A new AI weather prediction system, Aardvark Weather, can deliver accurate forecasts tens of times faster and using thousands of times less computing power than current AI and physics-based forecasting systems, according to research published in Nature.
Aardvark has been developed by researchers from the University of Cambridge supported by the Alan Turing Institute, Microsoft Research and the European Centre for Medium Range Weather Forecasting, providing a blueprint for a completely new approach to weather forecasting with the potential to transform current practices.
The weather forecasts that people rely upon are currently generated through a complex set of stages, each taking several hours to run on bespoke supercomputers. Aside from daily usage, the development, maintenance and deployment of these complex systems requires significant time and large teams of experts.
More recently, research by Huawei, Google, and Microsoft has demonstrated that one component of this pipeline, the numerical solver (which calculates how weather evolves over time), can be replaced with AI resulting in faster and more accurate predictions. This combination of AI and traditional approaches is now being deployed by the European Centre for Medium Range Weather Forecasts.
But with Aardvark, researchers have replaced the entire weather prediction pipeline with a single, simple machine learning model. The new model takes in observations from satellites, weather stations and other sensors and outputs both global and local forecasts. This fully AI driven approach means that predictions are now achievable in minutes on a desktop computer.
When using just 10% of the input data of existing systems, Aardvark already outperforms the United States national GFS forecasting system on many variables and it is also competitive with United States Weather Service forecasts that use input from dozens of weather models and analysis by expert human forecasters.
One of the most exciting aspects of Aardvark is its flexibility and simple design. Because it learns directly from data it can be quickly adapted to produce bespoke forecasts for specific industries or locations, be that predicting temperatures for African agriculture or wind speeds for a renewable energy company in Europe.
This contrasts to traditional weather prediction systems where creating a customised system takes years of work by large teams of researchers.
This capability has the potential to transform weather prediction in developing countries where access to the expertise and computational resources required to develop conventional systems is not typically available.
Professor Richard Turner, Lead Researcher for Weather Prediction at the Alan Turing Institute and Professor of Machine Learning in the Department of Engineering at the University of Cambridge, said: “Aardvark reimagines current weather prediction methods offering the potential to make weather forecasts faster, cheaper, more flexible and more accurate than ever before, helping to transform weather prediction in both developed and developing countries.”
“Importantly, Aardvark would not have been possible without decades of physical-model development by the community, and we are particularly indebted to ECMWF for their ERA5 dataset which is essential for training Aardvark.”
Anna Allen, lead author from the University of Cambridge, said “These results are just the beginning of what Aardvark can achieve. This end-to-end learning approach can be easily applied to other weather forecasting problems, for example hurricanes, wildfires, and tornadoes. Beyond weather, its applications extend to broader Earth system forecasting, including air quality, ocean dynamics, and sea ice prediction.”
Matthew Chantry, Strategic Lead for Machine Learning at ECMWF said: “We have been thrilled to collaborate on this project which explores the next generation of weather forecasting systems — part of our mission to develop and deliver operational AI-weather forecasting while openly sharing data to benefit science and the wider community. It is essential that academia and industry work together to address technological challenges and leverage new opportunities that AI offers. Aardvark’s approach combines both modularity with end-to-end forecasting optimisation, ensuring effective use of the available datasets.”
Dr. Chris Bishop, Technical Fellow and Director, Microsoft Research AI for Science, said: “Aardvark represents not only an important achievement in AI weather prediction but it also reflects the power of collaboration and bringing the research community together to improve and apply AI technology in meaningful ways.”
Dr Scott Hosking, Director of Science and Innovation for Environment and Sustainability at The Alan Turing Institute, said: “Unleashing AI’s potential will transform decision-making for everyone from policymakers and emergency planners to industries that rely on accurate weather forecasts. Aardvark’s breakthrough is not just about speed, it’s about access. By shifting weather prediction from supercomputers to desktop computers, we can democratise forecasting, making these powerful technologies available to developing nations and data-sparse regions around the world.”
Next steps for Aardvark include developing a new team within the Alan Turing Institute led by Professor Richard Turner, exploring the potential to deploy Aardvark in the global south and integrating the technology into the Institute’s wider work to develop high-precision environmental forecasting for weather, oceans and sea ice.
Journal Nature Reference: Allen, A., et al. 2025. ‘End-to-end data-driven weather prediction’, Nature, DOI: 10.1038/s41586-025-08897-0
DOI 10.1038/s41586-025-08897-0
Discover more from Watts Up With That?
Subscribe to get the latest posts sent to your email.
A faster [failed] BBQ summer prediction?
Thanks, but no thanks.
