Automatic identification of tokamak plasma confinement states (L-mode, ELM-free H-mode, and ELMy H-mode) with multi-task learning neural network

The identification of plasma confinement states (L-mode, ELM-free H-mode, and ELMy H-mode) is carried out using a multi-task learning neural network (MTL-NN) in EAST. The identification process can be divided into two tasks: identifying the operational modes and detecting the edge localized modes (ELMs). Dα and Mirnov coil measurements are selected as features for detecting the ELM. Parameters from scaling laws, which are related to thermal energy confinement time and heating threshold of L–H transition, are selected as features for identifying the operational modes. The data set used for supervised learning is collected from ELM control experiments in EAST. The MTL-NN comprises two task-specific layers and a shared layer. The multi-task learning framework allows for mutual error correction between tasks, resulting in higher accuracy and robustness compared to single-task models. Evaluation results demonstrate that the MTL-NN achieves an accuracy of 96.7% on the test set, representing a 3.6% improvement compared to single-task models.