A Machine Learning-based Real Time Disruption Predictor on DIII-D

Machine Learning-based disruption predictors have shown different performances on DIII-D and Alcator C-Mod. Nevertheless, it is important to develop predictors to avoid disruptions without empirical tuning in future devices, like ITER or SPARC. A new disruption prediction algorithm called DPRF (Disruption Prediction via Random Forests) is now embedded in the DIII-D plasma control system; it predicts impending disruptions with >100 ms warning time, and has a low false alarm rate. DPRF real-time disruptivity warning was exploited during an ITER baseline scenario DIII-D discharge to ramp down Ip and actively avoid an impending disruption. DPRF was trained on >5k disruptive and non-disruptive discharges, during flattop Ip and independent of their cause. DPRF average computation time is ~300 us, and input signals are mainly dimensionless or cast in dimensionless form, which facilitates the algorithm’s portability across different devices. DPRF’s novelty is the accessible interpretability of its predictions: by identifying the causes underlying disruption events, a better understanding of disruption dynamics is achieved, and a clear path toward the design of disruption avoidance strategies can be provided.