Multi-Modal Foundation · mirror-physics
Equiformer v3
An SE(3)-equivariant graph transformer predicting energies and forces for molecules and crystalline materials.
What Equiformer v3 is
Equiformer v3 (also designated EquiformerV3) is an SE(3)-equivariant graph attention transformer developed by mirror-physics and researchers including Yi-Lun Liao, Alexander J. Hoffman, Sabrina C. Shen, Alexandre Duval, Sam Walton Norwood, and Tess Smidt. It is designed for 3D atomistic modeling across chemistry, physics, and materials science.
The model addresses the computational challenge of accurately predicting molecular and crystal properties, such as total energies and atomic forces, while respecting 3D spatial symmetries. By maintaining strict SE(3) equivariance—ensuring predictions transform predictably under three-dimensional rotations and translations—it models atomic systems with high physical fidelity.
Researchers can utilise Equiformer v3 to assess structural stability and atomic interactions in diverse chemical systems, supported by pre-trained checkpoints evaluated on standard benchmarks including OC20, OMat24, and Matbench Discovery.
How it works
Building upon the EquiformerV2 architecture, Equiformer v3 models atomistic structures as geometric graphs. It introduces architectural refinements including equivariant merged layer normalisation, graph attention mechanisms featuring smooth cutoffs, and SwiGLU-S^2 activation functions that capture many-body atomic interactions while strictly preserving equivariance.
Training strategies vary by checkpoint. Available weights include models trained directly on the MPtrj dataset using gradient training, as well as a sequential training pipeline: direct pre-training on OMat24, followed by gradient fine-tuning on OMat24, and subsequent gradient fine-tuning on the MPtrj and sAlex datasets.
Key facts
- Architecture
- SE(3)-equivariant graph attention transformer
- Primary training data
- OMat24, MPtrj, and sAlex
- Benchmarks evaluated
- OC20, OMat24, and Matbench Discovery
What you can use it for
- Crystal Property Prediction
- Predicting total energies and atomic forces across inorganic crystal structures trained on materials databases such as MPtrj and OMat24.
- Molecular and Materials Modelling
- Evaluating the physical properties, stability, and configurations of 3D atomistic and molecular systems using equivariant graph representations.
- Materials Discovery Screening
- Assessing candidate structures within discovery pipelines, such as those evaluated on the Matbench Discovery benchmark.
Run it here
We list this model but cannot run it yet: no library and no architectures — nothing can load it.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Maintains strict SE(3) equivariance under 3D spatial rotations and translations.
- Captures complex many-body atomic interactions through SwiGLU-S^2 activation functions.
- Incorporates equivariant merged layer normalisation and attention mechanisms with smooth cutoffs to improve modeling efficiency and expressivity.
- Provides fine-tuned checkpoints trained across multiple large materials datasets including OMat24, MPtrj, and sAlex.
Watch out for
- Specialised for 3D atomic structures and requires coordinate-based geometric graph inputs.
- Requires matching configuration files and environment dependencies specific to equivariant tensor operations.
Inputs and outputs
You provide
Graph
Node and edge data in the format this model expects — check its model card.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · mit
The mit licence carries terms worth reading before you depend on it.
Questions
Can I use Equiformer v3 commercially?
Check the licence before commercial use. The mit licence carries terms worth reading before you depend on it.
What is the primary difference between Equiformer v3 and its predecessor?
Equiformer v3 introduces software efficiency optimizations, architectural modifications such as equivariant merged layer normalisation and attention with smooth cutoffs, and SwiGLU-S^2 activations that integrate many-body interactions whilst retaining strict SE(3) equivariance.
Which datasets were used to train the available checkpoints?
Checkpoints are provided for MPtrj alone via gradient training, as well as an OMat24 pre-trained model followed by gradient fine-tuning on OMat24 and further gradient fine-tuning on combined MPtrj and sAlex data.
What properties can Equiformer v3 predict?
The model is designed to predict energies, forces, and related physical properties for crystalline materials and molecular structures.
How to cite it
Yi-Lun Liao, Alexander J. Hoffman, Sabrina C. Shen, Alexandre Duval, Sam Walton Norwood, Tess Smidt. EquiformerV3: Scaling Efficient, Expressive, and General SE(3)-Equivariant Graph Attention Transformers. arXiv preprint arXiv:2604.09130, 2026.
Written from Equiformer v3'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.