AI-enabled Turbulent Transport Models for Stellarators

AI-enabled Turbulent Transport Models for Stellarators Part I: Validation of physics-based active learning for predictive turbulent transport surrogates on Wendelstein 7-X

Proxima Fusion is a Munich-based fusion startup whose long-term commercial goal is Stellaris — an optimized quasi-isodynamic (QI) stellarator power-plant concept with simple, buildable coils, designed to deliver clean, sovereign European energy through steady-state magnetic confinement. As a near-term step toward Stellaris, Proxima is developing Alpha, a stellarator demonstrator designed to achieve net energy gain, whose magnetic configuration belongs to the same QI class as Wendelstein 7-X (W7-X), the world's largest and most advanced stellarator, operated by the Max Planck Institute for Plasma Physics (IPP) in Greifswald, Germany and the experimental anchor for Proxima's design work under a formal IPP scientific collaboration. Building on W7-X's progress, stellarators have emerged as one of the most promising approaches to continuous fusion operation, but translating that progress into a commercial reactor depends critically on large-scale computational modeling, and in particular on predicting turbulent transport, one of the dominant unresolved challenges governing plasma confinement and power exhaust in next-generation QI stellarator reactors.

Predicting transport at reactor-relevant conditions requires nonlinear electromagnetic gyrokinetic simulations, where a single converged evaluation costs hundreds of GPU-hours. This makes systematic, first-principles exploration of reactor operating trajectories computationally prohibitive with conventional integrated modeling workflows. We propose a physics-based active learning framework that makes high-fidelity scenario design tractable by coupling probabilistic surrogate models (Gaussian processes) directly into a self-consistent transport solver.

This project is the first validation phase of a deliberately staged, three-part Large Scale AI Factory program. The structure is designed to allow EuroHPC to commit large-scale compute incrementally, ensuring confidence in the methodology; only after we have demonstrably delivered this validation milestone will we request the larger allocations needed for the subsequent phases in Proposal 2 and Proposal 3. For Proposal 1, we are currently requesting a 12-month allocation of approximately ~600,000 GPU-hours (which we round up from ~580,000 GPU-hours)  or ~150,000 node-hours to deliver this initial milestone.

This first proposal targets the validation step. We will train and converge GP turbulence surrogates for a well-diagnosed, pellet-fueled Wendelstein 7-X scenario using iterative active learning with electromagnetic GENE, embed them in Proxima’s FUS3 transport solver, and validate at the profile level by comparing the predicted density and temperature profiles against W7-X experimental measurements. The deliverable is a scientifically validated turbulence-surrogate methodology and the trained W7-X GP surrogates, which are the foundation upon which Proposals 2 and 3 are then built. These proposals involve applying the same methodology to Proxima Fusion’s Alpha stellarator reactor for surrogate construction across the startup-to-burn envelope (Proposal 2) and time-dependent actuator trajectory optimization for fusion gain Q > 1 (Proposal 3).