Investigating the robustness of machine learning-based disruption prediction with synthetic training data

Andrew Maris


Tuesday, November 15, 2022


PSFC Student Seminars

Data-driven algorithms have shown great promise in magnetic fusion applications, but these techniques can be unreliable when there are few examples for the algorithms to learn from. This is will be especially true for predicting disruptions on ITER, where the device will only be able to experience a small number of disruptions at full power over its lifetime. In this study, we investigate whether synthetically generated time series can augment small training sets of real data to achieve robust data-driven predictions of rare events. We perform numerical experiments by varying the amount of real and synthetic data in the training set of machine learning (ML) models tasked with predicting MHD-driven disruptions at DIII-D. We find that synthetic data appears to reduce underfitting and leads to higher prediction accuracies for standard neural networks.