Graph neural network-accelerated Lagrangian fluid simulation

Abstract

We present a data-driven model for fluid simulation under Lagrangian representation. Our model, Fluid Graph Networks (FGN), uses graphs to represent the fluid field. In FGN, fluid particles are represented as nodes and their interactions are represented as edges. Instead of directly predicting the acceleration or position correction given the current state, FGN decomposes the simulation scheme into separate parts — advection, collision, and pressure projection. For these different predictions tasks, we propose two kinds of graph neural network structures, node-focused networks and edge-focused networks. We show that the learned model can produce accurate results and remain stable in scenarios with different geometries. In addition, FGN is able to retain many important physical properties of incompressible fluids, such as low velocity divergence, and adapt to time step sizes beyond the one used in the training set. FGN is also computationally efficient compared to classical simulation methods as it operates on a smaller neighborhood and does not require iteration at each timestep during the inference.