Comparing Learned and Iterative Pressure Solvers for Fluid Simulation

Anonymous

Keywords: Fluid simulation, Pressure projection, Convolutional neural network, Jacobi, RBGS, PCG

TL;DR: We compare the performance of the neural network based pressure projection approach of Tompson et al. to traditional iterative solvers.

Abstract: This paper compares the performance of the neural network based pressure projection approach of Tompson et al. to traditional iterative solvers. Our investigation includes the Jacobi and preconditioned conjugate gradient solver comparison included in the previous work, as well as a red-black Gauss-Seidel method, all running with a GPU implementation. Our investigation focuses on 2D fluid simulations and three scenarios that present boundary conditions and velocity sources of different complexity. We collect convergence of the velocity divergence norm as the error in these test simulations and use plots of the error distribution to make high-level observations about the performance of iterative solvers in comparison to the fixed time cost of the neural network solution. Our results show that Jacobi provides the best bang of the buck with respect to minimizing error using a small fixed time budget.