Symmetry-preserving neural networks in lattice field theories
This work addresses the challenge of incorporating symmetries into neural networks for physicists and researchers in lattice field theory, offering a domain-specific improvement.
The paper tackles the problem of applying neural networks to lattice field theory by designing symmetry-preserving architectures, specifically Lattice Gauge Equivariant Convolutional Neural Networks (L-CNNs), which successfully regress physical observables like Wilson loops and outperform traditional non-symmetric architectures.
This thesis deals with neural networks that respect symmetries and presents the advantages in applying them to lattice field theory problems. The concept of equivariance is explained, together with the reason why such a property is crucial for the network to preserve the desired symmetry. The benefits of choosing equivariant networks are first illustrated for translational symmetry on a complex scalar field toy model. The discussion is then extended to gauge theories, for which Lattice Gauge Equivariant Convolutional Neural Networks (L-CNNs) are specifically designed ad hoc. Regressions of physical observables such as Wilson loops are successfully solved by L-CNNs, whereas traditional architectures which are not gauge symmetric perform significantly worse. Finally, we introduce the technique of neural gradient flow, which is an ordinary differential equation solved by neural networks, and propose it as a method to generate lattice gauge configurations.