The Joint recognition of Multi-MHD instabilities on HL-2A

Abstract

Listen

Translate article

Abstract In tokamak plasmas, various Magnetohydrodynamic (MHD) instabilities could be driven by free energy, which enhance the plasma's transportation, leading to a reduction in critical fusion parameters such as temperature and density, and in severe cases, can even cause major plasma disruptions.This study employs deep learning techniques to learn from 1000 shots manually labelled as three types of MHD instabilities: fishbone mode, long-lived mode, and tearing mode, enabling real-time automated recognition of the instabilities. High accuracies of 97.83%, 95.32%, 94.84% are obtained on 200 testing shots, which are measured by area under the receiver-operator characteristic curve (AUC). Data processing methods that conform to the intuition of physics experts, such as Short Time Fourier Transform (STFT), have been retained, and advanced Artificial Intelligence (AI) algorithms such as Resnet have been combined to achieve a high accuracy rate. It also demonstrated the robustness in fully automatic detections over thousands of discharges. Furthermore, this study explores multitask learning techniques. Instead of using three individual neural network to recognize the different instabilities, a joint recognition algorithm is proposed. The joint algorithm shares the encoder of the three networks and use separate decoder branches to output the result of different instabilities. An inspiring outcome is found that the joint algorithm outperforms the individual ones on all of the instability recognition tasks. Implementing multiple MHD recognition tasks jointly can comprehensively improve the model's performance on each task by sharing related information between intrinsically related tasks. This means that in the future, the model can further develop as more tasks are added, revealing a possible technique routine to build an accurate and comprehensive large-scale model for fusion applications.