Incorporating gauge-invariance in equivariant networks
Lists
Only you see this
Abstract
Gauge theories, which describe fundamental forces in nature, arise from the principle of locality in physical interactions. These theories are characterized by their invariance under local symmetry transformations and the presence of a gauge field that mediates interactions. While recent works have introduced gauge equivariant neural networks, these models often focus on specific cases like tangent bundles or quotient spaces, limiting their applicability to the diverse gauge theories in physics. We propose a novel architecture for learning general gauge invariant quantities by explicitly modeling the gauge field in the context of graph neural networks. Our framework fills a critical gap in the existing literature by providing a general recipe for gauge invariance without restrictions on the fiber spaces. This approach allows for the modeling of more complex gauge theories, such as those with $SU(N)$ gauge groups, which are prevalent in particle physics. We evaluate our method on classical physical systems, including the XY model on various curved geometries, demonstrating its ability to capture gauge invariant properties in settings where existing equivariant architectures fall short. Our work takes a significant step towards bridging the gap between gauge theories in physics and equivariant neural network architectures, opening new avenues for applying machine learning to fundamental physical problems.
BibTeX
@misc{theodosis2025incorporating,
author = {Manos Theodosis and Demba E. Ba and Nima Dehmamy},
title = {{Incorporating gauge-invariance in equivariant networks}},
year = {2025},
url = {https://openreview.net/forum?id=YAINolpm8n},
howpublished = {OpenReview},
note = {Submitted to ICLR 2025},
}