Learning to splash

Physics-based methods have been successfully applied to the production of realistic liquids simulation. Nowadays, a giant tide of floods sweeping through the New York city is no longer a surprise in movies. However, expensive computation prohibits applying the same visual effects to games, medical training, and other real-time applications. One way to achieve the real-time simulation is to simplify the model, so that computation cost can be reduced to an acceptable amount. However, among infinite number of possible simplification methods, how to choose the one that preserves the most visual credibility is a difficult task. Another way to reduce the computation cost while preserving visual quality is to use more complicated geometry primitives to represent liquids. For example, particle-based methods represent liquids as clusters of spherical particles, which are very simple primitives. This simplicity benefits human who can easily develop simulation methods based on those primitives, but it sacrifices efficiency of representation, and thus requires much more sampling and computation cost. This project aims at a real-time but visually plausible simulation system by applying machine learning methods learned from [Ng 2009]. To achieve this goal, the planned work is divided into three parts: first, data generation, which is to implement a full fluids simulator to generate training data; second, geometry primitives learning, which is to use an unsupervised learning method to derive a set of representative primitives from the full simulation results; third, dynamics rule learning, which is to use a supervised learning method to develop the updating rules of the representative primitives attained previously from the full simulation results.