Diffusion for Fusion: Designing Stellarators with Generative AI

Stellarators are a prospective class of power plants that confine a hot plasma with three-dimensional magnetic fields. Their design is typically framed as a PDE-constrained optimization problem that can take hours of computation to solve on a computing cluster. Recently, however, the stellarator community has started to explore data driven stellarator design, thanks to the development of large datasets of optimized stellarators. In this work, we leverage diffusion models for generating stellarator designs, effectively solving an ill-posed inverse problem. The diffusion models are used to learn a conditional probability distribution of stellarator design variables given the desirable stellarator characteristics. The diffusion models can be used to rapidly generate stellarators with desirable characteristics, which Is useful for warm-starting optimization routines.