When asking those in the VFX industry which processes are the slowest and most compute-intensive fluid simulation is bound to be somewhere towards the top of the list. Physical phenomena such as fire and water are notoriously difficult to control, even in the hands of the most experienced artists, and usually require a significant number of expensive iterations before the final result can be achieved. Often multiple versions of a single simulation are run over night, each a slight variation on the last and each taking multiple hours on possibly multiple render farm machines, with the expectation that one of them is either acceptable or a good starting point for subsequent iteration.
Refining a single shot can take weeks of artist time and hundreds or thousands of hours of CPU time. Deep learning techniques have the potential to break that workflow and allow for near-interactive manipulation of the most challenging effects.
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.
Source code: https://github.com/google/FluidNet
Are you aware of some research that warrants coverage here? Contact us or let us know in the comments section below!