Machine learning and CFD simulators
 

Hello everybody,

do you, and if yes, how do you leverage machine learning methods in your CFD codes or post-processing / analysis?

I’m looking to learn about use-cases / workflows for the platform we’re building at our company in the area of physics-informed ML in CFD.

I'm not using one yet, but I'm doing some research in this field.

It's exciting to see how ML may be applied to several branches of research, particularly physics! I'm interested in studying ML, albeit for more hedonistic reasons such as high-paying jobs in Data Science, but I constantly grin and get thrilled when I see how this burgeoning discipline is being used to addressing open challenges such as fluid computing, quantum, and even biology such as protein folding.

Maybe someone here is already using it and can help us out.

PINNs
 

I've been talking to bunch of people doing research around physics-informed neural networks (PINNs) and for last few months at DimensionLab we've also experimented with graph neural networks applied to CFD (mostly around work by DeepMind - https://sites.google.com/view/learning-to-simulate and https://sites.google.com/view/meshgraphnets). There's not much specialized tooling for such stuff so everybody is hacking around and experimenting. Great times! There's also been some interesting course on ML in CFD by Von Karman Institute recently (few months back).

We're currently trying to train few AI-based simulators on datasets from simulations of wind blowing at cloth, wind blowing at ship sail, vortex shedding, and two fluids mixing in a tank. We have some software tooling in works to streamline the process of creating such "learnable" simulators (https://www.siml.ai/) that we hope to start beta-testing in July/August.

It's very interesting to see how this will play out and what immediate and long-term benefits/changes will this bring into the CFD landscape.

Cool example of merging ML and CFD for automotive and energy use-cases is MonolithAI (https://www.monolithai.com/industry/reduce-simulations). Didn't yet try their product, but from the stuff they shared around the internet, it seems pretty solid.

Prof. Brunton has a nic introductory video on ML in CFD on YT: https://www.youtube.com/watch?v=IXMSOSEj14Q

nice to see fellow people interested in this field
 

My project involves ML too and i would like to see more people interacting in this thread.

where is the best place to learn ML content online? (paid/unpaid)

ML+CFD resources
 

Here are few resources (presentations and python codes) from the ML in CFD course my company participated in:

https://github.com/DimensionLab/vki-course

It was really great course, lot of examples. Definitely look at some of those.

Thanks a lot :)

Check out some videos by Polytechnique on CFD systems and learning methods here:
https://www.youtube.com/watch?v=4t7lYQ5Luvo
https://www.youtube.com/watch?v=TZvtnpJsDXw

If you understand French, there are more on the channel as well.
Best,
Poly

SciML and PINNs
 

Recently I've been researching more into various approaches for leveraging AI in physics and it seems that scientific machine learning (SciML) ecosystem around Julia is going pretty strong https://sciml.ai/.

They have loads of packages available for building physics-informed neural networks (PINNs), using neural operators and working with high-dimensional PDEs. I highly encourage everybody to look into this - it might be very useful for modeling and simulating CFD problems.

On the other hand, recently we started using NVIDIA's Modulus framework in SIML.ai for building PINNs. It's extremely powerful! They provide various examples, e.g. how to train AI simulator to simulate wave propagation, or automatic optimization of heatsink geometry based on heat transfer.

Hope some of these resources help you guys in navigating this exciting area!

Hexagon sells a code called "Odyssee" that uses ML in CFD