For me, all these questions are open. The challenges are exciting, but at the moment I feel lost. A lot to think about!

4. How to use PINNs, or should we use PINNs in climate modeling?

3. Related: How many models should we use to train AI models? One of the main strengths of the CMIP exercise is that it is a multimodel ensemble. If everyone uses the same dataset to train emulators (e.g. ERA5), will we be able to consider several emulators as a multimodel ensemble?

2. How to create the learning samples? Which parts should we emulate or improve with ML? Should we use observational data (if so, how?)? Since we don't have « observations » from warmer climates, we should also use model data (or at least physical constraints). But model data have biases.

This is a challenging issue that requires a lot of expertise in numerical modeling of the climate, which only a few people has worldwide. I’ve been using ARPEGE-climat since ~5 years now, but I think that this is not sufficient.
And things are changing fast, so it is difficult to make decisions.

1. How to adapt « legacy » Fortran codes to new hardwares? The DSL solution seems appealing (eg. Using GT4Py), but maybe using JAX in Python could be sufficient? Both options rely on packages that requires to be maintained (seems OK at the moment).

All this got me thinking about the use of ML/AI in climate science. In contrast to NWP, hybrid approaches still seem to be the best option. But there are tons of problems to solve, like:

🥇 ArchesWeather (Couairon et al., 2024).
Trained on ERA5 data using just 2 A100 GPUs for 2.5 day — an impressive achievement! This model, ArchesWeather, rivals other SoTA AI NWP models at 1.5° resolution, thanks to innovations in the attention layer.

arxiv.org/abs/2405.14527

🥈 Bano-Medina et al., Towards calibrated ensembles of neural weather model forecasts.
White the need to perturb model parameters can be debated, this paper tackles the challenge of sampling both model and input uncertainties in NN-based weather prediction.

essopenarchive.org/users/777909...

🥉 Hakim et al., Dynamical Tests of a DL Weather Prediction Model.
This short paper evaluates whether the dynamical behavior of PanguWeather aligns with expectations, by assessing the response of the model to small perturbations of the input.

journals.ametsoc.org/view/journal...

I also read a lot and love sharing papers, websites, and GitHub repos I find interesting — something I hope to continue here.
Excited to connect with other AI and NWP/Climate enthousiasts!
3/3

Previously, I worked on efficient Fortran/Python coupling for a full GCM (ARP-GEM1) on heterogeneous HPC resources (using both CPU and GPU nodes at the same time). Now, I’m back to the AI side, focusing on sparse physics-informed neural networks for climate modeling.
2/3

Hi Ferran! You’ve just been added!

Of course, you have been added!

Done!

Haha I wish there was an option for that too. Thanks for sharing!