Towards a surrogate unified robust generation engine for fusion multi-physics applications

We present advances in machine learning-based surrogate models to accelerate fusion energy research across diverse fusion physics domains. Building on validated surrogate work for ICRF heating at NSTX and WEST [1,2], we develop a unified ML methodology that streamlines surrogate implementation with uncertainty quantification, demonstrating the capabilities of automated hyperparameter tuning via Bayesian optimization. As a new application, we develop neural network surrogates to replace the ideal MHD electromagnetic closure ∂A∥/∂t in XGC [3,4], which can impose strict simulation timestep constraints. Trained on high-fidelity total-f simulations, the surrogates are tested to reduce statistical noise, improve numerical stability, and enable larger time steps. The demonstrated surrogate modeling capabilities establish a foundation for integrating surrogates into fusion energy applications, enabling not only reduced-order modeling, but also broader access to costly high-fidelity, physics-informed data.

[1] A. Sanchez-Villar et al., NF 64 096039 (2024)

[2] A. Sanchez-Villar et al., PoP 32 062504 (2025)

[3] S. Ku et al., PoP 25 056107 (2018)

[4] R. Hager et al., PoP 29 112308 (2022)