GE_at_TUM

Learning temporal predictions and reduced representations at EG’19

Our two papers on learning temporal predictions and reduced representations for fluids have been accepted to the CGF Journal and will be presented at Eurographics 2019 in Milano! Congratulations to Steffen, Moritz, Byungsoo and Vinicius!
Our work explores methods for the data-driven inference of temporal evolutions of physical functions with deep learning techniques. More specifically, we target fluid flow problems, and we propose a novel network architecture to predict the changes of the pressure field over time. The central challenge in this context is the high dimensionality of Eulerian space-time data sets. Key for arriving at a feasible algorithm is a technique for dimensionality reduction based on convolutional neural networks, as well as a special architecture for temporal prediction. We demonstrate that dense 3D+time functions of physics system can be predicted with neural networks, and we arrive at a neural-network based simulation algorithm with practical speed-ups. We demonstrate the capabilities of our method with a series of complex liquid simulations, and with a set of single-phase simulations. Our method predicts pressure fields very efficiently. It is more than two orders of magnitudes faster than a regular solver. Additionally, we present and discuss a series of detailed evaluations for the different components of our algorithm.
This paper presents a novel generative model to synthesize fluid simulations from a set of reduced parameters. A convolutional neural network is trained on a collection of discrete, parameterizable fluid simulation velocity fields. Due to the capability of deep learning architectures to learn representative features of the data, our generative model is able to accurately approximate the training data set, while providing plausible interpolated in-betweens. The proposed generative model is optimized for fluids by a novel loss function that guarantees divergence-free velocity fields at all times. In addition, we demonstrate that we can handle complex parameterizations in reduced spaces, and advance simulations in time by integrating in the latent space with a second network. Our method models a wide variety of fluid behaviors, thus enabling applications such as fast construction of simulations, interpolation of fluids with different parameters, time re-sampling, latent space simulations, and compression of fluid simulation data. Reconstructed velocity fields are generated up to 700x faster than traditional CPU solvers, while achieving compression rates of over 1300x.
GE_at_TUM

Deep Learning for Graphics Course at SIGGRAPH Asia 2018

We yesterday held our course on Deep Learning for Graphics Course at SIGGRAPH Asia 2018 in Tokyo. The slides are now available online at:

http://geometry.cs.ucl.ac.uk/creativeai/

Abstract: In computer graphics, many traditional problems are now better handled by deep-learning based data-driven methods. In an increasing variety of problem settings, deep networks are state-of-the-art, beating dedicated hand-crafted methods by significant margins. This tutorial gives an organized overview of core theory, practice, and graphics-related applications of deep learning.

 

GE_at_TUM

Video Super-resolution with Deep Learning

Our work on video super-resolution with GANs is online now as a preview. The main trick is a special discriminator CNN that learns to supervise in terms of detail as well as temporal coherence. In addition, we propose a novel set of metrics for quantifying temporal coherence in videos. Enjoy 🙂 !

Abstract: Adversarial training has been highly successful in the context of image super-resolution. It was demonstrated to yield realistic and highly detailed results. Despite this success, many state-of-the-art methods for video super-resolution still favor simpler norms such as L_2 over adversarial loss functions. This is caused by the fact that the averaging nature of direct vector norms as loss functions leads to temporal smoothness. The lack of spatial detail means temporal coherence is easily established. In our work, we instead propose an adversarial training for video super-resolution that leads to temporally coherent solutions without sacrificing spatial detail.

In our generator, we use a recurrent, residual framework that naturally encourages temporal consistency. For adversarial training, we propose a novel spatio-temporal discriminator in combination with motion compensation to guarantee photo-realistic and temporally coherent details in the results. We additionally identify a class of temporal artifacts in these recurrent networks, and propose a novel Ping-Pong loss to remove them. Quantifying the temporal coherence for image super-resolution tasks has also not been addressed previously. We propose a first set of metrics to evaluate the accuracy as well as the perceptual quality of the temporal evolution, and we demonstrate that our method outperforms previous work by yielding realistic and detailed images with natural temporal changes.

