Fluid Simulations and Machine Learning

In our research we leverage machine learning to address persistent problems of fluid simulations in graphics. We particularly aim at increasing the performance and supporting the design process of simulations. We have developed a real-time fluid solver that predicts the motion with Regression Forests, resulting in real-time simulations of up to 2 million SPH particles. Our most recent CNN-based DeepFluids method can efficiently reconstruct parameterized simulations, allowing quick previews when changing simulation parameters. Other applications of DeepFluids include time resampling and extrapolation and efficient compression of fluid data.


Particle-based Simulations

Our research on particle-based fluid simulation with SPH aims at improving the robustness and efficiency of the solver. We have developed predictor-corrector methods to efficiently enforce incompressibility in SPH, and presented coupling methods for the interaction with complex boundaries and multiple liquids. Due to the high computational costs of such simulations, we have developed a multi-scale algorithm that allocates computational resources only in visually important areas of the fluid.


Artistic Control

Novel methods are required to support the workflow of artists when designing physics simulations for visual effects or games. The iterative nature of the design process requires not only efficient and robust methods, but also specific control functionalities that provide maximal control to an artist over the resulting motion and appearance. Project images will be added soon.