolds","ko":"Harmonic Sampling: 콤팩트 다양체 위의 제로 파라미터 흐름","authors":"Isay Katsman, Anna Gilbert","abs":"

Sampling from probability distributions over compact Riemannian manifolds is a central challenge in the natural sciences, with critical applications ranging from directional statistics to lattice quantum field theory (QFT). Recent approaches rely on manifold continuous normalizing flows to model these densities; however, training these manifold flows via likelihood maximization or distribution matching is notoriously unstable and computationally expensive. In this paper, we bypass the optimization paradigm entirely by introducing Harmonic Sampling, a novel, zero-parameter flow-based sampling framework. By projecting the target density onto the manifold's harmonic basis (the eigenfunctions of the Laplace-Beltrami operator), we analytically derive a deterministic, closed-form vector field that exactly transports a uniform prior to the target distribution. We provide rigorous theoretical guarantees demonstrating that this vector field is globally well-posed, singularity-free, and bounded, with approximation errors that decay super-algebraically. Empirically, we validate our framework on hyperspheres ($S^2$, $S^3$) and the special unitary groups $SU(2)$ and $SU(3)$. Across all evaluated geometries, our generated samples demonstrate rapid spectral convergence, recovering exact ground-truth densities. For lattice QFT applications, we show how one can easily adapt Harmonic Sampling to natively enforce strict gauge (conjugation) invariances. In short, our paper provides a method to obtain exact, high-fidelity samples given only a density function over a compact manifold, entirely eliminating the need for neural network training in this context.

","absKo":"콤팩트한 Riemannian manifold 위의 probability distributions에서 sampling하는 문제는 자연과학의 핵심 도전으로, directional statistics부터 lattice quantum field theory (QFT)에 이르기까지 중요한 응용을 가진다. 최근 접근법들은 이러한 밀도를 모델링하기 위해 manifold continuous normalizing flow에 의존하지만, likelihood maximization이나 distribution matching을 통한 학습은 악명 높게 불안정하고 계산 비용도 크다. 본 논문에서는 optimization paradigm 자체를 완전히 우회하여, zero-parameter flow-based sampling framework인 Harmonic Sampling을 제안한다. 목표 밀도를 manifold의 harmonic basis(Laplace-Beltrami operator의 eigenfunctions)에 투영함으로써, 우리는 uniform prior를 target distribution으로 정확히 transport하는 deterministic한 closed-form vector field를 해석적으로 도출한다. 우리는 이 vector field가 전역적으로 well-posed이고, singularity-free이며, bounded이고, approximation error는 super-algebraically 감소함을 엄밀한 이론적 보장으로 제시한다. 실험적으로는 hypersphere ($S^2$, $S^3$)와 special unitary group $SU(2)$, $SU(3)$에서 본 프레임워크를 검증한다. 평가한 모든 geometry에서 생성된 샘플은 빠른 spectral convergence를 보이며 정확한 ground-truth density를 복원한다. lattice QFT 응용의 경우, Harmonic Sampling을 손쉽게 조정하여 strict gauge (conjugation) invariance를 네이티브하게 강제할 수 있음을 보인다. 요약하면, 본 논문은 compact manifold 위의 density function만 주어졌을 때 exact하고 high-fidelity한 샘플을 얻는 방법을 제시하며, 이 맥락에서 neural network training의 필요성을 완전히 제거한다.","link":"https://openreview.net/forum?id=Le0DXiZKlk"},{"id":"LD0n70PA0U","en":"Alpha-RF: Automated RF-Filter-Circuit Design with Neural Simulator and Reinforcement Learning","ko":"Alpha-RF: 신경 시뮬레이터와 강화학습을 이용한 자동 RF 필터 회로 설계","authors":"Nhat Tran, Chenjie Hao, Alexander Stameroff, Anh-Vu Pham, Yubei Chen","abs":"

Accurate, high-performance radio-frequency (RF) filter circuits are ubiquitous in radio-frequency communication and sensing systems for accepting and rejecting signals at desired frequencies. Conventional RF filter design process involves manual calculations of design parameters, followed by intuition-guided iterations to achieve the desired response for a set of filter specifications. This process is time-consuming due to time- and resource-intensive electromagnetic simulations using full-wave numerical PDE solvers, and requires many intuition-guided adjustments to achieve an practically usable design. This process is also highly sensitive to domain expertise and requires many years of professional training. To address these bottlenecks, we propose an automatic RF filter circuit design tool using neural simulator and reinforcement learning. First, we train a neural simulator to replace the PDE electromagnetic simulator. The neural-network-based simulator reduces each of the simulation time from 4 minutes on average to less than 100 millisecond while maintaining a high precision. Such dramatic acceleration enable us to leverage deep reinforcement learning algorithm and train an amortized inference policy to perform automatic design in the imagined space from the neural simulator. The resulted automatic circuit-design agent achieves super-human design results and exceeds specifications in several cases. The automatic circuit-design agent also reduces the on-average design cycle from days to under a few seconds. Even more surprisingly, we demonstrate that the neural simulator can generalize to design spaces far from the training dataset and in a sense it has learned the underlying physics--Maxwell equations. We also demonstrate that the reinforcement learning has discovered many expert-like design intuitions. This work marks a step in using neural simulators and reinforcement learning in RF circuit design and the proposed method is generally applicable to many other design problems and domains in close affinity.

