Weather advancements

[lustyd]

Thought this might interest a few after the recent conversation about AI and weather models. This one is Google working with ECMWF to make some pretty good progress. While it may seem like a small step, note that this also helps move to faster compute so we’ll likely see rapid iterations over the coming year along with more experimentation. They’ll also pull in more seemingly unconnected data and move away from big cubes.
It’s interesting to think that this isn’t more accurate, it’s just right more often.
The secret to better weather forecasts may be a dash of AI

 

[franksingleton]

I have not been watching YBW recently and missed this.
ECMWF and the Met Office are both studying AI forecasting. ECMWF used 40 years of historical analyses to create the starting point for the AI learning process. Their trial AI forecasts have started from the IFS analyses. Results so far suggest that AI could outperform the current physical models. The unknown is whether AI based on raw observational data will also perform better than physical modelling. In other words, can the IFS analysis stage be omitted.
To get to the starting point for a physical model takes about 3 hours from the nominal observation times of 00 and 12 UTC.(my guess). The forecast takes about another 3 hours. An AI forecast would be far faster. So, even now, using AI would produce earlier forecasts and save some 6 hours a day computing effort.

 

[lustyd]

Saw your article in the CA mag, good stuff

 

[franksingleton]

Saw your article in the CA mag, good stuff

I have put a copy at AI Forecasts - Franks-Weather - The Weather Window

 

[oldbloke]

The "AI" is only doing what weather forecasters have always done. Use the available data and previous experience, Red sky at night sailors delight. If you get wet when you go outside, it's raining. Nothing new, may be quicker, may be more accurate more often. The big difference from The Met Office etc is it's not trying to understand the mechanisms. Fortunately it won't be an either/or

 

[Roberto]

AI experimental weather products

 

[franksingleton]

The "AI" is only doing what weather forecasters have always done. Use the available data and previous experience, Red sky at night sailors delight. If you get wet when you go outside, it's raining. Nothing new, may be quicker, may be more accurate more often. The big difference from The Met Office etc is it's not trying to understand the mechanisms. Fortunately it won't be an either/or

At #6 Roberto reminded us of AI experimental weather products. In that, I said -

“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.”

AI may provide statistically better forecasts than NWP models but there will always be some uncertainty. The atmosphere will still be governed by the laws of physics. There will still be turbulence and consequent predictability limits.

Models may cease to be operational tools.Their future role might be in understanding the atmosphere.