Implementation of an AI-enabled Deep Learning Disruption Predictor into a Tokamak Plasma Control System

Abstract:  

The focus of the Fusion Energy Science (FES) mission has been on the Tokamak line because this approach represents the closest Magnetic Fusion Energy (MFE) track toward actually achieving the Lawson criterion “triple product” and associated ignition requirement for delivering fusion energy.  For risk mitigation purposes, alternative confinement concepts (stellarators, FRC’s, etc.) can of course also be pursued with AI-based approaches.  In order to meet the goals of the $25B international burning plasma ITER tokamak experiment, AI-enabled capabilities will need to be delivered in a timely way to help ensure real-time control of the plasma state in burning plasmas.  Overall, risk mitigation strategies must focus on “big data from observations/measurements” needed for the rigorous validation and uncertainty quantification analysis featuring realistic sensitivity studies.  

The specific requirement is to of recent studies at Princeton U/PPPL has been to develop and implement a true AI/DL-based prediction capability within an actual tokamak plasma control system – such as the DIII-D tokamak facility in S. Diego, CA.  Here, it is necessary not only to accurately predict disruptive activity but also to categorize the primary causes.  From a control perspective, the myriad (over 100) actuator possibilities make it essential to be able to properly design and effectively utilize the advanced AI/DL capabilities to automatically make the best choices.