","absKo":"정확하고 고성능인 radio-frequency(RF) filter circuit은 원하는 주파수에서 신호를 통과시키거나 차단하기 위해 radio-frequency communication 및 sensing system 전반에 널리 쓰인다. 전통적인 RF filter 설계 과정은 설계 파라미터를 수작업으로 계산한 뒤, 필터 사양 집합에 대해 원하는 응답을 얻기 위해 직관에 기반한 반복을 수행하는 방식이다. 이 과정은 full-wave numerical PDE solver를 사용하는 시간 및 자원 집약적 electromagnetic simulation 때문에 오래 걸리며, 실제로 사용할 수 있는 설계를 얻기 위해 많은 직관적 조정이 필요하다. 또한 이 과정은 domain expertise에 매우 민감하고, 수년간의 전문 훈련을 요구한다. 이러한 병목을 해결하기 위해 우리는 neural simulator와 reinforcement learning을 사용하는 자동 RF filter circuit 설계 도구를 제안한다. 먼저, PDE electromagnetic simulator를 대체하기 위해 neural simulator를 학습한다. 이 neural-network 기반 simulator는 높은 정밀도를 유지하면서 각 simulation 시간을 평균 4분에서 100밀리초 미만으로 줄인다. 이러한 극적인 가속 덕분에 deep reinforcement learning algorithm을 활용하여, neural simulator가 상상한 공간에서 automatic design을 수행하는 amortized inference policy를 학습할 수 있다. 그 결과 생성된 자동 회로 설계 agent는 인간을 능가하는 설계 결과를 달성하며 여러 경우에서 사양을 초과한다. 이 자동 회로 설계 agent는 또한 평균 설계 사이클을 며칠에서 몇 초 미만으로 줄인다. 더욱 놀랍게도, neural simulator가 training dataset에서 멀리 떨어진 design space로 일반화할 수 있음을 보이며, 어떤 의미에서는 underlying physics인 Maxwell equations를 학습했음을 보여준다. 또한 reinforcement learning이 전문가 수준과 유사한 많은 설계 직관을 발견했음을 시연한다. 이 연구는 RF circuit design에서 neural simulator와 reinforcement learning을 활용하는 데 한 걸음 나아간 것이며, 제안한 방법은 밀접한 관련이 있는 다른 많은 설계 문제와 도메인에도 일반적으로 적용 가능하다.","link":"https://openreview.net/forum?id=LD0n70PA0U"},{"id":"KqfHzWc2QS","en":"GC-500K: A Physics-Consistent Large-Scale Benchmark Dataset for Grating Coupler Design","ko":"GC-500K: 회절 결합기 설계를 위한 물리 일관적 대규모 벤치마크 데이터셋","authors":"Rahul D Ray, Ravikiran Yeleswarapu, Anuvansh Gera","abs":"

We introduce GC-500K, a physics-consistent syn-

thetic dataset of 500,000 grating coupler de-

signs for silicon photonics. Each sample com-

prises five geometric parameters, 100-point re-

flectance, transmittance, and absorbance spectra

(1200–1600 nm), derived scalar metrics, and en-

ergy conservation flags. The dataset is gener-

ated via a validated analytical model with con-

trolled noise and is rigorously validated for input

independence, output diversity, spectral smooth-

ness, and physical plausibility, with fidelity con-

firmed against finite-element simulations. A sys-

tematic benchmark evaluates 23 models span-

ning classical ML, deep learning, operator meth-

ods, and generative models across four tasks:

forward scalar/spectrum prediction and inverse

scalar/spectrum design. Key findings: (i) tree-

based models excel at forward scalar prediction

(R2=0.978); (ii) UNet1D achieves near-perfect

spectral reconstruction (MSE=3×10−5); and (iii)

inverse success rates remain < 4%, confirm-

ing fundamental non-uniqueness. GC-500K pro-

vides a standardized testbed for reproducible data-

driven photonic design. The GC-500K dataset

and benchmarking codebase will be made pub-

licly available upon acceptance.

","absKo":"우리는 silicon photonics를 위한 physics-consistent synthetic dataset인 GC-500K를 소개한다. 이 데이터셋은 50만 개의 grating coupler design으로 구성된다. 각 샘플은 5개의 geometric parameter, 1200–1600 nm 범위의 100-point reflectance, transmittance, absorbance spectrum, 파생 scalar metric, 그리고 energy conservation flag를 포함한다. 데이터셋은 제어된 noise를 포함하는 검증된 analytical model로 생성되었으며, input independence, output diversity, spectral smoothness, physical plausibility에 대해 엄격하게 검증되었고, finite-element simulation과의 비교로 fidelity가 확인되었다. 체계적인 benchmark는 classical ML, deep learning, operator method, generative model을 아우르는 23개 모델을 네 가지 task에서 평가한다. 즉 forward scalar/spectrum prediction과 inverse scalar/spectrum design이다. 핵심 결과는 다음과 같다. (i) tree-based model이 forward scalar prediction에서 뛰어난 성능을 보였고 (R2=0.978); (ii) UNet1D는 거의 완벽한 spectral reconstruction을 달성했으며 (MSE=3×10−5); (iii) inverse success rate는 4% 미만에 머물러 근본적인 비유일성(non-uniqueness)을 확인했다. GC-500K는 재현 가능한 data-driven photonic design을 위한 표준화된 testbed를 제공한다. GC-500K dataset과 benchmarking codebase는 채택 시 공개될 예정이다.","link":"https://openreview.net/forum?id=KqfHzWc2QS"},{"id":"YpzBAC0ZoI","en":"Loss Landscape Diagnosis for Gradient-Based Gray-Scott System Inversion: Disentangling the Roles of PINN Components","ko":"Gradient 기반 Gray-Scott 시스템 역문제의 손실 지형 진단: PINN 구성요소의 역할 분리","authors":"Yan Yang","abs":"

Gradient-based inversion of reaction-diffusion systems is typically approached via surrogate models or physics-informed neural networks (PINNs), while the most direct route, backpropagation through the PDE's structure itself, has largely been avoided. We pursue this direct route as a diagnostic probe, backpropagating a steady-state loss through unrolled Gray-Scott simulation to recover its parameters, with no surrogate or neural-network augmentation. Optimization fails to converge, and plotting the landscape directly locates the failure in its geometry—flat plateaus with no gradient signal, bounded by sharp cliffs that align with bifurcation boundaries—a structure that recurs across loss functions and is inherited however the gradients are routed to parameters. Reading this minimal setup as an ablation of PINN, we disentangle each component's role: with the neural network fixed, the residual loss is quadratic in the PDE parameters and yields a smooth landscape, so it alone already avoids the pathology, by implicitly encoding the full PDE dynamics across all initial conditions. The neural network, for its part, cannot repair an ill-posed parameter subspace, and so serves only to complete the observed data—a division of labor not previously made explicit. These findings carry concrete design implications for PINN-type methods and a broader heuristic on when added dimensions actually help.

