Data-Driven Particle-Based Liquid Simulation with Deep Learning Utilizing Sub-Pixel Convolution

Abstract

In recent years, the performance of neural network inference has been drastically improved. This rapid change has paved the way for research projects focusing on accelerating physics-based simulations by replacing solver with its approximation. In this paper, we propose several efficient architectures of neural networks, which can be used to exploit this idea. The purpose of our research was to specifically target a liquid simulation problem. The central challenge for us was to create an efficient solution capable of approximating Position Based Fluid [Macklin and Müller 2013] solver. It requires the network to produce an accurate output at particles located in a continuous space and be significantly faster than the GPU based simulation. We achieved this by using modern sub-pixel convolution techniques originally used for image super-resolution. In our experiments, our method runs up to 200 times faster than the reference GPU simulation.