Speaker
Description
An integrated sampling and optimization framework is introduced for generating cavity gradient configurations under strict energy constraints in a CEBAF digital twin. A constraint-aware iterative projection sampler produces feasible gradient vectors by adjusting random samples toward the 1050 MeV energy target while preserving per-cavity operating limits. Corrections are assigned according to each cavity’s available gradient margin, enabling a diverse and energy-consistent set of initial configurations.
A multi-objective reinforcement learning environment is then constructed to optimize beam energy accuracy, RF heat load, and cavity trip probability. A MOTD3-based agent with vector-valued critics and randomized preference weighting explores the high-dimensional control space and identifies Pareto-optimal trade-offs. Physics-informed penalties restrict deviations from the permitted energy band, guiding the learning process toward operationally viable and thermally stable configurations.
Results show improved feasibility and objective performance relative to unconstrained random sampling, along with expanded Pareto fronts and adherence to the underlying energy-constraint structure.
| In which format do you inted to submit your paper? | LaTeX |
|---|