[Paper Review] Accelerating Eulerian Fluid Simulation With Convolutional Networks

Jonathan Tompson, Kristofer Schlachter+2 more|arXiv (Cornell University)|Jul 13, 2016

Computer Graphics and Visualization Techniques25 references296 citations

TL;DR

The paper proposes a data-driven convolutional network to approximate the pressure projection step in Eulerian fluid simulations, enabling real-time 2D/3D simulations with competitive accuracy and stability. It frames learning as unsupervised divergence minimization and demonstrates significant speedups over Jacobi and competitive results vs PCG.

ABSTRACT

Efficient simulation of the Navier-Stokes equations for fluid flow is a long standing problem in applied mathematics, for which state-of-the-art methods require large compute resources. In this work, we propose a data-driven approach that leverages the approximation power of deep-learning with the precision of standard solvers to obtain fast and highly realistic simulations. Our method solves the incompressible Euler equations using the standard operator splitting method, in which a large sparse linear system with many free parameters must be solved. We use a Convolutional Network with a highly tailored architecture, trained using a novel unsupervised learning framework to solve the linear system. We present real-time 2D and 3D simulations that outperform recently proposed data-driven methods; the obtained results are realistic and show good generalization properties.

Motivation & Objective

• Motivate real-time fluid simulation by reducing the computational burden of solving the Poisson pressure equation in the incompressible Euler/Navier-Stokes framework.
• Develop a data-driven, convolutional network to approximate the pressure solve within an operator-splitting Eulerian solver.
• Ensure long-term stability and divergence-free velocity fields under a fixed computational budget.
• Leverage unsupervised learning to avoid ground-truth pressure targets and enable multi-frame loss augmentation.

Proposed method

• Formulate the pressure projection as a learned linear solver: p_t = f_conv( input divergence ∇·u*, geometry g_{t-1} ) and obtain û_t = u*_t − (1/ρ)∇p_t.
• Use a fully convolutional, multi-resolution architecture with a pressure bottleneck to prevent trivial solutions and capture long-range effects.
• Train with an unsupervised loss that minimizes the divergence of the predicted velocity, potentially extended to multiple future steps to improve stability.
• Scale-normalize inputs by the standard deviation of u*, enforcing global scale invariance.
• Maintain compatibility with standard advection, body forces, and vorticity confinement within the Eulerian framework.

Experimental results

Research questions

• RQ1 Can a CNN-based surrogate for the pressure solve achieve real-time performance while maintaining stable, divergence-free velocity fields over long sequences?
• RQ2 Does unsupervised divergence minimization, coupled with a domain-specific ConvNet architecture, generalize to unseen geometries and flows?
• RQ3 How does a data-driven pressure projection compare to Jacobi and PCG solvers in speed and visual fidelity for 2D/3D simulations?

Key findings

• The approach achieves real-time pressure projection faster than Jacobi and yields competitive results to PCG for complex plume and geometry scenarios.
• A multi-resolution ConvNet with a pressure bottleneck effectively models long-range pressure effects and preserves stability across time steps.
• Unsupervised training with multi-frame divergence minimization improves long-term divergence performance over single-frame training.
• Compared to learned baselines, the proposed method demonstrates better generalization to unseen geometries and maintains plausible physical behavior in 3D smoke simulations.
• The method enables a drop-in replacement for the traditional pressure solve, offering significant speedups on GPU hardware.