Machine Learning Cuts Compact QI Stellarator Test Loss by 87 Percent Using ConStellaration Data

The team extended prior conditional boundary generation to the low-aspect-ratio regime by first learning the shared geometric features of existing QI equilibria and then fine-tuning on a small high-fidelity compact dataset. This two-stage adaptation allowed the model to extrapolate beyond the original data distribution while preserving key isodynamic properties. Resulting candidates showed reduced neoclassical transport indicators and one configuration supplied a usable seed for subsequent optimization. The approach functions as an efficient front-end filter before expensive physics evaluations. Fusion reactor design has long been limited by the enormous configuration space of three-dimensional stellarator boundaries. Traditional optimization loops require thousands of costly equilibrium calculations, making exhaustive search impractical for compact devices needed in power plants. By learning statistical structure from prior QI solutions and steering generation with target properties, the method reduces wasted computation on non-viable shapes. This data-informed narrowing directly addresses the economic barrier of achieving reactor-relevant aspect ratios below 4 while maintaining good confinement. The work builds on Wendelstein 7-X results that demonstrated QI viability at larger aspect ratios and on earlier machine-learning surrogate efforts for stellarator coils. It also aligns with growing interest in stellarators from private fusion companies seeking simpler steady-state operation than tokamaks. If scaled, the technique could compress the timeline from concept to experimental test article by focusing high-fidelity codes on fewer, higher-quality starting points. Next steps include full finite-beta assessments and coil feasibility checks on the generated seeds.