The popular press has been saying that an AI weather prediction model is fifteen days better than traditional prediction models.
Nature article,Probabilistic weather forecasting with machine learning, reports three incredibly impressive results from Google DeepMind‘s #Gencast that aren’t quite that.
The results are on 15-day advanced predictions. The Machine-Learning model is better than the traditional numeric model on the 15-day advance prediction test.
The first of three ways that it gets better results, which is good.
The second of three ways is that it produces specific ranges for each prediction (where each prediction is a small piece of space over a small period of time), rather than just a single average range to be applied to all.
And the third way is that it takes less computation / less energy than the traditional numeric methods.
In 1984, when I started playing with super-computers, a top-line supercomputer of the day, a Cray-XMP could do a really accurate 12-hour forecast in 24 hours. As the length of the forecast got longer, that ratio quickly got worse, but it was easy to see that Moore’s law would come to our rescue.
Great efforts went into optimizing those programs, and the idea that “AI,” which I think of as “crazy computationally expensive,” might be cheaper caught me by surprise.
After a little thought, I remembered that the training of the AI/ML is expensive rather than the running the models. Still, it makes me smile. And apparently, I won’t need an umbrella anytime soon.
Ilan Price Alvaro Sanchez Gonzalez Ferran Alet Puig #tomrandersson Andrew El-Kadi Dominic Masters Timo Ewalds #jacklynnscott Shakir Mohamed Peter Battaglia Rémi Lam Matthew Willson Nature Magazine
