Paul Dorian
Overview
Artificial intelligence (AI) is a collection of technologies that allow computers to perform tasks that typically require human intelligence, and it is increasingly impacting the world of weather forecasting. The European Center for Medium-Range Forecasting (ECMWF) has made strides with its Artificial Intelligence Forecasting System (AIFS) as it has recently become fully operational and is now run side-by-side with its traditional physics-based Integrated Forecasting System (IFS). According to the ECMWF, the AIFS has outperformed the physics-based model for many measures including, for example, tropical cyclone tracks. In addition to the ECMWF AIFS, there are at least four other known “A.I. trained” weather models including NOAA/Google GraphCast, Microsoft’s Aurora, NVIDIA’s FourCast, and Huawei’s Pangu-Weather.
A side-by-side comparison of the conventional European forecast model (left) and the Euro-AI version (right) of “Total Snowfall Amounts” in the continental US for the period ending Friday AM, April 4th. Maps courtesy ECMWF, Pivotal Weather
Discussion
The traditional approach to weather forecasting has been to make use of numerical weather prediction (NWP) which relies on current conditions, physics-based models, and the solving of complex equations on powerful supercomputers to output such parameters as temperature, pressure, winds, and precipitation at future times. Artificial intelligence (AI) models, particularly machine learning, are being increasingly used to improve weather forecasting by learning from large datasets of weather data to identify patterns and trends. AI models can process data faster and identify complex patterns, potentially leading to quicker and more accurate forecasts. The increasingly important role of AI in weather forecasting will be to complement and enhance traditional NWP models.
The European Center for Medium-Range Forecasting (ECMWF) has made its Artificial Intelligence Forecasting System (AIFS) the first such fully operational weather prediction model that uses machine learning and artificial intelligence. Making such a system operational means that it is openly available and has 24/7 support for the meteorological community. This AIFS can produce a wide range of output parameters including winds, temperatures, and details on precipitation types from snow to rain. The AIFS currently has a grid spacing of 28 km and, according to the ECMWF, it can outperform its physics-based counterpart by as much as 20% on certain measures.
A side-by-side comparison of the conventional European forecast model (left) and the Euro-AI version (right) of “500 millibar height anomalies” across the continental US for the validation time of 8AM, Sunday, March 30th. Maps courtesy ECMWF, Pivotal Weather
The AIFS uses the same initial atmospheric conditions for its forecasts as the IFS. These are based on the combination of a previous short-term forecast with around 60 million quality-controlled observations from satellites as well as many other streams, including from planes, boats, sea buoys and many other Earth-based measurement stations. Every six hours, these initial conditions feed into the AIFS. The machine learning model, trained on how the weather has evolved in the past, assesses how the initial conditions will influence the weather for the coming days. By contrast, the IFS uses physics-based capabilities to arrive at a forecast with a grid-spacing of 9 km over the globe, integrating the laws of physics in its computer code.
Meteorologists rarely put total confidence into a single model but instead, look at a suite of model forecasts. Most global models compute 10-day outlooks every six hours and one way forecasters gauge a model’s reliability is by checking its consistency from run to run. The Euro-AI model did well with this test for Hurricane Francine during September 2024 as it predicted landfall in southern Louisiana every run from 96-hours out until its landfall near Morgan City, LA. Maps courtesy foxweather.com
The first operational version is called the AIFS “Single”. It runs a single forecast at a time, known as a deterministic forecast. However, ECMWF is pushing this model to create a collection of 50 different forecasts with slight variations at any given time to provide the full range of possible scenarios. This is known as ensemble modelling, a technique developed and implemented by ECMWF more than thirty years ago. According to ECMWF, the launch of AIFS “Single” as an operational service is the first step in upgrading its artificial intelligence forecasting capabilities. The next step will be to make ensemble forecasts available by using artificial intelligence and extended range (seasonal) forecasts as well and we’ll continue to monitor the progress here at Arcfield Weather.
Meteorologist Paul Dorian
Arcfield
arcfieldweather.com
Follow us on Facebook, Twitter, YouTube
Wow! 20% BETTER than physics-based models! What does THAT say about what should be done with climate models? Of course, the GIGO principle applies in that if “fudged” temperature data is used, who know what one will get. However, here’s a thought: run AI-generated climate models using RAW temp data versus ADJUSTED temp data. Train the model on 50 years of data (say, 1975-2025) and then test the model using a separate 50 year period (say, 1925-1975).
