Progress Toward Interpretable Machine Learning–Based Disruption Predictors Across Tokamaks

In this paper we lay the groundwork for a robust cross-device comparison of data-driven disruption prediction algorithms on DIII-D and JET tokamaks. In order to consistently carry on a comparative analysis, we define physics-based indicators of disruption precursors based on temperature, density, and radiation profiles that are currently not used in many other machine learning predictors for DIII-D data. These profile-based indicators are shown to well-describe impurity accumulation events in both DIII-D and JET discharges that eventually disrupt. The univariate analysis of the features used as input signals in the data-driven algorithms applied on the data of both tokamaks statistically highlights the differences in the dominant disruption precursors. JET with its ITER-like wall is more prone to impurity accumulation events, while DIII-D is more subject to edge-cooling mechanisms that destabilize dangerous magnetohydrodynamic modes. Even though the analyzed data sets are characterized by such intrinsic differences, we show through a few examples that the inclusion of physics-based disruption markers in data-driven algorithms is a promising path toward the realization of a uniform framework to predict and interpret disruptive scenarios across different tokamaks. As long as the destabilizing precursors are diagnosed in a device-independent way, the knowledge that data-driven algorithms learn on one device can be re-used to explain a disruptive behavior on another device.