Calculation of neutral source terms with deep learning to accelerate edge plasma simulations

Abstract Numerical simulations of edge plasmas are essential for optimizing divertor designs in fusion reactors. The neutral source terms in plasma fluid simulations are typically computed using the Monte Carlo method, which is computationally expensive and constitutes a bottleneck for simulation efficiency. To accelerate edge plasma simulations, this study investigates the feasibility of applying deep learning on the calculation of neutral source terms. Transformers excel at capturing sequence relationships and performing parallel computations, making them well-suited for modeling interactions between various nodes in complex data structures. A Transformer-based neural network (NN) is thus employed to learn the mapping from plasma backgrounds to neutral source terms. The model is trained and evaluated on a data set of approximately 600 samples obtained by SOLPS-ITER simulations of pure deuterium under a typical EAST upper single-null configuration. Subsequent tests show that it achieves relative errors of ~ 5% and ~ 3% at the peak values of particle and electron energy sources, respectively, and relatively large errors for momentum and ion energy sources, reaching up to ~ 20%. Coupled B2.5-NN simulations demonstrate acceptable accuracy for preliminary divertor design assessments, with relative errors below 5% in peak particle and heat flux densities at divertor targets. The computational time per simulation time step is reduced by 80%–90%, underscoring the potential of deep learning in enhancing the efficiency of edge plasma simulations. Nevertheless, the present NN model does not yet provide physical understanding or reliable extrapolation. This remains an important direction for future research.