This is incredibly cool stuff, it's great to see it being worked into actual production models - particularly combining it with deterministic physics-based predictions to balance out the strengths...
This is incredibly cool stuff, it's great to see it being worked into actual production models - particularly combining it with deterministic physics-based predictions to balance out the strengths and weaknesses of both.
DeepMind themselves are actually one iteration beyond that by now, using a technique they call functional generative networks - somewhere around 8% better performance, according to their own metrics - but they haven't open sourced the new implementation and it looks like they probably don't plan to.
Google seem to be selling access to that one under the name WeatherNext 2, but for the amount of serious research DeepMind performs, publishes in detail, and open sources I'm going to take off my "fuck corporate incentives" hat for once and say that I don't actually begrudge them monetising that one when they've already published the initial version that made the real serious leap from previous methods. The FGN paper actually gives a pretty thorough explanation of how to implement it in the appendix, too, so I'd expect the NOAA are more than capable of coding their own version from there.
The new suite of AI weather models includes three distinct applications:
AIGFS (Artificial Intelligence Global Forecast System): A weather forecast model that implements AI to deliver improved weather forecasts more quickly and efficiently (using up to 99.7% less computing resources) than its traditional counterpart.
AIGEFS (Artificial Intelligence Global Ensemble Forecast System): An AI-based ensemble system that provides a range of probable forecast outcomes to meteorologists and decision-makers. Early results show improved performance over the traditional GEFS, extending forecast skill by an additional 18 to 24 hours.
HGEFS (Hybrid-GEFS): A pioneering, hybrid "grand ensemble" that combines the new AI-based AIGEFS (above) with NOAA’s flagship ensemble model, the Global Ensemble Forecast System. Initial testing shows that this model, a first-of-its kind approach for an operational weather center, consistently outperforms both the AI-only and physics-only ensemble systems.
...
The team leveraged Google DeepMind's GraphCast model as an initial foundation and fine-tuned the model using NOAA's own Global Data Assimilation System analyses. This additional training with NOAA data improved the Google model's performance, particularly when using GFS-based initial conditions.
This is incredibly cool stuff, it's great to see it being worked into actual production models - particularly combining it with deterministic physics-based predictions to balance out the strengths and weaknesses of both.
The code for GraphCast and GenCast is available, if anyone's interested in poking around more deeply: https://github.com/google-deepmind/graphcast
DeepMind themselves are actually one iteration beyond that by now, using a technique they call functional generative networks - somewhere around 8% better performance, according to their own metrics - but they haven't open sourced the new implementation and it looks like they probably don't plan to.
Google seem to be selling access to that one under the name WeatherNext 2, but for the amount of serious research DeepMind performs, publishes in detail, and open sources I'm going to take off my "fuck corporate incentives" hat for once and say that I don't actually begrudge them monetising that one when they've already published the initial version that made the real serious leap from previous methods. The FGN paper actually gives a pretty thorough explanation of how to implement it in the appendix, too, so I'd expect the NOAA are more than capable of coding their own version from there.
From the article:
...