Neural-network based propagator approach for modeling neutral particles transport in plasma

Efficient and accurate modeling of neutral particles transport in fusion edge plasma remains an active area of research. For fusion boundary plasmas, Direct-Simulation Monte Carlo (DSMC) neutral modeling remains the main high-fidelity tool because unlike continuum methods it can easily include complex material wall geometry. The drawback of traditional DSMC-based methods for neutrals is poor coupling with continuum models for the plasma, due to the statistical nature of DSMC. In the present study, DSMC is used to calculate the propagator, for a set of plasma profiles, for evolution of the neutral distribution function after a single collision event. Next, a Neural Network (NN) is trained to approximate the propagator for a given plasma profile, and then the solution for steady-state neutral distribution is found using linear algebra. The resulting algorithm allows calculating the steady-state neutral distribution function and its moments accurately and very fast, compared to DSMC modeling. Importantly, the approximate neutral distribution output by the NN is differentiable with respect to perturbations of plasma parameters, which opens the possibility for using Jacobian-based methods, e.g., for efficient integration of the proposed neutral model in edge plasma transport simulations. Initial 1D testing of the model provides encouraging results, and the continuing investigation is extending now to larger 2D systems.