Project collaborators
Garrett Merz and Kyle Cranmer, Data Science Institute
Tianji Cai and Lance Dixon, SLAC National Accelerator Laboratory
François Charton and Niklas Nolte, FAIR Meta
Matthias Wilhelm, Niels Bohr Institute
Project start and end dates
3/2023-3/2026
Project summary
“We’re beginning to think this is the kind of [work] where we can help physicists by discovering or suggesting properties that seem to appear but are difficult to spot, because we’re not used to looking at this data. Transformers work in high dimensions, so we know they can do better than us, and if we can find those regularities, we can potentially help the physicists.” – François Charton
Scattering amplitudes — the mathematical quantities at the heart of many high-energy physics calculations — are major drivers of theoretical uncertainties in high-energy physics experiments such as those occurring at the Large Hadron Collider. However, increasing the precision of these calculations requires performing increasingly complex integrals and, after a certain point, these become too difficult to solve under the current paradigm. As part of a joint award with SLAC National Accelerator Laboratory, and in collaboration with researchers from Meta and the Niels Bohr Institute in Copenhagen, we explore the use of transformers — deep learning models that have revolutionized fields in AI from natural language processing to computer vision — in order to perform these calculations numerically. To develop the formalism of this approach, we work in a “toy universe” theory where integrals have special properties that “real-world” calculations do not — namely, that the proposed form of an amplitude must obey a complex set of rules that makes it difficult to guess, but easy to check for correctness. Thus, an AI model can generate a potential solution, which can then be checked analytically. In the future, this approach may be extensible to quantum chromodynamics — the theory that describes the strong nuclear force that binds atomic nuclei. This project, which is intended to run for at least three years, will deepen DSI’s connections with the theoretical physics and machine learning research communities.
Project outputs/deliverables
Primary deliverables so far:
- Publication: Currently under review. Preprint on ArXiv
- NeurIPS ML for the Physical Sciences paper and poster
Additional Talks:
- Harvard New Technologies in Mathematics Seminar: François Charton
- Origins Data Science Lab (Munich) Seminar: Garrett Merz
- SLAC Seminar: Tianji Cai
- N=4 Super-Yang-Mills Workshop, Simons Center SUNY Stony Brook: Tianji Cai
- Machine Learning at SLAC Seminar: Garrett Merz
- Hammers and Nails Workshop: Garrett Merz
- NeurIPS AI for Science Keynote: Kyle Cranmer
- CERN LPCC IML Working Group: François Charton
Forthcoming:
- IAIFI Workshop Invited Talk: Tianji Cai
- Amplitudes 2024 (IAS) : François Charton
- IAIFI Workshop Poster: Garrett Merz