Data-driven prediction of the: L–H transition power threshold in the EAST tokamak using ensemble learning regression

Accurate prediction of the L–H transition power threshold is essential for achieving steady-state high-performance burning plasmas in future fusion reactors. In this study, we develop a machine learning–based regression framework to predict the plasma power loss across the separatrix, , at the onset of the L–H transition. The model is trained on 3308 EAST discharges; from each discharge, we extract exactly one sample at the onset of its first L–H transition. Three ensemble learning models—random forest, XGBoost, and CatBoost—are systematically compared. As a benchmark, we compare our results with the empirical ITPA scaling law, which was derived from the high-density branch of the L–H power threshold in favorable ion drift configurations of ITER-like plasmas. On our present EAST database, which spans both high-density and low-density regimes and includes discharges with both favorable and unfavorable drift directions, the CatBoost model attains the highest predictive accuracy and yields significantly lower errors than the multivariate linear regression baseline. Model interpretation is performed using SHAP (SHapley Additive exPlanations). The SHAP rankings indicate that the key variables appearing in empirical L–H scaling relations (S, , and ) remain the dominant predictors, while the plasma current is consistently identified as an additional leading contributor. Equilibrium-related and geometry-related parameters (, , q95, κ) also exhibit non-negligible contributions. Underexplored spatial variables such as ZLX show a notable impact on the power threshold within the present EAST database. These results suggest that the machine-learning analysis can reproduce the main dependencies embodied in conventional scaling laws, while also indicating additional effects that are not explicitly encoded in the Martin scaling. The present framework provides an alternative data-driven approach for predicting the L–H transition, which may be particularly useful for exploring how multiple coupled control parameters jointly influence the L–H power threshold in complex, reactor-relevant operating scenarios.