ECMWF AI

[franksingleton]

ECMWF has recently announced that it is now running an AI forecast model operationally rather than as a trial. AIFS runs in parallel with the IFS (Integrated Forecast Sysren). Nesr future developments will include producing an ensemble.

This is not yet a full AI system in that the starting point is the same analysis, requiring the same computer resources, as used for the IFS.


Using raw observations rather than a pre-processed data set will be more difficult to implement in terms of the necessary learning. Such a system when implemented will, in principle, greatly reduce the computer resources needed. Work is in hand on AI-DOP, AI-Direct Observation Prediction.


However, the introduction of any new system in weather prediction has always been a case of "Softlee, softlee, catchee monkey." Both ECMWF and the Met Office are thinking in terms of hybrid operational systems in order to maximise the strengths of the two approaches.

 

[lustyd]

I look forward to seeing the results of this. Sadly no longer in the business but these sorts of developments are very interesting

 

[AntarcticPilot]

I think it's important that we recognise that the term AI covers a wide range of computational techniques which differ enormously in what they do. Neural nets are very different animals to Large Language Models, for example. And there are many different flavours of Neural Net! But you can think of Neural Nets being massively parallel multidimensional correlation machines, while LLMs look for statistical predictions of what the next term in a progression will be given the previous terms. The former is what is generally used in situations like weather forecasting.

 

[newtothis]

I think it's important that we recognise that the term AI covers a wide range of computational techniques which differ enormously in what they do. Neural nets are very different animals to Large Language Models, for example. And there are many different flavours of Neural Net! But you can think of Neural Nets being massively parallel multidimensional correlation machines, while LLMs look for statistical predictions of what the next term in a progression will be given the previous terms. The former is what is generally used in situations like weather forecasting.

If it's written in R, it's statistics
If it's written in Python, it's machine learning
If it's written in PowerPoint, it's AI

 

[AntarcticPilot]

If it's written in R, it's statistics
If it's written in Python, it's machine learning
If it's written in PowerPoint, it's AI

Isn't "written in PowerPoint" an oxymoron?

 

[lustyd]

Very few people are still using R, although Uni's are still pushing it for reasons only they know. It's a nightmare when people graduate and then have to learn Python to get a job.
No comment on which is better, just the reality of what's in use out there. For the record, Powerpoint is considerably more popular than either or the others, and often gets quicker results.

 

[franksingleton]

O

I look forward to seeing the results of this. Sadly no longer in the business but these sorts of developments are very interesting

So do I. No doubt incorrectly, I had gained the impression that AI and ML were, at least to some people, all about throwing masses of numbers into the pot and seeing what came out. As far as weather forecasting is concerned, it is good to see the involvement of experts in understanding weather with those who are AI experts. That will reduce the number of blind alleys explored.
I did not find it surprising that the ECMWF trial using the ERA5 data set was positive for deterministic forecasts. From my knowledge of the many data forms, my expectation is that the gains in AI forecasting using raw data will be in better ensembles and in reducing usage of computers. Deterministic forecast accuracy may not be greatly improved.

 

[AntarcticPilot]

Very few people are still using R, although Uni's are still pushing it for reasons only they know. It's a nightmare when people graduate and then have to learn Python to get a job.
No comment on which is better, just the reality of what's in use out there. For the record, Powerpoint is considerably more popular than either or the others, and often gets quicker results.

I think the reason is because R's statistical correctness is well validated. For complex statistical analyses, that's important.

 

[lustyd]

I think the reason is because R's statistical correctness is well validated. For complex statistical analyses, that's important.

There are certainly niches where it's great, and I've worked on such projects. My comment was just on the reality of 2025 AI/ML and right now almost everything and every organisation is Python. If I were learning today I'd learn Python and then move to R if I got a role that required it. Python isn't a good language by almost any measure, but it is easy, popular and flexible so for now it's won the race. Whether that lasts once budgets are reduced and efficiency becomes important again I don't know, but I'm certain it won't be R that replaces it when that happens.

 

[AntarcticPilot]

There are certainly niches where it's great, and I've worked on such projects. My comment was just on the reality of 2025 AI/ML and right now almost everything and every organisation is Python. If I were learning today I'd learn Python and then move to R if I got a role that required it. Python isn't a good language by almost any measure, but it is easy, popular and flexible so for now it's won the race. Whether that lasts once budgets are reduced and efficiency becomes important again I don't know, but I'm certain it won't be R that replaces it when that happens.

