A review of machine learning-driven studies of tearing modes in tokamaks

Abstract

Over the last decade, tearing research progress has been accelerated thanks to the demonstrated efficacy of machine learning (ML) techniques—specifically for predicting and controlling tearing modes (TMs) in tokamak plasmas. These approaches leverage extensive experimental data and sophisticated ML algorithms to enhance plasma stability and performance. Additionally, the threat of TMs degrading plasma confinement and leading to disruptions has driven for years research under the International Tokamak Physics Activity, aimed at designing a successful trigger for the ITER's disruption mitigation system. In this review, we systematically explain the challenge behind tearing onset prediction, and summarize the latest achievements of ML in the tokamak tearing literature. We focus on the advancements that large statistical data-analyses and ML have afforded in interpreting the physics governing tearing onset, predicting onset, and developing algorithms to avoid and suppress TMs. Open challenges are discussed, including the need for reliable and reproducible TM prediction and control, using a reduced set of pilot-plant appropriate diagnostics, and the identification of passively tearing-stable power plant scenarios.