AI Surrogates Enable Direct Turbulent Heat Flux Optimization in Stellarator Equilibria

The team trained surrogates on more than 200,000 nonlinear adiabatic-electron simulations spanning 23,000 distinct stellarator configurations and multiple radial locations. These models replace expensive flux-tube calculations inside the optimization loop, permitting direct minimization of ion heat flux profiles rather than reliance on proxy metrics such as magnetic well depth or quasisymmetry error. The resulting equilibria can be post-processed with a transport solver to confirm higher ion temperatures without ad-hoc scaling assumptions.

This approach directly targets the dominant loss channel in stellarators, where turbulent transport often exceeds neoclassical levels by an order of magnitude. By embedding the surrogate inside an outer AI-agent loop, the authors automatically explored objective-function and hyperparameter combinations that human designers rarely test, uncovering configurations with elevated critical gradients at multiple radii.

Prior stellarator optimization campaigns, including those for W7-X and HSX, used indirect proxies because full gyrokinetic evaluations were computationally prohibitive. The new capability removes that bottleneck and could shift design targets from neoclassical transport or coil complexity toward explicit turbulence suppression.

Next steps include coupling the surrogate to self-consistent electron dynamics and validating the optimized equilibria against existing stellarator discharges to quantify predictive accuracy before any new device construction.