Full Shot Predictions for the DIII-D Tokamak via Deep Recurrent Networks*

Although tokamaks are one of the most promising devices for making nuclear fusion energy a reality, there are still key obstacles when it comes to controlling and understanding the dynamics of the plasma. As such, it is crucial that we develop high quality models that we can employ to create better controllers and further our understanding. In this work, we take an entirely data driven approach to train a model. In particular, we use 7,884 historical shots from the DIII-D tokamak in order to train a deep recurrent network that is able to predict the full time evolution of plasma discharges. The model takes in a number of parameters for the shot (e.g. ip, Bt, etc.) and actuator information (e.g. injected power from neutral beams, shape controls, etc.) then predicts 17 scalar values (e.g. β_N, elongation, etc.) and 6 profiles (e.g. temperature, density, etc.) throughout time. Following this, we investigate different recurrent architectures as well as ensembling methods to create uncertainty estimates. We then evaluate these choices using explained variance and uncertainty calibration.