Slashdot is powered by your submissions, so send in your scoop

 



Forgot your password?
typodupeerror
×
AI Technology

AI-Driven Weather Prediction Breakthrough Reported (theguardian.com) 54

A new AI system called Aardvark could deliver weather forecasts as accurate as those from advanced public weather services but run on desktop computers, according to a project unveiled Thursday and published in Nature. Developed by the UK's Alan Turing Institute with partners including Cambridge University, the European Centre for Medium-Range Weather Forecasts and Microsoft, Aardvark aims to make sophisticated forecasting accessible to countries with fewer resources, particularly in Africa.

The system has already outperformed the US Global Forecast System on many variables in testing. Project leader Richard Turner noted the system is "completely open source" and not planned for commercialization by Microsoft.

AI-Driven Weather Prediction Breakthrough Reported

Comments Filter:
  • by Rufty ( 37223 ) on Thursday March 20, 2025 @03:26PM (#65248065) Homepage
    Aardvark? 'Coz it works aard???
  • The formulas for the algorithm may be genius, but the forecasts can only be as good as the data being used to make the forecasts. Getting good data has long been like the hardest part.
    • Re: Okay, but... (Score:4, Insightful)

      by jacks smirking reven ( 909048 ) on Thursday March 20, 2025 @04:37PM (#65248233)

      NOAA and NWS make all their data freely available and while I can't access the paper I have to assume part of its training and resources are all that data that comes from national agencies.

      Developments like this are all the more reason to maintain and expand those public systems.

      • Re: Okay, but... (Score:5, Insightful)

        by Zarhan ( 415465 ) on Thursday March 20, 2025 @04:41PM (#65248241)

        Considering Elon is axing thousands of people from those agencies I'd say you are soon going to need to run a local AI on your basement RTX 4090 to get that tornado alert during the windy season.

        • True, I was thinking more who is gonna be installing their own radar stations to feed those models and sure enough while I thought it was a joke you can do it, for cheaper than expected.

          Build a Small Radar System Capable of Sensing Range, Doppler, and Synthetic Aperture Radar Imaging [mit.edu]

          Is this our future, one which includes a resistance of pirate weather operators?

          • by Anonymous Coward

            Is this our future, one which includes a resistance of pirate weather operators?

            No.

            This is much more complicated than setting up your own microwave internet connection with a cantenna. When you start working with Doppler signals you get into voltage controlled oscillators, waveguides, mixers, and port circulators. SAR systems involve phased arrays, signal processors, and PPI displays. While technically feasible, it's about as easy as building your own tunneling electron microscope in your basement.

        • lol

          Probably the best geek burn of that whole shit-show I've seen.
        • by gtall ( 79522 )

          It will have to wait until the training data can accept alternate maps with Sharpie lines drawn on them. No one knows weather like la Presidenta.

      • NOAA and NWS make all their data freely available and while I can't access the paper I have to assume part of its training and resources are all that data that comes from national agencies.

        Developments like this are all the more reason to maintain and expand those public systems.

        In the recent past, supercomputers have tried to improve weather forecasts by increasing the resolution of weather models, i.e., by throwing more data at the problem. This approach does work, but it requires increasingly larger computers with more memory running for longer times, i.e., the approach doesn't scale well.

        That AI can improve the efficiency of this approach of using past data to predict future data is not surprising. Maybe the slightly surprising thing is that the approach scales so well, altho

    • Re: Okay, but... (Score:4, Informative)

      by drinkypoo ( 153816 ) <drink@hyperlogos.org> on Thursday March 20, 2025 @05:12PM (#65248331) Homepage Journal

      Wunderground shows that you can get people to contribute data at their expense. You can get a quality weather station under $100 now*, Vevor has a good one. (I'm not 100% on the accuracy of the wind gauge, but given that it tracks both baseline and gust speed it might be accurate. I need to get someone with a portable to come by and check it.) With more sensors, you can cross-check.

      * My last weather station was $80 on sale from $120 and that didn't even include what you needed to share or log the data. This one has WiFi built into the display unit. You can install a proxy if you want to log locally.

      • I have 2 coworkers that run them. I'm in a condo, so I can't really- but the more people who run these things, the better.
        The age on relying on shit like NOAA may be coming to a swift end. Even after the fucking shit-show at 1600 Penn is over, will the next administration have the mental bandwidth to try to rebuild it all?
        • Since I put one up there are two in my town, which is great because redundancy.

          I just happen to live in a house where it's convenient to mount one. It's got a flat roof (concave actually, what a dumb design, they made it dump the water onto the back patio like idiots too... 1930s deco in a logging town) and the sewage vent sticks up from a side wall, so it was an easy install. This unit is also zero penetration, it's solar powered. And it has light power (from solar) and a dedicated UV sensor, too.

