Machine Learning for Real-time Fusion Plasma Behavior Prediction and Manipulation

An overview is presented of a new multi-institution collaboration funded by DOE among researchers from Princeton University, Carnegie Mellon University, PPPL, SLAC, University of Wisconsin – Madison to build the fundamental basis for use of Artificial Intelligence (AI)/Deep Learning (DL)/Machine Learning (ML) to augment real-time fusion plasma modeling, prediction, and manipulation. The group is developing a hierarchy of AI/DL/ML (in short ML) algorithms to 1) enable real-time simultaneous analysis of high- resolution multi-channel ECE, BES diagnostics, 2a) label and predict proximity to instability limits (specifically Alfven Eigenmodes and edge instabilities), 2b) produce real-time control-relevant predictions of plasma evolution that are difficult to obtain from physics simulations alone, and 3) manipulate experimental actuators in real-time. We are installing and testing this system on the DIII-D National Fusion Facility Plasma Control System. This work brings the power of ML techniques to plasma control. This will set the stage for successful operation of ITER, which requires plasma control at a level beyond current capabilities and will also expand the scientific understanding of plasma evolution and instabilities.