Model for ALIGNN-FF Forces¶
- Description: This is an AI benchmark to evaluate how accurately a machine learning force-field (MLFF) can predict forces using the relaxation trajectories (energy and forces of intermediate steps) of the JARVIS-DFT (dft_3d) dataset, calculated with the OPTB88vdw density functional. The dataset contains different types of chemical formula and atomic structures. Here we use multi-mean absolute error (multi-MAE) to compare MLFFs with respect to DFT (OPT) accuracy.
Reference(s): https://github.com/materialsvirtuallab/m3gnet, https://pubs.rsc.org/en/content/articlehtml/2023/dd/d2dd00096b, https://doi.org/10.1039/D2DD00096B
Model benchmarks
| Model name | Dataset | Accuracy | Team name | Dataset size | Date submitted | Notes |
|---|---|---|---|---|---|---|
| m3gnet_pretrained | alignn_ff_db | 0.08448789649784973 | M3GNET | 307111 | 01-14-2023 | CSV, JSON, run.sh, Info |
| alignnff_pretrained_wt10 | alignn_ff_db | 0.5782337321792528 | JARVIS | 307111 | 01-14-2023 | CSV, JSON, run.sh, Info |