Presentation Information
[2C17]A Physics-Informed Graph Neural Network Approach for Viscous Fluid Dynamics
*Ryo Yokoyama1, Shuichiro Miwa1, Koji Okamoto1 (1. Univ.Tokyo)
Keywords:
thermal hydraulic,Computational Fluid Dynamics,Viscous Fluid,Deep Learning
Surrogate modeling using artificial intelligence has advanced rapidly in recent years, and computational fluid dynamics (CFD) is no exception. While CNN and PINN-based CFD surrogates have been actively studied, Lagrangian methods that can capture large deformation and complex free-surface dynamics still face challenges in both predictive accuracy and training cost. In this study, we are developing a “Physics-informed Graph Neural Network” approach that improves accuracy and interpretability by incorporating the Navier–Stokes equations together with particle interaction models specific to particle methods into a graph neural network (GNN). This report outlines the concept of the proposed method and our current progress, and discusses new possibilities for surrogate modeling of viscous flow simulations.