Paul Dorian
Surface forecast map for next Monday, April 1st, made by the 00Z “Artificial intelligence” version of the Euro model; Map courtesy ECMWF, tropicaltidbits.com
Overview
It was just a matter of time…artificial intelligence (AI) has hit the numerical weather prediction world with a strong emphasis on “pattern recognition” and there is no telling where this will lead in the world of weather forecasting. Numerical weather prediction is well suited for AI as – in its current form – it requires a tremendous amount of data crunching and super computing power to resolve the physical laws of fluid dynamics to produce weather conditions in the future. One of the most notable AI advances in recent years has come with the European Centre for Medium-Range Weather Forecasts which is generating experimental AI forecasts that are made available to the public.
Surface forecast map for next Monday, April 1st, made by the 00Z “conventional” version of the Euro model; Map courtesy ECMWF, tropicaltidbits.com
Details
Weather forecasts have improved in accuracy over the years with today’s 6-day forecasts about as good as the 3-day forecast from 30 years ago. This improvement in overall accuracy has come about for numerous reasons one of which has to do with the much better computing power in today’s world compared to three decades ago. Artificial intelligence is now spurring a new revolution in numerical weather prediction that many believe will produce model-based weather forecasts as good or even better than the best traditional models.
The European Centre for Medium-Range Weather Forecasts (ECMWF) is known for generating what is considered to be one of the top “traditional” computer forecast models in the world known to most as the “Euro”. In the fall of 2023, this agency began to generate its own experimental AI model-based forecasts known officially as the “ECMWF-AIFS” where AIFS is an acronym for “Artificial Intelligence Forecasting System”. This experimental forecast model, based on ECMWF initial conditions, has been made available in an alpha version to the general public for free and can be found at their own web site here. The resolution of the ECMWF-AIFS model is approximately one degree (111 km) with plans for this to be regularly increased in the future.
850 mb temperature anomaly forecast map for next Monday, April 1st, made by the 00Z “Artificial intelligence” version of the Euro model; Map courtesy ECMWF, tropicaltidbits.com
Traditional weather models start off by feeding a snapshot of current conditions, based on observations from satellites, weather stations and buoys, into a grid-like computer model that divides the atmosphere into millions of boxes. This snapshot is then run forward in time for each box by applying equations that are based on the physical laws of fluid dynamics and this requires great computational power. Indeed, this kind of data crunching requires supercomputers with 1 million processors and can take several hours to run…usually four times a day.
850 mb temperature anomaly forecast map for next Monday, April 1st, made by the 00Z “Artificial intelligence” version of the Euro model; Map courtesy ECMWF, tropicaltidbits.com
The new AI models play a role in weather prediction by simulating and analyzing past weather events, learning from historical data, and recognizing recurring weather patterns which enhances AI’s ability to predict future weather conditions. In other words, AI skips the expense of solving the equations in favor of “deep learning” after training on 40 years of ECMWF “reanalysis” data (a combination of observations and short-term model forecasts that best represents past weather) (source).
The European Agency is not alone in producing AI forecast models as numerous tech giants are getting involved. In a paper published recently in Science, Google introduced GraphCast and claims it can make weather predictions more accurately (and faster) than the ECMWF High-Resolution Forecast (HRES) on 90% of its verification targets up to 10 days in advance.
The advance in AI forecasting has been rapid during the past few years and one of the important next steps will be to produce ensemble results, which helps to capture uncertainty by running a model multiple times with slightly differing input parameters to create a range of outcomes. While few expect traditional model forecasts to disappear anytime soon, AI will likely approach the point in the near-term where it can be a very useful complement. And when it comes to artificial intelligence, the bottom line is that there is really no telling where this will lead us over the next five or ten years; therefore, as is usually the case when it comes to weather forecasting, stay tuned.
Meteorologist Paul Dorian
Arcfield
arcfieldweather.com
Follow us on Facebook, Twitter, YouTube
Does this program have any tendency to hallucinate, like ChatGPS?
