Learning Molecular Symmetry Breaking via Symmetry-adapted Neural Networks
Keywords: relaxed equivariance, equivariance, symmetry breaking, geometric deep learning, molecular modeling
TL;DR: We propose a new symmetry-adapted neural network model to learn molecular symmetry breaking.
Abstract: E(3)-equivariant neural networks have achieved remarkable performance in molecular modeling. However, the equivariance constraint limits the model's effectiveness in learning tasks involving symmetry breaking, particularly those that violate the celebrated Curie principle. Relaxing the equivariance constraint is essential for addressing these challenges. In this paper, we explore the intricate symmetry relationships between an object and its spontaneously symmetry-broken outcomes. We introduce a relaxed equivariance based on the molecule's inherent symmetries. Additionally, we develop SANN -- a symmetry-adapted neural network architecture that learns symmetry breaking through equivalence classes of atoms. SANN decomposes the molecular point cloud into sets of symmetry-equivalent atoms and performs message-passing both within and across these classes. We demonstrate the advantages of our method over invariant and equivariant models through synthetic tasks and show that SANN effectively learns both equivariance and symmetry breaking in various benchmark molecular modeling tasks.
Supplementary Material: zip
Primary Area: learning on graphs and other geometries & topologies
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11169
Loading