An unstructured adaptive mesh refinement for steady flows based on physics-informed neural networks

Abstract

Mesh generation is essential for accurate and efficient computational fluid dynamics simulations. To resolve critical features in the flow, adaptive mesh refinement (AMR) is routinely employed in certain regions of the computational domain, where gradients or error estimates of the solution are often considered as the refining criteria. In many scenarios, however, these indicators can lead to unnecessary refinement over a large region, making the process a matter of trial and error and resulting in slow convergence of the computation. To this end, we propose a heuristic strategy that employs the residuals of the governing partial differential equations (PDEs) as a novel criterion to adaptively guide the mesh refining process. In particular, we leverage on the physics-informed neural networks (PINNs) to integrate imprecise data obtained on a coarse mesh and the governing PDEs. Once trained, PINNs are capable of identifying regions of highest residuals of the Navier-Stokes/Euler equations and suggesting new potential vertices for the coarse mesh cells. Moreover, we put forth two schemes to maintain the quality of the refined mesh through the strategic insertion of vertices and the implementation of Delaunay triangulation. By applying the residuals-guided AMR to address a multitude of typical incompressible/compressible flow problems and comparing the outcomes with those of gradient-based methods, we illustrate that the former effectively attains a favorable balance between the computational accuracy and cost.