Accelerating the Pressure Projection Step in Fluid Simulation Using a Physics Informed CNN

Abstract

The rapid increase in overall computational power and developement of domain specific hardware has allowed previously offline only techniques to reach the real-time consumer market. Fluid simulations are one such problem that until recently still remained a purely offline endeavor. Extensive research has been done to accelerate the solution of the underlying equations that describe fluid flow. Machine learning methods have recently seen significant success in the realm of computational fluid dynamics. We target pressure projection, the most expensive part of the simulation as a target for optimization/acceleration. By using a CNN we hope to take advantage of highly optimzied libraries in addition to hardware designed specifically to operate efficiently on tensors. Combined with a carefully formulated physics informed loss, the model gains the ability generalize effectively. We build upon work by Tompson et al. By making some informed decisions about model arcthitecture and simplifications in the training process we are able to improve on the current state of the art by achieving faster execution times and higher accuracies across a wide range of scenes. These improvements have brought machine learning based methods closer to being a viable solution for real-time interactive fluid simulations.