Cross-validation of a machine learning disruption predictor on EAST and Alcator C-Mod

A disruption prediction algorithm based on the Random Forests method has been developed using large databases of both disruptive and non-disruptive discharges from EAST and Alcator C-Mod. The machine learning algorithm was trained on flattop data using plasma parameters that can be available in real time; most are dimensionless (e.g. l_i, β_p) or cast in a dimensionless form (e.g. n/n_G) to facilitate multi-machine analysis. To make a robust disruption predictor, the algorithm was trained and tested on all disruptions, independently of their cause. A binary classification scheme based on a time sample’s proximity to the disruption achieves F1 scores of 0.71 and 0.50 on EAST and C-Mod, respectively. However, individual time sample predictions must be mapped to a disruption warning alarm using optimized control thresholds, a time window, and a time threshold to distinguish between classes. This poster describes a cross-validation procedure to determine such an optimal mapping for triggering an alarm of an impending disruption. A comparative study of the algorithm’s performance on both tokamaks is shown, and machine learning methods are used to interpret the model predictions and determine the drivers of disruptive behavior.