Accelerating fluid simulations with graph convolution network predicted flow fields

The field of computational fluid dynamics (CFD) is integral to engineering disciplines, particularly for designing systems that operate under complex fluid flow conditions. Accurate simulation of flow fields is essential for optimizing performance across a variety of applications, including aviation, automotive, marine, and renewable energy sectors. Recent advancements in deep learning, particularly graph convolution networks (GCNs), offer promising alternatives for improving simulation processes. This work introduces a novel approach to accelerating fluid simulations using GCNs for flow field initialization. To this end, two different GCN models were employed, incorporating prior knowledge of the problem like its boundary conditions, as well as residual training. Extensive experiments using over 2000 sets of simulation results of various NACA airfoil shapes and flow conditions demonstrate that GCN-based initialization significantly reduces computational resources while maintaining high accuracy, achieving a 30% – 50% reduction in simulation time compared to conventional CFD initialization method. • Beneficial to integrate graph convolutional network (GCN) with a fluid solver. • GCN trained with >2000 Reynolds-averaged Navier–Stokes (RANS) airfoil results. • Lower residuals and faster convergence for RANS initialized with GCN predictions. • Time reduction of 30% – 50% observed for both low and high Reynolds number cases.