","absKo":"reaction-diffusion system의 gradient-based inversion은 일반적으로 surrogate model이나 physics-informed neural network(PINN)를 통해 접근하지만, 가장 직접적인 경로인 PDE 구조 자체를 통한 backpropagation은 대체로 회피되어 왔다. 우리는 이 직접 경로를 진단적 probe로 삼아, surrogate나 neural network augmentation 없이 unrolled Gray-Scott simulation을 통해 steady-state loss를 backpropagation하여 그 parameter를 복원한다. optimization은 수렴에 실패하고, landscape를 직접 그려 보면 실패의 원인은 기하학에 있음을 알 수 있다. 즉, gradient signal이 없는 평평한 plateau와 bifurcation boundary에 정렬된 가파른 cliff가 그 원인인데, 이러한 구조는 loss function 전반에 걸쳐 반복되며 gradient를 parameter로 어떻게 전달하든 그대로 상속된다. 이 최소 설정을 PINN의 ablation으로 읽으면 각 구성 요소의 역할이 분리된다. neural network를 고정하면 residual loss는 PDE parameter에 대해 quadratic하며 매끄러운 landscape를 내므로, 그것만으로도 pathology를 이미 피한다. 이는 모든 initial condition에 걸친 전체 PDE dynamics를 암묵적으로 인코딩하기 때문이다. 반면 neural network는 잘못 정식화된 parameter subspace를 고칠 수 없으므로, 관측된 데이터를 완성하는 역할만 수행한다. 이는 이전에 명시적으로 드러나지 않았던 역할 분담이다. 이러한 발견은 PINN-type method의 설계에 구체적인 함의를 제공하며, 어떤 경우에 추가 차원이 실제로 도움이 되는지에 대한 더 넓은 휴리스틱을 제시한다.","link":"https://openreview.net/forum?id=YpzBAC0ZoI"},{"id":"vuHHaqe0Hy","en":"FlowForge: A Staged Local Rollout Engine for Flow Field Prediction","ko":"FlowForge: flow field 예측을 위한 단계적 로컬 롤아웃 엔진","authors":"Xiaowen Zhang, Ziming Zhou, Fengnian Zhao, David L.S. Hung","abs":"

Learning-based surrogate simulation for time-dependent physical flows must be accurate, robust to noisy data, and fast enough for repeated use. Existing surrogates often rely on complex models and update the entire field in one pass, which can be slow and fragile to imperfect data. We introduce FlowFoege, a staged local rollout surrogate for fixed-horizon velocity-field prediction that orchestrates local updates within a single prediction. This design allows FlowForge to simulate flow fields using lightweight models while maintaining strong accuracy. Across incompressible, compressible, and multiphase benchmarks from CFDBench, PDEBench, and BubbleML, FlowForge matches or improves upon strong baseline families, while reducing fixed-horizon prediction latency and showing consistently higher robustness to noise and missing observations.

","absKo":"시간 의존적인 physical flow에 대한 learning-based surrogate simulation은 정확해야 하고, noisy data에 robust해야 하며, 반복 사용이 가능할 만큼 충분히 빨라야 합니다. 기존 surrogate는 종종 복잡한 모델에 의존하고 전체 field를 한 번에 갱신하므로, 느리고 imperfect data에 취약할 수 있습니다. 우리는 fixed-horizon velocity-field prediction을 위한 staged local rollout surrogate인 FlowFoege를 제안합니다. 이 방법은 단일 prediction 안에서 local update를 조율합니다. 이러한 설계는 FlowForge가 경량 모델을 사용해 flow field를 시뮬레이션하면서도 강한 정확도를 유지하게 합니다. CFDBench, PDEBench, BubbleML의 incompressible, compressible, multiphase benchmark 전반에서 FlowForge는 강력한 baseline family와 대등하거나 더 나은 성능을 보이면서, fixed-horizon prediction latency를 줄이고 noise 및 missing observation에 대해 일관되게 더 높은 robustness를 나타냅니다.","link":"https://openreview.net/forum?id=vuHHaqe0Hy"},{"id":"QGXY4FeSff","en":"Revisiting the Form of Attention with Positional Encoding for Molecular Structures","ko":"분자 구조를 위한 positional encoding이 있는 attention의 형식 다시 보기","authors":"Yusei Ito, Aidar Alimbayev, Klea Ziu, Deepak Kumar, Kanta Ono, Martin Takáč","abs":"

Molecular property prediction is a core task in machine learning for physics, with applications to interatomic potential learning and materials discovery. Recent transformer-based models achieve strong performance, but their gains often conflate two coupled choices: the design of SE(3)-aware positional encodings and how these encodings are incorporated into attention. Consequently, it remains unclear whether the empirical improvements stem from the positional encoding itself or from the attention formulation used to exploit it. Here, we present a controlled study of attention design in molecular transformers that fixes the positional encoding and only varies the attention formulation. Our results show that attention design alone has a significant effect on performance. Under a shared positional encoding, vector-valued positional encodings are most effective when used to modulate attention logits, while the best overall results are obtained when geometric information is introduced through a separate pathway that is adaptively gated by the main attention message. Our study isolates the effect of attention design, identifies key ingredients for accurate prediction of molecular properties, and offers practical guidelines for future research.

","absKo":"분자 물성 예측은 physics를 위한 machine learning의 핵심 과제로, interatomic potential learning과 materials discovery에 응용된다. 최근 transformer 기반 모델들은 강한 성능을 보이지만, 그 향상은 종종 두 가지 결합된 선택, 즉 SE(3)-aware positional encoding의 설계와 이 encoding을 attention에 어떻게 통합하는지에 대한 선택이 뒤섞인 결과다. 따라서 경험적 성능 향상이 positional encoding 자체에서 비롯된 것인지, 아니면 이를 활용하기 위해 사용된 attention formulation에서 비롯된 것인지 여전히 불분명하다. 여기서 우리는 positional encoding을 고정하고 attention formulation만 바꾸는 방식으로 molecular transformer의 attention 설계를 통제된 조건에서 연구한다. 우리의 결과는 attention 설계 자체가 성능에 유의미한 영향을 미친다는 사실을 보여준다. 공통의 positional encoding 하에서, vector-valued positional encoding은 attention logit을 조절하는 데 사용될 때 가장 효과적이며, 전체적으로 가장 좋은 결과는 geometric information이 main attention message에 의해 적응적으로 gating되는 별도의 경로를 통해 도입될 때 얻어진다. 이 연구는 attention 설계의 효과를 분리해 내고, 분자 물성의 정확한 예측에 중요한 요소를 식별하며, 향후 연구를 위한 실용적 지침을 제공한다.","link":"https://openreview.net/forum?id=QGXY4FeSff"},{"id":"u9boDlH65n","en":"Learning Pore-scale Multiphase Flow from 4D Velocimetry","ko":"4D 속도계측으로 pore-scale 다상 유동 학습하기","authors":"Chunyang Wang, Linqi Zhu, Yuxuan Gu, Robert van der Merwe, Xin Ju, Catherine Spurin, Samuel Krevor, Rex Ying, Tobias Pfaff, Martin J Blunt, Tom Bultreys, Gege Wen","abs":"

Multiphase flow in porous media underpins subsurface energy technologies, including geological $CO_{2}$ and underground hydrogen storage, yet pore-scale dynamics in realistic three-dimensional materials remain difficult to predict. Here we introduce a multimodal machine learning framework that learns multiphase pore-scale flow directly from time-resolved four-dimensional micro-velocimetry measurements. The model couples a graph network simulator for Lagrangian tracer-particle dynamics with a 3D U-Net for voxelized fluid--fluid interface evolution, using the imaged pore geometry as a shared boundary constraint. Trained autoregressively on capillary-dominated drainage sequences ($Ca \\approx 10^{-6}$), the framework captures transient, nonlocal flow perturbations and abrupt interface rearrangements (Haines jumps) over rollouts spanning seconds of physical time, compressing hour-to-day-scale direct numerical simulations to seconds of inference. The framework generalizes to distinct experimental boundary conditions and, zero-shot, to a structurally different rock type, demonstrating that experimental 4D velocimetry can underpin physically grounded surrogate models for rapid digital experiments in subsurface energy applications.