I’d be really interested to see how Aardvark does forecasting our weather in Northern Virginia. We seem to be on the cusp of two atmospheric cells, here, and the weather forecasts are completely unrelated to reality. In fact, they sometimes can’t even tell what the weather was an hour ago.
I’m also a little skeptical of AI in general, having had Chat GPT BS me when it didn’t know an answer, and even lie. But that’s a general purpose AI, and not the best (Grok is pretty good).
So …. With Aardvark, we can replace most of the National Weather Service staff with a few hundred or so desktop computers. Color me skeptical.
It could, but could it be better than what we get on channel 5 now?
I don’t know which channel 5 you mean. I use Flowx on my Android phone, and it lets me switch among 8 different models. The differences are fascinating. I live in the Sierra foothills, and they seldom are useful in detail even a day ahead. Sometimes they predict snow, and the sky is clear. Sometimes they predict clear, and I get snow. And that’s just one day ahead. As a general trend of temperatures rising and falling, they are usually good for 5 days or a week ahead. They even are decent with potential storms that far ahead. But reliable? No.
Aardvark sure can’t be any worse.
I see what you mean… 🙂
It might be correct that this AI approach will consistently yield better weather forecasts than before. If so, congratulations on a series of fine technical achievements and advances.
However, the big problem involves what will be done with these forecasts if they are used by the wrong hands, such as activists for net zero world.
For example, we have seen claimed advances in measurement of sea level change such as by satellite altimetry versus tide gauges, but in the wrong hands such information is being used to formulate policy on land use, with some draconian outcomes requiring compulsion to avoid building in some localities, loss of resale value of properties deemed threatened by sea level rise and so on. These are leading to increased legal and Court challenges at cost to the community.
It is easy to picture activism fooling around with forecasts to stop this or push that.
There are also hazards ahead for normal people. e.g. What happens if an AI forecast is badly wrong, such as with an outcome that farmers planted seeds at a recommended good weather time, only to have them harmed by unforecast high rain? Do they go to Court for compensation? One can imagine many consequences of wrong forecasts, also a real cost of eventual forecast hesitancy when too many forecasts turn out wrong.
But I hope that it all turns out well.
Geoff S
It sounds interesting and promising … but Nature? They have covered themselves in septic effluent for so long now on so many topics that I’m reluctant to believe anything they publish.
I am looking forward to Elephant – the climate prediction AI version of Aardvark.
One input that Elephant won’t need is atmospheric CO2. I could not imagine anything labelled :”intelligent” concluding that atmospheric CO2 was a vital input. Greening and leaf coverage certainly but absolutely nothing to do with the climate otherwise.
I now have my own sunspot predictor. I am certain that AI let loose on this could derive something even better than my 87% fit.
This is the kind of thing ‘Artificial Intelligence’ (really just neural networks and pattern recognition*, no actual ‘intelligence’!) is useful for. A limited set of data inputs, matched with a limited set of results. It can learn what typically will occur given similar inputs, ie meteorological conditions in this case. I have to say that such technology hasn’t significantly improved since I was programming such things 30 years ago. The processing is obviously a great deal faster, and data throughput potential is massive in comparison to then.
Helping you with day-to-day life, though? Not so much!
(*pattern recognition is what software is very bad at compared to human and other brains. We are extremely good at it, to the point where ‘intuitive’ decisions are often entirely unconscious. Neural networks are a crude attempt at replicating such skills. They are still in an incredibly ‘brute force’ stage of evolution. Millions of scenarios need to be processed, whereas a human brain can manage it in dozens, or even less. Our own data inputs and processing powers are immense, and poorly understood.)
“researchers have replaced the entire weather prediction pipeline with a single, simple machine learning model… takes in observations from satellites, weather stations and other sensors … fully AI driven”
I wonder where the satellites and sensors come from, and who decided where they’d be?
From the seventh paragraph of the above article:
Yeah, I’m excited . . . I’ll soon become a bazillionaire by being the first person to apply Aardvark to the US stock market.
After all, I’ve been told repeatedly that week-to-week predictions of stock market performance are stochastically equivalent to week-to-week predictions of weather.
/sarc
Related to this, does Aardvark falsify the old adage: “Past performance is no guarantee of future results.“?
In fact, neural networks can be very good at predictions of share prices. It was actually tried, in simulation on real data. The main problem was that if applied to any significant volumes of purchasing/selling, the results would sway the market itself, this led to increasingly erratic swings, a bit like an inexperienced plane or boat driver trying to correct direction. It got really ugly, really fast, because of the massive reduction in response times. Think Gamestop (I think it was?).
