Predicting airfoil pressure distribution using boundary graph neural networks

Abstract

Surrogate models are essential for fast and accurate surface pressure and friction predictions during design optimization of complex lifting surfaces. This study focuses on predicting pressure distribution over two-dimensional airfoils using graph neural networks (GNNs), leveraging their ability to process non-parametric geometries. We introduce boundary graph neural networks (B-GNNs) that operate exclusively on surface meshes and compare these to previous work on volumetric GNNs operating on volume meshes. All of the training and evaluation is done using the airfRANS (Reynolds-averaged Navier-Stokes) database. We demonstrate the importance of all-to-all communication in GNNs to enforce the global incompressible flow constraint and ensure accurate predictions. We show that supplying the B-GNNs with local physics-based input-features, such as an approximate local Reynolds number $\mathrm{Re}_x$ and the inviscid pressure distribution from a panel method code, enables a $83\%$ reduction of model size and $87\%$ of training set size relative to models using purely geometric inputs to achieve the same in-distribution prediction accuracy. We investigate the generalization capabilities of the B-GNNs to out-of-distribution predictions on the S809/27 wind turbine blade section and find that incorporating inviscid pressure distribution as a feature reduces error by up to $88\%$ relative to purely geometry-based inputs. Finally, we find that the physics-based model reduces error by $85\%$ compared to the state-of-the-art volumetric model INFINITY.