LAUSR

Abstract Robust disruption avoidance techniques are critical for the development of reliable fusion reactor devices. A viable reactor will require non-disruptive, long pulse operation where simply shutting down a discharge is undesirable. To achieve such performance, the plasma must be controlled to continuously avoid hazardous regimes instead of asynchronously aborting. A recent experiment on DIII-D demonstrated for the first time real-time control of proximity to a disruptive instability boundary. In particular, the vertical growth rate, an eigenvalue that characterizes the degree of instability of the plasma's vertical position, was regulated so as not to exceed DIII-D's vertical controllability limit. The open-loop growth rate was estimated in real time on the DIII-D plasma control system using a neural network model trained with tens of thousands of DIII-D shots. The model was trained to replicate the results of RZRIG [1] , a rigid displacement code for calculating the growth rate. Once trained, producing an estimate using the neural network is multiple orders of magnitude faster than RZRIG, thereby making the calculation suitable for real-time execution. The control system regulated the estimated growth rate by adjusting plasma elongation and distance to the inner wall of the vessel, and this regulation was shown to reliably avoid vertical displacement event disruptions (i.e. uncontrolled vertical oscillations) of the plasma. This work presents these experimental results, including the dynamic performance and the effectiveness of the control technique. Details are presented on the training of the neural network model, including concerns such as hyperparameter tuning and uncertainty quantification. Additionally, the methodology for embedding the neural network into the control system is discussed.