","absKo":"다공성 매질에서의 multiphase flow는 geological $CO_{2}$ 및 underground hydrogen storage를 포함한 subsurface energy technology의 기반을 이룬다. 그러나 현실적인 3차원 재료에서 pore-scale dynamics를 예측하는 일은 여전히 어렵다. 여기서는 time-resolved four-dimensional micro-velocimetry 측정으로부터 직접 multiphase pore-scale flow를 학습하는 multimodal machine learning framework를 제안한다. 이 모델은 Lagrangian tracer-particle dynamics를 위한 graph network simulator와 voxelized fluid--fluid interface evolution을 위한 3D U-Net을 결합하며, 촬영된 pore geometry를 공유 boundary constraint로 사용한다. capillary-dominated drainage sequence($Ca \\approx 10^{-6}$) 위에서 autoregressive하게 학습된 이 framework는 물리적 시간 수 초에 걸친 rollout 동안의 일시적인 비국소적 flow perturbation과 급격한 interface 재배열(Haines jump)을 포착하며, 시간당에서 일 단위에 이르는 direct numerical simulation을 수 초의 inference로 압축한다. 이 framework는 서로 다른 experimental boundary condition과, zero-shot으로 구조적으로 다른 rock type까지 일반화하며, experimental 4D velocimetry가 subsurface energy application에서 빠른 digital experiment를 위한 물리적으로 grounded surrogate model의 기반이 될 수 있음을 보여준다.","link":"https://openreview.net/forum?id=u9boDlH65n"},{"id":"yb1tvNHDhl","en":"Efficient Surrogate Modeling for Fast Hydrodynamic Evolution of the Quark-Gluon Plasma","ko":"Quark-Gluon Plasma의 빠른 유체역학적 진화를 위한 효율적 Surrogate Modeling","authors":"Seungjun Lee, Yeonju Go, Bjoern Schenke, Chun Shen, Yihui Ren, Jin Huang, Maria Chamizo-Llatas","abs":"

We propose a neural operator-based surrogate model for fast simulation of 2+1D and 3+1D quark-gluon plasma (QGP) hydrodynamic evolution. Accurate event-by-event simulations require solving computationally expensive PDEs, often taking hours per event. While neural operators enable efficient operator learning, existing approaches assume static domains and do not explicitly capture dissipative dynamics, evolving spatial support, or scale efficiently to spatiotemporal settings. We introduce a time-conditioned neural operator that jointly models field evolution, energy dissipation, and dynamic spatial support. Our approach combines a factorized spatial neural operator with a time-conditioned decoder that adaptively modulates latent features over proper time, along with a gated spatial masking module to capture the evolving support of the plasma fireball. Experiments on simulation data generated by the IP-Glasma+MUSIC framework show accurate predictions of key physical observables while achieving orders-of-magnitude speedup over traditional solvers. These results highlight the potential of our proposed model as a scalable and high-fidelity surrogate for QGP hydrodynamics.

","absKo":"우리는 2+1D 및 3+1D quark-gluon plasma(QGP) hydrodynamic evolution의 빠른 시뮬레이션을 위한 neural operator 기반 surrogate model을 제안한다. 정확한 event-by-event simulation은 계산 비용이 큰 PDE를 풀어야 하며, event당 몇 시간씩 소요되는 경우가 많다. Neural operator는 효율적인 operator learning을 가능하게 하지만, 기존 접근법은 static domain을 가정하고 dissipative dynamics, evolving spatial support를 명시적으로 포착하지 못하며, spatiotemporal setting으로의 확장성도 제한적이다. 우리는 field evolution, energy dissipation, dynamic spatial support를 공동으로 모델링하는 time-conditioned neural operator를 도입한다. 우리의 접근법은 factorized spatial neural operator와 proper time에 따라 latent feature를 적응적으로 조절하는 time-conditioned decoder를 결합하며, plasma fireball의 evolving support를 포착하기 위한 gated spatial masking module도 포함한다. IP-Glasma+MUSIC framework로 생성한 simulation data에 대한 실험에서, 핵심 물리 관측량을 정확하게 예측하면서 전통적 solver 대비 수십 배에서 수백 배의 속도 향상을 달성한다. 이러한 결과는 제안 모델이 QGP hydrodynamics를 위한 확장 가능하고 고충실도의 surrogate로서 가질 잠재력을 보여 준다.","link":"https://openreview.net/forum?id=yb1tvNHDhl"},{"id":"5tlXAcM9gv","en":"Data-Driven Transport Geometry for Rare Events in Metastable Systems","ko":"준안정 시스템의 희귀 사건을 위한 데이터 기반 transport geometry","authors":"Victor Kawasaki-Borruat, Pierre Vandergheynst","abs":"

Rare-event transitions in metastable systems are expensive to sample because direct simulation spends most of its time in long-lived basins, and only rarely undergo transitions. We study whether Implicit Manifold-valued Diffusions (IMDs), learned directly from trajectory point clouds, can act as data-driven surrogates of the transport geometry underlying such transitions. Our central hypothesis is that when the training cloud is drawn from a biased dynamical ensemble (such as successful reactive trajectories or steady-state Langevin samples) the learned operator inherits both the support of the data and its sampling bias. On a curved transition tube and on the Müller-Brown landscape, IMDs trained on reactive data recover the dominant transition corridor and yield substantially higher short-horizon crossing probabilities than unbiased reference simulations, without access to the ambient drift. When trained on steady-state data, IMDs reproduce the coarse basin occupancy structure of the reference process. These results support IMDs as lightweight surrogates for transport geometry in metastable systems, while clarifying that they approximate coarse transport structure rather than exact kinetics or invariant laws.

