Advanced Deep Learning for Physics (IN 2298)

hold by Nils Thuerey, Kiwon Um, Marie-Lena Eckert

Content

Deep learning algorithms for physical problems are a very active field of research. The group of Prof. Thuerey has published a series of papers in this area, you can find a summary here: physics-based deep learning resarch. This course targets the corresponding deep learning and physics modeling foundations.

Specifically, it targets deep learning techniques and numerical simulation
algorithms for materials such as fluids and deformable objects in the context
of computer animation. The lecture and exercises will all be in English. The following topics are discussed:

• Generative neural networks & temporal network architectures
• Physically-based animation, fluid modeling
• Discretizations, and partial differential equations
• Exercises to gain hands-on experience with CNN training and fluid simulation algorithms

Prerequisites

• Introduction to Deep Learning (Previously called: Deep Learning for Computer Vision) 
• Computer Gaphics Fundamentals, and Game Physics highly recommended

Literature

Machine Learning

• Goodfellow, Bengio, Courville: Deep Learning, 2016. http://www.deeplearningbook.org
• M. Nielsen: Neural Networks and Deep Learning, 2016.http://neuralnetworksanddeeplearning.com/

Fluid Simulation

• R. Bridson, M. Mueller-Fischer: Fluid Simulation for Computer Graphics;http://www.cs.ubc.ca/~rbridson/fluidsimulation/fluids_notes.pdf
• Griebel, Dornseifer, Neunhoeffer: Numerical Simulation in Fluid Dynamics: A Practical Introduction, Soc for Industrial & Applied Math

General Background

• Introduction to Linear Algebra: Gilbert Strang, Wellesley-Cambridge Press
• Computer Animation: Algorithms and Techniques, Parent, Morgan Kaufmann

Research

If you're interested, you can also check out some of our previous work on deep learning algorithms for fluid flow. These papers demonstrate several ways of training physics-aware neural networks for different parts of Navier-Stokes solvers. 

• Descriptor learning for fluid flow problems
• tempoGAN: Temporally coherent generative adversarial networks (GANs)
• Learning to generate 4D deformations for real-time liquids

Lecture Slides

Will be made available on the moodle page.