Steerable E(3)-Equivariant Graph Neural Networks

A new publication by Johannes Brandstetter, Rob Hesselink, Elise van der Pol, Erik J Bekkers and Max Welling introduces the steerable E(3)-equivariant graph neural network, or SEGNN. This model uses irreducible representations of the orthogonal group O(3) to condition messages on relative positions and other geometric attributes and achieves excellent results on QM9 and the Open Catalyst Project. Accompanying the paper is a blog post that explains how to build O(3) steerable networks and showcases some of the benefits of equivariant message passing. The paper was awarded a spotlight at ICLR 2022.