AI experimental weather products

In case anyone wants to have a look at weather forecast charts from some experimental products like Ecmwf AI Forecasting System, Machine Learning from GoogleDeepMind, FuXi, FourCastNet, etc :

ECMWF | Charts

Hopefully the link works, otherwise left column select Experimental AIFS and Experimental Machine Learning Models

Roberto said:

In case anyone wants to have a look at weather forecast charts from some experimental products like Ecmwf AI Forecasting System, Machine Learning from GoogleDeepMind, FuXi, FourCastNet, etc :

ECMWF | Charts

Hopefully the link works, otherwise left column select Experimental AIFS and Experimental Machine Learning Models

Click to expand...

ECMWF and the Met Office are both looking at AI seriously. The Met Office in collaboration with the Turing institute. Initial ECMWF results are promising but it is still early days. As I understand it, the starting point is the IFS analysis and the learning is based on ECMWF ERA5 data covering the period 1979-2018 (I might be adrift on the dares). In a limited trial AI is a little better than the IFS. However, arguably the biggest problem in NWP is the initial data analysis. This uses fairly small amounts of in situ ;data of high quality plus vast amounts of various types of satellite data. Some satellite data provide wind at high level and at the surface. Much satellite data from infrared soundings, microwave soundings and GPS Radio Occultation do not measure temperatures or humidities. They measure the effects of the atmosphere on the various radiations. The equations cannot be inverted and that is a problem dealt with by treating the radiances as model variables.

The next stage will be to see how AI works with the observations rather than with analyses. One of the problems with the current data analysis schemes is how you weight the different volumes of data - small amounts of highly accurate in situ data versus vast amounts of low spatial but high temporal resolution satellite data. It could well be that AI makes a significant improvement there. (Ed. As put, that is not quite correct. I was really trying to highlight the differences in in situ instrumental data and data with quite different characteristics in terms of precision.)

As a scientist, I dislike the black box approach but have to admit that if it really does work, then so be it. ECMWF and the Met Office will, for the time being, follow both lines.