Mr. Halla, Well, sort of, if they programmed it to predict ocean-boiling. (If they haven’t yet, they will get around to it).
ChatGPT 3.5 hallucinated an answer to me that didn’t seem correct. I asked it for some references and it hallucinated up them as well.
I decided to switch to claude.ai, it gives better answers, but it is Climate Crazy!
Fascinating.
Have to say it doesn’t surprise me. Pattern recognition is probably yielding better results in protein folding than the models built from the bottom up.
Computational physical models do have severe limitations in complex systems. Of course the people who do it never like to advertise the known failings too loudly.
a strong emphasis on “pattern recognition”
Did it recognize any new patters so far? Is El Nino a pattern?
Recognise new patterns? Probably not is my guess.
Protein folding is the same in some ways. It can only recognise patterns that have been selected for by billions of years of evolution.
As far as I know they don’t run their protein folding competitions on random protein sequences that don’t fold into anything selectable.
Artificially eliminate the vast majority of the negatives and you end up with a lot of positives that make the program look better than it really is.
Are you saying that protein folding is not deterministic (i.e. dependent on local geometry and charge distribution)? Sounds like you are implying it is a hysteresis effect, dependent on a cell’s history. Is that correct?
Deep learning models also have severe limitations in complex systems.
And simple systems as well.
IPCC CMIP updates seem to be developing a pattern . . . a pattern of ever increasing error in climate model predictions versus actual data measurements.
Should we be worried if AI “consults” these updates?
Yes.
“The advance in AI forecasting has been rapid during the past few years and one of the important next steps will be to produce ensemble results…”
Isn’t this what climate modelers have been doing without much success?
What is the collective noun for climate scientists?
An Ensemble….
A herd
A consensus, if you’re an alarmist.
A failure
Successful propaganda.
The AI “contact us” answering at companies is pretty stupid.
Any AI solution to weather forcasting could also beat the stock market and win enough to give everyone a free EV and free electricity forever.
And free beer!
Artificial Intelligence will kill the “butterfly effect”.
I can book my vacation a year in advance.
“Artificial intelligence and weather forecasting…a quiet revolution is taking place in numerical weather prediction”
Filtered output: Advanced super fast pattern-matching and weather forecasting… will prove just as wrong as the models.
One form of processing data (tortured or otherwise) against another form of… processing data.
The scientific understanding of climate doesn’t change even if the answers do.
Back in July 2014, when the Met Office unveiled their latest update:[Even Newer Dynamics for General atmospheric modelling of the environment (ENDGame)] they mistakenly made this comment public.
New Dynamics has served us well over more than a decade: not only have we continued to improve the skill of our large scale forecasts at the rate of 1 day lead time per decade (so for example today’s 3 day forecast is as accurate as the 2 day forecast was 10 years ago) but we have seen the introduction of a very high resolution (1-1/2 km) model over the UK which provides unprecedented levels of detail to our forecasters.
So at this rate they will be able to get a 7 day forecast just a accurate as todays 3 day forecast in only 40 more years and 100 years to get to 14 day forecasts. 2014 to 2024 & Still on the 1 day per decade. Now if they could just get an accurate 2 day forecast they might have something to sell.
And yet strangely their forecasts change overnight such that what they forecast yesterday for today is not the same as the forecast this morning.
Hmmm, I would have bet that computer driven pattern recognition for weather prediction has been around for a long time.
I’d also bet that the climate crusaders abuse it to produce propaganda.
Actually, forecasting via human pattern recognition has a long past. Going back to Irving Krick and the 1930s. But it’s success was limited and abandoned once NWP got going.
https://en.wikipedia.org/wiki/Irving_P._Krick
Joe Bastardi uses pattern recognition. He calls is analogs.
They still talk about cold fronts and warm fronts in the weather forecasts.
Why stop at weather forecasting? AI that can forcast chaos would find random chance child’s play.