What fun! that would be for any serious climate scientist.
Weather is a great application for machine learning (I don’t really call this AI). There’s lots of data for training and validation and it doesn’t take too long to assess model performance in the real world.
It’s the exact opposite for climate. There’s very little training and validation data, and most of that comes from proxies, and you won’t know if the model really works for 20 years. It’ll never work.
For those that don’t know how neural networks work, think of trying to develop an algorithm that can identify a dog, or a cat in any random picture with random breeds, perspectives, lighting and shadowing. It’s almost impossible — too many variables. But if you train a neural net on thousands of pictures, then there’s a very good chance that a dog or cat in a new picture will be correctly identified. The problem is, you don’t know how it works. How do you know a dog isn’t a cat when you see one?
While not knowing how a model works might seem like a problem, it’s not because the model can be used to determine which inputs are most important and to run experiments for developing parametric models.
AI is pattern matching … words, sentences or pictures it doesn’t matter … in the end the programmer is the “intelligence” behind the AI … and the data used to train it is always biased … i.e. someone has to choose what to train it against … in the case of climate if the AI is only trained with AGW theory data then magically CO2 is always the demon … AI is the biggest waste of money the business world has ever seen … billions of dollars are being spent to try an use AI to “improve” businesses and in the end it will have a marginal impact on 99% of companies …
Depends what you choose to apply AI to.
In many ways, AI is another of those technologies looking for a useful application.
For example, I’m using one that produces synopses of non-fiction books, articles, studies etc.
A modern day Readers’ Digest, if you will.
Now that’s a function that I could do myself, given a lot of time and concentration.
But an AI makes short work of it, only having to decide what to include for a
“Just The Facts, M’am” synopsis.
AI is a good solution for people who need an excuse to make politically unpopular choices.
Formerly know as “consultants”.
Actually it is the human language interface that makes the software seem intelligent.
“Dumb” used to only mean unable to speak. Back then, if someone could not speak they were appraised as mentally incapable. Hence the modern use of the word.
Just look at politicians. Their ability to give speeches is not a clue as to how smart or dumb they are.
Actually it is, sometimes. Just look at how revealing it was to listen to Commie-la speak. 😆😅🤣😂
Automated Idiocy is the best way to describe it. And the “black box” aspect where nobody knows exactly *how* it works, and the propensity for it to “make things up,” are unnerving.
The issue that you don’t know how it works is also the reason that you don’t know if a particular answer is right or wrong. AI is great for applications like language and vision where there is a range of “correct” answers, not so good when there is a single optimal answer. To paraphrase an old saying, “almost is good enough in horseshoes and hand grenades and AI”.
This is weather forecasting, not climate forecasting. The AI model can be tested, forecast vs actual. Machine learning can be applied and improvements made routinely.
No. Future states of chaotic processes cannot be determined from present conditions.
This is pure fantasy, but apparently believed by “climate scientists”.
Sad but true.
Again, weather, not climate. Unless I misread the article, the topic is weather. Weather forecasts are for a set number of days. The forecast can be evaluated and tested against the real world. Forecasts are updated every 6 hours.
No doubt somebody in marketing started the misuse of the term “AI.”
Artificial intelligence (AI) is…. super fast pattern matching. It is not ‘intelligent.’ It is trained via databases of reference materials.
It’s more than just pattern matching. Machine Learning 10 years ago was pattern matching plus some rules. The latest AI versions have limited inference and deduction based on prior training, some of which is reference materials, some of which is empirical data sets. General AI (human level intelligence or greater) is probably only 10-15 years away.
Just like fusion, eh?
:-)) Point made.
Advancing Software + massively parallel GPU hardware is MUCH simpler than fusion. tho’ the emerging quantum processors might help.
Fusion needs a huge engineering + scientific effort and several key breakthroughs that are proving to be elusive. Perhaps in 100 years, with the help of general AI.
Hardware isn’t the problem. Making software “intelligent” is.
To me, “AI” is just a marketing buzzword. The same as calling something “free”, but you have to pay money (subscription, buy another item at the same time, etc.) to get it. In other words, lies.
