Bayesian Optimization for Sheared-Flow Stabilized Z-Pinches

The search for optimal operation settings for a fusion device requires searching a high dimensional parameter space. To complicate matters, the nonlinearity of plasma dynamics causes some shot-to-shot variation, despite macroscopically invariant parameters. Bayesian optimization is a reinforcement learning technique for searching a high dimensional, continuous parameter space efficiently for the optimal value, while taking into consideration that the samples taken on the manifold are generated by a distribution. This optimization technique is more efficient than other search techniques such as gradient descent, or grid search, and is designed for continuous parameter space making it better suited than Bandit algorithms. For a sheared-flow-stabilized Z pinch the search space includes gas puff timings and pressures which create the initial distribution of gas in the vacuum chamber, and capacitor bank voltages, and discharge timings which sets how much and how the energy is put into the system. This framework allows for the optimization of various objective functions such as neutron yield, ion temperature, plasma density, or quiescent period duration. Here the results of initial deployment on the FuZE device are presented.