Neural Network-Based Confinement Mode Prediction for Real-Time Disruption Avoidance

Reliable disruption avoidance techniques are critical for the development of safe and economically viable fusion reactors. This incipient need has driven the fusion community to spend substantial effort to develop machine learning (ML) models that aim to assist in this area. One area of concern is a sudden loss in global confinement associated with an unintended back-transition from H-mode confinement to L-mode confinement ("H-L back-transition"), which can sometimes lead to a disruption and increased risk of damage to a reactor-grade fusion device. A recent experiment on DIII-D demonstrated a real-time control method able to steer away from such unwanted back transitions. The present confinement mode of the plasma was inferred in real-time using a temporal convolutional neural network that outputs a continuous scalar estimate suitable for input to a closed-loop controller. This scalar estimate was then regulated to a value consistent with H-mode by adjusting power associated with the ratio of the plasma pressure to magnetic pressure (beta(N)). The model was trained on hundreds of thousands of discrete time steps of expert-labeled data of periods of H-mode and L-mode confinement, taken from hundreds of DIII-D tokamak discharges. This work presents these experimental results, including the method of collecting expert-labeled H- and L-mode confinement data, and the ML model performance and efficacy on the mode prediction task. In addition, the first application of this tool for H-L back-transition prevention is demonstrated in this experiment. Details are presented on the training of the neural network model, including concerns such as hyperparameter tuning and the network architecture. In addition, the methodology for embedding the neural network into the control system for real-time inference and closed-loop control is discussed.