A New Perspective on Fluid Simulation: An Image-to-Image Translation Task via Neural Networks

Roman Lehmann, Markus Hoffmann, Simon Leufen, Wolfgang Karl

Keywords: cGAN, CFD, Image-to-Image, Fluid Simulation

Abstract: Standard numerical methods for creating simulation models in the field of fluid dynamics are designed to be close to perfection, which results in high computational effort and high computation times in many cases. Unfortunately, there is no mathematical way to decrease this correctness in cases where only approximate predictions are needed. For such cases, we developed an approach based on Neural Networks that is much less time-consuming but nearly as accurate as the numerical model for a human observer. We show that we can keep our results stable and nearly indistinguishable from their numerical counterparts over tenth to hundreds of time steps.

One-sentence Summary: A new perspective on fluid simulation is presented which works directly on the image representation of the solution.