Person:

Tang, William

Nuclear fusion power delivered by magnetic confinement tokamak reactors carries the promise of sustainable and clean energy for the future [1]. The avoidance of large-scale plasma instabilities called disruptions [2, 3] is one of the most pressing challenges [4, 5] towards this goal. Disruptions are particularly deleterious for large burning plasma systems such as the multi- billion dollar international ITER project [6] currently under construction, where the fusion reaction aims to be the first to produce more power from fusion than is injected to heat the plasma. Here we present a new method, based on deep learning, to forecast disruptions and extend considerably the capabilities of previous strategies such as first-principles-based [5] and classical machine-learning approaches [7, 8, 9, 10, 11]. In particular, our method delivers for the first time reliable predictions on machines other than the one on which it was trained – a crucial requirement for large future reactors that cannot afford training disruptions. Our approach takes advantage of high-dimensional training data to boost the predictive performance while also engaging supercomputing resources at the largest scale in order to deliver solutions with improved accuracy and speed. Trained on experimental data from the largest tokamaks in the US (DIII-D [12]) and the world (JET [13]), our method can also be applied to specific tasks such as prediction with long warning times: this opens up possible avenues for moving from passive disruption prediction to active reactor control and optimization. These initial results illustrate the potential for deep learning to accelerate progress in fusion energy science and, in general, in the understanding and prediction of complex physical systems.