An In-depth Study of Stochastic Stellarator Coil Optimization
We extend stochastic coil optimization for stellarators and analyze the effect of optimizing for coil manufacturing errors, assembly errors represented by centroid translations and angular offsets, and errors in the coil currents. We further analyze the impact of different perturbation methods and use them in conjunction with the Augmented Lagrangian algorithm. For three stellarators, we show that stochastic methods regularize the coil optimization landscape and improve robustness in every case, with the magnitude of the improvement varying substantially across different plasmas. Our metric for coil robustness, $Q$, is improved by $4.50\%$ to $729\%$ across different configurations. Lastly, we perform numerous hyperparameter scans, including e.g. scanning the training perturbation size and the out-of-sample (OOS) evaluation perturbation size and find coupled and de-coupled training tolerance gains ranging from $0.09$ mm to $1.35$ mm depending strongly on the particular stellarator.