Improved training framework in a neural network model for disruption prediction and its application on EXL-50

Abstract

A neural network model with a classical annotation method has been used on the EXL-50 tokamak to predict impending disruption. However, the results revealed issues of overfitting and overconfidence in predictions caused by inaccurate labeling. To mitigate these issues, an improved training framework has been proposed. In this approach, soft labels from previous training serve as teachers to supervise the further learning process; this has lead to a significant improvement in predictive model performance. Notably, this enhancement is primarily attributed to the coupling effect of the soft labels and correction mechanism. This improved training framework introduces an instance-specific label smoothing method, which reflects a more nuanced model assessment on the likelihood of a disruption. It presents a possible solution to effectively address the challenges associated with accurate labeling across different machines.