Software

The Department is also the home of the MACE software, developed in collaboration with Ilyes Batatia. MACE is a mildly nonlinear equivariant message passing neural network (Batatia et al. Neurips 2022) that unites the geometric description of local neighbourhoods defined by the Atomic Cluster Expansion (ACE, Drautz 2019), and the equivariant message passing idea of Cormorant (Kondor, 2019) and Nequip (Kozinsky, 2022). One of many equivariant message passing network architectures, MACE is notable for its smoothness (or more precisely, regularity), and because it was the first model in which the wide out-of-domain generalisation was discovered that lead to so-called “foundation models”.