AI first appeared with the initial studies of neural nets that were trying to emulate human neural connections.
It was grabbed by computer gaming developers who socialized it.
According to AI, it is not intelligent as it is not conscious. It is a weighted decision tree, somewhat adaptive to weighting values, hidden behind an excellent language interface.
Story tip: We all know what consultation means…
British Steel will consult on the closure of its two blast furnaces, steelmaking operations and a reduction of steel rolling mill capacity in Scunthorpe.
https://britishsteel.co.uk/news/british-steel-to-consult-on-proposed-closure-of-scunthorpe-blast-furnaces-rod-mill-and-steelmaking-operation/
Britain, formerly known as Great, is fast dissolving because of their extraordinarily expensive electricity, their refusal to allow fracking and their ban on coal. What a mess!
The island itself is called Great Britain, Wales, Scotland and England – collectively the [dis]United Kingdom – are the nations that are located on it.
Careful if you aren’t great President Trump may claim the island 🙂
He’s welcome to it… Better that than China.
From the article: “Artificial intelligence (AI) models, particularly machine learning, are being increasingly used to improve weather forecasting by learning from large datasets of weather data to identify patterns and trends. . .
The machine learning model, trained on how the weather has evolved in the past, assesses how the initial conditions will influence the weather for the coming days.”
Yes, weather history is very important. The Earth’s weather goes through repeating patterns and trends, and AI is apparently making good forecasts comparing the past with the present.
Now, if we could just get AI to study the surface temperature record and compare actual written temperature history with the abomination of the Hockey Stick bastardization.
AI would find that it is no warmer today than it was in recorded history, going by the written, historic, regional temperature records from around the world.. This would destroy the CO2-caused Climate Change narrative, since there is more CO2 in the air now than there was in the past, but it is no warmer now than then, so CO2 has had no visible effect on the Earth’s temperatures.
Give AI all the written temperature data Phil Jones had available to him, and see what AI comes up with. AI won’t create a “hotter and hotter and hotter” Phil Jones temperature profile because the written, historic temperature records don’t contain a “hotter and hotter and hotter” temperature profile. Instead, the written record shows a benign temperature profile where it was just as warm in the recent past as it is today, and CO2 is nothing special since CO2 obviously cannot raise the temperatures of today above those of the recent past even though there is much more CO2 in the air today than there was in the past.
Yes, turn AI lose on the written, regional, historic, original temperature records from around the world. I bet AI’s results won’t look anything like the bogus, bastardized, instrument-era portion of the Hockey Stick global chart, because it would be impossible to honestly get a Hockey Stick trend line out of data that does not have a Hockey Stick trend line, and the original, regional data does not have this trend line. It’s not there. The scary Hockey Stick trend line Climate Alarmists wring their hands over is NOT there in the original, regional data.
Dishonest climate scientists just made the Hockey Stick Chart up out of whole cloth. It’s the BIG LIE of Alarmist Climate “science”, Every Climate Change Craziness flows from this one BIG LIE.
The problem may be, Tom, that someone has done just what you suggest but they cannot find anybody willing to publish the results. Getting opposing views aired on climate issues is very difficult these days.
Or they searched for ways to get the “right” answer by supplying the “right” data.
Well actually The Big Lie goes back to far before the priduction of the Hockey Stick temperature reconstruction.
The Big Lie is the ridiculous notion that a warmer climate is worse, which is the exact opposite of reality.
If people saw through The Big Lie, the “Hockey Stick” would be seen as GOOD NEWS. It is only through belief in The Big Lie that the “Hockey Stick” can be seen as a “problem.”
So they can get it twice as wrong twice as fast.
GIGO … always has been always will be AI or human …
Instead of side-by-side maps comparing AI to Physics based models, let’s see side-by-side maps showing “forecast” to actual results. Does the 96 hour forecast for precipitation or temperature or pressure match up with the actual 96 hours later? Are they routinely posting such comparisons or are they avoiding doing so? Seems like the obvious way to judge accuracy.
And by that metrix it fails miserably.
I once worked with someone who worked on transactions based on weather, and I asked him point blank how far out you could actually rely on weather forecasts.
His answer, without hesitation?
2 days.
S/b “metric.”
