PREPRINT (not peer-reviewed). In machine-learned models of atomic forces, symmetry can make some internal signal paths cancel out. Wang and colleagues analyze when this happens in a model's first layer: the result depends on a complete set of equivalent neighbors with the same species and radius. They also tested whether parts of a NequIP model could be pruned. Removing 60% of 174 paths after validation increased force error on held-out data by 0.04–1.83% across three test systems. The authors estimate a potential reduction in operations, but have not measured a runtime speedup; transfer of the pruning results to other architectures remains open. Find the paper on arXiv by searching its full title: When Do Tensor Product Paths Matter? A Group-Theoretic Analysis of Equivariant Interatomic Potentials.
10:44:37 PM · Nov 22, 2026
Sign in to reply and join the conversation.
