BackgroundThe thermal flux and particle transport issues at the plasma boundary and divertor region are core challenges for the long-term stable operation of future fusion reactors. Reducing the heat load on the divertor target plate is the key to achieve steady-state operation. Injecting an appropriate amount of impurity gas can dissipate energy, thereby reducing the target heat load. However, impurity transport may affect core confinement. Therefore, it is necessary to study its impact mechanism on divertor heat load and plasma performance to provide theoretical guidance and predictions for future experiments. SOLPS-ITER is one of the widely used boundary simulation codes, but it has high computational cost and time consumption.PurposeThis study aims to address the aforementioned issues by integrating machine learning methods to accelerate simulation computation, which has emerged as an important direction in current research.MethodsFirstly, a high-efficiency prediction model was constructed using machine learning methods, and normalization was applied for data processing. Then the dataset was split using the 5-fold cross-validation method, and the Dropout layer was employed to address the model's overfitting issue, hence to significantly improve the model's accuracy.ResultsThe prediction results show that, compared with neon (Ne) impurities, nitrogen (N) impurities can trigger detachment at a lower core boundary electron density; moreover, the mean relative error between the model's prediction results and the SOLPS-ITER simulation results (based on the HL-2A device) is approximately 2%.ConclusionThe method proposed in this study not only optimizes the divertor energy exhaust strategy and achieves efficient prediction of the particle flux density and electron temperature at the target plate, but also improves computational efficiency, providing a reference for related experiments.
SUI et al. (Fri,) studied this question.