Methodology for surrogate modeling implementation: application to the ICRF wave absorption forward problem

Traditional approaches to radio-frequency actuator design and operation are limited by the lack of real-time capable predictive models. A robust machine learning based methodology for forward surrogate model implementation is presented, applied to the ICRF absorption at the hot plasma core. The methodology is proven for two ICRF heating schemes: HHFW and hydrogen minority. A latin hypercube sampling method [1] allows to reduce the dataset statistical bias maximizing the covered variance of input physical parametric space. Effective absorption predictions are obtained using the random forest regressor (RFR) [2] and the multilayer perceptron (MLP) [3]. MLP shows higher sensitivity to outliers but outperforms the RFR when both trained in a refined dataset and with adequate hyperparameter tuning. Principal component analysis allows to simplify the neural network by a reduction of the dataset dimensionality, resulting in surrogate scoring improvement (e.g. R²=0.6 to 0.85). RFR models are applied to obtain physical predictions in outlier scenarios for the original model.

[1] M. Stein, Technometrics 29(2), 143-151 (1987)

[2] L. Breiman, Machine Learning 45, 5-32 (2001)

[3] M. W. Gardner, and S. R. Dorling, Atmospheric environment 32, 14-15 (1998)