The only way it could work is to introduce random errors to avoid all programs doing the same thing, which is ironic.
You appear to be claiming that the US stock market, as it currently exists (i.e., without “random errors”) is completely deterministic. I disgree.
Also, all sorts of alarms bells go off when I see the phrase “simulation of real data” . . . that is, what’s wrong with using real (past) data?
As I recall, those wild swings occurred when the first computer software used for tracking and purchasing occurred. So many computers had the same software detecting the same trends and executing sells and buys at the same time, the market almost imploded. Regulation was put in place limiting the volume of purchases to keep the effects from devastating the economy.
I feel another Mencken quote coming up..
So far, the Atlantic Hurricane Season has been fairly quiet. What does Aardvark predict for the rest of the season?
I have a sneaky suspension that these investigators check the Old Farmer’s Almanac website for info and data for use in their computer programs.
Nope. At least not without many millennia worth of data covering many experienced scenarios to train on.
Its one thing to predict the trajectory of weather before chaos dominates but its entirely another thing to predict the trajectory of average weather without any useful historical training data.
And no, they cant use model data to train with although I’m sure they’ll try.
A fairly short answer for a fairy article:
“in your dreams”
I could have shortened it to “no” but then the wordplay with fairy and dreams would have gone overboard.
Waihing for someone to claim that AI will significantly improve modelling…yeah garbage input into a garbage model will still be garbage, irelevant if it’s artificial or even “intelligent”.
Let’s keep tossing coins like always, that at least requires some skill to predict the weather and is completely energy free.
Could a fully AI driven weather prediction system start a revolution in forecasting?
Yes, it seems quite plausible. I am judging from the example of chess. The same thing is true of Go also, but I don’t know that game well enough to be able to comment sensibly on it.
Alpha Zero turned out to be able decisively to defeat stockfish, the previously strongest chess program, which was already far stronger than the strongest human player. But its how it happened that may be indicative.
Alpha Zero had as input just the rules of the game. It then was trained by playing against itself millions or times, after which it had arrived at some method of finding strong moves. Its unlike stockfish in this respect, which seems to be the result of lots of human inputs, tunings, and openings and endgame knowledge.
It turned out that Alpha Zero could do pretty much what the forecasting program in the piece seems to be doing. That is, develop from its own iterative analysis moves which led to the goal. Something similar seems to be happening with AI analysis of X-Ray images.
So, within the limits of the chaotic nature of weather, surely it is quite plausible that an AI program could indeed teach itself to forecast far better than the current system of human and algorithmic forecasting.
If you are a reasonably strong chess player, the book on Alpha Zero is worth careful reading. Particularly the game analysis. It will change your view of AI.
The other thing that may happen is something else that happened in chess. The availability of very strong programs meant that analysis could happen much faster. Minutes rather than years. At the moment any A player can get instant analysis from his laptop from the equivalent of a super-grandmaster. Well that has meant that lost of previously received wisdom in chess, particularly the openings, has been overturned.
So the use of AI may not simply result in improved forecasts, it may also change and improve the way humans forecast as the human forecasters acquire better understanding of weather.
This is the way neural networks work, unlike previous chess programs. My brother was a National Master and programmer, and worked in a team to develop programs to play against each other, probably 40 years ago. Neural networks weren’t up to the immense possibilities in chess, which quickly escalates into trillions. Now they obviously are.
Chess is extremely simple, however. The weather isn’t! We don’t even know all of the factors, let alone what effects they have. Because of this, neural networks will still be far more erratic at predicting weather than playing chess.
BTW, most chess masters know all the openings, and will normally speed through them at an incredible pace, faster than most can follow, in timed matches. Neural networks probably have an edge here with greater numbers of examples to match to.
Yes, only within the limits of the chaotic nature of weather.
Chess, and probably Go also, and probably also diagnosis of X-rays, are much more limited fields than weather. But still, I do think it quite plausible that AI could do a lot better than any humans plus current methods and systems are doing now. And with far less equipment and cost.
Agreed. My only sticking point would be that we will probably not gain a greater understanding of weather. By their very nature, neural networks don’t reveal their ‘workings’. As such, the reasons they ‘predict’ something are not clear, and they have zero ability to attempt an explanation, unlike humans. This is their major flaw in predictions.
I’ve got another sticking point – the “AI” propensity to “hallucinate” aka “make shit up.”
Plus, of course, the high likelihood that the “training” will bake in “climate” BS, in particular given the political bent of most of the tech company leaders.
