model_name: TACE-v1-OAM-M model_key: tace-v1-oam-m model_version: '0.1.0' date_added: '2026-01-06' date_published: '2026-01-06' authors: - name: Zemin Xu affiliation: ShanghaiTech University, Nanjing University email: xv_chana@foxmail.com github: https://github.com/xvzemin corresponding: false - name: Wenbo Xie affiliation: ShanghaiTech University email: xiewb1@shanghaitech.edu.cn orcid: https://orcid.org/0000-0001-5321-2494 corresponding: true - name: P. Hu affiliation: ShanghaiTech University email: hupj@shanghaitech.edu.cn orcid: https://orcid.org/0000-0002-6318-1051 corresponding: true trained_by: - name: Zemin Xu affiliation: ShanghaiTech University, Nanjing University email: xv_chana@foxmail.com github: https://github.com/xvzemin corresponding: false repo: https://github.com/xvzemin/tace checkpoint_url: https://huggingface.co/xvzemin/tace-foundations/resolve/main/TACE-v1-OAM-M.pt url: https://github.com/xvzemin/tace doi: https://doi.org/10.57967/hf/7458 paper: https://arxiv.org/abs/2509.14961 pr_url: https://github.com/janosh/matbench-discovery/pull/301 openness: OSOD # see `Open` enum in matbench_discovery/enums.py train_task: S2EFS # see `Task` enum in matbench_discovery/enums.py test_task: IS2RE-SR # see `Task` enum in matbench_discovery/enums.py targets: EFS_G # see `Targets` enum in matbench_discovery/enums.py model_type: UIP # see `ModelType` enum in matbench_discovery/enums.py model_params: 18_779_248 trained_for_benchmark: false status: superseded superseded_by: tace-oam-l n_estimators: 1 license: code: MIT code_url: https://github.com/xvzemin/tace/blob/main/LICENSE.md checkpoint: CC-BY-4.0 # URL that points to the license file for the model checkpoint, not the checkpoint file itself. checkpoint_url: https://creativecommons.org/licenses/by/4.0/legalcode hyperparams: # strongly recommended to list relaxation hyperparams max_force: 0.05 max_steps: 500 ase_optimizer: GOQN # faster than FIRE with same results; see SI of https://doi.org/10.1088/2515-7655/ade916 cell_filter: FrechetCellFilter optimizer: AdamW graph_construction_radius: 6.0 max_neighbors: .inf epochs: 4 # 4 for OMat24, 2 for sAlex+MPtrj loss_weights: energy: 1 force: 8 stress: 8 huber_delta: 0.01 initial_learning_rate: 0.004 batch: 512 # OMat24: 512, sAlex+MPtrj: 256 num_channel: 48 # same for every l Lmax: 2 lmax: 3 l1l2: <= # when tp, only l1 < l2 are allowed num_layers: 5 correlation: 2 training_cost: # list any hardware used to train the model and for how long Nvidia H100 GPUs: { amount: 8, hours: 120, cost: 960 } requirements: # strongly recommended python: '3.12.11' torch: '2.9.1' torch-geometric: '2.7.0' pytorch-lightning: '2.5.5' tace: '>=0.1.0' # recommended newest github commit training_set: [OMat24, sAlex, MPtrj] notes: Description: | Based on irreducible Cartesian tensors, described in https://arxiv.org/abs/2509.14961 and https://arxiv.org/abs/2512.16882. Steps: | Training performed by: (1) pre-training on OMat24 (lr-4e-3, loss-1:8:8); (2) fine-tuning on MPtrj+sAlex (lr-1e-4, loss-1:1:2). metrics: phonons: kappa_103: pred_file: models/tace/tace-v1-oam-m/2026-01-06-kappa-103-FIRE-dist=0.03-fmax=0.0001-symprec=1e-05.json.gz pred_file_url: https://figshare.com/files/60807073 # # OMat24, before finetune # κ_SRME: 0.1575 # κ_SRE: 0.0699 # # sAlex + MPtrj, after finetune # κ_SRME: 0.1729 # κ_SRE: 0.0765 κ_SRME: 0.1729 κ_SRE: 0.0765 geo_opt: pred_file: models/tace/tace-v1-oam-m/2026-01-06-wbm-IS2RE-GOQN.jsonl.gz pred_file_url: https://figshare.com/files/60806854 struct_col: tace_structure symprec=1e-2: rmsd: 0.0651 # unitless n_sym_ops_mae: 1.7372 # unitless symmetry_decrease: 0.0544 # fraction symmetry_match: 0.8167 # fraction symmetry_increase: 0.1219 # fraction n_structures: 256963 # count analysis_file: models/tace/tace-v1-oam-m/2026-01-06-wbm-IS2RE-GOQN-symprec=1e-2-moyo=0.7.4.csv.gz analysis_file_url: https://figshare.com/files/60807367 symprec=1e-5: rmsd: 0.0651 # unitless n_sym_ops_mae: 1.828 # unitless symmetry_decrease: 0.0559 # fraction symmetry_match: 0.7327 # fraction symmetry_increase: 0.2061 # fraction n_structures: 256963 # count analysis_file: models/tace/tace-v1-oam-m/2026-01-06-wbm-IS2RE-GOQN-symprec=1e-5-moyo=0.7.4.csv.gz analysis_file_url: https://figshare.com/files/60807373 discovery: pred_file: models/tace/tace-v1-oam-m/2026-01-06-wbm-IS2RE.csv.gz # tace-filtered_preds.csv.gz tace-preds.csv.gz pred_file_url: https://figshare.com/files/60806971 pred_col: e_form_per_atom_tace full_test_set: F1: 0.868 # fraction DAF: 4.986 # dimensionless Precision: 0.855 # fraction Recall: 0.881 # fraction Accuracy: 0.954 # fraction TPR: 0.881 # fraction FPR: 0.031 # fraction TNR: 0.969 # fraction FNR: 0.119 # fraction TP: 38839.0 # count FP: 6561.0 # count TN: 206310.0 # count FN: 5253.0 # count MAE: 0.022 # eV/atom RMSE: 0.068 # eV/atom R2: 0.859 # dimensionless missing_preds: 2 # count most_stable_10k: F1: 0.986 # fraction DAF: 6.36 # dimensionless Precision: 0.972 # fraction Recall: 1.0 # fraction Accuracy: 0.972 # fraction TPR: 1.0 # fraction FPR: 1.0 # fraction TNR: 0.0 # fraction FNR: 0.0 # fraction TP: 9722.0 # count FP: 278.0 # count TN: 0.0 # count FN: 0.0 # count MAE: 0.021 # eV/atom RMSE: 0.077 # eV/atom R2: 0.872 # dimensionless missing_preds: 0 # count unique_prototypes: F1: 0.889 # fraction DAF: 5.749 # dimensionless Precision: 0.879 # fraction Recall: 0.9 # fraction Accuracy: 0.965 # fraction TPR: 0.9 # fraction FPR: 0.023 # fraction TNR: 0.977 # fraction FNR: 0.1 # fraction TP: 30042.0 # count FP: 4141.0 # count TN: 177973.0 # count FN: 3332.0 # count MAE: 0.022 # eV/atom RMSE: 0.068 # eV/atom R2: 0.865 # dimensionless missing_preds: 0 # count