Recently, the Sun Youwen Three-dimensional Physics Group of the EAST Major Science and Engineering Team at the Institute of Plasma Physics, Chinese Academy of Sciences, has made a series of new advances in tokamak artificial intelligence physics research. The team successfully developed intelligent prediction and recognition models for key physical phenomena using advanced machine learning and neural network technologies. The related research results were published in the core journals of nuclear fusion under the titles "Automatic identification of tokamak plasma confinement states (L-mode, ELM-free H-mode, and ELMy H-mode) with Multi-task Learning Neural Network" and "Interpretability analysis and real-time prediction of locked mode-induced disruptions in EAST." Dr. Wang Huihui, associate researcher, assisted in guiding the PhD student Deng Guohong, who is the first author of both papers.

In future large-scale fusion devices (such as ITER), plasma major disruptions, due to the instantaneous release of enormous thermal and magnetic energy, are considered the greatest potential threat to the safe operation of the device. Therefore, establishing a reliable disruption mitigation system (DMS) is crucial, and the prerequisite for this is the ability to accurately and timely predict major disruptions ("locked mode" phenomenon is one of the main triggers leading to major disruptions). In addition to avoiding catastrophic disruptions, achieving precise and intelligent recognition and control of plasma operating states is another core issue for future fusion reactors to achieve high-performance steady-state operation. High-confinement mode (H-mode) is the standard operating mode for ITER, but its accompanying edge-localized modes (ELMs) may cause excessive heat loads on the divertor target plates; automatic and timely recognition of them is an important step toward intelligent control in future fusion reactors.

In the research on locked-mode disruption prediction, the group adopted an "ante-hoc" machine learning approach. This method leverages the inherent transparency of decision tree (Decision Tree) models, not only to "know what it is" but also to "know why it is." The "interpretable prediction model" developed by the team achieved an area under the receiver operating characteristic curve (AUC) of up to 0.997 on the test set and successfully revealed the key physical quantities leading to locked-mode disruptions. On this basis, to meet real experimental needs, the team further developed a "real-time prediction model," achieving a 94% successful early warning rate and an average warning time of 137 milliseconds, sufficient to meet ITER's requirements for disruption early warning. This research not only provides reliable disruption early warning for EAST but also offers references for a deeper understanding of the disruption physical process through interpretability analysis.

In another study on automatic identification of plasma confinement states, the group innovatively adopted a multi-task learning neural network (MTL-NN). This method ingeniously integrates two closely related physical tasks—operating mode recognition (determining L-mode or H-mode) and edge-localized mode (ELM) detection—into one model for collaborative learning. Through shared network layers, the model achieves mutual error correction between tasks, significantly improving the model's accuracy and robustness. To reduce signal noise interference, the model uses scalar parameters in the physical normalization rate as input features. Experimental results show that the multi-task learning model achieves an identification accuracy of up to 96.7%, an improvement of 3.6% compared to single-task models on the same database. This result provides an efficient and precise real-time "diagnostic instrument" for tokamak plasma operating states, which is a key step in developing advanced plasma feedback control systems and achieving high-performance steady-state operation.

The above work benefits from the collaborative efforts of the EAST Major Science and Engineering Device team members. These research results demonstrate the enormous potential of artificial intelligence in solving key problems in nuclear fusion: not only do they have direct application value for the future efficient operation of the EAST device but also provide important scientific references and technical reserves for the intelligent control and stable operation of ITER. Currently, the group is conducting in-depth research on artificial intelligence integrated control for multiple physical processes. The related research is funded by the National Magnetic Confinement Fusion Energy Development Research Special Project, National Natural Science Foundation of China, Chinese Academy of Sciences Pioneer B Project, Anhui Provincial Natural Science Foundation, and the Director's Fund of the Hefei Institutes of Physical Science.

Figure 1. Decision Path for Locked-mode Disruption Prediction and Model Early Warning in a Single Discharge Experiment

Figure 2. Importance Ranking of Features in the Locked-mode Disruption Model

Figure 3. Structural Diagram of Multi-task Learning Neural Network for Automatic Identification of Plasma Confinement States