Machine Learning in Lattice QCD with Fixed Point Actions

This award funds the research activities of Professor Kieran Holland at the University of the Pacific.

Decades of intensive experimental and theoretical work in particle physics has led to a deeper understanding of how nature behaves at the most fundamental level, with an intricate mathematical description of observed effects between subatomic particles, as well as raising many important and unanswered questions. Modern theoretical physics research takes advantage of a wide variety of computational resources and algorithms, which sometimes provide the only possible way to precisely use the underlying mathematical model, through what are known as “simulations” of how particles behave.

In the last few years, the booming field of artificial intelligence or more specifically machine learning has found new applications in science, where enormous datasets can be scoured to give more insight into scientific models. This research is important for the national interest, not only by continuing the nation’s leading role in advancing fundamental science, but also because artificial intelligence is a completely new tool which has already had an enormous impact and could revolutionize societies, much as personal computing has done in the last forty years.

By supporting research in this area, the US will continue to be one of the leaders of future technological developments and applications. There will also be a broader impact of the research, through Professor Holland’s mentoring of students in undergraduate research, developing their computing skills for future industry and academic careers, and by sharing the results of the research through public talks.

Being more specific, Professor Holland will work on Monte Carlo simulations of Quantum Chromodynamics (QCD), the fundamental model for interactions between quarks and gluons, binding them together to form nuclei. Simulations can be systematically improved, for example by using very high resolution when discretizing a physical system, analogous to pixelating an image, but the computational cost increases very rapidly, and can quickly demand more computing power than any supercomputer can provide.

An alternative is to find a more complex discretization of a physical system, where even a coarse resolution can be used to obtain correct physical predictions, an approach known as improved lattice actions. Professor Holland will work with machine learning algorithms and software platforms to design a highly improved lattice action for QCD, which he will also test through Monte Carlo simulations, to compare with existing lattice actions in terms of accuracy and computing cost.

Part of the project involves parametrization of an improved lattice action, a task ideally suited to machine learning methods. Professor Holland will first work on the dynamics of gluons without quarks, known as pure SU(3) Yang-Mills theory, where the overall computing cost is lower and can serve both as a testbed for the method, as well as have possible phenomenological impact for predicting the QCD coupling at the Electroweak energy scale, where Yang-Mills theory can connect to the full theory including quarks.

As many research groups perform lattice QCD simulations, this work is broadly relevant in exploring ways to maximize efficiency when using large-scale supercomputers to study the properties of QCD.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.