Data-driven approach to adaptive mesh refinement in PeleC

We study the viability of data-driven adaptive mesh refinement in PeleC, a fully compressible reacting flow solver that utilizes the AMReX library for structured mesh management. The current strategy for grid cell tagging for refinement employs ad hoc thresholding criteria on a select subset of flow variables and their gradients. We demonstrate that the neural network trained to classify cells based on a spatial discretization error threshold outperforms the existing heuristic tagging in PeleC. Various architectures including fully connected networks and convolutional neural networks are tested for efficacy and universality across regimes of a 3D turbulent CO_2 jet. Extensive testing is carried out to determine the optimal feature tensor to be input to the neural network by comparing localized flow features and global inputs such as 2D flow field slices and by employing feature importance studies.