Date of Award

2022

Publication Type

Thesis

Degree Name

M.A.Sc.

Keywords

Graph neural networks, Machine learning, Computational Fluid Dynamics problems

Supervisor

R.Barron

Supervisor

R.Balachandar

Rights

info:eu-repo/semantics/openAccess

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Abstract

Graph neural networks provide a framework for learning on unstructured data, such as meshes used for solving Computational Fluid Dynamics problems. However, current applications do not take advantage of known physical laws in the training process. This thesis addresses that gap by introducing graph convolution layers to calculate the divergence and gradient operator. The convolutions are valid on any 2D or 3D graph storing spatial data, and can be added to existing graph architectures. Using these convolutions, the residuals of the conservation of mass and momentum equations are computed and minimized through a physics-aware loss function. Two classical fluid dynamics problems are studied: the 2D flow past a NACA 0012 airfoil, and the 3D flow in the wake of an Ahmed body. In each study, a baseline and physics-aware Graph U-Net model is trained to predict the pressure and velocity fields at varying operating conditions. Despite achieving similar mean squared error, the physics-aware model has an order of magnitude smaller error in the residuals of the conservation equations. Further, the physics-aware model predicts flow fields with smaller error in the gradient, making them appear smoother. The same methodology can be applied to any general graph learning problem which requires minimizing a quantity composed from divergence or gradient operations.

Share

COinS