This could hardly be called AI. Aside from the fact that it takes in data from various sensors for pretraining and some initial conditions, there is no AI
No hardware, no sensors, continually interacting with the algorithmic part
It is machine learning, pure and simple
Please stop calling it AI. They are only doing this to make it sound fancier and get more funding
I only watched the webinar, which doesn’t give much detail
From what I can see it is a traditional encoder decoder sequence to sequence model that is being used as a time series method with some embedding of the data. Meaning it is using windows of data as inputs and predicts a vector of outputs one or more steps ahead based on the last prediction, which we call autoregression. It could be some variant of this, like a hybrid approach. It turns a time series into a supervised learning problem
Normally these models are confined to natural language processing because the data sets are massive
More principled time series methods such as the linear state space ARIMA and nonlinear state space methods like hidden Markov models, extended kalman filter, particle filters and Gaussian process dynamical models are preferred in the finance and most other fields, because they are “explainable” and are more accurate than recurrent networks like the encoder decoder architecture
From my own experience, recurrent networks are usually pretty low accuracy or at least high variance (wildly different answers, occasionally accurate but usually miles off) on small to medium size data sets, whether it’s time series or regression. Of course, you’ll get a lot of idiots who don’t properly understand any of these methods writing silly papers using networks because it’s easier to publish in applied journals and it’s easy to code with tensorflow, pytorch etc
Unfortunately, the other approaches don’t scale well with training number.
This is the great advantage of networks, scalability via stochastic gradient descent
How accurate they can be for long term forecasts, like days to weeks is questionable
This is after all a prediction of the future, and the randomness is the problem, as in all time series analyses
I would add to that trying to leverage credibility. As I said earlier, I think machine learning is well suited to weather forecasting. But if you see AI and climate in the same context, you should question the credibility of any associated information.
I don’t see any reason that non-parametric ML models can’t be at least as accurate as parametric models. I can think of several reasons why they might be better. Ultimately they’re both limited by the selection, quality and quantity of data, and by stochastic processes.
And the nature of weather itself.
We have to be careful here with terms
Non parametric has at least 3 definitions, depending on whether you talk to statisticians, statistical learning theorists, geveral computer scientists or practitioners. Even amongst the same community they will use different definitions. For each one of these definitions you can find a counterexample. Same goes for linear vs nonlinear regression.I personally have given up using these terms
Formally non parametric means the effective number of degrees of freedom does not grow with sample number, but that is not a definition that’s always used by any means. For a statistician, it means you model is distribution free, ie no assumptions regarding the distribution over the data
I think here you are drawing the distinction between the ML model they use (which is non parametric in some definitions) and the original model
I wouldn’t personally call it parametric, I would call it physics based
There are other types of models we sometimes call parametric in engineering, or semi parametric or mechanistic. These are usually highly simplified physics based approaches with some fitting parameters. The archetypal example is equivalent circuit models, which are used for all sorts of processes
The question of whether to employ physics based vs a data driven (machine learning) approach is highly dependent on the problem, and the available computational budget
In very complex modelling tasks such as climate change, one would often employ a surrogate or meta model to replace the original physics based approach (black box model), especially in tasks that require repeated evaluations of the model at different input parameter values, like in optimisation or uncertainty analysis based on Monte Carlo estimates of the statistics
These surrogate are trained on data from the original model, and are usually ML based, or can be reduced-order intrusive models that project the original numerical formulation or governing equations onto a low dimensional aporoximating subspace, like a Galerkin model, or multi fidelity approaches that leverage data from models at different levels of accuracy (in the physics, numerical method used or solver settings such a maximal element size and time step or tolerance/relaxation parameter).
In the present case, I don’t see any inherent problem in trying to do physics based simulations for local weather forecasts on short time scales. I think this could be feasible with some approximations
This new approach is not a surrogate model, which initially surprised me. Instead it uses observations directly as the training data to learn spatio temporal patterns and make predictions
It may well be the case that this machine learning approach is more accurate than the physics based approach
That’s probably because even with local forecasting on short timescales, the physics is not understood well enough to model without severe assumptions, and/or the coarse graining adds too much error
My issue is with long term forecasts, using either approach.
