This project aims to enable calculations of nuclear structure and reactions that are currently computationally intractable, such as the interactions between nuclei and a large class of dark matter candidate particles. By coupling advanced machine learning and state-of-the-art physics simulations, it will provide critical input for experimental searches aiming to unravel the mysteries of dark matter while simultaneously giving insight into fundamental particle physics.

The common method of studying such physical processes is lattice quantum chromodynamics (QCD). This method simulates particle interactions using the generation of random configurations for some physical distribution. In the project, this approach is enforced with machine learning techniques that have proven successful in generating samples for other domains, such as images. However, unlike for image generation, the scientific study must guarantee the ​“correctness” of new knowledge. In order to satisfy this requirement, we used a special architecture of neural networks, namely normalizing flows.

We built high-quality models to incorporate physical symmetries. Much like incorporating translation symmetries helped build convolutional neural networks and drastically improve performance, incorporating physical ​“gauge symmetry” improved the performance by an order of magnitude.

As a result it was possible to build models for some lattice QCD distributions that outperform traditional state-of-the-art algorithms. Following this proof-of-principle success, next steps involve the development of models for theories with fermions, four-dimensional lattice gauge theories, and scaling to bigger lattices. Through the Aurora Early Science Program, we are preparing to make use of Aurora, the coming exascale supercomputer, as soon as it is available