Machine Learning for Fluid Rendering in Video Games

Data-driven Fluid Simulations using Regression Forests from SoHyeon Jeong on Vimeo.

Water is one of elements that 3D animators put a lot of effort to get the perfection. It is really difficult to simulate dynamic movements of millions of particles and it requires huge amount of computational resources, specially for real time video rendering in 3D games.

However a group of researchers from Swiss Federal University of Technology and Disney Research has published a paper that shines some light on above problem with the aid of Machine Learning.

Rather than calculating the movements of each particle or reducing the particle count, this method could predict the dynamic movements of particles accurately. Presented algorithm needs to be trained with a random set of videos which have animated fluid particles with accurate calculations. Furthermore, the algorithm doesn't calculate movements in real-time, rather it predicts according to prior knowledge.

According to researches, this algorithm can render fluid animations 3 times faster than existing methods and can animate nearly 2 million particles in real-time.