BayesOpt, a Fully Integrated Optimization Framework for the Magnetic Mirror

We present an automated optimization framework, BayesOpt for the engineering design of axisymmetric magnetic mirrors. Built around RealTwin, Realta Fusion’s integrated modeling suite for high-field axisymmetric magnetic mirrors, the framework leverages Bayesian optimization with Gaussian processes (GP) to efficiently map large parameter spaces and maximize user-defined performance metrics. RealTwin includes self-consistently coupled models for transport (CQL3D-m), MHD (Pleiades), heating (FREYA, GENRAY-c), and neutrals (KN1D-c) [1]. Given a performance metric f(x) and a design space X, the framework solves max_x∈X f(x), where common choices of f include plasma density, triple product, and stability conditions, and common choices of X include mirror throat magnetic field, midplane magnetic field, neutral beam power, energy, and injection angle, mirror throat radius, mirror length, and neutral wall density. BayesOpt is built for integration with NERSC and automates RealTwin simulation submission, data retrieval, and analysis. The framework has been instrumental in the design process of Anvil-DD, Realta Fusion’s next step magentic mirror, and the Bayesian optimization with GP scheme is shown to be incredibly effective as compared to uniform scanning. BayesOpt has successfully identified Anvil-DD operating scenarios capable of triple products in excess of 1· 10²⁰ keV s m⁻³ using realistic engineering specifications.

[1] Frank, S. J. et al., “Confinement performance predictions for a high field axisymmetric tandem mirror,” accepted to J. Plasma Phys., preprint available at arXiv 2411.06644 (2025).