3 ms·
Google's is initialized with a gridded dataset, ERA5, from ECMWF. Using ERA5 is the current standard here, and ECMWF themselves build on that mostly now. Meanwh
by scellus 2y ago
Google's is initialized with a gridded dataset, ERA5, from ECMWF. Using ERA5 is the current standard here, and ECMWF themselves build on that mostly now. Meanwhile, Aardvark tries to do the same directly from observations.
- Onavo 2y agoWhat do you mean directly from observations?
- counters 2y agoGDM's GraphCast/GenCast require an "analysis state" of the atmosphere. This is a 3D, gridded dataset with key variables like temperature, humidity, and winds. Generally speaking, an analysis is produced by a physics-based weather model through "data assimilation", an optimization process which tries to create a 3D state that is consistent with observations over some window of time. "Observations" is overloaded; it's really any "raw" weather data product, like satellite imagery, surface station measurements, or weather balloon traces. I'm handwaving away a lot of complexity here. The AardvarkWeather model is a significant new development and paradigm shift - it's one of the first models of a new class which do not require an analysis, and can directly use the "raw" weather data observations that are typically used to perform data assimilation.
- Onavo 2y agoThanks