On the development of robust real-time capable ICRF modeling via machine learning

Robust real-time capable ICRF heating predictions are demonstrated via application of ML-based surrogate models trained on a large curated dataset generated using the TORIC [1] spectral solver. Our model [2, 3] can predict accurately both high harmonic fast wave (HHFW) and ion cyclotron (IC) minority heating, even beyond the validity range of the underlying physics code. The surrogates architecture is based on the random forest ensemble of decision trees regressor (RFR) and the multilayer perceptron regressor (MLP). The surrogates achieve a regression accuracy of R² = [0.94–0.97] and [0.7–0.81], for the HHFW on NSTX and the IC minority heating on WEST, respectively. The surrogates’ profile average inference times of O(2–50) μs represent a six to seven order of magnitude acceleration with respect to TORIC’s O(1–5) min simulation time. The datasets cover a wide range of physical parameter space including unexplored regions. Deep analysis of these datasets revealed a HHFW parametric subspace where the spectral solver presents spatially localized numerical instabilities [1]. In contrast, RFR surrogates are capable of predicting physical ICRF heating profiles even within this subspace. Additionally, we extend this methodology to the edge using TORIC-Petra-M coupled simulations to include the SOL effects and predict coupling efficiency. Overall, the achieved regression accuracy, computational efficiency and extended applicability of TORIC-based ICRF heating predictions make these surrogate models ideal for integrated modeling frameworks, and open the possibility for real-time control applications.

[1] M. Brambilla, Plasma Phys. Controlled Fusion 41, 1 (1999).

[2] Á. Sánchez-Villar et al, EPS Conf. Proc. 47A, o5.104 (2023).

[3] Á. Sánchez-Villar et al, Nucl. Fusion (2024, submitted).