Tearing mode avoidance using reinforcement learning and classical delta prime stability analysis on DIII-D

Using a large database of labeled tearing modes (TMs) in DIII-D shots, we developed a machine learning-based model that predicts "Tearability," the probability of a TM occurring in the next time interval. This model utilizes actuator information and profile data to predict the effect of a given action on our Tearability metric. This model provides a real-time estimate of Tearability that was used to train reinforcement learning-based controllers that were tested in experiment in the ITER baseline scenario. Additionally, we have used our database to compute Δ′ in toroidal geometry to see if the classical metric is a relevant predictor for TM occurrence. Using the STRIDE code, we calculate most unstable mode, typically m,n=2/1, and compare the calculated Δ′ values from standard equilibria EFIT and rtEFIT, as well as the consistent kinetic equilibria generated by CAKE, and the new real-time capable RTCAKENN. This analysis shows that we can provide a physics-based metric that can be used for real-time TM control.