When training EquiformerV2 (31M) on only MPtrj we found that adding an auxiliary denoising objective (see DeNS arxiv.org/abs/2403.09549) substantially improved performance due to effective data augmentation and by preventing overfitting on a smaller dataset. 5/x

Our EquiformerV2 (86M) model pre-trained on OMat and fine-tuned on MPtraj and a subset of Alexandria, is state-of-the-art on Matbench discovery across all metrics with an F1 score > 0.9 and an accuracy of 20 meV/atom. 4/x

OMat24 contains over 100 million Density Functional Theory calculations focused on structural and compositional diversity. The force and stress distributions are substantially more broad than other open datasets for training ML potentials. 2/x

Introducing Meta’s Open Materials 2024 (OMat24) Dataset and Models! All under permissive open licenses for commercial and non-commercial use! Paper: arxiv.org/abs/2410.12771 Dataset: huggingface.co/datasets/fair… Models: huggingface.co/fairchem/OMAT… 🧵1/x

Before using an adsorption energy model, one should be aware of surface reconstructions that can impact results. Alternatively, total energy models are more robust models to surface reconstructions that still work on par with existing adsorption energy models

Perhaps the most notable impact on errors are surface reconstructions. This can pose an ill-defined adsorption energy, making it difficult for models to train and evaluate.

Interestingly, they also show that the performance of models trained on data with the tighter DFT settings is comparable to those trained with the original OC20 settings.

While tighter convergence settings have a considerable impact on DFT total energy; when they look at energy differences like the adsorption energy, they see significantly smaller convergence errors due to a cancellation of errors. Results shown for the OC20-200k dataset.

A DFT convergence study shows that there is room to improve the calculations for OC20 by tightening settings like the plane-wave cutoff, k-point sampling, and smearing width.

We have completely overhauled our documentation and tutorials to make finding the answers to your questions a lot easier. fair-chem.github.io/