Physics simulations for virtual smoke, explosions or water are by now crucial tools for special effects. Despite their wide spread use, it is still difficult to get get these simulations under control, and they are still far too expensive for practical interactive applications. In this talk I will outline research directions to alleviate these inherent difficulties with machine learning techniques based on neural networks. These networks can learn and represent highly non-linear functions, which turns out to be highly useful in the flow simulation context.
A central part of this talk will be devoted to methods to enable the data-driven synthesis of fluid effects. In particular, I will outline methods to synthesize smoke volumes with pre-computed libraries of data, and a method the synthesizes liquid animations by deforming space-time data-sets of liquid surfaces. In both cases convolutional neural networks play a central role to make the approaches efficient and tractable.
I will show several examples of smoke and liquid animations generated with these approaches, and explain challenges encountered during data generation and training. In addition, I will discuss limitations and future directions of this new field of research.
13:30 - 14:00