GenCast is out in @Nature! It's the first high res ML ensemble weather forecast which outperformed the operational state of the art. And few more things have happened since the preprint was first released ⬇️🧵.

I had a great time speaking at Cambridge yesterday. Thanks @PetarV_93 @pl219_Cambridge for the invite!.

RT @PetarV_93: I'm excited to share that we'll have @IlanPrice giving a talk @Cambridge_CL on GenCast -- a state-of-the-art model for proba….

RT @MaxTsaparis: An AI-powered weather forecast model has just been unveiled by @GoogleDeepMind, and it's outperforming the current best pr….

RT @nytimes: A new A.I. tool from DeepMind, a Google company in London that develops A.I. applications, has achieved what its makers call u….

@Nature Work with a remarkable team: Alvaro Sanchez-Gonzalez, @FerranAlet, @tom_r_andersson, @andrew_elkadi, @dominic_masters, @TimoEwalds, Jacklynn Stott, @Shakir_za, @PeterWBattaglia, Remi Lam, & Matthew Willson.

@Nature 5. Check out the paper: The code: And the blog:

@Nature 4. Overall, GenCast marks something of an inflection point in the advance of AI for weather prediction, with SOTA raw forecasts now coming from AI. I think we can expect them to be increasingly incorporated operationally alongside traditional models (and to continue to improve!).

@Nature 3 cont) *caveats on this example* a) individual examples always to be taken with a pinch of salt - need rigorous evaluation over an extended period, see the cyclone track section of the paper. b) these are track predictions, not intensity predictions.

@Nature 3 cont) GenCast predicted 60-80% probability of landfall in Florida already from 8.5 days before landfall eventually happened - a couple of days before Milton even formed - and more than 90% from 5.75 days before.

@Nature 3. We also recently fine-tuned the model to work on operational inputs so that it can be run live. We conducted a retrospective analysis of this model's forecasts of the track Hurricane Milton 🌪️.

2. Alongside publication in @Nature, we are making the code and model weights available to the community (incl. a mini version of the model which gives reasonable results and can run in a free colab). And soon we'll share an archive of the model's historical & current forecasts.

1. The paper shows GenCast provided better probabilistic weather forecasts, including better forecast of extreme weather, than operational gold standard over our year-long evaluation period. This could mean earlier preparation for extreme events, more reliable wind power, & more

RT @PeterWBattaglia: We're hiring a Research Scientist in AI for Sustainability @GoogleDeepMind (in Mountain View or Cambridge MA). Seeking….

RT @flimsin: These are fun to record - we answer your climate Qs. Not sure how often I'll give a shout out in one programme to @GoogleDeepM….

Work with my brilliant colleagues: Matthew Willson, Alvaro Sanchez-Gonzalez, @FerranAlet, @tom_r_andersson, @andrew_elkadi, @dominic_masters, @TimoEwalds, Jacklynn Stott, @shakir_za, Remi Lam, @PeterWBattaglia 8/8.

Read the newly updated paper for full details and even more evaluations and ablations (and an 80-page appendix!). 7/8.

Importantly, GenCast ensembles have well-calibrated uncertainty, allowing users to trust that the model generally has the right level of confidence in its own predictions (flat rank histograms & spread/skill ≈ 1). 6/8.

GenCast ensembles capture spatial structure and correlations, which is necessary for important applications like renewable energy planning and grid management. In a simplified regional wind power prediction task, GenCast improves on ENS by >20% at 24 hours ahead. 5/8

GenCast’s forecasts are overall more skillful, providing more value to decision makers faced with the risk of extreme heat/cold/wind or the chance of being hit by a tropical cyclone. 4/8