Met Office model upgrade

[franksingleton]

Reading the Met O announcement at the OP and associated references, it is clear that they are proceeding cautiously and carefully as is only sensible with an operational system that is directed toward safety and public service. In the near future, 2, 3,4 years, some form of hybrid system is likely. Forecasts on the large scale for the next few days or weeks, the UK UM could become ERA5/6 based, possibly using raw data. I have no idea how long it will take to develop a reliable total AI/ML system. I would anticipate significantly longer than for am LLM but really have no idea.

Detailed modelling, generating forecasts on a UKV, or smaller scale, seems much more likely to be a physical model with AI input at the initial data input and output stages. One of the problems with NWP is putting in sensible values for heat transfer processes, ie conduction, convection, radiation, latent heat effects. These are not measurable at present operationally but are derived from a system known as parameterisation, intelligent guesses based on internationally run observing studies. AI/ML could be used to improve those guesses.

Models do provide forecasts at the nominal 10 m height. The lowest level, for computational reasons, is a height where the pressure is 0.998 of the surface pressure. Heights above that are multiples of the surface pressure such as 0,7, 0.5, 0.3 etc. these equate roughly to 700. 500. 300 hPa. Models heights are terrain following. Algorithms are used to derive values at specific heights and locations AI could be derived to generate better, more realistic values.

We live in interesting times meteorology will not stand still and the UK will, no doubt, be a leader as it has been since the 1960s.

 

[lustyd]

how long it will take to develop a reliable total AI/ML system. I would anticipate significantly longer than for am LLM but really have no idea.

My estimate is two years to train the first version (pure number crunching time on a very large cluster, no human involvement). Following that maybe a year to make it operational in terms of experts using it, then probably two years to real world usage by average customers. Early access for power users and developers is very likely.

That's assuming very little change in terms of tooling though, as I said above there's no sensible reason the output would be grid based as it is now since the processing won't be grid based and it probably won't output every grid point. It's possible that they'll make the output compatible but I'd expect changes. I don't know for sure what's happening here though or the planned outputs.

 

[franksingleton]

I was only trying to read between the lines of their statement to try and understand their thinking. Progress will be incremental because of all the many organisations and individuals that are weather dependant. We have come a long way beyond the scenarios painted in this evening’s BBC4 offering. They still had to show dear old Mike Fish!