PhD student in computer science, bringing physics into machine learning models. Researcher @know_center, Physics-Informed Deep Learning for CFD
Graz, Austria
Joined February 2020
- Tweets33
- Following90
- Followers23
- Likes231
Franz Martin Rohrhofer retweeted
Are Physics-Informed Neural-Networks (#PINNs) able to predict chaotic motion? We investigated this in our paper "How PINNs cheat: Predicting chaotic motion of a double pendulum" that will be presented in the DLDE workshop at #NeurIPS22: openreview.net/forum?id=shUb…
TL;DR below
1 🧵
Still impressed with how modern mobile phones and some easter traditions can be used to nicely capture turbulent combustion - a topic which needs to be further studied in order to decrease pollution and carbon emissions!
Why is it precarious to use Physics-Informed Neural Networks (#PINNs) for simulations of dynamical systems?
Often, we encounter poor convergence behavior and lack of physical correctness...
Let me show you what we've found out!
(Spoiler alert 📢 arxiv.org/abs/2203.13648)
1/🧵
Interestingly, these nonphysical solutions can even conform to the global minimum of the physics loss optimization, as we found out.
What helps in this regard, is reducing the computational domain, i.e., the simulation time.
7/🧵
If you want to read up more on this topic, please have a look here 👉 arxiv.org/abs/2203.13648
Special thanks to @BernhardGeiger 🙏!
If you want to play around with the PINN code by yourself, download the basic version here 👉 github.com/frohrhofer/PINN_T…
end/🧵
Franz Martin Rohrhofer retweeted
Vector calculus & PDEs!
youtu.be/Jt5R-Tm8cV8
Powerful & beautiful intersection of math and physics. How to encode physical conservation laws as partial differential equations.
And still relevant for modern machine learning!
(~20 vid series... how I wish I'd learned it)
Franz Martin Rohrhofer retweeted
1/12 We - Russia - want to be a nation of peace. Alas, few people would call us that now.
Finally received today! Looking forward to interesting insights and new methods for tackling our challenges in turbulent fluid flow! 💨
Also great to see such valuable content freely available on YouTube! 🙏 @eigensteve
Franz Martin Rohrhofer retweeted
Danke für diesen Artikel @derStandardat! Extrem gute Aufarbeitung (graphisch wie auch inhaltlich) unseres Einflusses auf das Klima in den nächsten Jahrzehnten. Definitive Lese-empfehlung derstandard.at/story/2000132…
Physics-informed neural networks (#PINNs) have become a powerful tool to tackle physical-related problems; their convergence, however, is a vague issue.
In our recent work, we shed light on the multi-objective optimization in PINNs: arxiv.org/abs/2105.00862
In a nutshell:
1/
Adaptive training methods, such as adaptive activation functions or adaptive loss weighting, improve the convergence but still are not able to overcome the infeasible region of the multi-objective optimization.
4/
Thanks a lot to all the collaborators in this project! @BernhardGeiger @Know_Center and Stefan Posch @LargeLec.
The authors also acknowledge the financial support of the Austrian #COMET program, managed by the Austrian Research Promotion Agency (#FFG).
end/