Master-Seminar - Deep Learning in Physics (IN2107, IN0014)

Prof. Dr. Nils Thuerey Patrick Schnell, Björn List 

Content

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.

Paper Selection

The paper list for this seminar can be found below. Please send us an email with your 5 ranked preferences by Tuesday, October 8th. 

We will assign papers based on these preferences, while remaining papers will be distributed randomly. The paper-matching results will be made available before the Kick-off meeting.

Requirements

Report

When:

• A semi-final version is due one week before your talk (Wednesday by 23:59)
• Send your final report within one week after your talk (Wednesday by 23:59).

Style:

• Please use ACM SIGGRAPH TOG format (acmtog) precompiled latex template. You can also download the precompiled latex template.
• Maximum 6 pages excluding references.

Guidelines:

• You can begin with writing a summary of the work you present as a starting 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 include your own reasoning about the work, such as pros and cons, limitations, possible future work, your own ideas for the issues, etc. The final version of your report must cover a summary of the assigned paper and your own work.
• The semi-final version must contain a complete summary of the paper. Failing to meet this requirement will result in cancellation of the talk and failing the seminar.

Slides

When:

• Send semi-final slides at least one week before your presentation, together with the semi-final report. Otherwise the presentation will be cancelled. Please also make an appointment with your advisor when you send your slides. There is an mandatory discussion with your advisor in the week before your presentation. Your advisor will give your feedback on your slides. 
• Send final slides within one week after your presentation to us (Wednesday by 23:59).

Style:

• Any slide style you like, prepare slides as PDF file.

Guidelines:

• Ensure readability (colors, images and font size).
• Avoid using too much text.
• Highly encouraged to do some paper-related experiments and show some results in the presentation.
• The final version of your slides must cover a summary of the assigned paper and your own work.
• The semi-final version must contain a complete summary of the paper. Failing to meet this requirement will result in cancellation of the talk and failing the seminar.

Presentation

• Present your topic in English.
• You have 25 minutes for presentation and 10 minutes for questions and discussion.
• Please actively participate in the discussion for other presentations.
• Please test your setup (laptop/connection to projector) before giving your presentation!

Attendance

• Missing one session is allowed, if you let us know in advance and write a short summary of the papers (ca. 1 page) in your own words.
• Missing another session means failing the seminar (special rules for severe issues as appropriate).

Paper list

Nb.	Title
1	Physics Informed Deep Learning: Data-driven Solutions of Nonlinear Partial Differential Equations
2	DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators
3	Fourier Neural Operator for Parametric Partial Differential Equations
4	Learning to control PDEs with differentiable physics
5	Do Differentiable Simulators Give Better Policy Gradients?
6	Accelerated Policy Learning With Parallel Differentiable Simulation
7	On the difficulty of learning chaotic dynamics with RNNs
8	The curse of unrolling
9	low variance gradient optimization in unrolled computation graphs with es single
10	Automatic Symmetry Discovery with Lie Algebra Convolutional Network
11	Lagrangian Fluid Simulation with Continuous Convolutions
12	Message Passing Neural PDE Solvers
13	PDE-Refiner: Achieving Accurate Long Rollouts with Neural PDE Solvers
14	Convolutional Neural Operators for robust and accurate learning of PDEs
15	Learning skillful medium-range global weather forecasting
16	Machine learning–accelerated computational fluid dynamics
17	Deep learning for universal linear embeddings of nonlinear dynamics
18	Learned coarse models for efficient turbulence simulation
19	Learning data-driven discretizations for partial differential equations
20	DiscretizationNet: A machine-learning based solver for Navier–Stokes equations using finite volume discretization

 

You can access the papers through TUM library's eAccess.

Seminar Schedule

Date	Paper Number	Presenter	Advisor
16.10.24	Kickoff
23.10.24	no seminar
30.10.24	no seminar
06.11.24	1	Cortes Cendales	Patrick
06.11.24	2	Ortiz Del Angel	Patrick
06.11.24	19	Kocabasoglu	Bjoern
13.11.24	3	Tegeltija	Patrick
13.11.24	16	Ganahl	Bjoern
13.11.24	20	Nguyen	Bjoern
20.11.24	4	Uyar	Patrick
20.11.24	5	Liu	Patrick
20.11.24	18	Hertel	Bjoern
27.11.24	6	Bucher	Patrick
27.11.24	11	Grgič Meško	Bjoern
27.11.24	12	Apostolidou	Bjoern
04.12.24	7	Heckel	Patrick
04.12.24	8	Aslan	Patrick
04.12.24	15	Dojcsák	Bjoern
11.12.24	9	Chou	Patrick
11.12.24	14	Goroncy	Bjoern
11.12.24	17	Stefanov	Bjoern
18.12.24	10	Hattler	Patrick
18.12.24	13	Vidulejs	Bjoern