I created the Virtual Wind Tunnel to transform students’ use of Computational Fluid Dynamics (CFD) software using HPC and Mesh Graph Neural Networks. I believe that CFD should be an easy-to-use design analysis tool and aerodynamic tutor instead of a challenge in numerical methods.

Students can submit their 3D model on a website and specify a Reynolds number for the flow. After submitting their design, a fully automated CFD analysis is completed using a finite volume solver (OpenFOAM) on a local compute cluster. This analysis is completed at a level of discretization not possible on their personal laptops. Additionally, hundreds of geometric variations are programmatically created, a portion of which are also run through the automated CFD process to create a training dataset.

The CFD solution dataset is used to train a Mesh Graph Neural Network for inference on the remaining geometric variants. Results from all of the simulations are then compiled to provide students with simple force reports and suggestions on where their geometry can be modified to meet different aerodynamic goals, such as decreasing forces and moving the center of pressure.

Software used: OpenFOAM, NVidia Modulus, SLURM
Languages used: Python, C++, Ruby, Bash, Javascript

This is a link to a paper I wrote about an earlier version of my research in this area.