I am an old fuddy duddy when it comes to AI and neural networks. My experience tells me that forecasting weather with any accuracy is like forecasting the position of an electron. It can only be probability based and includes the effects of from outside an atom, i.e,, an EM field. In comparison with chess, it would be like moving your pawn last move and then on the next move, you find that pawn has changed to a bishop or worse, your king! Chaos exists. The further in the future you look, the less likely the result will be what you think it will be!
its interesting. I’m not completely sure how different weather is from chess for this purpose. The number of different games is estimated at very high levels, I have come on estimates between 10^50 and 10^several hundred. Go seems to be several times higher. At that level how much difference is there if you are approaching it with AI?
I guess the big difference is the chaotic nature of weather, where very small changes in inputs can produce large and essentially unpredictable variations in the weather. There is probably nothing to correspond to that in either chess or Go.
But surely the important thing is whether the predictions that come out of the machine learning are better (= more accurate) than those currently being done by the combination of humans and conventional programming? I think they may well be.
We shall see. A lot of money to be made by someone, if they are!
Poor choice: chess is based on strictly defined rules of logic/game piece movement and subject to miscalculation by the individual players responding to each other’s moves and mistakes . . . weather, on the other hand, is based on physics (including random variability in Earth’s hydrologic cycle).
It has been shown that modern supercomputers and their programming can now easily defeat the best human chess players . . . for weather, the best supercomputers and their programming are no better than humans in predicting weather beyond about 7 days into the future.
If humans didn’t get to look at the computer forecasts before making their predictions the performance of human forecasts would fall far short of the computer.
Strange comment.
The first successful computerized weather forecast was generated in 1950 using the ENIAC computer by a team of scientists led by meteorologist Jule Charney.
One can only wonder how during World War II, weather forecasters within the US military were able to make accurate weather predictions without ANY computer forecasts to “look at”.
From Google’s “AI overview”:
“The most critical and accurate weather forecast during World War II was the Allies’ prediction for the Normandy invasion on D-Day, June 6, 1944. Its accuracy allowed Allied Supreme Commander Dwight D. Eisenhower to choose a narrow, favorable window in an otherwise stormy period, catching the Germans by surprise.
“Other important forecasts profoundly influenced key battles, including the Battle of the Bulge and the campaigns in the Pacific.”
Rather, it was the prediction that the weather would be too difficult to launch the invasion on the originally planned date that was the key to success. During the war my mother translated ENIGMA decodes of diplomatic traffic. The most dramatic was a telegram from the German embassy in Dublin whose staff had managed to secure the details of the D-Day plans. The small team she worked in (located in Berkeley Street rather than Bletchley Park) were nervous wrecks on the morning of 5th June, thinking about the troops and the stormy conditions: although they had seen the detailed plans they were not aware of the last minute deferral for the weather.
It may seem bizarre, but the telegram was routed via the UK, a process that required it be rekeyed for onward transmission. The intercept and decode was a couple of weeks ahead of D-Day. The decision was taken to forward the telegram after the invasion, so as to avoid compromising the fact that diplomatic traffic was being read. The intercept itself was at least as important as the weather in ensuring the success of D-Day. Without it there was a high risk that Rommel would have redeployed defences, and the invasion might have required major delay and replanning.
“the best supercomputers and their programming are no better than humans in predicting weather beyond about 7 days into the future.”
True, but the claim is that the new method does produce better forecasts at far lower computing and human costs. I find it plausible that it could.
Reading Sadler’s book on Alpha Zero, and playing through the games, will give a different perspective. What you say about chess (and I think Go is the same) isn’t really correct. Yes, its deterministic in the way you describe. But the issue is that when the number of choices and possibilities get as large as they are with chess (and even more so with Go), the fact that it is in principle deterministic is no help in applying AI to it.
You can see this in the example of Stockfish: you can run it on a moderate spec desktop machine. Nevertheless this, and the open source variation on Alpha Zero, will probably beat any human grandmaster on even time limits. Alpha Zero is not doing brute force on a deterministic system. I don’t know quite what its doing, but play through the games and you will see that very clearly.
” . . . but the claim is that the new method does produce better forecasts . . .”
Ahhhh, so many promising claims, so many unfulfilled hopes.
Story tip:
From Mark Morano’s https://www.climatedepot.com/
Funerals should comply with net zero – The Telegraph
Death is Net Zero.
Net Zero is death.
A much better summation!
Following the policies prescribed by the Climate Fascists will leave people freezing and starving in the dark.
I do hope the AI colours predicted temperatures of 20ºC+ as angry red on the weather maps, otherwise the UK Met office will claim that it’s unrealistic.
viz:
Amusingly, the red starts at 15°C, in August. The horror!
