Reinforcement Learning Improves Target Polarization Control in Nuclear Physics Experiments

Summary

Dynamically polarized targets in nuclear physics experiments require continuous microwave-frequency tuning as radiation damage and material properties change, a process traditionally handled through expert trial and error. This study presents a data-driven control framework that combines surrogate modeling with reinforcement learning and uses operational data from the APOLLO cryogenic target system. Multilayer perceptrons and Gaussian process regression models predict polarization from microwave frequency, beam current, and accumulated radiation dose. Gaussian-process models provide calibrated uncertainty estimates and can identify regions outside the training distribution, while the MLPs show limited sensitivity to distributional shift. To support control across multiple target samples, the researchers introduce a Gaussian-process approximation and embed the surrogate model in a standardized simulation environment. The reinforcement-learning agent uses a lower-confidence-bound reward that balances higher polarization against uncertainty. In the reported evaluation, the agent achieves an almost twofold improvement over operators’ actions.