Stephan Hoyer

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…

Check out our new API for seasonal aggreggations, including support for custom seasons! `ds.resample(time=SeasonResampler(["DJF", "MAM", "JJAS", "ON"]).mean()` and `ds.groupby(time=SeasonGrouper(["DJFM", "MAMJ", "JJAS", "SOND"]).mean()` xarray.dev/blog/season-gr…

"What's the Future of AI Weather Forecasting?" - Video of #SFClimateWeek Panel with @shoyer of @Google , Karthik Kashinath from @nvidia, Tim Janssen of @Sofarocean, and Joan Creus-Costa of @WindBorneWx youtu.be/QIPvMRgVdgE #AIWeather #AIWP #GenerativeAI

As someone who is both an AI and climate expert, let me point out that AI and climate risk have many qualitative differences. To start, there is much higher degree of concensus about climate change than about AGI.

At AGU I talked to NASA people about how agencies could better support open-source tools they rely on. I argued that our recent collaboration between Xarray and NASA ESDIS on xarray.DataTree was a good model to copy - read about how it happened here! xarray.dev/blog/datatree

Twitter was fun, but I think I'm done. X has too much clickbait and not enough science. Pretty exciting to see how quickly my network is replicating itself over on the other app!

xarray.DataTree is @xarray_dev's first major new data structure since it's initial release baxk in 2014!

This recording of this public talk on NeuralGCM is now available on Youtube: youtube.com/watch?v=n4Rw3R…

A nice summary of why I moved on from ML for PDEs. The literature on weather & climate modeling isn't perfect, but the baselines are certainly much more compelling and easier to find.

I'm giving a public talk next week in downtown Palo Alto on "AI-based Weather and Climate Modeling with NeuralGCM." Would love to see you there! It's the evening of Thursday Oct 3: lu.ma/bx8t93rc

COMMUNITY NOTE: Humans are unlikely to step foot on Mars until the 2040s/2050s at the earliest, and there’ll never be a significant human settlement there because Mars is basically uninhabitable. Except for scientific research, there’s no compelling reason for humans to be there.

At long last Xarray can now group by multiple arrays. The culmination of 3 years of on/off work xarray.dev/blog/multiple-… Try it!

Genuine question: What is the best evidence for the value of physics-based regional weather/climate modeling? WRF certainly "looks" more realistic, but when is it actually consistently quantitatively superior to statistical post-processing of global models?

NeuralGCM is the result of persistence and long-term vision from @shoyer and the team. 4 years ago I wrote a blog post on the difficulties of building hybrid weather/climate models as all approaches back then were trained "offline". raspstephan.github.io/blog/optimizat… 1/

🌍We've developed the first conditional generative model for accurate & efficient ensemble climate simulations! 📄Read the full paper: arxiv.org/abs/2406.14798 🎉Excited to share that a preliminary version was accepted as an oral presentation @ml4esmworkshop at #ICML2024! 🧵

New @nature paper: nature.com/articles/s4158… NeuralGCM results (all are a "first"): 1) A differentiable hybrid atmospheric model 2) Competitive with ECMWF ensemble 3) Competitive with a GCRM in a year-long simulation 4) 40-year AMIP-like runs. Smaller bias than AMIP runs. 1/

In @Nature today: NeuralGCM, a breakthrough in climate modeling. It combines physics-based modeling with AI, and is up to 100K times more efficient than other models for simulating the atmosphere, providing scientists with new tools for predicting climate change.…