Multi-Modal Foundation · facebook
OMat24
Inorganic material property prediction models using EquiformerV2 and eSEN architectures.
What OMat24 is
OMat24 is a collection of machine learning potentials developed by Meta's FAIR Chemistry team for predicting inorganic material properties. The suite provides neural network models capable of evaluating energies, forces, and stresses for solid-state crystal structures and material simulations.
The models address the computational bottleneck of ab initio simulations by emulating density functional theory (DFT) calculations at significantly lower computational cost. Researchers can incorporate these checkpoints directly into standard atomistic modelling workflows, such as the Atomic Simulation Environment (ASE), to perform crystal structure relaxations and material discovery tasks.
How it works
OMat24 comprises models based on two primary neural network architectures: EquiformerV2 and eSEN. EquiformerV2 models come in three sizes—31M (S), 86M (M), and 153M (L) parameters—and are evaluated with standard training or with denoising augmentation objectives (DeNS). eSEN models contain 30M parameters and are initially pretrained using direct force prediction, followed by fine-tuning for gradient-based predictions.
Training and fine-tuning leverage several datasets of inorganic materials: the Open Materials 2024 (OMat24) dataset, the Materials Project Trajectory dataset (MPtrj), and sAlexandria (sAlex). Checkpoints are provided either trained exclusively on single datasets (OMat or MPtrj) or pre-trained on OMat and fine-tuned on combined MPtrj and sAlexandria data.
Key facts
- Architectures
- EquiformerV2 (eqV2), eSEN
- Parameter count
- 30M (eSEN), 31M (eqV2-S), 86M (eqV2-M), 153M (eqV2-L)
- Training datasets
- OMat24, MPtrj, sAlexandria
- Outputs
- Energy, direct/gradient forces, stress predictions
What you can use it for
- Crystal structure relaxation
- Execute unit cell and atomic coordinate relaxations using stress and force predictions via the Atomic Simulation Environment (ASE).
- Inorganic material property screening
- Predict total energies and forces for novel crystal structures across high-throughput inorganic material discovery campaigns.
- Interatomic potential evaluation
- Use gradient-based and direct-force predictions as fast surrogate potentials to assess thermodynamic stability and trajectory points.
Run it here
We list this model but cannot run it yet: gated on the Hub — never run past a gate the user did not accept.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Provides varied model capacities ranging from 30M to 153M parameters to balance speed and predictive performance.
- Integrates into standard computational chemistry pipelines through fairchem's custom ASE-calculator interface.
- Supports full cell relaxation through automated stress tensor predictions.
Watch out for
- Total energy labels from OMat24 models are derived from DFT and DFT+U calculations that are not directly compatible with Materials Project standard calculations.
- Comparing energy differences (such as formation energies or energy above hull) to Materials Project data requires specific unary reference compounds and MP2020-style mixing corrections due to differing pseudopotentials and magnetic ground states.
- OMat-only checkpoints require specific fairchem correction schemes rather than direct MP2020 correction parameters.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · other
This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Questions
Can I use OMat24 commercially?
Check the licence before commercial use. This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Can OMat24 energies be compared directly with Materials Project values?
No. Models trained on OMat24 use distinct DFT and DFT+U pseudopotentials and may feature different magnetic ground states. To calculate formation energies or energies above the convex hull relative to the Materials Project, researchers must apply MP2020-style corrections and unary references provided in the fairchem repository rather than using Materials Project references directly.
Which model architectures are available in OMat24?
The suite provides EquiformerV2 models in three scales (31M, 86M, and 153M parameters) trained with or without denoising objectives (DeNS), alongside 30M-parameter eSEN models trained with direct force pre-training and gradient-based fine-tuning.
How are cell relaxations performed with these checkpoints?
Relaxations can be run using the custom OCPCalculator within the Atomic Simulation Environment (ASE) alongside optimisers such as FIRE and cell filters like FrechetCellFilter, relying on the model's predicted stresses and forces.
How to cite it
Luis Barroso-Luque, Muhammed Shuaibi, Xiang Fu, Brandon M. Wood, Misko Dzamba, Meng Gao, Ammar Rizvi, C. Lawrence Zitnick, Zachary W. Ulissi. Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models. arXiv:2410.12771 [cond-mat.mtrl-sci], 2024.
Written from OMat24's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.