A Pioneering Neural Network Method for Efficient and Robust Fluid Simulation

TL;DR

This work addresses the computational burden of traditional Navier-Stokes-based fluid simulation by treating fluids as point clouds and introducing a neural network designed for robust performance in complex environments. The core method integrates two continuous convolution pathways (CConv and ASCC) with a lightweight global stability branch through Triangle Feature Fusion, balanced to respect fluid dynamics, physical laws, and global stability. A key contribution is the Fueltank dataset, comprising 320k frames across four tank types to benchmark realistic fuel sloshing under flight maneuvers. Empirically, the approach achieves substantial speedups over Flow3D and traditional SPH while maintaining high accuracy and long-term stability, and it generalizes to aircraft takeoff scenarios, offering practical impact for real-time CG and engineering simulations.

Abstract

Fluid simulation is an important research topic in computer graphics (CG) and animation in video games. Traditional methods based on Navier-Stokes equations are computationally expensive. In this paper, we treat fluid motion as point cloud transformation and propose the first neural network method specifically designed for efficient and robust fluid simulation in complex environments. This model is also the deep learning model that is the first to be capable of stably modeling fluid particle dynamics in such complex scenarios. Our triangle feature fusion design achieves an optimal balance among fluid dynamics modeling, momentum conservation constraints, and global stability control. We conducted comprehensive experiments on datasets. Compared to existing neural network-based fluid simulation algorithms, we significantly enhanced accuracy while maintaining high computational speed. Compared to traditional SPH methods, our speed improved approximately 10 times. Furthermore, compared to traditional fluid simulation software such as Flow3D, our computation speed increased by more than 300 times.

Figures (9)

• Figure 1: Position-based fluids scheme. It first computes external forces to obtain intermediate fuel particle states. Then, our neural network with trainable parameters $\theta$ predicts position and velocity changes induced by internal forces.
• Figure 2: The architecture of our network and Triangle Feature Fusion (TFF). The three types of TFF modules share the same architecture but serve three distinct roles in different positions within the network. The Type-TFF handles type-aware input for fuel and tank particles. The Main-TFF integrates three pathways to balance fluid dynamics modeling, physical constraints, and global stability. The Res-TFF adds a residual connection between the second and fourth layers.
• Figure 3: Construction strategy of Fueltank dataset. We use an iterative generation strategy, producing 400 frames per iteration. The final frame of the current iteration serve as the initial state for the next iteration.
• Figure 4: Examples of four tank types from the Fueltank dataset. The fuel tank undergoes random pitch and roll rotations at frame 0 and flows over the next 400 frames. The thumbnails depict the fuel surface from the SPH particle perspective.
• Figure 5: Qualitative experiments on Tank II show that existing neural network methods fail in complex scenarios, whereas our method provides stable fluid simulations comparable to traditional methods.