","absKo":"Metastable system에서의 rare-event transition은 직접 시뮬레이션이 대부분의 시간을 오래 지속되는 basin 안에서 보내고, 전이도 드물게 일어나기 때문에 샘플링 비용이 높다. 우리는 trajectory point cloud에서 직접 학습한 Implicit Manifold-valued Diffusions(IMDs)가 이러한 transition을 지배하는 transport geometry의 data-driven surrogate로 작동할 수 있는지 연구한다. 우리의 핵심 가설은, 훈련 cloud가 편향된 dynamical ensemble(예: 성공한 reactive trajectory 또는 steady-state Langevin sample)에서 추출될 때, 학습된 operator가 데이터의 support와 샘플링 bias를 함께 계승한다는 것이다. 곡면 transition tube와 Müller-Brown landscape에서 reactive data로 학습한 IMDs는 지배적인 transition corridor를 복원하고, ambient drift에 접근하지 못한 상태에서도 unbiased reference simulation보다 훨씬 높은 short-horizon crossing probability를 제공한다. steady-state data로 학습된 IMDs는 reference process의 coarse basin occupancy structure를 재현한다. 이러한 결과는 metastable system에서 transport geometry의 경량 surrogate로서 IMDs를 지지하는 한편, 그것들이 exact kinetics나 invariant law가 아니라 coarse transport structure를 근사한다는 점을 분명히 한다.","link":"https://openreview.net/forum?id=5tlXAcM9gv"},{"id":"nUki4FsrIe","en":"Microlensing Detection and Inference via Learned Bayes Factors","ko":"학습된 Bayes factor를 통한 microlensing 검출 및 추론","authors":"Nolan Smyth, Laurence Perreault-Levasseur, Yashar Hezaveh","abs":"

We present a unified framework for gravitational microlensing event detection and parameter inference. Traditional pipelines use deterministic hard cuts on photometric statistics, systematically missing low-magnification events in the finite-source regime. We instead frame detection as Bayesian model comparison using Evidence Networks, which learn calibrated Bayes factors from binary-labeled simulations, and combine this with Neural Posterior Estimation (NPE) for amortized parameter inference. Both share a transformer encoder that handles irregularly-sampled time series without imputation. On simulated Roman Space Telescope data, our Evidence Network achieves $99.9\\%$ detection efficiency with zero false positives. Gains are most dramatic in the extreme finite-source regime ($\\rho \\gtrsim 5$), where detection rates reach ${\\sim}95\\%$ versus ${\\sim}50\\%$ for hard cuts, precisely the short-duration free-floating planet events most constraining for formation scenarios. Our NPE provides calibrated posteriors, working towards real-time analysis at survey scale.

","absKo":"우리는 gravitational microlensing event detection과 parameter inference를 위한 통합 framework를 제시한다. 전통적인 pipeline은 photometric statistics에 대한 deterministic hard cut을 사용하여 finite-source regime의 low-magnification event를 체계적으로 놓친다. 우리는 이를 대신 binary-labeled simulation에서 calibrated Bayes factor를 학습하는 Evidence Network를 사용한 Bayesian model comparison으로 detection을 정식화하고, 이를 Neural Posterior Estimation (NPE)과 결합해 amortized parameter inference를 수행한다. 두 구성요소는 모두 imputation 없이 불규칙하게 샘플링된 time series를 처리하는 transformer encoder를 공유한다. 시뮬레이션된 Roman Space Telescope data에서 우리의 Evidence Network는 false positive 없이 99.9%의 detection efficiency를 달성한다. 향상은 극단적인 finite-source regime($\\rho \\gtrsim 5$)에서 가장 두드러지며, detection rate가 약 95%에 이르는 반면 hard cut은 약 50%에 그친다. 이는 formation scenario에 가장 강한 제약을 주는 짧은 지속시간의 free-floating planet event와 정확히 일치한다. 우리의 NPE는 calibrated posterior를 제공하여 survey 규모에서의 real-time analysis를 향해 나아간다.","link":"https://openreview.net/forum?id=nUki4FsrIe"},{"id":"EXfZVTq8Yu","en":"PLAID: A Unified Data Model for Machine Learning on Heterogeneous Physics Simulations","ko":"PLAID: 이질적 물리 시뮬레이션을 위한 통합 데이터 모델","authors":"Fabien Casenave, Xavier Roynard, Brian Staber, Alexandre Devaux-Rivière, William PIAT, Michele Alessandro Bucci, Nissrine Akkari, Abbas Kabalan, Xuan Minh Vuong Nguyen, Luca Saverio, Raphael Carpintero Perez, Anthony Kalaydjian, Samy Fouché, Thierry Gonon, Ghassan Najjar, Thomas Daniel, Emmanuel Menier, Matthieu Nastorg, Giovanni Catalani, Christian Rey","abs":"

Machine learning-based surrogate models have emerged as a powerful tool to accelerate simulation-driven scientific workflows, but their adoption is limited by the lack of large-scale, diverse, and standardized datasets for physics-based simulations. Existing benchmarks often focus on narrow domains or rely on simplified data models, and fail to capture the heterogeneity arising from variable geometries, meshes, and topologies, which is critical for assessing generalization in realistic settings. We introduce PLAID (Physics-Learning AI Data model), a unified and extensible data layer for heterogeneous physics simulations. It preserves the full complexity of simulation data while enabling efficient and scalable machine learning workflows, together with a library for dataset construction and manipulation [LINK REMOVED]. We release six datasets covering structural mechanics and computational fluid dynamics, designed to reflect realistic industrial scenarios and provide standardized benchmarks. The framework includes reproducible evaluation protocols and is integrated with Hugging Face to enable open, community-driven benchmarking with active user participation [LINK REMOVED].

","absKo":"Machine learning 기반 surrogate model은 simulation-driven scientific workflow를 가속하는 강력한 도구로 부상했지만, physics-based simulation을 위한 대규모, 다양성, 표준화를 갖춘 dataset의 부재로 인해 채택이 제한되어 왔다. 기존 benchmark는 종종 좁은 domain에 초점을 맞추거나 단순화된 data model에 의존하며, 다양한 geometry, mesh, topology에서 비롯되는 heterogeneity를 포착하지 못한다. 이는 현실적인 setting에서 generalization을 평가하는 데 매우 중요하다. 우리는 이질적인 physics simulation을 위한 통합 가능하고 확장 가능한 data layer인 PLAID (Physics-Learning AI Data model)를 소개한다. PLAID는 simulation data의 전체 복잡성을 보존하면서도 효율적이고 확장 가능한 machine learning workflow를 가능하게 하며, dataset 구성 및 조작을 위한 library도 함께 제공한다 [LINK REMOVED]. 우리는 structural mechanics와 computational fluid dynamics를 포괄하는 여섯 개의 dataset을 공개하며, 이는 현실적인 industrial scenario를 반영하고 표준화된 benchmark를 제공하도록 설계되었다. 이 framework는 재현 가능한 evaluation protocol을 포함하며 Hugging Face와 통합되어, 활발한 사용자 참여를 통한 개방형 community-driven benchmarking을 가능하게 한다 [LINK REMOVED].","link":"https://openreview.net/forum?id=EXfZVTq8Yu"},{"id":"8mkX2WU0hC","en":"Taming Gradient Perturbations: Probabilistic Reparameterization for Robust Physics-Informed MRI Protocol Optimization","ko":"gradient perturbation 다루기: 강건한 physics-informed MRI protocol 최적화를 위한 확률적 재매개변수화","authors":"Beomgu Kang, Hye Young Heo, Hyunseok Seo","abs":"

