Ryan Keisler

📢 Time to share a project I’ve been working on: Forecasting Global Weather with Graph Neural Networks 🧵 (1/N)

@jcamdr70 @benedictk__ Agreed, it's a good and under-explored direction. Here they used an AI weather model to discover a problem with the gold-standard record of the atmospheric state: arxiv.org/abs/2601.04701

@RyanKeisler @benedictk__ Very interesting, congratulations! This raises a fascinating question: how do we identify what the AI ​​model possesses that the physical model lacks?

I'm excited to finally open-source the model from my 2022 paper, “Forecasting Global Weather with Graph Neural Networks”. Highlights: • 10-day forecast in <1 min • Initialize forecasts from ERA5 or IFS analysis • Scripts for eval, sensitivities, & Hurricane Sandy

@benedictk__ It has skill, i.e. better than the avg weather, out to 9 or 10 days. When I released the paper in 2022, this was the SOTA AI model. Fast forward to 2026, and the top (open) AI models are from Google DeepMind and ECMWF, and are generally better than the best physics-based model.

@RyanKeisler how good is it?

@TeksEdge Good question! I haven't profiled it, but I bet it will work on most laptops. I would start with CPU only (the default). You'll need to first install git and uv. If you try it and run into any issues, please LMK via a github issue. #installation" target="_blank" rel="nofollow noopener">github.com/rkeisler/keisl…

@RyanKeisler Congratulations! What are the minimum requirements for model/GPU support? I'd love to predict my weather.

@dcxStep Yea, you've got it. The only detail I would add is red = positive derivative = a perturbation to U at some point x reinforces a same-signed perturbation to U at the cursor point 24h later; blue = negative derivative = reverse-sign perturbation at cursor point.

@RyanKeisler And I guess blue is east and red is west

@shoyer Thank you, Stephan!

@RyanKeisler This paper was such a breakthrough! Reading it was the first time I believed that SOTA pure-AI weather prediction was possible. Thanks for sharing, Ryan.

@paul_skeie I knew that I wanted to highlight the forecast sensitivities because they're so fast to compute (I think this was 5 minutes?) and under-visualized. Then I happened to land on d(u500)/d(u500) as a nice one.

@RyanKeisler It is very cool what you did there. How did you come up with the idea to do that?

My hope is that, as a relatively lightweight model, this can serve as a starting point for anyone interested in ML weather prediction. github.com/rkeisler/keisl…

The video above illustrates an example of forecast sensitivity: how does tomorrow’s wind at a certain location depend on today’s winds everywhere else?

@shoyer I thought OmniCast by @tungnd_13 and collaborators was an interesting step away from autoregressive training. arxiv.org/abs/2510.18707

Genuine question for those who train LLMs -- are there any effective strategies for going beyond next-few-token prediction for pre-training? This is serious issue for building AI climate models, which are currently trained on weather forecasting (predicting forward a few days).

Our internship program at Polymathic is open for opportunities from now through fall 2025! I believe our program provides an opportunity to work alongside some of the best researchers and engineering experts in the world — exploring the unknown of building foundation models for science, together. These are full-time and paid positions in the Big Apple! Interested? 👇 forms.gle/Jm9v4VxJrna7VR… Deadline Nov 5!

I'm incredibly proud to share NeuralGCM, our new AI and physics based approach to weather and climate modeling with state-of-the-art accuracy, published today in @Nature: nature.com/articles/s4158…

@shoyer @yuqirose @PierreGentine @DWatsonParris @AstroVivi @StephanMandt @SciPritchard Yes! I remember seeing these window-to-window jumps most strongly in the low-pressure levels of ERA5. I was initially surprised but it makes sense in hindsight since there is no overlap in the underlying obs data (except indirectly, as it informs initial guess for the next state)

@yuqirose @PierreGentine @DWatsonParris @AstroVivi @StephanMandt @SciPritchard Amusingly ERA5 fails this test -- if you look at high enough resolution with hourly output you can see the forecasts reset every 12 hours (corresponding to a new data assimilation window).

1/2 We hear a lot about "Earth twins" these days. There are still major challenges on the way before we can get to Earth's twins whether using high-resolution Earth system models or AI-based models but progress is happening fast, especially on the AI front. Some comparisons:

New open source release from my team at Google: Dinosaur, a differentiable dynamical core for global atmospheric modeling, written in JAX: github.com/google-researc… Dinosaur is a core component of NeuralGCM and we hope it is useful for the weather/climate research community.

Anyway, I'd love to learn more! @IlanPrice @FerranAlet @shakir_za @PeterWBattaglia

Or a mix? E.g. the EDA perturbations are a little under-dispersed, some additional stochasticity (flow-dependent SVD, gaussian process, etc) helps, & the DM stochasticity goes in this bucket? And then the DM sample variance was (lightly) tuned to optimize performance/calibration?

Or, alternatively, is this DM-only, 1-initial-condition ensemble *under*-dispersed, and the good ensemble calibration that is observed in Section 5.4 comes primarily from using initial conditions based on ERA5 EDA perturbations?

(Sounds nice qualitatively, but would the amplitude of the sample variance depend on the details of DM training process, e.g. noise schedule? I'm not sure.)

The idea would be, when the DM generates samples conditioned on current state, there is higher sample variance in the nodes that are more difficult to predict (eg Fig 5 of arxiv.org/abs/2309.01745 @thuereyGroup). If you squint, this starts to look like a well calibrated ensemble.

The DM is trained on (deterministic) ERA5. This raises an interesting question: if you generate an ensemble from the *same initial conditions*, such that all stochasticity is sourced from the DM rather than the spread in the initial conditions, is the ensemble well calibrated?

From Section A.4.1 of the paper, the diffusion-model-only stochasticity "works relatively well". Interesting. If I put on my optimist hat, I can imagine a scenario where the DM learns to generate well calibrated samples "for free", even when learning from deterministic ERA5.