GE_at_TUM

Physics-based Deep Learning at NIPS 2018

We will be presenting our recent works on physics-based deep learning for fluid flow at the NIPS 2018 workshop on “Modeling the Physical World: Learning, Perception, and Control“, organized by Jiajun Wu, Kelsey Allen, Kevin Smith, Jessica Hamrick, Emmanuel Dupoux, Marc Toussaint, and Joshua Tenenbaum.

NIPS Conference: https://nips.cc

NIPS 2018 Workshop “Modeling the Physical World: Learning, Perception, and Control”: https://nips.cc/Conferences/2018/Schedule?showEvent=10931

Workshop homepage: http://phys2018.csail.mit.edu/submission.html

In particular we will discuss our works on:

Detailed abstracts:

Latent-space Physics: Towards Learning the Temporal Evolution of Fluid Flow: Our work explores methods for the data-driven inference of temporal evolutions of physical functions with deep learning techniques. More specifically, we target fluid flow problems, and we propose a novel network architecture to predict the changes of the pressure field over time. The central challenge in this context is the high dimensionality of Eulerian space-time data sets. Key for arriving at a feasible algorithm is a technique for dimensionality reduction based on convolutional neural networks, as well as a special architecture for temporal prediction. We demonstrate that dense 3D+time functions of physics system can be predicted with neural networks, and we arrive at a neural-network based simulation algorithm with practical speed-ups. We demonstrate the capabilities of our method with a series of complex liquid simulations, and with a set of single-phase simulations. Our method predicts pressure fields very efficiently. It is more than two orders of magnitudes faster than a regular solver. Additionally, we present and discuss a series of detailed evaluations for the different components of our algorithm.

Temporally Coherent, Volumetric GAN for Super-resolution Fluid Flow: We propose a temporally coherent generative model addressing the super-resolution problem for fluid flows. Our work represents the first approach to synthesize four-dimensional physics fields with neural networks. Based on a conditional generative adversarial network that is designed for the inference of three-dimensional volumetric data, our model generates consistent and detailed results by using a novel temporal discriminator, in addition to the commonly used spatial one. Our experiments show that the generator is able to infer more realistic high-resolution details by using additional physical quantities, such as low-resolution velocities or vorticities. Besides improvements in the training process and in the generated outputs, these inputs offer means for artistic control as well. We additionally employ a physics-aware data augmentation step, which is crucial to avoid overfitting and to reduce memory requirements. In this way, our network learns to generate advected quantities with highly detailed, realistic, and temporally coherent features. Our method works instantaneously, using only a single time-step of low-resolution fluid data. We demonstrate the abilities of our method using a variety of complex inputs and applications in two and three dimensions.

Coupled Fluid Density and Motion from Single Views: We present a novel method to reconstruct a fluid’s 3D density and motion based on just a single sequence of images. This is rendered possible by using powerful physical priors for this strongly under-determined problem. More specifically, we propose a novel strategy to infer density updates strongly coupled to previous and current estimates of the flow motion. Additionally, we employ an accurate discretization and depth-based regularizers to compute stable solutions. Using only one view for the reconstruction reduces the complexity of the capturing setup drastically and could even allow for online video databases or smart-phone videos as inputs. The reconstructed 3D velocity can then be flexibly utilized, e.g., for re-simulation, domain modification or guiding purposes. We will demonstrate the capacity of our method with a series of synthetic test cases and the reconstruction of real smoke plumes captured with a Raspberry Pi camera.

 

GE_at_TUM

Reynolds-averaged Navier Stokes Simulations with Deep Learning

Our paper on Reynolds-averaged Navier Stokes simulations with deep learning (i.e. convolutional neural networks) is online now. You can read it on arXiv:
https://arxiv.org/abs/1810.08217

The full source code and a first training data set with 6400 RANS simulations as ground truth data is available here:
https://github.com/thunil/Deep-Flow-Prediction

Deep Flow Prediction is a pytorch framework for fluid flow (Reynolds-averaged Navier Stokes) predictions with deep learning. It contains code for data generation, network training, and evaluation for the aforementioned paper. A brief manual for the ode can be found in the github repository.

