Disruption Prediction in Fusion Plasmas Using Recurrent Neural Networks
Julian Kates-Harbeck, Harvard University
Disruptions in fusion plasmas are destructive instabilities that lead to the deposition of large amounts of energy into the reactor walls in short amounts of time. They are particularly dangerous for large machines like ITER and are one of the main challenges in stable reactor operation facing the fusion community today. Effective disruption mitigation requires accurate predictions and sufficient warning time.
We train deep recurrent neural networks (RNNs) for disruption prediction. Deep architectures are able to extract salient features from diverse and higher-dimensional data and recurrent nets have an inherent notion of time and memory, making them a natural candidate to analyze temporal data. Our results for models trained on a single GPU architecture are competitive with state-of-the-art disruption prediction using SVMs at the JET machine.
The vast quantities of training data further motivate an accelerated data-parallel training architecture, where distinct worker machines ingest batches of training data in parallel and the gradient descent on the model parameters is synchronized by a master machine. We provide scaling and performance results for this data-parallel training scheme.
Abstract Author(s): J. Kates-Harbeck, B. Tang, E. Feibush