Sure - of course R is a niche language. I quit before Python started to dominate; no doubt I'd have learnt it if I hadn't retired! At the moment, I think I'd be looking at Rust. But I've used most things from the 1970s on, when I started with IBM Basic!

 

[lustyd]

To be fair, with Github copilot, learning specific languages is becoming largely redundant anyway. Rust seems a good move but there's so much code in other languages I'm not sure it'll really catch on as so much will need rewriting. I've said that many times over the years though and usually wrong

 

[AntarcticPilot]

To be fair, with Github copilot, learning specific languages is becoming largely redundant anyway. Rust seems a good move but there's so much code in other languages I'm not sure it'll really catch on as so much will need rewriting. I've said that many times over the years though and usually wrong

Big war going on in the Linux kernel between those who want to write in Rust and the traditional C/C++ writers!

But you can write Fortran in anything

 

[lustyd]

Yes that Linux thing is funny to watch play out. Linus seems to not care and just doesn’t want to generate more work.

I never tried Fortran but tempted to have a play as there are quite a few consultancy opportunities there which may work from a boat

 

[AntarcticPilot]

Yes that Linux thing is funny to watch play out. Linus seems to not care and just doesn’t want to generate more work.

I never tried Fortran but tempted to have a play as there are quite a few consultancy opportunities there which may work from a boat

FORTRAN 77 (or at least, the IBM Fortran H) was my main language for many years. Shared Common was fun - forget about memory protection; you could do amazing things with it! I once wrote code to convert from one floating point representation to another using it...

But for mathematical work, it's hard to beat Fortran for efficiency and rich libraries (NAG did the heavy lifting for that!)

 

[capnsensible]

F.U.N.E.M.N.X ?

 

[RobbieW]

For interest, I came across this recent thread on Cruisers Forum earlier...
AI Powered Weather Forecasting - good news - Cruisers & Sailing Forums

 

[franksingleton]

For interest, I came across this recent thread on Cruisers Forum earlier...
AI Powered Weather Forecasting - good news - Cruisers & Sailing Forums

As ever, I preach caution. Although just published, that paper was submitted last July. A recent, spring 2025, ECMWF Newsletter has an update on progress. This ends up by saying -
“Predictions using only observations is a highly significant milestone in the field of AI data-driven forecasting. AI–DOP represents a radical departure from using observations in data assimilation to create initial conditions for physics-based models or analysis-based data-driven systems. It remains to be seen, of course, to what extent the skill of these new observation-based forecasts, either in the pure form described here or possibly hybridised with other approaches, will challenge other more conventional methods. This activity remains an extremely exciting area of research for ECMWF.”

AI-DOP is a trial using raw observations rather than a NWP data analysis.

AI is clearly on the way and could be implemented now using NWP analyses with a significant saving of computer resources. To get to the stage of using observations directly is still on the horizon.

 

[lustyd]

I'd agree, even if the model was perfect on day one, the "plumbing" to get new data to it to get results will take a while to establish. It's a motivated area of research and development so I don't think it will be that long for things to change. Personally I think 2 years will see significant change in weather forecasting assuming the world holds itself together and funding/teams remain in place. Obviously the model isn't perfect on day one, and they're not even tackling the main problem in this instance, but those things will come quickly and iterate often as we've seen with other AI tech. In the case of ChatGPT those iterations were just training on more data, for the most part (they're now working on efficiency, having consumed all data sets). For weather, the bigger problem is collating existing and new data sufficiently to do that training

 

[franksingleton]

I am sure that, in the future, forecasts will be produced with far less computer resources and far quicker. Those are givens. Ensembles will be better. What is far less certain is whether there will be a significant improvement in the quality of the forecast. Whether it is a physical model or an AI construct the end product will depend on the data input and whether there are sufficient data to cover all or enough possibilities. NWP and AI models will be limited by the infinite variability in the atmosphere itself. To me, it is inconceivable that an AI model will cope with the random element in the atmosphere.

 

[lustyd]

What you call “infinite variability” and “random” are simply elements too large for the human brain to comprehend, or the number of variables becomes too big. This is where AI and ML shine and they can deal with as many parameters as required to do the job.
They’re also entirely unconcerned with the “why” so will produce results decades before we understand the systems they model. I’m sure before science took over humans knew where to find a rainbow a long time before we knew why it was there.