Paper Details:
Title: Well, how accurate is it? A Study of Deep Learning Methods for Reynolds-Averaged Navier-Stokes Simulations

Authors: Nils Thuerey, Konstantin Weissenow, Harshit Mehrotra, Nischal Mainali, Lukas Prantl, Xiangyu Hu

Abstract: With this study we investigate the accuracy of deep learning models for the inference of Reynolds-Averaged Navier-Stokes solutions. We focus on a modernized U-net architecture, and evaluate a large number of trained neural networks with respect to their accuracy for the calculation of pressure and velocity distributions. In particular, we illustrate how training data size and the number of weights influence the accuracy of the solutions. With our best models we arrive at a mean relative pressure and velocity error of less than 3% across a range of previously unseen airfoil shapes. In addition all source code is publicly available in order to ensure reproducibility and to provide a starting point for researchers interested in deep learning methods for physics problems. While this work focuses on RANS solutions, the neural network architecture and learning setup are very generic, and applicable to a wide range of PDE boundary value problems on Cartesian grids.

Details

GE_at_TUM

Real-time Fluid Simulations at the TUM Open House 2018

We’are currently (as of Oct. 13) demoing our deformation neural network app and real-time fluid solver at the TUM open house 2018.

If you have time today – stop by at Garching Forschungszentrum 🙂

Otherwise you can check out the corresponding papers:

You can of course also check out our “Neural Liquid Drop” App on the Android app store.
GE_at_TUM

SIGGRAPH 2018 tempoGAN talk on YouTube

You Xie and Rachel Chu just successfully presented their paper “tempoGAN: A Temporally Coherent, Volumetric GAN for Super-resolution Fluid Flow” at SIGRGAPH 2018. You can enjoy their full presentation via the YouTube video below.

Presentation:

More info regarding the SIGGRAPH 2018 technical papers:

Technical Papers

 

 

tempoGAN paper abstract:

We propose a temporally coherent generative model addressing the super-resolution problem for fluid flows. Our work represents the first approach to synthesize four-dimensional physics fields with neural networks. Based on a conditional generative adversarial network that is designed for the inference of three-dimensional volumetric data, our model generates consistent and detailed results by using a novel temporal discriminator, in addition to the commonly used spatial one. Our experiments show that the generator is able to infer more realistic high-resolution details by using additional physical quantities, such as low-resolution velocities or vorticities. Besides improvements in the training process and in the generated outputs, these inputs offer means for artistic control as well. We additionally employ a physics-aware data augmentation step, which is crucial to avoid overfitting and to reduce memory requirements. In this way, our network learns to generate advected quantities with highly detailed, realistic, and temporally coherent features. Our method works instantaneously, using only a single time-step of low-resolution fluid data. We demonstrate the abilities of our method using a variety of complex inputs and applications in two and three dimensions.

 

GE_at_TUM

tempoGAN source code now online

The source code for our tempoGAN project, which aims for super-resolution inference of Navier-Stokes transport phenomena (such as smoke clouds) is online now at: https://github.com/thunil/tempoGAN

It comes with a readme, data generation scripts, and should give an easy starting point for training generative adversarial nets for fluids. If you try it, let us know how it works!

Project page

Here’s again the full abstract of the paper: We propose a temporally coherent generative model addressing the super-resolution problem for fluid flows. Our work represents the first approach to synthesize four-dimensional physics fields with neural networks. Based on a conditional generative adversarial network that is designed for the inference of three-dimensional volumetric data, our model generates consistent and detailed results by using a novel temporal discriminator, in addition to the commonly used spatial one. Our experiments show that the generator is able to infer more realistic high-resolution details by using additional physical quantities, such as low-resolution velocities or vorticities. Besides improvements in the training process and in the generated outputs, these inputs offer means for artistic control as well. We additionally employ a physics-aware data augmentation step, which is crucial to avoid overfitting and to reduce memory requirements. In this way, our network learns to generate advected quantities with highly detailed, realistic, and temporally coherent features. Our method works instantaneously, using only a single time-step of low-resolution fluid data. We demonstrate the abilities of our method using a variety of complex inputs and applications in two and three dimensions.

GE_at_TUM

Best-paper award for “Coupled Fluid Density and Motion from Single Views” at SCA 2018

Congratulations to Marie-Lena Eckert for winning the best-paper award at SCA 2018 for her submission “Coupled Fluid Density and Motion from Single Views”.

