Recent progress on deep learning-based disruption prediction algorithm in HL-2A tokamak

Disruption prediction and mitigation is a crucial topic, especially for future large-scale tokamaks, due to disruption’s concomitant harmful effects on the devices. On this topic, disruption prediction algorithm takes the responsibility to give accurate trigger signal in advance of disruptions, therefore the disruption mitigation system can effectively alleviate the harmful effects. In the past 5 years, a deep learning-based algorithm is developed in HL-2A tokamak. It reaches a true positive rate of 92.2%, a false positive rate of 2.5% and a total accuracy of 96.1%. Further research is implemented on the basis of this algorithm to solve three key problems, i.e., the algorithm’s interpretability, real-time capability and transferability. For the interpretability, HL-2A’s algorithm gives saliency maps indicating the correlation between the algorithm’s input and output by perturbation analysis. The distribution of correlations shows good coherence with the disruption causes. For the transferability, a preliminary disruption predictor is successfully developed in HL-2M, a newly built tokamak in China. Although only 44 shots are used as the training set of this algorithm, it gives reasonable outputs with the help of data from HL-2A and J-TEXT. For the real-time capacity, the algorithm is accelerated to deal with an input slice within 0.3ms with the help of some adjustments on it and TFLite framework. It is also implemented into the plasma control system and gets an accuracy of 89.0% during online test. This paper gives a global perspective on these results and discusses the possible pathways to make HL-2A’s algorithm a more comprehensive solution for future tokamaks.