Disruption Prediction via the Fusion Recurrent Neural Network

Abstract

Power production via fusion is a long sought-after goal with a large set of impediments continuing to block its path to commercial viability. Plasma disruptions, a fast and anomalous loss of stability that can cause severe damage to plasma facing components, is one of these obstacles. To date, many of these disruptions have been avoided using empirical scaling laws to determine regimes of stable and unstable operation [1]. Nonetheless, it is necessary for fusion power output to operate close to the boundaries of these regimes, where disruptions are an approximately 5% probability event [2]. Online disruption prediction systems, which analyze plasma experiments as they progress and indicate whether or not a disruption is impending, are necessary to call and ultimately respond to these remaining disruptions. This paper analyzes the performance of a new prediction system called the “Fusion Recurrent Neural Network”, or FRNN, developed in the Tang group at the Princeton Plasma Physics Laboratory. We begin with motivation for fusion energy and a brief outline of the necessary features for a disruption prediction system. We then describe the basic physics behind disruptions and the empirical scalings that have shown to separate the stable from unstable regimes. Next, we look at the history of disruption prediction systems and compare them, including the Jet Protection System, APODIS, and the FRNN. Finally, we go into the key results from this study. The primary takeaway is that the FRNN represents the first machine learning disruption predictor capable of consistently outperforming, on all metrics that matter, a simple “locked-mode” based predictor.