The release blog post is here[1], model code released under MIT license here[2], along with weights on Huggingface, and some documentation here[3].
As a layman I've been following deep neural nets being used to solve quantum physics problems, where they do quite well for certain classes of hard problems, so perhaps not terribly surprising they do well with weather prediction as well I suppose.
Impressive. However, I don't like how AI foundation models are always advertised as alternatives to "traditional" (physics based) forecasting. Virtually all AI weather models are trained on ERA5 reanalysis, which is a blend of observations and numerical model forecasts. Without a good global numerical model of the atmosphere there would be no AI model. I wish this synergy were emphasised more, rather than always going straight for the easy "AI beats physics!!1!" headline.
Things are changing quickly in this area. Several projects/companies working on AI data assimilation (an alternative route to creating analysis data like ERA5)[0].
Also a lot of companies working on the data collection side, replacing/augmenting government data collection. Spire's an example of this in the space domain, and Windborne and Sorcerer (my company) do weather balloons.
I agree… the public and leaders need to know how the training data is generated: a combination of sensors and physics-based simulation models. Lacking this context could lead to poor decisions around research prioritization and funding.
Very cool project. There are some presentations by the PI on youtube that I recommend searching for. One of the interesting takeaways I had was that they were able to do better with mesoscale phenomena and extreme weather prediction than the other players (like Graphcast and Pangu and FourCastNet), in part due to their technique for training a higher resolution data space (0.1 deg vs 0.25 or 0.5). I also found it interesting that they were able to show a scaling relationship where performance increased by 5% every time they doubled the model size - and their loss was still improving when they had to cut it off due to cost constraints.
The are many great things about Aurora, here are a few as I've been using it since it came out.
1. Its open source & open weights and free to use non-commercially.
2. Its configurable to easily fit on my local gpu for development purposes.
3. I've also gotten great engagement from the repo owners.
Two of the authors of the original Aurora system left Microsoft to found https://silurian.ai/ - interesting to keep tabs on if you're interested in this space!
AI weather is making great progress with the likes of GraphCast, Aardvark, NeuralGCM, Aurora, etc. It seems like the teams that produce these models often include folks from Microsoft and Google, which makes me wonder if there's much cross pollination within those companies which is helping these advances or if researchers are siloed and the development of these models is entirely independent of one another?
Does this title make anyone else's Asimov senses tingle? In the book "Foundation and Earth", the protagonists travel from planet Aurora to planet Earth.
https://archive.md/vl87I
The release blog post is here[1], model code released under MIT license here[2], along with weights on Huggingface, and some documentation here[3].
As a layman I've been following deep neural nets being used to solve quantum physics problems, where they do quite well for certain classes of hard problems, so perhaps not terribly surprising they do well with weather prediction as well I suppose.
[1]: https://news.microsoft.com/source/features/ai/microsofts-aur...
[2]: https://github.com/microsoft/aurora
[3]: https://microsoft.github.io/aurora/intro.html
Also, research paper in Nature:
https://www.nature.com/articles/s41586-025-09005-y
Impressive. However, I don't like how AI foundation models are always advertised as alternatives to "traditional" (physics based) forecasting. Virtually all AI weather models are trained on ERA5 reanalysis, which is a blend of observations and numerical model forecasts. Without a good global numerical model of the atmosphere there would be no AI model. I wish this synergy were emphasised more, rather than always going straight for the easy "AI beats physics!!1!" headline.
Things are changing quickly in this area. Several projects/companies working on AI data assimilation (an alternative route to creating analysis data like ERA5)[0].
Also a lot of companies working on the data collection side, replacing/augmenting government data collection. Spire's an example of this in the space domain, and Windborne and Sorcerer (my company) do weather balloons.
[0]: E.g. Brightband's AIDA (https://www.brightband.com/blog/aida/) and Project Aardvark (https://www.turing.ac.uk/blog/project-aardvark-reimagining-a...)
I agree… the public and leaders need to know how the training data is generated: a combination of sensors and physics-based simulation models. Lacking this context could lead to poor decisions around research prioritization and funding.
Very cool project. There are some presentations by the PI on youtube that I recommend searching for. One of the interesting takeaways I had was that they were able to do better with mesoscale phenomena and extreme weather prediction than the other players (like Graphcast and Pangu and FourCastNet), in part due to their technique for training a higher resolution data space (0.1 deg vs 0.25 or 0.5). I also found it interesting that they were able to show a scaling relationship where performance increased by 5% every time they doubled the model size - and their loss was still improving when they had to cut it off due to cost constraints.
Very cool stuff!
The are many great things about Aurora, here are a few as I've been using it since it came out. 1. Its open source & open weights and free to use non-commercially. 2. Its configurable to easily fit on my local gpu for development purposes. 3. I've also gotten great engagement from the repo owners.
Is this similar to Google's model? https://research.google/blog/zooming-in-efficient-regional-e...
Two of the authors of the original Aurora system left Microsoft to found https://silurian.ai/ - interesting to keep tabs on if you're interested in this space!
Nature paper: https://www.nature.com/articles/s41586-025-09005-y
AI weather is making great progress with the likes of GraphCast, Aardvark, NeuralGCM, Aurora, etc. It seems like the teams that produce these models often include folks from Microsoft and Google, which makes me wonder if there's much cross pollination within those companies which is helping these advances or if researchers are siloed and the development of these models is entirely independent of one another?
Does this title make anyone else's Asimov senses tingle? In the book "Foundation and Earth", the protagonists travel from planet Aurora to planet Earth.
As superficial an application as this sounds, I wonder if this be used to make a fun sandbox simulation game…
Been following wesselb on GitHub for a while, great to see his work getting more attention!
We do we advertise for paywalled content here?
I'd suggest to take a look at the guidelines: https://news.ycombinator.com/newsfaq.html