Speaker
Description
In the boundary of magnetic fusion devices, plasma is strongly coupled with recycling
neutrals. Consequently, an integrated model for fusion boundary transport must encompass
neutral physics. A recently proposed propagator-based approach has the potential to
dramatically enhance the accuracy and efficiency of neutral transport modeling in plasma,
across the range of neutral collisionality regimes [1]. However, the cost of calculating the
propagator may still be prohibitive for efficient coupling with the plasma in an integrated
time-dependent model. To address this, a machine learning (ML) algorithm has been
developed for approximating the propagator. This allows for fast and accurate calculation
of neutral density for a given plasma background and neutral sources [2]. A salient feature
of the ML-predicted propagator is its smooth dependence on the plasma parameters. This
property facilitates the integration of the ML neutral model with plasma models using
Newton-based time integration methods. The present study is focused on implementation
and investigation of the ML propagator-based neutral transport algorithm in the edge
plasma transport code UEDGE. The propagator is calculated using the Monte Carlo neutral
transport code DEGAS-2 for a large number of samples corresponding to different plasma
profiles. These samples are used to train an ML model that approximates the propagator.
The performance of the resulting neutral calculation algorithm, integrated with the UEDGE
plasma model, is investigated for explicit and implicit time-stepping, using ODE and
Newton-Krylov solvers in the SciPy library.