Analysis of MHD Modes in Predicting ELM Onset using Machine Learning Techniques

Detection and prevention of edge localized modes (ELMs), which are destructive disturbances in the edge regions of H-mode toroidal plasmas, pose a significant roadblock to the development of stable, efficient fusion reactors. This problem is compounded by the fact that while models have been proposed to explain mechanisms behind them, ELMs are widely considered to be beyond the state-of-the-art predictive modeling, resulting in limited theoretical understanding of ELM onset to date. Machine learning (ML) has demonstrated effectiveness on other difficult plasma physics problems, such as predicting turbulent fluxes and estimating thermodynamic profiles from sensing data. Here we apply two machine learning models to predict ELM onset directly from Mirnov coil sensor data taken from a General Atomics DIII-D tokamak. The ML models support a prediction horizon long enough for realistic real-time ELM detection and mitigation. The models (multivariate polynomial regression and multilayer perceptron neural network) perform better than chance, with mean R² values of 0.39 and 0.45, respectively. In addition, we extract analytical representations from the polynomial regression model, yielding relationships between ELM occurrence and the presence of resonance modes in the magnetic flux density of fusion reactors.