Optimizing Permanent Magnet Stellarators with Machine Learning

Optimizing stellarators with attractive confinement properties while maintaining engineering feasibility remains a core challenge. MUSE is the world’s first quasi-axisymmetric optimized stellarator constructed with permanent magnets and planar coils, a feature that allows for simplified design and construction. Successful stellarator design traditionally uses numerical optimization techniques, consisting of iteratively exploring the space of possible design parameters for physics constraints to determine the best possible plasma shape and or coil design. Here, we conduct study using stochastic optimization and machine learning techniques that yields optimized stellarator configurations at scale for future iterations of the MUSE-like reactors including additional design goals and metrics. We investigate multiple attributes for these new configurations and find promising properties for future experiments.