Predicting Disruptions in Magnetic Confinement Fusion Reactors Via Deep Learning at the Largest Scale

Nuclear fusion power via magnetic confinement tokamak reactors carries the promise of sustainable and clean energy production. The avoidance of large-scale plasma instabilities called disruptions represents the most pressing obstacle to this goal. Disruptions are particularly deleterious for large reactors such as the multi-billion dollar international ITER project currently under construction, which for the first time aims to produce more energy than it consumes. Here, we present a new approach to forecast disruptions based on deep learning that for the first time overcomes the crucial limitations of past work, such as first- principles based and classical machine-learning approaches. In particular, our method for the first time (i) delivers reliable predictions on machines other than the one it was trained on – a crucial requirement for large future reactors that cannot afford "training" disruptions; (ii) utilizes high-dimensional training data such as profiles to boost predictive performance; and (iii) engages supercomputing at the largest scale to deliver solutions with unprecedented accuracy and speed. Trained on experimental data from the largest tokamaks in the United States (DIII-D) and the world (JET), our method moreover provides new insights into fusion physics and associated discovery science, can be tuned for physics-specific tasks such as prediction with long warning times, and opens up promising avenues for advancing the goal from passive disruption prediction to active reactor control and optimization.