          If there

  • COULD deliver (Score:5, Informative)

    by Viol8 ( 599362 ) on Thursday March 20, 2025 @04:10PM (#65248161) Homepage

    Ie doesnt yet. Right now I'd prefer my forecasts from a program that's done hard fluid dynamics maths rather than some statistical blender that's doing little more than guided guesswork. The humans in the chian already do that.

    • Can I post what ChatGPT turned my troll-hearted comment into?

      ""What if electricity plays a more significant role in the weather than fluid dynamics? If atmospheric electrical forces are more influential than traditionally assumed, then models relying purely on fluid dynamics might be missing a key piece of the puzzle."

    • Re:COULD deliver (Score:4, Informative)

      by DamnOregonian ( 963763 ) on Thursday March 20, 2025 @08:44PM (#65248683)
      All fluid dynamics atmospheric modeling is statistical. Who the fuck moderated this informative?
      An NN can fit any curve under the sun, dependent only upon the amount of hidden layers it has.
      Until you're modeling every atom in the atmosphere, which sadly requires a machine with the mass of the atmosphere at a physical minimum, you're engaging in statistics, and as long as you're engaging in statistics, there is a neural network that is going to outperform what you're doing.
      • " there is a neural network that is going to outperform what you're doing."

        Does the attention mechanism allow for long-range context that neural networks alone were never able to provide to solve context-sensitive grammars?

        • Yes, attention layers are interspersed between each feed-forward layer to apply context to the transform... obligatory reading: "transformer block"
          But I feel like you probably know that.
      • by Viol8 ( 599362 )

        Since you apparently don't know the difference between statistics and interpolation I think we can assume you're an idiot who knows some impressive sounding buzzwords but doesn't actually have a fucking clue.

        • lol!
          I love it.

          CFD has to account for turbulent flows, which are inherently stochastic. Interpolation in this case fits the inherently stochastic terms into neat functions, which is why different interpolation functions are more or less accurate for different sets of data. Because it's fundamentally statistical.

          The only person who threw around a buzzword here, is you- because you didn't know what CFD was. You dumbshits crack me up.
          • by Viol8 ( 599362 )

            Nice meaningless word soup. Did chatgpt write it for you? LOL :)

            • It was a spelled out chain of logic to help you understand why you're being stupid... I mean, I can ask ChatGPT to explain it to you, if you want... Not sure you'd get it then, either.
              • by Viol8 ( 599362 )

                Save it pal. I've seen your posts before. You're a grade A know nothing bullshitter. You've demonstrated you have zero clue how weather forecasting models work but feel free to continue making a fool of yourself. I'll get the popcorn.

    • Even if it does end up being as accurate it's still not a useful breakthrough. I don't care whether my forecast comes from a national weather service or an AI algorithm running on a desktop, what I care about is the accuracy. What would be a lot more useful is an AI algorithm that increases the the accuract of a forecast, not one that runs on a smaller computer.
  • The jokes just write themselves. Oh, wait, maybe that's an AI writing the jokes.

  • by Local ID10T ( 790134 ) <ID10T.L.USER@gmail.com> on Thursday March 20, 2025 @05:03PM (#65248301) Homepage

    To replace NOAA/NWS...

  • Unless and untill that gets changed to "will" and there is liability when it fails in ways no actual weaterperson would have failed, this is worthless.

  • Likely just another reason to gather data from people's machines.
  • AIs doing things cheaper, faster, with less CPU.

    That we hear a lot.

    AIs doing novel stuff...

    Not so much.

    Hmm.
  • Our weather can turn on a DIME! Just this week, we had temperatures in the 80's (26C) along with tornado, thunder/lightning/hail. Then 50 mph wind & a bit of snow with temperatures bottoming out below freezing. And in 24-48 hours, it will be in the 70's again. We can get all four seasons within 24 hours LOL.
  • by PPH ( 736903 ) on Thursday March 20, 2025 @09:19PM (#65248717)

    Aardvark passes when I can't tell the difference between it and the cute weather lady on Channel 37.

  • by Bumbul ( 7920730 ) on Friday March 21, 2025 @01:09AM (#65248885)
    For those interested, The European Centre for Medium-Range Weather Forecasts (ECMWF) has a good selection of these Deep Learning models (e.g. from Microsoft, Deepmind) fed with real data - you can check and compare the resulting forecasts here. Weather forecasts [ecmwf.int]
    • It seems the AI research from Huawei, Google (GenCast), and Microsoft is already being harnessed by ECMWF to enhance forecasting accuracy and speed by replacing the numerical solver component. However, Aardvark sets its sights higher, aiming to overhaul the entire forecasting system rather than just optimizing one part. The Singularity is approaching.
  • The forecasts that are most important to businesses are the ones that are two seasons in advance. The local daily and weekly forecasts are for broadcast ratings, with the exception for local severe weather, which are needed by those that could be affected.
  • A comparison between this and Google Gencast seems in order. https://deepmind.google/discov... [deepmind.google]

Almost anything derogatory you could say about today's software design would be accurate. -- K.E. Iverson

Working...