Boilerplate for integrating ML models into ICON weather/climate model

Announcements
1

Since 2022, there has been a surge in integrating and developing machine learning models for numerical weather prediction and climate models [1]. Here in Germany, several organisations (DWD, DLR, IOW, IAP, etc.) are actively exploring ML integration into the Icosahedral Nonydrostatic (ICON) model.

We are using FTorch at work with the ICON model, and recently I have been contributing to FTorch to simplify its integration into ICON and also wrote boilerplate configuration for ICON so scientists can quickly get their ML models running. Check out my GitHub repository with the public ICON source code and support for FTorch here:

If you find it useful, feel free to star :star: my repo and FTorch on GitHub. Even if you aren’t interested in ICON, I think it’s still good for the people here to know about FTorch since you can develop your models in PyTorch directly then easily export them to a format that can be consumed in Fortran. Even when alternative approaches exist like neural fortran, I think it’s always useful to spread the word :))

[1] W. Dong, “AI foundation models for whole atmosphere climate: Development and applications,” presented at the CCMC Workshop, Maryland, USA, June 5, 2024. Available: poster.

6 Likes
2

Thanks, @jaredfrazier. This is very interesting to me, as one of my PhD students is working on stats/ML based models for parallelisable GPU-optimised downscaling of extremes. Later in her PhD, the direct integration into ICON as a module will be explored.

1 Like
3

Glad one of your students is working on this :)) I plan to add more documentation on running ICON for people who don’t have direct access to ICON consortium HPC systems (e.g., DKRZ’s Levante, DWD’s NEC-Aurora, etc), but my repo should help get your student pointed in the right direction. ICON has been open-sourced fairly recently, so unfortunately such documentation for external use is still a bit lacking, but we’re working on it. Cheers!

1 Like
4

That would be great. I’m happy to forward any feedback (what’s understandable, what barriers remain for easy adoption). I plan to run it on our GPU cluster here too at some point. (Previously, I had just run ECHAM on CPUs). I’m glad to see this open and inviting approach by the ICON community!

1 Like