Development of surrogate models for the TORIC ICRF spectrum solver using ML algorithms

Machine learning (ML) approaches could enable real-time capable high-fidelity radio-frequency actuator modeling. Recently Wallace et al. [1] demonstrated ML-based surrogate models applied to the RF lower hybrid current drive providing 1D current and power predictions. In this work we present power absorption predictions in the ion cyclotron range of frequencies for two heating schemes: high harmonic fast wave (HHFW) at NSTX and hydrogen minority at WEST. Accurate (i.e. R²=0.93-0.96) and fast (i.e. O(50μs)) predictions are obtained for the HHFW 1D power absorption profiles. Datasets are generated using the full-wave code TORIC [2,3]. Even though minority dataset variance is higher than that of HHFW scenario, application of principal component analysis allowed us to also reach fast predictions of deuterium, hydrogen and electron power absorption profiles, of R²=0.71, 0.85, and 0.85, respectively. Preliminary results of surrogate predictions for the 2D are also presented. Overall, these surrogates show promising results for their incorporation in integrated modeling and control algorithms.

[1] Wallace et al. JPP 88, 895880401 (2022).

[2] M. Brambilla, PPCF 41, 1, (1999).

[3] M Brambilla, PPCF 44, 2423 (2002).