It could quickly pay off the national debt and start paying taxpayers money every year rather than taking it. Who wouldn’t vote AI in as PM or President on a promise to pay back everyone’s taxes plus interest?
As Dr Patrick Moore, once said. We’d be better off, using a statistical analysis of historical trends, rather than using the climate models to postulate future trends.
How long until someone compares AI results to the Farmers Almanac ?
I’ll take The Farmer’s Almanac at 3:1 odds.
Back in the 1970’s my parents actually bought a calendar that predicted the daily weather based on something-in-the-past.
I don’t recall it being worse than the regular UK weather forecast. Of course, we payed no attention to either.
AI might actually be an improvement over the flood of DEI hires that are now slowing many government systems to a crawl. However, how are they going to teach the AI computer to accomplish the first step in forecasting: look out the window.
‘…the flood of DEI hires that are now slowing many government systems to a crawl.’
Is that necessarily a bad outcome?
It is when you have to wait for governmental permission to perform even the most basic of business functions.
story tip
Hail storm in Damon TX destroys huge PV system:
https://twitter.com/Roughneck2real/status/1772339177264148491?s=20
PV modules are designed to withstand only 25mm ice balls, falling vertically with no wind velocity component, this is the result.
I tried to point this out 20-25 years ago, without success.
This crop has been damaged badly. Could it possibly withstand a sonic boom?
Oh dear.
Regular FF power plants last an awful lot longer. In fact dynamite is often employed to remove them.
We were predicted to get at least a foot of snow from the last storm and we got less than an inch. I don’t think that could be classified as a win for AI.
Did the weather service that you were relying on use AI?
I assume that computer models are a form of AI.
Got the snow part right, eh wot?
From the above article:
“Artificial intelligence is now spurring a new revolution in numerical weather prediction that many believe will produce model-based weather forecasts as good or even better than the best traditional models.”
Well, when has the consensus ever been wrong in the past?
/ sarc off
As I understand it, AI “mines” all available electronic information to determine a consensus of average “correct” information on a given subject, which it then may use in in its own algorithms to develop a response to a question, or even a “prediction” if asked for such.
IMHO, the fundamental
problemfailing of AI is that it has no independent means for distinguishing incorrect or intentionally-misleading information (i.e., falsehood) from correct, valid information (i.e., truth) other than by the quantity of each that is available. Given that there is very likely more incorrect, misleading information posted on-line than there is correct, accurate information, AI is doomed to be essentially useless in making forecasts . . . the ol’ GIGO effect.Just one example: heaven help us if AI interprets publications (e.g., assessment reports or CMIPx reports) from the IPCC as being factual. 😜
Bingo, +1000.
“AI training” means that a lie repeated 1000 times becomes truth.
IMHO, the fundamental failing of people is that they have no independent means for distinguishing incorrect or intentionally-misleading information (i.e., falsehood) from correct, valid information (i.e., truth) other than by the quantity of each that is available.
I disagree . . . most people, albeit not all, have an innate ability to recognize a untruth (a falsehood) when they see it. Among other things, this is the basis of the Magna Carta, the US Constitution and the jury system in the USA.
It is also a problem facing both Joe Biden and Donald Trump as the US approaches the next presidential election.
Psychologists can debate the source of this human ability. Computer programmers can debate if this ability can ever be developed in “artificial intelligence”.
As to the latter, I am very doubtful.
They may think they do but they don’t. If I was to ask you whether the Baltimore bridge recently collapsed, unless you had direct experience of it which is unlikely, your belief of it comes entirely from large numbers of reports.
Actually just the opposite! My belief in reports of things I didn’t directly observe comes a relatively few news and science sources that I’ve come to trust for accurate news reports, NOT from “large numbers of reports.” You know, quality over quantity.
I don’t automatically believe any single source, but when three or more (of the group that I trust) are reporting essentially the same story then I’m onboard with it.