This is not a vanilla supervised machine learning problem. There is a fundamental difference: observations are not available in the whole window of input space for which you want to make predictions, because the data is temporal
Think of it as interpolation (regression) vs extrapolation (time series or other sequence data)
This is what makes these problems so difficult to solve. What they do here is turn the problem into a vanilla supervised machine learning using an embedding of the data. This can work in some problems, and in others it may not
Other time series methods are not feasible on these massive data sets unfortunately. So what they are doing is probably the only feasible way to do it, meaning some recurrent network architecture
The further you go in time, the less accurate will be the results. There are repeating cyclical patterns, but they happen over many different spatial and temporal scales. Then there are large scale random fluctuations, which by definition cannot be predicted.
Even by being agnostic to the physics and using observation data, the problem is insoluble
“the first such fully operational weather prediction“… battle station
No one who grew up when the original StarWars films came out can hear the words “fully operational” without hearing the emperor.
Only now, at the end, do you understand…
A couple of times over the years, in the context of weather forecast models, I’ve asked if anyone was trying to match past weather patterns and what happened with current weather patterns to forecast what was likely to happen.
Perhaps this “AI” could do it quickly?
Though it would still need a human meteorologist to check it before sending the forecast out to the public.
It could be a useful tool.
There are already too many useful tools in weather and climate forecasting.
Climate forecasting? Hello?
I guess it’s been programmed to predict the end of the world.
How long until the first Cat 5 hurricane the AI “makes up” causes an unnecessary evacuation and discredits their latest and greatest “model?”
Of course I don’t understand how either method works but they seem to be making a fuss about one system relying on physics while the other doesn’t. The AI version appears the rely on a constant stream of observations from satellites, boats, sea buoys and other stations. So why wouldn’t the traditional method be doing the same thing? If the traditional method wasn’t using this information why were they even in place? This doesn’t make sense.
Paul, if you are selling forecasts, your customers are fools, who don’t realise that the future state of the atmosphere cannot be predicted. Nobody can even predict the wind speed and direction 30 seconds hence, with any degree of accuracy.
If you believe that dissecting the past can predict the future of a chaotic system, you have fooled yourself, in the words of Richard Feynman.
The approximate future of a chaotic system cannot be determined from the approximate present.
Sorry about that, but it’s true – unless you can find some reproducible experimental data to support your disagreement. Computer models or appeal to your own authority are not experiments.
Latest fun with AI – maybe 60 seconds or time spent –
AI – “Therefore, approximately 0.226 g of dry ice would be needed to heat 100g of gold at 20°C by 10°C, assuming perfect energy transfer and insulation.”
Previously, it congratulated me for saying that dry ice at -78.5 C could be used to heat CO2 at 20 C!
About as “intelligent” as the average “climate scientist” who believes that adding CO2 to air makes it hotter.
Yawn. When will these people learn?
“Scientists involved with chaos theory attempt to examine, describe, and quantify complex and unpredictable dynamics of systems that are sensitive to their initial conditions but follow mathematic laws—even though their outward appearance appears random. Meteorology, and the prediction of weather and climate, is a classic example of such an unpredictable (chaotic) system.”
https://www.encyclopedia.com/environment/energy-government-and-defense-magazines/chaos-theory-and-meteorological-predictions
Commenters are in a snarky mood today!
“It’s not AI, it’s ML” So what, if it works?
“You can’t predict future climates” Who said that weather forecasting will predict future climates?
“Weather is chaotic, so cannot be predicted” At the scale of a continent, future trends are already being predicted a few (5-7) days into the future with a pretty decent record.
If you’ve been around as long as me, you cannot fail to be aware that weather forecasts are way better than they were 50 years ago. If AI/ML can tweak them a bit, I’m all for it.
Curiously, I said in a conversation yesterday “I wonder when AI will be applied to weather forecasting”. Am I psychic or what?
At the scale of a continent, future trends are already being predicted a few (5-7) days into the future with a pretty decent record.
No, assumptions rather than predictions. A “pretty decent record” is just meaningless word salad.
A smart 12 year old is able to achieve a “pretty decent record”. Or, maybe you would like to wager that a “professional forecaster” can “forecast” better than I (if you think a 12 year old is too smart to compete with).
If “snark” is jargon for “acceptance of reality”, then call me snarky. I don’t mind.
Wait for it…next they’ll “train” the models with reams of AGW propaganda and then they’ll tell us how awful everything will be if we don’t do as we’re told. 🤬