DDPS | Data-driven methods for fluid simulations in computer graphics

Описание: Fluid phenomena are ubiquitous to our world experience: winds swooshing through trembling leaves, turbulent water streams running down a river, and cellular patterns generated from wrinkled flames are some few examples. These complex phenomena capture our attention and awe due to the beautifully materialized complex patterns and become crucial elements to artistically support storytelling. In virtual environments, however, sophisticated manipulation of animated flow structures is still a burdensome task.

Given the amount of available fluid simulation data, data-driven approaches have emerged as attractive solutions. In this talk, I will introduce our recent works on data-driven methods for fluid simulations in computer graphics, including the first generative deep learning architecture that successfully synthesizes plausible and divergence-free 2D and 3D fluid simulation velocities from a set of reduced parameters, and its extension for predicting the complex dynamics of fluid flows with high temporal stability. Lastly, I will discuss our novel CNN-based reference frame algorithm for vortex extraction.


Short bio: Byungsoo Kim is a joint doctoral student at Computer Graphics Lab at ETH Zurich and Disney Research Studios, where Prof. Markus Gross has advised him. His research mainly focuses on deep learning methods for art-directable fluid simulations. His works have been published in premier venues such as ACM SIGGRAPH, SIGGRAPH Asia, and Eurographics and attracted media attention in outlets such as Two Minute Papers and Shiropen.


LLNL-VIDEO-818718