Fakultät Informatik

Pre-computed Liquids with Neural-Networks and Optical Flow

Authors: Lukas Prantl, Boris Bonev, Nils Thuerey

Liquids exhibit highly complex, non-linear behavior under changing simulation conditions such as user inputs. The goal of this project is to develop novel ways to pre-compute and simulate liquid effects. Our emphasis here lies on convolutional neural networks, and high-dimensional optical flow. In the future, our approach could enable stable, and highly efficient physics effects for interactive applications. In settings such as digital games, a typical range of interactions could be pre-computed, and synthesized with very moderate computational costs.

 Android App - Neural Liquid Drop

You can directly try a demo of our full approach with the free Android app above. It employs both a neural network and pre-computed 5D-optical flow correspondences to make interactive, three dimensional liquids possible with very little computational resources. To the best of our knowledge, it is the first  fully three-dimensional liquid simulator that runs on a regular mobile phone.

Our method maps the behavior of liquid effects over a chosen input parameter range onto a reduced representation based on space-time deformations.  In order to represent the complexity of the full space of inputs, we leverage the power of generative neural networks to learn a reduced representation.  We introduce a novel deformation-aware loss function, which enables optimization in the highly non-linear space of multiple deformations.

Our representation makes it possible to generate implicit surfaces of liquids very efficiently, which allows us to very efficiently display the scene from any angle, and to add secondary effects such as particle systems.  We have implemented a mobile application for our full output pipeline to demonstrate that real-time interaction is possible with our approach.



In a predecessor work we have demonstrated that 5D optical flow with SDF inputscan be used to robustly establish correspondences between varying space-time surfaces of fluid flow. 

An even earlier version with colleagues from Georgia-Tech and IST Austria employed non-rigid ICP in 4D to make mesh-based blends of liquid surfaces possible. 






Matthias Niessner, our new Professor from Stanford University, offers a number of interesting topics for  master theses.


PhD positions on   Computational Fabrication and 3D Printing and  Photorealistic Rendering for Deep Learning and Online Reconstruction are available at the Computer Graphics & Visualization group.