Bayesian Optimization of Sheared Flow Stabilized Z-Pinches

For high dimensional search spaces, such as a fusion reactor with a number of set parameters chosen before taking a shot, finding the optimum value for a given objective is difficult using gradient descent methods. With the reality of shot-to-shot variability the problem becomes even more difficult. Using Gaussian processes, the algorithm forms a surrogate model based on a handful of points in the search space formed by different shot input parameters. Beginning with multiple shots taken at the same set of input parameters, a mean and standard deviation is found for each of these points. With these points the algorithm forms a hypothesis about the search space. This hypothesis then forms an acquisition function which suggests the next point to sample. From this sample, a new hypothesis is generated, and the algorithm iterates until there is minimal improvement from each successive sample. Here, we discuss the optimization of the device for neutron yield, Q for SFS Z-Pinches as defined by Shumlak et al., and other objectives using the output of multiple diagnostics and Bayesian Optimization. Constraints are developed such that they deter the algorithm from sampling parameters that could damage the device.