Neoclassical Tearing Mode Anti-Lock Control With DECAF-Guided Reinforcement Learning

Neoclassical tearing modes (NTMs) arise in tokamaks when the magnetic topology at resonant surfaces rearranges into helical island chains, which degrades confinement, imposes limits on attainable plasma beta, and often leads to disruptions. The avoidance and stabilization of NTMs presents a challenging nonlinear control problem. Inspired by recent advances in reinforcement learning (RL)-based plasma control, we aim to develop a novel “physics-guided” RL framework using Disruption Event Characterization and Forecasting (DECAF) [1] to augment more standard avoidance approaches with artificial intelligence. This hybrid approach combines machine learning with analytical first-principles and semi-empirical physics modeling (e.g., for magnetic island width evolution via the modified Rutherford equation). By providing the state-value function, the physics model will guide the learning process in its search for a robust and reliable control policy. Analysis begins offline using multiple years of the KSTAR database with subsequent deployment on KSTAR in real time. Our approach aims to understand and validate NTM physics models, to predict NTM onset with high accuracy and sufficient warning time, and to develop a high-performance multi-input multi-output NTM controller.

[1] S.A. Sabbagh, et al., Phys. Plasmas 30, 032506 (2023). https://doi.org/10.1063/5.0133825