Reliable estimation of tissue characteristics in quantitative MRI depends on optimizing acquisition protocols defined by combinations of control parameters. Existing physics-informed optimization methods rely on gradient-based updates, but are highly sensitive to stochastic noise arising from both the imaging system and stochastic gradient descent (SGD), often leading to unstable updates and suboptimal solutions in combinatorial control parameter spaces. To address this, we present Robust Physics-Informed Optimization (RPO), a noise-robust framework that replaces direct optimization of control parameter values with probabilistic optimization over candidate control parameters. Specifically, RPO introduces a probabilistic reparameterization of the control parameters via the Gumbel-Softmax mapping, thereby changing how gradient perturbations propagate through the Jacobian of the mapping. Theoretically, this reparameterization attenuates gradient perturbations, yielding an operator-norm bound proportional to $1/\\tau$ and providing a mechanistic explanation for robustness to system noise and SGD-induced stochasticity. Furthermore, the temperature parameter $\\tau$ shapes the optimization dynamics by balancing broad exploration of candidate configurations in the early iterations with progressively sharper focus on high-probability solutions. Extensive experiments on both simulated and in vivo MRI data demonstrate that RPO consistently improves optimization robustness and tissue-estimation accuracy under realistic noise conditions, with successful translation to actual MRI systems.

","absKo":"정량적 MRI에서 조직 특성을 신뢰성 있게 추정하려면 제어 파라미터들의 조합으로 정의되는 acquisition protocol을 최적화해야 한다. 기존의 physics-informed 최적화 방법은 gradient-based update에 의존하지만, 영상 시스템과 stochastic gradient descent(SGD) 모두에서 발생하는 stochastic noise에 매우 민감하여 combinatorial control parameter space에서 불안정한 업데이트와 비최적 해로 이어지기 쉽다. 이를 해결하기 위해, 우리는 제어 파라미터 값을 직접 최적화하는 대신 후보 control parameter들에 대한 확률적 최적화를 수행하는 noise-robust framework인 Robust Physics-Informed Optimization(RPO)을 제안한다. 구체적으로 RPO는 Gumbel-Softmax mapping을 통해 control parameter의 probabilistic reparameterization을 도입함으로써, gradient perturbation이 mapping의 Jacobian을 따라 전파되는 방식을 바꾼다. 이 재매개변수화는 이론적으로 gradient perturbation을 완화하여, $1/\\tau$에 비례하는 operator-norm bound를 도출하고 system noise 및 SGD-induced stochasticity에 대한 robustness를 기계적으로 설명한다. 또한 temperature parameter $\\tau$는 초기 반복에서는 후보 configuration의 폭넓은 탐색과 점차적으로 높은 확률의 해에 더 날카롭게 집중하는 동적 균형을 통해 optimization dynamics를 형성한다. simulated 및 in vivo MRI data에 대한 광범위한 실험은 RPO가 현실적인 noise 조건에서 optimization robustness와 tissue-estimation accuracy를 일관되게 향상시키며, 실제 MRI system으로의 성공적인 전이도 가능함을 보여준다.","link":"https://openreview.net/forum?id=8mkX2WU0hC"},{"id":"iW8zWqe13T","en":"Mat-Pref: Verifiable-Reward Training Improves Compositional Reasoning in Inorganic Materials","ko":"Mat-Pref: 무기 재료에서의 구성적 추론을 향상시키는 검증 가능 보상 학습","authors":"Sarrah Mikhail Leung, Taehan Kim, Jeongbin Park","abs":"

Reinforcement learning from verifiable rewards (RLVR) has driven rapid progress in mathematical and code reasoning, but when extended to science, existing benchmarks do not decompose what generalizes: do gains reflect structural transfer, property transfer, or memorization? We introduce \\textsc{Mat-Pref}, a benchmark of 10,837 ionic-substitution questions across 11 inorganic structure families, grounded in density functional theory calculations from the Materials Project, with three evaluation splits that isolate in-distribution performance, generalization to entirely held-out structure families, and cross-property transfer: applying band-gap reasoning to hosts seen during training only through formation-energy supervision. Four zero-shot frontier models (70--671B parameters) remain in the 33--54\\% range on every split, confirming that scale alone does not resolve the compositional chemical reasoning this task demands. A two-stage pipeline of supervised fine-tuning followed by Group Relative Policy Optimization (GRPO) lifts Qwen3-8B to 65.2\\% in-distribution and 71.6\\% on held-out families, exceeding zero-shot Qwen3-235B by over 20 percentage points on both structural-generalization splits, at a at a total pipeline cost under 50 dollars. Self-consistency sampling shows that the SFT policy can already produce correct answers but cannot reliably surface them as the modal response; GRPO reshapes the distribution so that correct answers become modal rather than merely reachable, and this sharper commitment is visible mechanistically: logit lens analysis reveals a ${\\sim}$20pp advantage in answer crystallization at the critical decision layer. We formalize this observation as a distractor-permutation consistency metric under which GRPO narrows the gap between lenient scoring (at least one permutation correct) and strict scoring (all permutations correct) from 24.0 to 14.3 percentage points. All results replicate on Llama-3.1-8B with identical hyperparameters, confirming the findings are not architecture-specific.

