Prediction of macroscopic plasma parameters in EAST using machine learning models

Machine learning offers a data-driven approach for rapid pre-shot prediction of key plasma parameters in tokamak experiments. In this study, random forest, support vector regression (SVR), and independent single-output multilayer perceptron (MLP) models were developed to predict macroscopic and dimensionless plasma parameters in EAST under H-mode operation. Ten experimentally accessible parameters—plasma current, toroidal magnetic field, line-averaged density, radiative power loss, confinement enhancement factor, Greenwald density ratio, safety factor, normalized beta, triangularity, and elongation—were used as inputs. Five targets were predicted: normalized ion gyroradius ρ*, normalized collisionality ν*, internal inductance li, poloidal beta βP, and central ion temperature Ti0. To avoid data leakage from correlated samples within the same discharge, a discharge-based splitting strategy was adopted, assigning complete discharges exclusively to either the training or testing set. Performance was evaluated using R2 and mean relative percentage error. SVR provides the most balanced overall performance, achieving high testing accuracy for most targets, especially ρ* and βP. The independent MLP models show strong nonlinear modeling capability, particularly for ν*, whereas RF provides acceptable baseline performance but is limited by weak extrapolation when training and testing distributions differ. Robustness tests with 5% Gaussian input perturbations show that SVR maintains stable predictions for most variables, especially ρ*, ν*, and βP. MLP also remains robust for ν* and βP, while Ti0 and li are more challenging because of their stronger dependence on current profile, temperature profile, transport behavior, and equilibrium state. Compared with conventional zero-dimensional scaling-law predictions, the machine learning models show improved agreement with experimental measurements and greater nonlinear mapping flexibility. These results indicate that SVR is a suitable and robust model for multi-parameter pre-shot prediction in EAST, while MLP may benefit from larger datasets.