Her work aims for reconstructing fluid flow phenomena with strong Navier-Stokes priors. This makes it possible to compute dense flow fields based on only a monocular video, i.e., an image sequence from a single viewpoint.

Paper Abstract: We present a novel method to reconstruct a fluid’s 3D density and motion based on just a single sequence of images. This is rendered possible by using powerful physical priors for this strongly under-determined problem. More specifically, we propose a novel strategy to infer density updates strongly coupled to previous and current estimates of the flow motion. Additionally, we employ an accurate discretization and depth-based regularizers to compute stable solutions. Using only one view for the reconstruction reduces the complexity of the capturing setup drastically and could even allow for online video databases or smart-phone videos as inputs. The reconstructed 3D velocity can then be flexibly utilized, e.g., for re-simulation, domain modification or guiding purposes.

More information about the ACM SIGGRAPH / Eurographics Symposium on Computer Animation: (SCA): http://sca2018.inria.fr

Coupled Fluid Density and Motion from Single Views

 

GE_at_TUM

Latent-space Physics source code

The source code for our latent space physics paper is online now:
https://github.com/wiewel/LatentSpacePhysics

It contains both the Navier-Stokes solver for data generation (based on mantaflow http://mantaflow.com), and the keras code (for tensorflow https://www.tensorflow.org) for training the autoencoder and LSTM networks.

The preprint can be found here: https://arxiv.org/pdf/1802.10123

Paper Abstract

Our work explores methods for the data-driven inference of temporal evolutions of physical functions with deep learning techniques. More specifically, we target fluid flow problems, and we propose a novel network architecture to predict the changes of the pressure field over time. The central challenge in this context is the high dimensionality of Eulerian space-time data sets. Key for arriving at a feasible algorithm is a technique for dimensionality reduction based on convolutional neural networks, as well as a special architecture for temporal prediction. We demonstrate that dense 3D+time functions of physics system can be predicted with neural networks, and we arrive at a neural-network based simulation algorithm with practical speed-ups. We demonstrate the capabilities of our method with a series of complex liquid simulations, and with a set of single-phase simulations. Our method predicts pressure fields very efficiently. It is more than two orders of magnitudes faster than a regular solver. Additionally, we present and discuss a series of detailed evaluations for the different components of our algorithm.

GE_at_TUM

Droplets with Neural Networks

The final version of our paper on learning droplet formation models with neural networks is online now. It will be presented at SCA 2018 in Paris.

Our paper proposes a new data-driven approach to model detailed splashes for liquid simulations with neural networks. Our model learns to generate small-scale splash detail for the fluid-implicit-particle method using training data acquired from physically parameterized, high resolution simulations. We use neural networks to model the regression of splash formation using a classifier together with a velocity modifier. For the velocity modification, we employ a heteroscedastic model. We evaluate our method for different spatial scales, simulation setups, and Navier-Stokes solvers. Our simulation results demonstrate that our model significantly improves visual fidelity with a large amount of realistic droplet formation and yields splash detail much more efficiently than finer discretizations.

https://arxiv.org/abs/1704.04456

Further information

 

GE_at_TUM

Volumetric Fluid Flow from a Single Video

Our SCA 2018 paper on Coupled Fluid Density and Motion from Single Views is online now! You can check it out here. We’re reconstructing a dense fluid flow field from a single video stream using a strong Navier-Stokes prior.

To be presented at: http://sca2018.inria.fr

Full Abstract
We present a novel method to reconstruct a fluid’s 3D density and motion based on just a single sequence of images. This is rendered possible by using powerful physical priors for this strongly under-determined problem. More specifically, we propose a novel strategy to infer density updates strongly coupled to previous and current estimates of the flow motion. Additionally, we employ an accurate discretization and depth-based regularizers to compute stable solutions. Using only one view for the reconstruction reduces the complexity of the capturing setup drastically and could even allow for online video databases or smart-phone videos as inputs. The reconstructed 3D velocity can then be flexibly utilized, e.g., for re-simulation, domain modification or guiding purposes. We will demonstrate the capacity of our method with a series of synthetic test cases and the reconstruction of real smoke plumes captured with a Raspberry Pi camera.

