New AI model TECE-OAM-RRA-1.0 achieves SOTA in atomic potential predictions
Radial Rotary Attention lets AI predict atomic interactions 40% more accurately than prior methods.
A team led by Zemin Xu, Wenbo Xie, and P. Hu has published a paper introducing a novel approach to machine learning interatomic potentials (MLIPs). They systematically investigate SO(2) theory limitations relative to SO(3) Clebsch-Gordan Tensor Products, and propose two innovations. First, the Edge Complex Product Basis uses Generalized Asymmetric Contraction to directly build higher-order atomic interactions on edges via complex-valued equivariant multiplications. Second, Radial Rotary Complex Attention (RRA) enhances extrapolation performance beyond existing attention mechanisms. The model, TECE-OAM-RRA-1.0, also incorporates improvements to the Atomic Cluster Expansion module.
Training on three major materials datasets—OMat24, sAlex, and MPTrj—the team demonstrates state-of-the-art performance on the Matbench Discovery benchmark. This breakthrough enables more accurate predictions of material properties like formation energy and band gaps, which are critical for designing new catalysts, batteries, and semiconductors. By improving both accuracy and extrapolation, the method promises to accelerate computational materials discovery.
- Introduces Edge Complex Product Basis and Radial Rotary Complex Attention (RRA), outperforming conventional SO(3) tensor product methods.
- Model TECE-OAM-RRA-1.0 achieves state-of-the-art results on Matbench Discovery benchmark.
- Trained on three large datasets (OMat24, sAlex, MPTrj) covering diverse inorganic crystals and molecular dynamics.
Why It Matters
Enables faster, more accurate screening of materials for batteries, solar cells, and catalysts directly from atomic structure.