Prediction of Radiative Collapse in the Large Helical Device Plasma Discharges using Convolutional Neural Networks

Abstract

Predicting and preventing abrupt plasma termination incidents pose considerable challenges in nuclear fusion research. In the Large Helical Device (LHD), this occurrence is referred to as radiative collapse. During radiative collapse, impurity particles induce energy dissipation via radiation, hindering the maintenance of plasma discharges. Our approach aims to predict radiative collapse by analyzing the visible light emitted during such events. LHD uses approximately ten cameras to continuously observe plasma discharges, resulting in the accumulation of substantial video data from previous experiments. Using these images, convolutional neural network (CNN) models were trained to identify discharge states and subsequently applied to plasma discharge videos of the plasma discharges as a predictor. As a result, a determination model was developed, capable of discerning between stable and collapsed plasma discharge states with an accuracy of 91.5% ± 4% using plasma discharge images. Notably, this model demonstrated the potential to predict radiative collapse approximately three frames (66–132 ms) in advance. An examination of the model’s focal points revealed consistency with findings from prior research.