Exploring One-Cell Inversion Method for Transient Transport on GPU

To find deterministic solutions to the transient $S_N$ neutron transport equation, iterative schemes are typically used to treat the scattering (and fission) source terms. We explore the one-cell inversion iteration scheme to do this on the GPU and make comparisons to a source iteration scheme. We examine convergence behavior, through the analysis of spectral radii, of both one-cell inversion and source iterations. To further boost the GPU parallel efficiency, we derive a higher-order discretization method, simple corner balance (in space) and multiple balance (in time), to add more work to the threads and gain accuracy. Fourier analysis on this higher-order numerical method shows that it is unconditionally stable, but it can produce negative flux alterations that are critically damped through time. We explore a whole-problem (in all angle and all cell) sparse linear algebra framework, for both iterative schemes, to quickly produce performant code for GPUs. Despite one-cell inversion requiring additional iterations to convergence, those iterations can be done faster to provide a significant speedup over source iteration in quadrature sets at or below $S_{128}$. Going forward we will produce a two-dimensional implementation of this code to experiment with memory and performance impacts of a whole-problem framework including methods of synthetic acceleration and pre-conditioners for this scheme, then we will begin making direct comparisons to traditionally implemented source iteration in production code.