Disruption Prediction via Deep Recurrent Neural Networks

Disruption prediction and mitigation is a crucial challenge in pushing the performance limits of tokamaks while maintaining safe operation of the device. This is an even more critical challenge for future operations on the ITER tokamak. A predictive model which can reliably output the likelihood or imminence of a disruption can be incorporated into a controller, which can either initiate preventative measures or safe termination of the discharge. In this work, we take a fully data-driven approach to train a predictive model for disruption, and in particular, we utilize a deep recurrent neural network with historical data from DIII-D 2013-2017 campaigns. We investigate the effects of model choice (e.g. classifier or regressor, model size, and presence of recurrent units), label generation (i.e. how the presence of a disruption is labeled for model training), and training procedure (e.g. how the loss function is applied to recurrent predictions) on model performance. We evaluate model performance via a suite of metrics in binary classification (accuracy, Brier score, AUROC, and calibration).