Kiwon Um from the Technical University of Munich will visit us next Wednesday and will give a talk entitled Physics-based Simulation with Deep Learning.
Abstract: Physics-based simulation has been dominant in recreation of a variety of natural phenomena in digital world such as computer-generated imagery. Accordingly, developing an efficient and effective simulation method has been gaining significant attention in the computer graphics community. Despite its significant technical advances, simulations of complex phenomena such as fluid dynamics still remain challenging particularly when it urgently requires both efficiency and accuracy. Recently, a huge amount of accumulated data and its effective use with machine learning have opened the door to many practical solutions for a lot of long-lasting unresolved problems. Along the line of the direction, this talk discusses the use of deep learning techniques with physics-based approaches particularly for fluid simulations in computer graphics. The presenter will introduce an efficient and effective data-driven, i.e., deep learning, method that improves small scale details on top of a traditional liquid simulation method. The experiments of the method will demonstrate how the machine learning successfully improve the target simulation problem together with the physics-based method and carefully refined data. This approach will be further discussed in follow-up projects. Additionally, the presenter will introduce a perceptual evaluation of liquid simulations and its extension to engineering applications. Investigating into its potential, this will be further discussed in conjunction with machine learning.