Machine Learning Could Help Video Game Water Look Better Than Ever
Watch 2 million water particles move in real time.
Water, like hair or cloth, is one of those things that video games are still struggling to get right. Simulating so many little particles moving dynamically requires an incredible amount of computational resources, especially if they need to react in real-time, like a video game, as opposed to being pre-rendered, like a Pixar movie.
However, a group of researchers from the Swiss Federal Institute of Technology in Zurich working in collaboration with Disney Research (Mickey Mouse's international network of skunkworks labs) have published a paper and hypnotic video that suggests machine learning might be the solution to the problem.
Rather than computing each particle's trajectory or simplifying the simulation with less particles, this new machine learning method is able to accurately predict how the particles are supposed to move based on data collected from a large number of random, previous videos showing water particles generated with more resource intensive methods. Basically, the machine learning method doesn't have to compute every particle's trajectory in real-time. It can predict the trajectory based on previous data.
The authors of the paper said that their method worked up to three times faster than previous water physics rendering methods, and that it managed to render up to 2 million particles in real-time.
For now, all we have is the mesmerizing video above, but one can only hope that what we're looking at will one day lead to an incredible Wave Race 64 sequel.