A machine-learning surrogate model for ion cyclotron heating in SPARC

SPARC will rely on ion cyclotron range of frequency (ICRF) heating to achieve the relevant temperatures for a burning plasma. Surrogate models enable fast and accurate prediction of ICRF heating in SPARC for use in integrated modeling, pulse planning, and real-time control, similar to previous work for neutral beam [Boyer, Nuc. Fus. (2019)] and ICRF [Sanchez-Villar, Nuc. Fus. (2024)] heating. The process begins by creating a database of 104 TORIC full-wave ICRF simulations using latin hypercube sampling to optimally distribute samples within a 10-dimensional input parameter space corresponding to the anticipated operating space of high performance L-Mode pulses [Rodriguez-Fernandez, Phys. Plasmas 31, 062501 (2024)]. The output features of the database consist of the absorbed power density profiles for the fundamental minority species (He3), electrons, and second-harmonic T. Of the surrogate model architectures tested, the multi-layer perceptron neural network combined with principal component analysis to reduce dimensionality of the output profiles performed best. Inference times drop from O(min) for the underlying model to O(µs) for the surrogate model, making the surrogate model sufficiently fast for integrated transport frameworks (e.g. ASTRA) and the plasma control system.