model_name: Nequip-OAM-L # required (this must match the model's label which is the 3rd arg in the matbench_discovery.preds.Model enum) model_key: nequip-OAM-L-0.1 # this should match the name of the YAML file and determines the URL /models/ on which details of the model are displayed on the website model_version: '0.1' date_added: '2025-09-08' date_published: '2025-08-28' authors: - name: Seán R. Kavanagh affiliation: Center for the Environment, Harvard University & MIR Group, Harvard University email: skavanagh@seas.harvard.edu orcid: https://orcid.org/0000-0003-4577-9647 url: https://seankavanagh.com corresponding: true - name: Chuin Wei Tan affiliation: MIR Group, Harvard University - name: Albert Musaelian affiliation: MIR Group, Harvard University & Mirian Technologies - name: William C. Witt affiliation: MIR Group, Harvard University - name: Gabriel de Miranda Nascimento affiliation: MIR Group, Harvard University & MIT - name: Ulrik Unneberg affiliation: MIR Group, Harvard University & MIT - name: Marc L. Descoteaux affiliation: MIR Group, Harvard University - name: Boris Kozinsky affiliation: MIR Group, Harvard University trained_by: - name: Seán R. Kavanagh affiliation: Center for the Environment, Harvard University & MIR Group, Harvard University email: skavanagh@seas.harvard.edu orcid: https://orcid.org/0000-0003-4577-9647 url: https://seankavanagh.com corresponding: true repo: https://github.com/mir-group/nequip url: https://nequip.readthedocs.io/en/latest/ doi: https://doi.org/10.5281/zenodo.16980200 paper: https://arxiv.org/abs/2504.16068 # To be updated... checkpoint_url: https://www.nequip.net/models/mir-group/NequIP-OAM-L:0.1 pr_url: https://github.com/janosh/matbench-discovery/pull/282 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: 9_600_000 trained_for_benchmark: true n_estimators: 1 license: code: MIT code_url: https://github.com/mir-group/nequip/blob/main/LICENSE checkpoint: CC-BY-4.0 checkpoint_url: https://creativecommons.org/licenses/by/4.0/legalcode 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 weight_decay: 1e-8 graph_construction_radius: 6.0 sph_harmonics_l_max: 3 n_layers: 6 n_features: 128 (l=0 scalars), 64 (l=1 vectors), 32 (l=2,3 tensors) parity: false zbl_potential: true type_embed_num_features: 48 polynomial_cutoff: 5 n_radial_bessel_basis: 8 loss: Huber - delta=0.01 for energy, delta=0.1 for stress, stratified delta (0.01, 0.007, 0.004, 0.001) for force loss_weights: energy: 1.0 force: 5.0 stress: 0.01 batch_size: 640 # 4 (gpus) * 160 (batch per gpu) = 640 (total batch size) initial_learning_rate: 0.005 gradient_clip_val: 1 learning_rate_schedule: ReduceLROnPlateau - factor=0.1, patience=10, min_lr=1e-6 epochs: 30 max_neighbors: .inf training_cost: Nvidia H100 GPUs: { amount: 4, hours: 100, cost: 400 } requirements: torch: '2' nequip: '>=0.7.0' # >=0.14.0 recommended training_set: [OMat24, sAlex, MPtrj] notes: Description: Large NequIP foundation potential; see for details and for model/training infrastructure. Steps: | Training performed by: (1) pre-training on OMat24; (2) fine-tuning on MPtrj+sAlex, with a reduced learning rate (1e-4), energy-loss-upweighting (1:1:0.01 instead of 1:5:0.01) and `StochasticWeightAveraging` (SWA). metrics: phonons: kappa_103: pred_file: models/nequip/nequip-OAM-L-0.1/2025-08-27-kappa-103-FIRE-dist=0.03-fmax=0.0001-symprec=1e-05.json.gz pred_file_url: https://figshare.com/files/57475504 κ_SRME: 0.1657 κ_SRE: 0.0812 geo_opt: pred_file: models/nequip/nequip-OAM-L-0.1/2025-08-27-wbm-IS2RE-GOQN.jsonl.gz pred_file_url: https://figshare.com/files/57731005 struct_col: nequip_structure # same for NequIP/Allegro symprec=1e-5: rmsd: 0.0647 # unitless n_sym_ops_mae: 1.9352 # unitless symmetry_decrease: 0.0792 # fraction symmetry_match: 0.7125 # fraction symmetry_increase: 0.2018 # fraction n_structures: 256963 # count analysis_file: models/nequip/nequip-OAM-L-0.1/2025-08-27-wbm-IS2RE-GOQN-symprec=1e-5-moyo=0.4.4.csv.gz analysis_file_url: https://figshare.com/files/57731032 symprec=1e-2: rmsd: 0.0647 # unitless n_sym_ops_mae: 1.7483 # unitless symmetry_decrease: 0.0543 # fraction symmetry_match: 0.816 # fraction symmetry_increase: 0.1226 # fraction n_structures: 256963 # count analysis_file: models/nequip/nequip-OAM-L-0.1/2025-08-27-wbm-IS2RE-GOQN-symprec=1e-2-moyo=0.4.4.csv.gz analysis_file_url: https://figshare.com/files/57731035 discovery: pred_file: models/nequip/nequip-OAM-L-0.1/2025-08-27-wbm-IS2RE.csv.gz pred_file_url: https://figshare.com/files/57574462 pred_col: e_form_per_atom_nequip # same for NequIP/Allegro full_test_set: F1: 0.87 # fraction DAF: 5.06 # dimensionless Precision: 0.868 # fraction Recall: 0.872 # fraction Accuracy: 0.955 # fraction TPR: 0.872 # fraction FPR: 0.027 # fraction TNR: 0.973 # fraction FNR: 0.128 # fraction TP: 38456.0 # count FP: 5835.0 # count TN: 207036.0 # count FN: 5636.0 # count MAE: 0.022 # eV/atom RMSE: 0.068 # eV/atom R2: 0.858 # dimensionless missing_preds: 2 # count unique_prototypes: F1: 0.893 # fraction DAF: 5.823 # dimensionless Precision: 0.89 # fraction Recall: 0.895 # fraction Accuracy: 0.967 # fraction TPR: 0.895 # fraction FPR: 0.02 # fraction TNR: 0.98 # fraction FNR: 0.105 # fraction TP: 29867.0 # count FP: 3684.0 # count TN: 178430.0 # count FN: 3507.0 # count MAE: 0.022 # eV/atom RMSE: 0.068 # eV/atom R2: 0.865 # dimensionless missing_preds: 0 # count most_stable_10k: F1: 0.985 # fraction DAF: 6.344 # dimensionless Precision: 0.97 # fraction Recall: 1.0 # fraction Accuracy: 0.97 # fraction TPR: 1.0 # fraction FPR: 1.0 # fraction TNR: 0.0 # fraction FNR: 0.0 # fraction TP: 9698.0 # count FP: 302.0 # count TN: 0.0 # count FN: 0.0 # count MAE: 0.021 # eV/atom RMSE: 0.083 # eV/atom R2: 0.854 # dimensionless missing_preds: 0 # count