Development of Fluid Dynamics Prediction Models using Graph Neural Network with Physical Equations

Abstract

In the development of aerodynamic-related products in the automotive manufacturing industry, evaluations must be fast and highly accurate with minimal deviation from the physical equations. On the other hand, Computational Fluid Dynamics and existing AI technologies had general problem with long processing time or low reliability of solution.

Therefore, in this study, we developed high-fidelity fluid prediction system based on Physics-Informed Graph Neural Network (PI-GNN), and design optimization system by using Multi-Objective Bayesian Optimization. We incorporated fluid equations into the loss function for consideration of physical equations when developing surrogate models. As a result, it was confirmed that the system enables highly accurate inference of fluid phenomena and fast optimization of the aerodynamic-related products.