Data-informed PCA and quantile transforms enable efficient Bayesian optimization of stellarator alpha confinement

The paper starts from an existing stellarator boundary dataset, applies either per-coefficient quantile transforms or PCA followed by quantiles, then treats the resulting [0,1] variables as bounded parameters for asynchronous Bayesian optimization. This replaces ad-hoc Fourier bounds that previously caused MHD solver failures or restricted shape expressiveness. Resulting configurations achieve excellent fast-ion confinement without enforcing quasisymmetry or quasi-isodynamicity, directly addressing the engineering bottleneck of alpha losses in stellarator reactors. The approach therefore shifts the design paradigm from symmetry constraints toward data-driven exploration of practically buildable coils. Related work on W7-X confirms that even modest improvements in alpha confinement reduce required heating power by tens of MW; the new parameterization makes such gains routine rather than exceptional. A key limitation remains the reliance on guiding-center rather than full-orbit tracing and the modest dataset size used for the PCA basis; larger, more diverse boundary libraries plus GPU-accelerated full-orbit verification would strengthen the claim that these optima remain accessible to real coils.