Global Bayesian Methods for Stage 1 Stellarator Optimization

Numerical optimization of stellarators has proved a promising first step in the pipeline of constructing efficient, fusion-capable plasma containment devices. Stage 1 optimization searches for stellarator shapes which target desirable physical properties for containment such as quasisymmetry and omnigeneity. However, these physical objectives are highly nonconvex and the complicated landscape of local minima is not well-suited for local optimization algorithms. Nevertheless, the literature on Stage 1 optimization exclusively makes use of deterministic and stochastic local optimization. In this work, we utilize global Bayesian optimization algorithms TuRBO and DTuRBO to search for optimal stellarator shapes not easily found by local algorithms. We find that global optimization widens the pool of potentially viable stellarator designs. We demonstrate that global optimization can be used to produce a collection of candidates to be fed as initial conditions into local algorithms to hone their effectiveness, identifying stellarators highly optimized over the whole design space.