Prediction of Radiative Collapse in Large Helical Device Using Feature Extraction by Exhaustive Search

A predictor model of radiative collapse of stellarator-heliotron plasmas has been developed by means of a machine learning technique and the feature of radiative collapse has been extracted with sparse modeling. The dataset used for training the model is constructed based on density ramp-up experiments in the Large Helical Device. As a result of feature extraction, the line averaged electron density, visible line emissions of CIV and OV, and the electron temperature at the edge have been selected as key parameters of radiative collapse. The likelihood of occurrence of radiative collapse has been quantified by using these parameters and this likelihood has been assessed in terms of predicting capability of the occurrence of radiative collapse. The collapse likelihood also implies the underlying physics of radiative collapse, therefore, the knowledge obtained by this data-driven study is expected to facilitate elucidation of the physics of the radiative collapse. In validation with discharges outside of the dataset, the predictor based on the likelihood has predicted over 85% of radiative collapse about 100 ms prior to this event on average while about 5% of stable discharges have been detected falsely as collapse discharges. The discharges in which the predictor made faults are discussed in order to investigate the cause of failure.