Efficient and accurate AI-forecasting of magnetic probe signals in tokamak

Tearing modes (TMs) are detrimental magnetohydrodynamic instabilities in tokamaks that degrade plasma confinement and obstruct fusion. Traditional approaches, including theoretical analysis, numerical simulations, and experimental diagnostics, face challenges in accuracy, cost, or forecasting time length. In this study, we develop a toolbox called Magnetic Probe Prediction, which leverages deep learning algorithms for time-series to process magnetic probe data from experimental diagnostics for forecasting tearing-modes-related diagnostic signals and providing early-warning detection of TMs activity. By appropriate training strategies and network architectures, our method achieves high prediction accuracy (10\% pointwise error on signal; 95\% statistical correctness at the decision level for TMs-related activity) and efficiency (3.3 ms inference time) for forecasts 10–100 ms ahead, with physical interpretability. Comparison across different AI models in the toolbox is provided.