Speaker
Description
Third-integer resonant extraction in Fermilab’s Delivery Ring relies on controlled evolution of the stable transverse phase-space region to provide uniform beam intensity for Mu2e. Spill quality depends on the interplay between quadrupole-driven tune evolution, nonlinear particle dynamics, and machine disturbances. We present progress on machine-learning methods for optimizing spill regulation in this setting, combining controller benchmarks with learned surrogate models. Direct optimization on a semi-analytical extraction simulator demonstrates that gated recurrent unit (GRU) controllers can outperform a proportional-integral-derivative (PID) baseline, with supervised training providing good sample efficiency and compact networks retaining competitive performance. A complementary surrogate framework combines generative modeling of stochastic disturbances with a GRU representation of the deterministic response to quadrupole actuation. Under the assumption of known disturbances during surrogate training, model-based reinforcement learning also produces policies that improve simulated spill duty factor over PID. We investigate model constraints and iterative training, including studies with linear time-varying surrogates, to address discrepancies caused by controller optimization exploiting surrogate inaccuracies. We also describe the automated data pipeline used to characterize machine response and disturbances during beam commissioning, and discuss ongoing quantization studies aimed at deploying compact GRU controllers on the existing spill-regulation FPGA hardware for validation with beam.