A high-fidelity and explainable machine learning (ML) method for neoclassical toroidal viscosity (NTV) torque prediction is investigated in this work. The ML surrogate model is trained based on NTV torque datasets obtained from first-principles numerical modeling, utilizing the extreme gradient boosting algorithm and the Bayesian hyper-parameter optimization strategy. The trained model exhibits high statistical accuracy regarding the coefficient of determination R2, the mean squared error and the relative error ɛs for NTV torque prediction, and a computational speed-up of about two orders of magnitude compared with physics-driven simulations. The prediction accuracy is further improved by applying a K-fold cross-validation averaging method. In particular, the probability of the occurrence of large local errors is reduced. Due to the radial locality of data samples, the trained model is able to make predictions for arbitrary numerical resolutions that can be different from the training data. An explainable artificial intelligence technique (the Shapley additive explanations method) is used to quantify the global and local contributions of input features to model output and shows physically meaningful feature relationships, indicating alignments with domain knowledge and thus improving model reliability. The fast, high-fidelity and explainable features of the ML surrogate model make self-consistent nonlinear simulations and other real-time applications possible for future tokamak physics studies.
Prediction of macroscopic plasma parameters in EAST using machine learning models