A new AI architecture advancing the frontier of chemical simulation
Mirror Physics and MIT announce EquiformerV3.
Today, the team at Mirror is proud to introduce EquiformerV3: the next generation of the SE(3)-equivariant graph Transformer architecture, and a family of the world’s most accurate AI models for chemical physics, developed in collaboration with the Atomic Architects Group of Professor Tess Smidt at MIT. Across domains of chemistry from materials physics to catalysis, EquiformerV3 models are significantly more accurate and training-efficient than the current state of the art, constituting a substantial advance in AI capabilities in chemical simulation.
EquiformerV3 obtains a new state of the art on the most competitive benchmark in chemical physics, Matbench Discovery. Our top model, EquiformerV3+DeNS-OAM, achieves the highest performance on the leaderboard while using less than a third of the training compute of the next-best methods (3.6x less training compute than EPFL’s PET-OAM-XL; 22.6x less than Meta’s UMA-M-1.11). On Matbench’s “compliant” evaluation, which constrains all competing models to use the same training dataset, EquiformerV3+DeNS-MP improves over the previous state of the art set by Meta’s eSEN in all metrics - including a leaping 19% improvement in thermal conductivity prediction (κSRME), an especially difficult extrapolative measure of physical accuracy.
For precision-critical chemistry, EquiformerV3 provides leading performance. In heterogeneous catalysis, accurately predicting chemical adsorption energies across candidate material surfaces can dramatically reduce the amount of experimental effort required in industrial R&D, with every increment of computational accuracy yielding better ability to prioritize experimental work. Trained on Open Catalyst 2020 (OC20) 2M, a standard chemical physics dataset for catalysis, EquiformerV3 makes strides - netting 17% more accurate energy predictions and 8% more accurate force predictions while using 40% less compute in training than the previous state of the art.
In the coming months, we’ll be releasing checkpoints trained on more data, including the Open Molecules 2025 (OMol25) dataset, which we expect to provide extremely high quality simulation of organic chemistry and biochemistry. Checkpoints under an MIT license are available at huggingface.com/mirror-physics/equiformer_v3.
Read the paper: arxiv.org/abs/2604.09130v1
The UMA models are designed to address many domains of chemistry beyond solid-state materials physics, complicating this comparison: EquiformerV3+DeNS-OAM is not trained on data containing charge and spin labels, for example. We highlight that despite the much larger amount of diverse training data used to pre-train UMA-M-1.1, EquiformerV3+DeNS-OAM obtains better results on Matbench Discovery.
 | A guest post by
|