model_name: ORB v3 model_key: orb-v3 model_version: v0.5.5 date_added: '2025-04-05' date_published: '2025-04-05' authors: - name: Ben Rhodes affiliation: Orbital Materials corresponding: true - name: Sander Vandenhaute affiliation: Orbital Materials - name: Vaidotas Šimkus affiliation: Orbital Materials - name: James Gin-Pollock affiliation: Orbital Materials - name: Tim Duignan affiliation: Orbital Materials - name: Jonathan Godwin affiliation: Orbital Materials email: jonathan@orbitalmaterials.com - name: Mark Neumann affiliation: Orbital Materials email: mark@orbitalmaterials.com corresponding: true repo: https://github.com/orbital-materials/orb-models url: https://orbitalmaterials.com/post/orb-v3-atomistic-simulation-at-scale doi: https://doi.org/10.48550/arXiv.2504.06231 paper: https://arxiv.org/abs/2504.06231 pr_url: https://github.com/janosh/matbench-discovery/pull/240 checkpoint_url: https://orbitalmaterials-public-models.s3.us-west-1.amazonaws.com/forcefields/orb-v3/orb-v3-conservative-inf-mpa-20250404.ckpt pypi: https://pypi.org/project/orb-models license: code: Apache-2.0 code_url: https://github.com/orbital-materials/orb-models/blob/3e9580809/LICENSE checkpoint: Apache-2.0 checkpoint_url: https://github.com/orbital-materials/orb-models/blob/3e9580809/LICENSE openness: OSCD trained_for_benchmark: true train_task: S2EFS test_task: IS2RE-SR targets: EFS_G model_type: UIP model_params: 25_510_582 n_estimators: 1 hyperparams: max_force: 0.02 max_steps: 500 ase_optimizer: FIRE cell_filter: FrechetCellFilter optimizer: Adam loss: MAE loss_weights: energy: 10.0 force: 1.0 stress: 1.0 batch_size: 300 # (avg, as batch size is dynamic) initial_learning_rate: 0.0003 learning_rate_schedule: CosineAnnealingLR(T_max=100) gradient_clip: 0.5 ema_decay: 0.999 n_layers: 5 features: 512 latent dim, 1024 MLP dim num_rbfs: 8 graph_construction_radius: 6.0 # Å max_neighbors: 120 requirements: orb-models: '0.5.0' training_set: [MPtrj, Alex, OMat24] training_cost: Nvidia H100: { amount: 8, hours: 125, cost: nan } notes: Description: | ORB-v3 is a pretrained model for atomic simulations. ORB-v3 uses a pretrained diffusion model backbone which is fine-tuned on MPtrj + Alexandria + OMAT. Data Leakage: | Note that the Alexandria dataset contains structure prototypes that overlap with the WBM test set (~97k overlapping prototypes). Models trained on this dataset benefit from some amount of train/test leakage, though testing on Orb v2 showed that brief (less than 1 epoch) fine-tuning resulted in identical metrics whether fine-tuning on Alexandria or sAlex. Orb v3 was fine-tuned for less than an epoch (10k steps with ~70x8 batch size), meaning it saw only a small part of the Alexandria dataset during fine-tuning. Testing with earlier Orb v2 showed minimal performance difference when using models specifically trained to exclude overlapping prototypes. metrics: phonons: kappa_103: κ_SRME: 0.2102 pred_file: models/orb/orb-v3/2025-04-04-kappa-103-disp0.03-float64-orb-v3-con-inf-mpa.json.gz pred_file_url: https://figshare.com/files/53461061 κ_SRE: 0.1258 geo_opt: pred_file: models/orb/orb-v3/2025-04-04-wbm-geo-opt-orb-v3-con-inf-mpa.jsonl.gz pred_file_url: https://figshare.com/files/53461040 struct_col: orb_structure symprec=1e-5: rmsd: 0.075 # unitless n_sym_ops_mae: 10.169 # unitless symmetry_decrease: 0.869 # fraction symmetry_match: 0.130 # fraction symmetry_increase: 0.0001 # fraction n_structures: 256963 # count analysis_file: models/orb/orb-v3/2025-04-04-wbm-geo-opt-symprec=1e-5-moyo=0.4.2-orb-v3-con-inf-mpa.csv.gz analysis_file_url: https://figshare.com/files/53461055 symprec=1e-2: rmsd: 0.075 # unitless n_sym_ops_mae: 2.609 # unitless symmetry_decrease: 0.159 # fraction symmetry_match: 0.728 # fraction symmetry_increase: 0.101 # fraction n_structures: 256963 # count analysis_file: models/orb/orb-v3/2025-04-04-wbm-geo-opt-symprec=1e-2-moyo=0.4.2-orb-v3-con-inf-mpa.csv.gz analysis_file_url: https://figshare.com/files/53461058 discovery: pred_file: models/orb/orb-v3/2025-04-04-wbm-IS2RE-orb-v3-con-inf-mpa.csv.gz pred_file_url: https://figshare.com/files/53461037 pred_col: e_form_per_atom_orb full_test_set: F1: 0.887 # fraction DAF: 5.159 # dimensionless Precision: 0.885 # fraction Recall: 0.888 # fraction Accuracy: 0.961 # fraction TPR: 0.888 # fraction FPR: 0.024 # fraction TNR: 0.976 # fraction FNR: 0.112 # fraction TP: 39145.0 # count FP: 5075.0 # count TN: 207796.0 # count FN: 4947.0 # count MAE: 0.023 # eV/atom RMSE: 0.076 # eV/atom R2: 0.82 # dimensionless missing_preds: 2 # count most_stable_10k: F1: 0.982 # fraction DAF: 6.307 # dimensionless Precision: 0.964 # fraction Recall: 1.0 # fraction Accuracy: 0.964 # fraction TPR: 1.0 # fraction FPR: 1.0 # fraction TNR: 0.0 # fraction FNR: 0.0 # fraction TP: 9642.0 # count FP: 358.0 # count TN: 0.0 # count FN: 0.0 # count MAE: 0.021 # eV/atom RMSE: 0.08 # eV/atom R2: 0.86 # dimensionless missing_preds: 0 # count unique_prototypes: F1: 0.905 # fraction DAF: 5.912 # dimensionless Precision: 0.904 # fraction Recall: 0.907 # fraction Accuracy: 0.971 # fraction TPR: 0.907 # fraction FPR: 0.018 # fraction TNR: 0.982 # fraction FNR: 0.093 # fraction TP: 30261.0 # count FP: 3222.0 # count TN: 178892.0 # count FN: 3113.0 # count MAE: 0.024 # eV/atom RMSE: 0.078 # eV/atom R2: 0.821 # dimensionless missing_preds: 0 # count