science

14983 readers

552 users here now

A community to post scientific articles, news, and civil discussion.

rule #1: be kind

<--- rules currently under construction, see current pinned post.

2024-11-11

founded 2 years ago

MODERATORS

DeepMind AI weather forecaster beats world-class system (this time with Bayesian statistics) (www.nature.com)

submitted 1 week ago by zlatiah to c/science

4 comments fedilink hide all child comments

Google DeepMind has developed the first artificial intelligence (AI) model of its kind to predict the weather more accurately than the best system currently in use... The system, called GenCast, is described today in Nature.

Conventional forecasts, including those from ENS, are based on mathematical models that simulate the laws of physics governing Earth’s atmosphere... GenCast, by contrast, has been trained only on historical weather data...

So yeah DeepMind is fucking going at it again.

Interestingly the model architecture seems to heavily integrate Bayesian maximum likelihood estimation in addition to their usual GNN-based deep learning approaches, which I didn't know is even possible. Their methods section states "[o]ur innovation in this work is an MLWP-based Forecast model, and we adopt a traditional NWP-based State inference approach

I'm not super familiar with Bayesian methods though so if anyone can add some more information I'd appreciate it

References:

The paper: Price I, Sanchez-Gonzalez A, Alet F et al. Probabilistic weather forecasting with machine learning. Nature (2024). https://doi.org/10.1038/s41586-024-08252-9
The codebase: google-deepmind/graphcast. GitHub, accessed 2024-12-05. https://github.com/google-deepmind/graphcast

you are viewing a single comment's thread
view the rest of the comments

[–] [email protected] 3 points 1 week ago

It's hard to make a good weather forecast tool without theoretical elements incorporated in it.

I'm sure the model produces higher accuracy results on historical data. I read the abstract and it's not mentioned if they tested it on new data.

With ml, the most difficult part of the work is making sure not to overfit historical data. Is target have a less accurate model (on the training data) than a model that can be justified using theoretical reasoning. This way, I can be much more confident that it will work in the long run.

Let's see where this will go in the coming years!