Relaxed Equivariant Graph Neural Networks
Elyssa Hofgard, Massachusetts Institute of Technology
Identifying symmetry breaking is crucial for understanding physical systems. For instance, symmetry breaking allows us to understand phase transitions in materials and the behavior of subatomic particles. It is thus desirable to have models that can identify and parametrize symmetry breaking. We introduce a versatile framework for relaxed E(3) graph equivariant neural networks capable of learning and representing symmetry breaking for continuous groups. By using non-scalar weights to relax strict equivariance constraints, we enable controlled symmetry breaking while ensuring the model retains the highest level of equivariance consistent with the input data. Our empirical results demonstrate that our formulation accurately learns the mathematical form of symmetry breaking parameters.