We investigate a data-driven approach for tuning the filtered spherical harmonics method (\fpn) when solving the radiation transport equation (RTE). The \fpn method extends the classical spherical harmonics approach (\pn) by introducing regularization through a filter operator, which mitigates spurious oscillations caused by Gibbs' phenomenon. This filter includes a tunable parameter, the {filter strength}, that controls the degree of smoothing applied to the solution. However, selecting an optimal filter strength is nontrivial, often requiring inaccessible information such as the true or a high-order reference solution. To overcome this limitation, we model the filter strength as a neural network whose inputs include local state variables and material cross-sections. The optimal filter strength is formulated as the solution to a PDE-constrained optimization problem, and the neural network is trained using a discretize-then-optimize formulation in PyTorch. We evaluate the learned filter strength across a suite of test problems and compare the results to those from a simple, but tunable constant filter strength. In all cases, the neural-network-driven filter substantially improves the accuracy of the \pn approximation. For 1-D test cases, the constant filter outperforms the neural-network filter in some cases, but in 2-D problems, the neural-network filter generally performs better.
翻译:暂无翻译