Acceleration of the NIMROD code to address future fusion challenges

Statistical validation of codes is a required next step for confidence in their predictive capabilities for fusion energy simulation and for the training of effective reduced models. This requires a large sample of simulations run for long periods of time that is not practical with present algorithms. We report on two approaches to accelerate the NIMROD [Sovinec, JCP, 195, 355-386 (2004)] extended-MHD code through the Center for Edge of Tokamak OPtimization SciDAC. The first approach is to use GPU acceleration and explore GPU-amenable approximate preconditioners such as those available in the Ginkgo code. We describe the performance of these approximate preconditioners relative to a direct solve on a tearing mode test case and implications for linear and nonlinear modeling. The second approach is to improve the time-discretization itself to take large implicit time steps. In this regard, we are presently exploring multi-rate integration for local atomic rate effects with the SUNDIALS code and the Kokkos backend for performance portability. The local ODE solves are integrated with MHD PDE solves through a Strang-split time with Douglas-Rachford-inspired coupling.