Authors: Ilyes Batatia, David P. Kovacs, Gregor N. C. Simm, Christoph Ortner, Gabor Csanyi (University of Cambridge & University of British Columbia) · arXiv: 2206.07697
Constructs 4-body (and higher-order) equivariant atomic features via spherical harmonic tensor products and Clebsch-Gordan coefficients in a single local atomic neighborhood, reducing required message-passing depth from 5-6 layers to just 2 layers.
Key Quantitative & Theoretical Takeaway: Higher body-order equivariant messages resolve local geometric degeneracies directly, enabling parallelizable multi-GPU molecular dynamics at ab-initio DFT accuracy.
Creating fast and accurate force fields is a long-standing challenge in computational chemistry and materials science. We introduce MACE, an equivariant message passing neural network architecture that uses higher body order messages to achieve state-of-the-art accuracy with only 2 message passing layers.
MACE first sums 2-body spherical harmonic projection features over neighbors to form the atomic basis A_i, and then takes symmetrized tensor products of A_i with itself using generalized Clebsch-Gordan coefficients to obtain nu-body features in O(N_neighbors) cost.
Replace arxiv.org with nujan.app on any arXiv URL (for example, https://nujan.app/abs/2006.11239) to open the PDF and AI research partner side-by-side.