GE_at_TUM

New paper online: Deep Fluids – A Generative Network for Parameterized Fluid Simulations

Our paper “Deep Fluids: A Generative Network for Parameterized Fluid Simulations” in collaboration with the CGL of ETH Zurich is online now!

You can check it out here:
http://www.byungsoo.me/project/deep-fluids/
https://arxiv.org/abs/1806.02071

The goal is a novel generative model to synthesize fluid simulations from a set of reduced parameters. A convolutional neural network is trained on a collection of discrete, parameterizable fluid simulation velocity fields. We also demonstrate that we can handle complex parameterizations in reduced spaces, and advance simulations in time by integrating in the latent space with a second network.

GE_at_TUM

Computer Game Laboratory, SS’18 up and running

The current instance of our computer game laboratory, a practical course for TUM games engineering students, is ongoing at the moment, you can check out the latest development of the four game projects on our wiki main page.

Look out for the final presentations and playable demos during the SS’18 demo day…

One interesting new development here is a first joint team between TUM and students from the Media Design Hochschule Muenchen (MDH).

You can also find a full list of previous games here.

 

GE_at_TUM

Ferienakademie 2018 – Accelerating Physics Simulations with Deep Learning

Computer simulations are a powerful method to study physical and engineering systems such as fluids, collections of molecules, or social agents. Traditionally, differential equations such as the Navier-Stokes equations form the basics of performing simulations as they dictate the time evolution. A recent, promising development is to use machine learning for model-free prediction of a systems’ behavior and to identify the appearance of spatio-temporal patterns. Deep learning with neural networks is a particularly interesting and powerful machine learning method that can be employed for this task.

More info can be found here.

GE_at_TUM

TEDx talk by Nils Thuerey

Nils recently gave a talk titled “Deep Learning Beyond Cats and Dogs” at a TEDx event organized at the Technical University of Munich.

Abstract: Deep learning, which is seemingly everywhere these days, is well-known for its capability to recognize cats and dogs in internet images, but it can and should be used for other things too. It can be used to figure out the complicated physics that dictate fluid behavior. Actually, simulating turbulence is not only a million dollar problem (really, google it!) but it can help us create more realistic virtual worlds. It can even help us understand medical and physiological behaviors like blood flowing through our body. Nils performs cutting-edge research and explains how neural networks are well on their way to becoming the fourth pillar of science.

Biography: Nils Thuerey’s work is in the field of computer graphics: he models physical behaviors of fluids such as water and smoke to enable computer created virtual effects to look like the real thing. These phenomena are very expensive to simulate computationally, so Nils’ research explores the use of deep learning methods to generate the effects more quickly and more realistically. Before assuming his assistant professor position at TUM, Nils studied in Erlangen, held a post-doc position in Zurich, and worked in the visual effects industry. He was awarded a technical Oscar for the development of an algorithm which aids in editing explosion and smoke effects for film.

This talk was given at a TEDx event using the TED conference format but independently organized by a local community. Learn more at https://www.ted.com/tedx , or visit the TEDxTUM website.

 

GE_at_TUM

Recent research on physics-based deep learning

In the following we give an overview of our recent publications on physics-based deep learning methods. In particular, we focus on solving various aspects of fluid problems modeled with the Navier-Stokes (NS) equations. These topics are a central theme of the research work in our group. While, naturally, the long term goal would be to simply give the initial conditions of a problem as input to a neural network, and then rely on the network to infer the solution with appropriate accuracy, the complexity of the NS equations makes this an extremely challenging problem.
Thus, we typically consider constrained solution spaces for sub-problems of the NS equations, that we believe are nonetheless very interesting, and useful for their respective domains. In this way, we are also working towards improving the state of the art in order to tackle more and more general problems in the future.
Our publications have targeted different aspects of a typical simulation pipeline, and differ in terms of how deeply integrated they are into the Navier-Stokes solve. The following list is order from loose to tight couplings. E.g., the last entry completely replaces a regular solver.
GE_at_TUM

mantaflow in Blender at BCon17

Sebastian Barschkis, CS-student at TUM, has just presented his latest progress regarding the integration of our mantaflow solver into Blender. You can check out his full presentation including insights about code structure as well as using the solver here:

There are admittedly still some rough edges, but mantaflow should give Blender users a significant step forward in terms of visual quality and performance.