Year: 2026
Type: preprint
Abstract: Computational research requires repeated decisions about models, calculations and the interpretation of results. We introduce Brilliance, an agent that combines large language models, reusable physics expertise and numerical simulation to construct and revise investigations in accelerator physics. Starting from a detailed expert-specified request, Brilliance develops an ideal storage-ring lattice meeting the specified emittance ceiling and sampled nonlinear-acceptance criteria after revising optics that fail particle-tracking tests. Further scientist-guided exploration produces a candidate with a natural emittance of 6.51 pm rad and millimeter-scale transverse acceptance, while exposing unresolved momentum-acceptance limitations. Using measured responses from the High Energy Photon Source, Brilliance also performs offline optics calibration consistent with an independent specialist analysis. An expert-module ablation examines how reusable domain knowledge influences research choices and outcomes. Here, we demonstrate a practical route to scientist-agent collaboration in sustained numerical research.
Source: arXiv (Cornell University)
Language: English
Subfield: Nuclear and High Energy Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
Award: G834293954
OpenAlex ID: W7221801703