Based on “from a large number of reports”, I take that to mean you’ve come to believe the IPCC claims of pending catastrophic global warming . . .after all, the IPCC has continuously issued report-after-report of such since its founding in 1988, 35 years ago, and most of the MSM outlets have endlessly parroted those same IPCC claims.
I’m sure that’s true for you but it doesn’t alter the fact there are large numbers of reports out there. If just one of your sources and only one over the entire internet had reported the Baltimore bridge had collapsed, would you have believed it?
Right. It takes more than one. Yes, reputable is definitely better.
No. Nor do I believe the many claims the earth is flat. Or that perpetual motion is a thing.
But the point is, far too many people do believe those things despite the claim that people are somehow better at spotting falsehoods. At the end of the day, we all spot falsehood through our trained knowledge. There is nothing magical about the way we do it and AI does that too.
Sorry, TTTM, I couldn’t disagree more. IMHO, there is something truly magical about the way intelligent humans (the majority of which, at least) are able to distinguish truth from falsehood.
I’ve never seen an adequate psychological explanation for how this is possible (just some vague references to it being the result of trial and error) but even children as young as six years old learn that their parents can catch them if they state a falsehood (i.e., a lie).
I don’t think we need go into further discussion as to AI “hallucinations” (a proven fact) being equivalent, or not, to lies.
Each to their own. IMO consciousness emerges from a running neural network and our brain is nothing but a biological machine. There is no magic, nothing supernatural involved.
A lie means it’s said with intention. Do you believe ChatGPT intends to lie when it hallucinates?
I don’t think we need go into further discussion as to AI “hallucinations” (a proven fact) being equivalent, or not, to lies.
Further to hallucinations, if I were to ask you to answer the following question
Describe how Azacitidine and Carboplatin can be effective in priming for anti-immunotherapy (Avelumab) in patients with advanced ICB-resistant melanoma.
And you were expected to answer purely on what you know, would you be “telling lies”? Or would you be simply following instruction to the best of your ability? Or both?
Its all very well for you to start an answer with something like: I dont know anything about this, but…
But what if you did know something about it because you read something at some point but aren’t sure of the details?
Now you might start an answer with something like: If I remember correctly, …
And what if its something you’ve read lots on?
Now you might answer with confidence even if you get parts wrong because you misremembered a detail or two.
So what is AI really missing? Humility? Afterall its typically extensively read on a subject but by its very nature can never have practical experience.
We don’t need AI for this, a woke government is enough.
AI is a false term coined years ago in computer games.
It is not intelligent. It lacks self-awareness. It lacks judgment. It lacks the ability to visualize alternate futures.
Hmmm…. Something to try. Ask an AI if it is intelligent. Curious to see the answer.
Expected response: “I am not programmed to answer that question.”
Norman coordinate!
Now if only AI could be packaged into those forms that Norman “coordinated”
😜
+42 intergalactic credits for the superb—albeit oblique—reference.
Try ChatGPT for yourself. Everything you think you know about AI is incorrect.
Are you intelligent?
ChatGPT response:
As an AI language model, I’m designed to process and generate human-like text based on the input I receive. My responses are generated through pattern recognition and statistical analysis of vast amounts of text data I was trained on. While I can provide helpful and relevant responses to a wide range of inquiries, I don’t possess emotions, consciousness, or subjective experiences like humans do, so the term “intelligence” as applied to me differs significantly from human intelligence. However, I strive to be as helpful and informative as possible within the scope of my programming and training.
That’s a whole lot of flatulence as a response
Exactly! And I find it fascinating that ChatGPT identified “emotions, consciousness and subjective experiences” as being consistent with human intelligence . . . while at the same time overlooking important things like reasoning, logic, objective thinking, critical analysis, memory capability and communication skills as contributing to the definition of “intelligence”.
Also, I was shocked—well, not really on second thought—to see ChatGPT admit that it was trained on “vast amounts of text data” . . . and apparently only that!
No wonder its called artificial.
