Master-Seminar - Deep Learning in Physics (IN2107, IN0014)
|Time, Place|| |
Wednesdays 14:00-16:00 Online
20th., October 2021
Using deep learning methods for physical problems is a very quickly developing area of research. The research group of Prof. Thuerey has studied learning-based methods for Navier-Stokes problems and fluid flow applications in recent years, examples of which include learning latent-spaces for physical predictions, generative adversarial networks with temporal coherence, and the inference of Reynolds-averaged Navier-Stokes flows around airfoils. Beyond these physics-based deep learning works of the Thuerey group, this seminar will give an overview of recent developments in the field.
In this course, students will autonomously investigate recent research about machine learning techniques in the physical simulation area. Independent investigation for further reading, critical analysis, and evaluation of the topic are required.
Participants are required to first read the assigned paper and start writing a report. This will help you prepare for your presentation.
- It is only allowed to miss a single time-slot. Missing a second one means failing the seminar. If you have to miss any, please let us know in advance.
- An advisor is assigned to each one with the paper.
- Two weeks before the talk there will be a mandatory meeting with your advisor to review the report and discuss the structure of the presentation.
- A short report (4 pages max., excluding references in the ACM SIGGRAPH TOG format (acmtog) - you can download the precompiled latex template) should be prepared before the meeting with the advisor.
- Guideline: You can begin with writing a summary of the work you present as a start point; but, it would be better if you focus more on your own research rather than just finishing with the summary of the paper. We, including you, are not interested in revisiting the work done before; it is more meaningful if you make an effort to put your own reasoning about the work, such as pros and cons, limitation, possible future work, your own ideas for the issues, etc.
- The participants have to present their topics in a talk (in English), which should last 25 minutes. Don't put too many technical details into the talk, make sure the audience gets the paper's main idea. Be prepared to answer questions regarding the technical details, you could prepare backup slides for that.
- Afterwards, a short discussion session will follow.
- Plagiarism is important; please do not simply copy the original authors' slides. You can certainly refer to them.
- The semi-final slides (PDF) should be sent one week before the talk; otherwise, the talk will be canceled. We strongly encourage you to finalize the semi-final version as far as possible. We will take a look at the version and give feedback. You can revise your slides until your presentation.
- Be ready in advance. We suggest testing the machines you are going to use before the lecture starts. You can bring your laptop or ask us one (also any converter you need for the projector) in advance. A laser pointer will be provided, so you can use if you want.
- The final slides and report should be sent after the talk.
|Paper list sent via Email|
|Send preferred 5 topics/papers by 05 September 2021 (Sunday 23:59)|
- 20.10 Kickoff (Introduction lecture)
- 27.10 No meeting
- 3.11 No meeting
- 10.11 Meeting #1 (First presentation)
- 17.11 Meeting #2
- 24.11 Meeting #3
- 1.12 Meeting #4
- 8.12 Meeting #5
- 15.12 Meeting #6
- 22.12 No meeting
- 12.1 Meeting #7
- 19.1 Meeting #8
- 26.1 Meeting #9 (if necessary!)
|No.||Paper||Date||First name||Last name||Supervisor|
|1||Combining Differentiable PDE Solvers and Graph Neural Networks for Fluid Flow Prediction||10 November||Louis||Dorge de Vacher de Saint Géran||Chen|
|2||Machine learning accelerated computational fluid dynamics||10 November||Tianhao||Lin||Holzschuh|
|3||Assessment of unsteady flow predictions using hybrid deep learning based reduced-order models|| |
|4||Solver-in-the-Loop: Learning from Differentiable Physics to Interact with Iterative PDE-Solvers|| |
|5||Deep learning methods for super-resolution reconstruction of turbulent flows|| |
|6||Transfer learning for nonlinear dynamics and its application to fluid turbulence https://arxiv.org/pdf/2009.01407.pdf|| |
|7||Learning to control PDEs with differentiable physics https://arxiv.org/pdf/2001.07457.pdf|| |
|8||Discovering physical concepts with neural networks https://arxiv.org/pdf/1807.10300.pdf|| |
|9||Neural Ordinary Differential Equations https://arxiv.org/pdf/1806.07366.pdf|| |
|10||Hamiltonian Neural Networks|| |
|José Miguel|| |
|11||Physics Informed Deep Learning: Data-driven Solutions of Nonlinear Partial Differential Equations https://arxiv.org/pdf/1711.10561.pdf||8 December||Christina||Nuss-Brill||Holzschuh|
|12||Model identification of reduced order fluid dynamics systems using deep learning|| |
|13||Deep learning for universal linear embeddings of nonlinear dynamics https://www.nature.com/articles/s41467-018-07210-0.pdf|| |
|14||tempoGAN: A Temporally Coherent, Volumetric GAN for Super-resolution Fluid Flow https://arxiv.org/pdf/1801.09710|| |
|15||Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations||15 December||Stefan||Frisch||Holzschuh|
|16||Lagrangian Fluid Simulation with Continuous Convolutions https://openreview.net/pdf?id=B1lDoJSYDH|| |
|17||Solving high-dimensional partial differential equations using deep learning https://www.pnas.org/content/115/34/8505|| |
|18||Learning data-driven discretizations for partial differential equations https://www.pnas.org/content/pnas/116/31/15344.full.pdf|| |
|19||SPNets: Differentiable Fluid Dynamics for Deep Neural Networks https://arxiv.org/pdf/1806.06094.pdf|| |
12 Jan. 2022
|20||Augmenting Physical Models with Deep Networks for Complex Dynamics Forecasting https://arxiv.org/abs/2010.04456|| |
12 Jan. 2022
|21||Discovering Symbolic Models from Deep Learning with Inductive Biases https://arxiv.org/pdf/2006.11287.pdf|| |
19 Jan. 2022
|22||Differentiable Strong Lensing: Uniting Gravity and Neural Nets through Differentiable Probabilistic Programming https://arxiv.org/pdf/1910.06157.pdf|| |
19 Jan. 2022
|23||DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators arxiv.org/abs/1910.03193|| |
- Thuerey group: List of Publications (including Physics-based Deep Learning works)
- Book: Bishop, Pattern Recognition and Machine Learning
- Book: Hastie et al., The Elements of Statistical Learning
- Online: Nielsen, Neural Networks and Deep Learning
- Online: Ruder, An Overview of Gradient Descent Optimization Algorithms