Data-driven approach to plasma profile evolution for plasma control

High-current disruptions must be avoided in ITER to prevent prohibitively expensive damage. Most modern Plasma Control Systems partially solve this problem by predicting disruptions in real time and shutting down the experiment altogether when necessary. However, a more useful system would automatically vary controllable parameters (such as ECH and neutral beam systems) to guide the shot along a path that circumvents instabilities before disruptions are imminent. This study tackles the first step toward such a system: predicting the future evolution of a shot based on the history of the shot up to the current moment. Using first principles simulations for this task is notoriously slow and unreliable, so our approach is instead data-driven. A statistical / machine learning algorithm takes in the history of the pressure profile, current profile, plasma rotation, and a few other parameters at a time t. It then outputs the prediction for the profile at time t+1. By repeating this process iteratively on predicted steps with various options for the controllable parameters at each step, a Plasma Control System could set the controllable parameters to maintain safety with optimized performance.