HyMeshAI: Deep learning enabled three-dimensional adaptive mesh generator for high-resolution atmospheric simulations

Abstract

Adaptive mesh refinement (AMR) plays an important role in achieving seamless multi-scale simulations within numerical weather prediction (NWP) models. However, the practical implementation of three-dimensional AMR techniques faces significant challenges due to the mesh’s dynamic nature, which requires frequent mesh reconstruction and dynamic topology adjustments-features that are absent in traditional NWP models. Thus, AMR introduces entirely new implementation difficulties in NWP models. In this study, a hybrid deep learning model HyMeshAI is developed, which integrates CNN-based mesh density prediction with ANN-driven nodal positioning through a mesh generation algorithm, achieving an implementation of end-to-end dynamic mesh generation in an AMR atmospheric model. HyMeshAI preserves traditional acceleration benefits that reduce the computational load from higher spatial dimensionality and extended mesh refinement iterations, while also addressing the critical challenges of dynamic AMR. A key limitation of most data-driven artificial intelligence models is their reliance on feature matrices with fixed dimensionality and feature order, which is inherently incompatible with dynamic AMR. HyMeshAI overcomes this constraint by distilling two essential static features of AMR: the mesh point cardinality and the spatial probability distribution of mesh generation. The HyMeshAI’s performance has been estimated in idealized advection and rising bubble scenarios. In the advection test, HyMeshAI achieved a 20%-40% reduction in mesh count and a significant improvement in mesh quality, while maintaining comparable accuracy. In the rising bubble test, HyMeshAI reproduced the key features of buoyancy-driven convection and successfully resolved fine-scale structures, including Kelvin-Helmholtz vortices with a 30%-40% improvement in mesh quality.