Computational Fluid Dynamics (CFD) methods are essential for determining aerodynamic properties of objects across a broad range of disciplines, from motorsport to aviation and civil engineering. CFD approaches fundamentally rely on discritizing fluid volume into a number of very small sections and iteratively solving the Navier-Stokes equations. This process can often take hours or days even with advanced hardware and constitutes a major bottleneck for the design process.
For our Machine Learning (CIS 520) final project, my partners and I decided to see if we could train a neural network to estimate pressure fields around airfoil cross sections based on ground truth data from the open-source CFD solver OpenFOAM. Since pressure fields are scalar quantities, we could represent both the input and output of our models as single-channel images. For the inputs to the model, we test both a binary mask with the cross section of the airfoil and the signed distance function on the same geometry. The two model types considered were a fairly standard 7-layer convolutional neural network and a U-Net architecture common in biomedical imaging.
Across both input representations and models, performance seemed to be relatively similar (see full pdf for details). Qualitatively, the models seemed to capture general trends in low pressure / high pressure regions, as seen in the sample renders below. Despite not being entirely accurate, these pressure field predictions could still provide a valuable initialization for proper CFD solvers and help accelerate solve times.