model_name: eqV2 S DeNS model_key: eqV2-s-dens-mp model_version: v2024.10.18 date_added: '2024-10-18' date_published: '2024-10-18' authors: - name: Luis Barroso-Luque affiliation: FAIR at Meta email: lbluque@meta.com orcid: https://orcid.org/0000-0002-6453-9545 github: https://github.com/lbluque corresponding: true - name: Muhammed Shuaibi affiliation: FAIR at Meta - name: Xiang Fu affiliation: FAIR at Meta - name: Brandon M. Wood affiliation: FAIR at Meta - name: Misko Dzamba affiliation: FAIR at Meta - name: Meng Gao affiliation: FAIR at Meta - name: Ammar Rizvi affiliation: FAIR at Meta - name: C.~Lawrence Zitnick affiliation: FAIR at Meta - name: Zachary W. Ulissi affiliation: FAIR at Meta email: zulissi@meta.com orcid: https://orcid.org/0000-0002-9401-4918 corresponding: true repo: https://github.com/FAIR-Chem/fairchem doi: https://doi.org/10.48550/arXiv.2410.12771 paper: https://arxiv.org/abs/2410.12771 url: https://huggingface.co/facebook/OMAT24#model-checkpoints pypi: https://pypi.org/project/fairchem-core pr_url: https://github.com/janosh/matbench-discovery/pull/146 # checkpoint URL copied from https://huggingface.co/fairchem/OMAT24#model-checkpoints checkpoint_url: https://huggingface.co/fairchem/OMAT24/blob/main/eqV2_dens_31M_mp.pt license: code: MIT code_url: https://github.com/FAIR-Chem/fairchem/blob/aa160789e1/LICENSE.md checkpoint: Meta Research checkpoint_url: https://huggingface.co/facebook/OMAT24/blob/main/LICENSE requirements: fairchem-core: 1.2.1 openness: OSOD trained_for_benchmark: true train_task: S2EFS test_task: IS2RE-SR targets: EFS_D model_type: UIP model_params: 31_207_434 # 31M n_estimators: 1 training_set: [MPtrj] training_cost: missing hyperparams: max_force: 0.02 max_steps: 500 ase_optimizer: FIRE cell_filter: FrechetCellFilter loss: MAE loss_weights: energy: 20 forces: 20 stress: 5 dens: 10 dens_probability: 0.5 noise_std: 0.1 optimizer: AdamW learning_rate_schedule: Cosine warmup_epochs: 0.1 warmup_factor: 0.2 max_learning_rate: 0.0002 min_learning_rate_factor: 0.01 grad_clip_threshold: 100 ema_decay: 0.999 weight_decay: 0.001 dropout_rate: 0.1 stochastic_depth: 0.1 batch_size: 512 epochs: 150 graph_construction_radius: 12.0 # Å max_neighbors: 20 # see table 7 in arXiv:2410.12771 notes: Description: | EquiformerV2 is an equivariant transformer that uses graph attention, attention re-normalization, and separable S^2 activations and layer normalization. Denoising Non-Equilibrium Structures (DeNS) uses a denoising generalized to structures with non-zero forces by encoding forces to obtain a well-posed denoising problem. Training: | Training was done from scratch using MPtrj only and a 50% probability for denoising training structures. metrics: phonons: kappa_103: κ_SRME: 1.6763 # eqV2 S without denoising (no DeNS) achieves slightly worse κ_SRME=1.772 pred_file: models/eqV2/eqV2-s-dens-mp/2024-11-08-kappa-103-FIRE-dist=0.01-fmax=1e-4-symprec=1e-5.json.gz pred_file_url: https://figshare.com/files/52151558 κ_SRE: 1.3866 geo_opt: pred_file: models/eqV2/eqV2-s-dens-mp/2024-10-18-wbm-geo-opt.jsonl.gz pred_file_url: https://figshare.com/files/53153468 struct_col: eqV2-86M-omat-salex-mp_structure symprec=1e-5: rmsd: 0.0757 # unitless n_sym_ops_mae: 10.1045 # unitless symmetry_decrease: 0.8671 # fraction symmetry_match: 0.1322 # fraction symmetry_increase: 0.0006 # fraction n_structures: 256614 # count analysis_file: models/eqV2/eqV2-s-dens-mp/2024-10-18-wbm-geo-opt-symprec=1e-5-moyo=0.4.2.csv.gz analysis_file_url: https://figshare.com/files/53504534 symprec=1e-2: rmsd: 0.0757 # unitless n_sym_ops_mae: 3.7426 # unitless symmetry_decrease: 0.3724 # fraction symmetry_match: 0.5396 # fraction symmetry_increase: 0.0653 # fraction n_structures: 256614 # count analysis_file: models/eqV2/eqV2-s-dens-mp/2024-10-18-wbm-geo-opt-symprec=1e-2-moyo=0.4.2.csv.gz analysis_file_url: https://figshare.com/files/53504540 discovery: pred_file: models/eqV2/eqV2-s-dens-mp/2024-10-18-wbm-IS2RE.csv.gz pred_file_url: https://figshare.com/files/52057568 pred_col: e_form_per_atom_eqV2-31M-dens-MP-p5 full_test_set: F1: 0.798 # fraction DAF: 4.362 # dimensionless Precision: 0.748 # fraction Recall: 0.855 # fraction Accuracy: 0.926 # fraction TPR: 0.855 # fraction FPR: 0.059 # fraction TNR: 0.941 # fraction FNR: 0.145 # fraction TP: 37687.0 # count FP: 12665.0 # count TN: 200206.0 # count FN: 6405.0 # count MAE: 0.035 # eV/atom RMSE: 0.084 # eV/atom R2: 0.785 # dimensionless missing_preds: 351 # count most_stable_10k: F1: 0.983 # fraction DAF: 6.326 # dimensionless Precision: 0.967 # fraction Recall: 1.0 # fraction Accuracy: 0.967 # fraction TPR: 1.0 # fraction FPR: 1.0 # fraction TNR: 0.0 # fraction FNR: 0.0 # fraction TP: 9670.0 # count FP: 330.0 # count TN: 0.0 # count FN: 0.0 # count MAE: 0.031 # eV/atom RMSE: 0.091 # eV/atom R2: 0.823 # dimensionless missing_preds: 0 # count unique_prototypes: F1: 0.815 # fraction DAF: 5.042 # dimensionless Precision: 0.771 # fraction Recall: 0.864 # fraction Accuracy: 0.939 # fraction TPR: 0.864 # fraction FPR: 0.047 # fraction TNR: 0.953 # fraction FNR: 0.136 # fraction TP: 28842.0 # count FP: 8576.0 # count TN: 173538.0 # count FN: 4532.0 # count MAE: 0.036 # eV/atom RMSE: 0.085 # eV/atom R2: 0.788 # dimensionless missing_preds: 309 # count