Accelerating Progress Toward Controlled Fusion via Deep Learning at the Largest Scales
Julian Kates-Harbeck, Harvard University
The prediction and avoidance of disruptions in tokamak fusion plasmas represents a key challenge on the way to stable energy production from nuclear fusion. Using time series of sensory diagnostic data from past plasma shots with both disruptive and non-disruptive outcomes, we train a deep recurrent neural network to predict the onset of disruptions. We describe predictive performance on the JET and DIII-D tokamaks, using both scalar signals as well as one-dimensional profile information in our training data. Predictive performance on unseen data, including cross-tokamak prediction, is benchmarked against traditional machine-learning methods (SVM, Random Forest). To deal with very large amounts of data and the need for iterative hyperparameter tuning, we describe a distributed training algorithm that can take advantage of parallelism at the scale of the largest supercomputers. Finally, we describe promising directions for future work, ultimately shifting the objective from prediction to control.
Abstract Author(s): Julian Kates-Harbeck, Alexey Svyatkovskiy, Kyle Felker, Eliot Feibush, Bill Tang