","absKo":"검증 가능한 보상(reinforcement learning from verifiable rewards, RLVR)은 mathematical 및 code reasoning에서 빠른 진전을 이끌었지만, science로 확장하면 기존 benchmark는 무엇이 일반화되는지 분해하지 못한다: 성능 향상이 structural transfer, property transfer, 아니면 memorization의 결과인가? 우리는 Materials Project의 density functional theory 계산에 기반한 11개 inorganic structure family 전반의 10,837개 ionic-substitution 질문으로 구성된 benchmark인 \\textsc{Mat-Pref}를 제안한다. 이 benchmark는 in-distribution 성능, 완전히 보류된 structure family로의 일반화, 그리고 cross-property transfer를 분리하는 세 가지 evaluation split을 포함한다. 여기서 cross-property transfer는 training 중에는 formation-energy supervision만으로 접한 host에 대해 band-gap reasoning을 적용하는 설정이다. 네 개의 zero-shot frontier model(70--671B parameters)은 모든 split에서 33--54\\% 범위에 머물러, 이 task가 요구하는 compositional chemical reasoning은 scale만으로 해결되지 않음을 확인한다. supervised fine-tuning 뒤에 Group Relative Policy Optimization(GRPO)을 적용하는 2단계 pipeline은 Qwen3-8B를 in-distribution 65.2\\%, held-out family에서 71.6\\%까지 끌어올리며, 두 구조 일반화 split 모두에서 zero-shot Qwen3-235B를 20 percentage point 이상 앞선다. 또한 총 pipeline 비용은 50달러 미만이다. Self-consistency sampling은 SFT policy가 이미 정답을 생성할 수는 있지만 그것을 modal response로 안정적으로 드러내지는 못함을 보여준다. 반면 GRPO는 분포를 재형성하여 정답이 단지 도달 가능한 수준이 아니라 modal이 되게 한다. 이 더 선명한 commitment는 mechanistically도 관찰되며, logit lens analysis는 critical decision layer에서 answer crystallization에 약 20pp의 우위를 드러낸다. 우리는 이를 distractor-permutation consistency metric으로 정식화하며, 이 metric 아래에서 GRPO는 lenient scoring(적어도 하나의 permutation이 정답)과 strict scoring(모든 permutation이 정답) 사이의 격차를 24.0에서 14.3 percentage point로 좁힌다. 모든 결과는 동일한 hyperparameter로 Llama-3.1-8B에서도 재현되어, 이 발견이 architecture-specific하지 않음을 확인한다.","link":"https://openreview.net/forum?id=iW8zWqe13T"},{"id":"zb78O5J0B7","en":"Koopman Operator Enhanced 3D Voxcel Auto-Encoder to Predict 3D Seismic Waves Propagation","ko":"3D Voxcel Auto-Encoder와 Koopman Operator를 이용한 3D 탄성파 전파 예측","authors":"Takayuki Shinohara","abs":"

Neural surrogates for seismic simulation are typically either transformer-heavy or depend on hand-crafted physics losses. We introduce VoxelKoop, a lightweight alternative that pairs a 3-D voxel auto-encoder with a single complex Koopman matrix. The encoder compresses a $64^3$ S-wave–velocity voxel data and seismic source parameter, the Koopman operator advances the latent state linearly in Fourier space, and the decoder reconstructs the three ground-motion components. This strictly linear core yields a numerically stable, eigen-interpretable model that is an order of magnitude smaller than transformer-based neural operators. Trained on 30 k SEM3D scenarios, VoxelKoop matches or exceeds baseline accuracy on peak-velocity metrics and executes $2.1\\times$ faster on one GPU, demonstrating its potential for real-time, site-specific hazard prediction.

","absKo":"Seismic simulation을 위한 neural surrogate는 대체로 transformer에 크게 의존하거나, 수작업으로 설계한 physics loss에 의존한다. 우리는 3-D voxel auto-encoder와 하나의 complex Koopman matrix를 결합한 경량 대안인 VoxelKoop를 소개한다. encoder는 $64^3$ S-wave–velocity voxel data와 seismic source parameter를 압축하고, Koopman operator는 Fourier space에서 latent state를 선형적으로 전진시키며, decoder는 세 개의 ground-motion component를 복원한다. 이 엄격히 선형적인 core는 수치적으로 안정적이고 eigen-interpretable한 model을 만들어내며, transformer-based neural operator보다 한 자릿수 이상 더 작다. 30 k SEM3D scenario로 학습한 결과, VoxelKoop는 peak-velocity metric에서 baseline accuracy와 동등하거나 그 이상을 달성하고 단일 GPU에서 $2.1\\times$ 더 빠르게 실행되어, real-time의 site-specific hazard prediction에 대한 잠재력을 입증한다.","link":"https://openreview.net/forum?id=zb78O5J0B7"},{"id":"4rgtkoCNnA","en":"Generative Priors Accelerate Inverse Design for Regularized, High Performance Integrated Photonics","ko":"Generative Prior가 정규화된 고성능 Integrated Photonics의 역설계를 가속한다","authors":"Taehan Kim, Junho Park, Mohammad Ali, SooHyuk Cho, Di Liang","abs":"

Adjoint-based inverse design has revolutionized integrated photonics, yet existing workflows remain computationally inefficient due to a heavy dependence on high-quality initialization. We propose that archives of past optimization traces, typically discarded, should be repurposed as $\\textbf{reusable scientific memory}$. We introduce a physics-compatible generative framework that learns a conditional prior over optimized device layouts to seed downstream adjoint refinement. Our model, a conditional Variational Autoencoder (cVAE), is trained on archived designs and incorporates a differentiable filter-project operator to enforce fabrication-feasible geometries. At inference, the model samples candidate templates which are ranked via a reduced-cost electromagnetic screen before final adjoint optimization. On arbitrary-ratio silicon power splitters with a matched 60-step refinement budget, our cVAE-seeded approach reduces median insertion loss by $3.2\\times$ (to $0.127$ dB) with no floor objective function and from $5.07$ dB to $1.94$ dB in the with-floor setting. Crucially, the learned prior also improves both convergence speed by 50\\% and the geometry-performance trade-off. Here, we demonstrate $\\textbf{amortization of historical optimization}$ into reusable priors transforms restart-heavy global search into efficient, template-guided refinement.