Note that ChatGPT told you what it couldn’t do, not what it could do. Perhaps you missed that subtlety in the answer.
This is what it said it could do:
While I can provide helpful and relevant responses to a wide range of inquiries
It also says
so the term “intelligence” as applied to me differs significantly from human intelligence.
If you use it, you’ll learn that it indeed can do
It has relatively rudimentary memory right now limited to the context of earlier queries. And it will also display faulty logic and reasoning at times. But people never do that, right?
Your post, with my bold emphasis added.
Perhaps you missed the contradiction in your statements.
The is from your response to expecting a list of things it could do. It didn’t answer that way and so you went on to believe, and I quote
That’s your answer, not ChatGPTs.
Yes, and I stand by my answer.
That’s not how this type of AI works. In this case, the AI is being trained on the specific weather data that they have. It is trained on a specific data set and learns from its mistakes, but it is only THAT data that it works on.
Can you please define who is “training” the AI on that “specific weather data”?
Or are they just letting the AI bot “train” itself . . . after all, what could possible go wrong with that considering the proven fact that AI is prone to “hallucinations”?
ROTFL.
BTW, love your phrase “this type of AI” . . . we already have a caste system setup for bots?
Trained on the Farmer’s Almanac!
Best weather forecaster is the one who takes the time to look out the window.
The ECMWF correctly forecast in 2021 the deluge in the Ardennes and Taunus mountains of eastern Belgium and western Germany, predicting the massive flooding of in particular the Ahr valley a few days in advance. They informed the authorities of the coming disaster, who did … nothing to warn the population. Some 200 people lost their lives.
Afterwards the bureaucrats blamed, you guessed it, climate change for the mishap.
I wonder if AI meteorology will be better inducing fear….? I am sure it will know when to make sure the forecast penetrates our soul and changing our outlook from better to worse. LOL.
Anyone one remember the song by Steve Reich “it’s gonna rain”…..and endless loop of it’s gonna rain!
Meteorology is fascinating science…consuming it is garbage.
Looks to me like they are more interested in the act of forecasting than the weather.
“Weather forecasts have improved in accuracy over the years with today’s 6-day forecasts about as good as the 3-day forecast from 30 years ago.”
Not in my town. Just saying.
Yeah. That statement is complete BS. The day’s forecast in the morning is often completely wrong by afternoon.
First, you have to understand the motivation for using AI in this way. It isn’t the desire to help people by making better forecasts. The real aim is to keep the gatekeepers in control of the gate. But it should end up with better mid-term weather forecasts anyway, ie, more than one week.
Isn’t pattern matching similar to what Weatherbell (Bastardi,D’Aleo) have been promoting?
“ learning from historical data,”
Years ago I asked/wondered if forecast models (Not Climate Model Abominations) couldn’t be programmed with past weather patterns (The Blizzard(s) of ’78) and recognize and then “forecast” based on what they did historically if the same pattern emerged.
If that’s all it does, the extra computing power of AI might make for an improvement in weather forecasting.
(If they include the hockey stick as “historical” data, best wet your finger and stick it out the window!)
A PS
Weather models’ forecast’s accuracy or lack thereof is quickly evident.
If the input is just a bit off, they can still hit close to the target because the time is so short. And they can learn from what they got wrong and then tweet the input values for the next forecast.
Climate Models’ forecast? They never learn.
“And when it comes to artificial intelligence, the bottom line is that there is really no telling where this will lead us over the next five or ten years; therefore, as is usually the case when it comes to weather forecasting, stay tuned.”
No telling at all. It’s likely that it will lead us down the garden path – as usual.
Anybody who believe that the future can be predicted by “simulating and analyzing past weather events, learning from historical data, and recognizing recurring weather patterns” is away with the fairies.
I don’t believe I’ll gain anything by staying tuned, just yet.
Artificial Intelligence = Programmed Conclusions.
Change my mind.
https://chat.openai.com/auth/login
Cliff Mass Weather Blog: Should I Let AI Write My Blogs?