I would actually regard those temperatures as pretty cold for August. Although my Geordie girlfriend might regard 16°C as a tad warmish for Newcastle….
“Importantly, Aardvark would not have been possible without decades of physical-model development by the community, and we are particularly indebted to ECMWF for their ERA5 dataset which is essential for training Aardvark.”
This is interesting in a very good way. I too am indebted to ECMWF for the ERA5 reanalysis, because its computed values of energy conversion show the absurdity of ever supposing that incremental CO2 will drive harmful “warming.”
More here, using the hourly parameter “vertical integral of energy conversion” for all of 2022. The minor radiative effect of 2XCO2 is massively overwhelmed by energy conversion within the general circulation. A well-trained AI system should be able to rapidly confirm this.
https://drive.google.com/drive/folders/1PDJP3F3rteoP99lR53YKp2fzuaza7Niz?usp=drive_link
I’ve been saying this for years. Convection vastly outweighs radiation everywhere except in a vacuum. The hydrological cycle is what regulates the climate, not radiation.
Thanks for your reply. It’s true that the mass and energy transfer delivered by convection and by water’s phase change properties outweigh the radiative influence of incremental CO2. But energy conversion is not the same thing. Rather, it directly demonstrates that the increased IR absorbing power from rising concentrations of CO2 cannot be uniquely realized as sensible heat gain as a final result. The active conversion of “heat” [internal energy + potential energy] to “wind” [kinetic energy] and vice versa in units of W/m^2 is what the ERA5 reanalysis model computes as the “vertical integral of energy conversion.”
Yes but I’m sure any such conclusion will be rapidly “memory holed” and the AI “reprogrammed” to express the “correct” (NOT) conclusion.
What I found interesting about this AI discussion is that Ardvark uses real world data and processes it looking for things that have happened before. It’s not taking the output of some computer game and then saying it’s the real world. If they add more data feeds and more history they just might detect useful patterns in the chaotic weather system.
This is indeed how neural networks are trained. Using a model would produce pointless results.
The proof of the pudding is in the…
… heating?
Put AI up against someone whose made a living off of watching the weather for decades, like a farmer, and see how it does. I grew up on a farm and still have roots in the farming community, farmers outperform the weatherman every single day good chance they’ll outperform AI too.
If they do, they’re taking in data that isn’t being entered into the neural network for training. We would need to find out what that data is.
How much do models tell the AI what to do, and how much straight observation and dsta,?
I ask this because a note was made regarding sea ice. If sea ice growth and decay over the year we’re reimagined as clouds (say), with an annual time frame reimagined as (say) weeks, could 46 years of sea ice satellite data be used to predict a 30 year, “day” to the computer, sea ice “weather”?
I suppose it will be interesting.
If the system uses the same data as current forecasters, will it produce better or worse results? Modelling based on the present system is just the adoption of the current orthodox expert opinion and is not likely to find new and better predictions. Maybe we will find out.
Observing the data patterns still doesn’t provide provide proof of causation or accurate prediction of the future.
They should try this on the financial markets.
“Aardvark already outperforms the United States national GFS forecasting system on many variables and it is also competitive with United States Weather Service forecasts that use input from dozens of weather models and analysis by expert human forecasters. “
A pretty low bar though, don’t you think?
If they spent time feeding in past weather patterns and what actually happened, then the AI might be able to quickly match patterns, calculate differences between past patterns and make a forecast.
Worth a shot but I think it will still need a skilled meteorologist to check it before it’s released.
Matching patterns and trends? Been going on in the stock market for years. The big difference is humans and the weather’s physical propensities. Are they that different? Review the charts on https://jessescrossroadscafe.blogspot.com/ and weep.
Actually I have long thought that a number of aspects of weather systems could be forecast at least for relatively short timescales using a neural network approach. A prime example is forecast precipitation based on weather radars and other inputs The simple forecasts we have now just consider how the rain bearing airmasses are moving to guesstimate their future positions. Training neural networks would allow implicit modelling of complex interactions such as local topology, clouds running out of rain to dump, vortices that lead to intense downpour in parts of rainstorms, etc.
In reality a lot of human forecaster skill is the result of human pattern recognition and ability to learn from similar patterns of historic measurement. Brute force modelling by supercomputer has well known distinct limitations because it ends up with inaccurately interpolated data feeding a three dimensional grid model of inadequate resolution. In the end weather forecasting is limited by the nature of complex chaotic systems, and an AI/neural network approach will also be subject to limitations with diverging accuracy for longer term forecasts. But AI could well take us another step further from Red sky at night, shepherd’s delight; red sky in the morning, shepherd’s warning.
I would hope the AI occasionally looks out the window.