","absKo":"Adjoint 기반 inverse design은 integrated photonics를 혁신했지만, 기존 workflow는 고품질 initialization에 크게 의존하기 때문에 계산적으로 비효율적이다. 우리는 과거 optimization trace의 기록, 즉 보통은 버려지는 기록이 재사용 가능한 scientific memory로 활용되어야 한다고 제안한다. 우리는 downstream adjoint refinement를 초기화하기 위해 optimized device layout에 대한 conditional prior를 학습하는 physics-compatible generative framework를 도입한다. 우리의 모델은 conditional Variational Autoencoder(cVAE)로, 아카이브된 design으로 학습되며 제조 가능한 geometry를 강제하기 위해 differentiable filter-project operator를 통합한다. inference 시에는 모델이 candidate template를 샘플링하고, 최종 adjoint optimization 전에 reduced-cost electromagnetic screen으로 순위를 매긴다. 임의 비율(silicon power splitter)에서 60-step refinement budget을 맞춘 설정에서, 우리의 cVAE-seeded 접근은 floor objective function이 없는 경우 median insertion loss를 3.2배(0.127 dB까지) 낮추고, floor가 있는 설정에서는 5.07 dB에서 1.94 dB로 낮춘다. 핵심적으로, 학습된 prior는 convergence speed를 50% 향상시키고 geometry-performance trade-off도 개선한다. 여기서 우리는 역사적 optimization의 amortization이 재사용 가능한 prior로 전환되어, restart-heavy global search를 효율적인 template-guided refinement로 바꾼다는 점을 보인다.","link":"https://openreview.net/forum?id=4rgtkoCNnA"},{"id":"NjCAj1hZGY","en":"StarEmbed: Benchmarking Time Series Foundation Models on Astronomical Observations of Variable Stars","ko":"StarEmbed: 변광성 천문 관측에서 시계열 foundation model 벤치마킹","authors":"Weijian Li, Hong-Yu Chen, Nabeel Rehemtulla, Ved G. Shah, Dongho Kim, Dennis Wu, Qinjie Lin, Adam A. Miller, Han Liu","abs":"

Current time series foundation model (TSFM) training corpora largely omit data with certain complexities like irregular temporal sampling. Astronomical time series of stellar fluxes (``light curves'') are available in immense quantities and exhibit irregular sampling, multiple variates, and heteroskedasticity. We introduce $\\texttt{StarEmbed}$, the first public benchmark for light curves comprised of real observations of $\\sim$40,000 stars across seven classes and evaluations in clustering, classification, and out-of-distribution (OOD) source detection. We benchmark TSFMs with differing architecture and training strategies as well as domain-specific transformers. Our results demonstrate that the $\\texttt{Chronos}$ family, despite being pre-trained on regularly sampled non-astronomical data, yields state-of-the-art (SOTA) performance in light curve clustering and OOD detection. While no TSFM strictly surpasses the classification performance of the long-established domain baseline, they do demonstrate excellent generalization abilities. $\\texttt{StarEmbed}$ marks a step toward universal light curve embeddings and improved TSFM performance on challenging data.

","absKo":"현재의 time series foundation model(TSFM) 학습 말뭉치는 불규칙한 temporal sampling과 같은 특정 복잡성을 지닌 데이터를 대체로 포함하지 않는다. 별의 flux로 이루어진 천문 time series(\"light curve\")는 매우 방대한 양으로 उपलब्ध하며, 불규칙한 sampling, 다변량성, heteroskedasticity를 보인다. 우리는 $\\texttt{StarEmbed}$를 제안하는데, 이는 실제 관측된 약 40,000개의 별을 7개 class로 구성하고 clustering, classification, out-of-distribution(OOD) source detection 평가를 포함하는 최초의 공개 light curve benchmark이다. 우리는 architecture와 training strategy가 서로 다른 TSFM과 domain-specific transformer를 평가한다. 우리의 결과는 $\\texttt{Chronos}$ family가, 비록 규칙적으로 샘플링된 비천문 데이터로 pre-training되었음에도 불구하고, light curve clustering과 OOD detection에서 state-of-the-art(SOTA) 성능을 보임을 보여준다. 어떤 TSFM도 오래전부터 확립된 domain baseline의 classification 성능을 엄밀히 뛰어넘지는 못하지만, 매우 우수한 generalization 능력은 확인된다. $\\texttt{StarEmbed}$는 universal light curve embedding과 어려운 데이터에서의 향상된 TSFM 성능을 향한 한 걸음이다.","link":"https://openreview.net/forum?id=NjCAj1hZGY"},{"id":"KI9UxQtJlM","en":"ChemHyperMag: Physics-informed magnetic hypergraph learning improves molecular ADMET prediction","ko":"ChemHyperMag: 물리 기반 magnetic hypergraph learning이 분자 ADMET 예측을 향상시킨다","authors":"Hexiao Ding, Hongzhao Chen, Jing Lan, Yufeng Jiang, Zihong Luo, Zehua Xiong, Tianlong Ruan, Yunlin Mao, Nga Chun Ng, Gwing Kei Yip, Gerald W.Y. Cheng, Kate Inyoung OH, Jing Cai, Liang-Ting Lin, Jung Sun Yoo","abs":"

Accurate prediction of ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) is important for drug discovery. Most predictors use undirected molecular graphs and pairwise edges. This choice misses asymmetric interactions, nonreversible dynamics, and motif level effects from functional groups and ring systems. We propose ChemHyperMag for multitask ADMET prediction under missing labels. ChemHyperMag builds a functional group hypergraph from rings, BRICS fragments, Bemis-Murcko scaffolds, and bonds. It also defines a potential driven nonreversible flow guided by electronegativity and Gasteiger partial charges. The resulting circulation is encoded by a Hermitian magnetic Laplacian and processed with a magnetic Chebyshev encoder. We perturb magnetic phases to form stochastic views and train with an InfoNCE objective. Experiments on multiple ADMET benchmarks show improvements over recent methods with fewer labeled samples and no conformers. ChemHyperMag is scalable and provides interpretable directional signals through its magnetic phases.

","absKo":"ADMET(Absorption, Distribution, Metabolism, Excretion, and Toxicity)의 정확한 예측은 drug discovery에서 중요하다. 대부분의 predictor는 무방향 molecular graph와 pairwise edge를 사용한다. 이러한 선택은 비대칭 상호작용, 비가역 동역학, 그리고 functional group과 ring system에서 오는 motif 수준의 효과를 놓친다. 우리는 missing label 하에서의 multitask ADMET 예측을 위해 ChemHyperMag를 제안한다. ChemHyperMag는 ring, BRICS fragment, Bemis-Murcko scaffold, bond로부터 functional group hypergraph를 구성한다. 또한 electronegativity와 Gasteiger partial charge에 의해 유도되는 potential-driven nonreversible flow를 정의한다. 이렇게 얻어진 circulation은 Hermitian magnetic Laplacian으로 인코딩되고 magnetic Chebyshev encoder로 처리된다. 우리는 magnetic phase를 perturb하여 stochastic view를 만들고 InfoNCE objective로 학습한다. 여러 ADMET benchmark에 대한 실험에서, 더 적은 labeled sample과 conformer 없이도 최근 방법들을 능가하는 성능을 보였다. ChemHyperMag는 확장 가능하며 magnetic phase를 통해 해석 가능한 방향성 신호를 제공한다.","link":"https://openreview.net/forum?id=KI9UxQtJlM"}];window.dispatchEvent(new CustomEvent("papers-loaded",{detail:"54069"}));