Skip to content

AI weather forecasting swamps the competition

The New York Times has a pretty interesting piece today about Google's latest and greatest AI weather forecasting tool, GenCast, which extends reliable forecasting from 10 days to 15:

A new artificial intelligence tool from DeepMind, a Google company in London that develops A.I. applications, has smashed through the old barriers and achieved what its makers call unmatched skill and speed in devising 15-day weather forecasts.

....The world leader in atmospheric prediction is the European Center for Medium-Range Weather Forecasts. Comparative tests regularly show that its projections exceed all others in accuracy.... The new agent outdid the center’s forecasts 97.2 percent of time. The A.I. achievement, the authors wrote, “helps open the next chapter in operational weather forecasting.”

Generative language AIs rely on training themselves with petabytes of written internet content. This helps them learn to speak and understand the real world, but obviously doesn't allow them to forecast the weather. So how does GenCast do it?

The DeepMind agent runs on smaller machines and studies the atmospheric patterns of the past to learn the subtle dynamics that result in the planet’s weather.

The DeepMind team trained GenCast on a massive archive of weather data curated by the European center. The training period went from 1979 to 2018, or 40 years. The team then tested how well the agent could predict 2019’s weather.... Mimicking how humans learn, it spots patterns in mountains of data and then makes new, original material that has similar characteristics.

Here's a chart that shows errors in tracking tropical cyclones. GenCast is about 25 km better at every date range:

Plus GenCast eliminates the need for huge, expensive supercomputers:

Instead, the DeepMind agent runs on smaller machines and studies the atmospheric patterns of the past to learn the subtle dynamics that result in the planet’s weather.

....Dr. Price of DeepMind said, the new agent can generate a 15-day forecast in minutes compared with hours for a supercomputer. That can make its projections much timelier — an advantage in tracking fast-moving storms.

I can imagine that different training sets will allow GenCast and its successors to predict weather events of all sorts at lower cost and with greater accuracy than ever before. Next up: Earthquakes. Let's get cracking, geo-boffins.

28 thoughts on “AI weather forecasting swamps the competition

  1. JohnH

    I'm surprised, which is to say really dubious. I'd have said that the problem can't e shortage off data, when we can see the entire planet, or the inability to interpret them. It's just that things change over time. More like Kevin's chart making and questionable conclusions that AI will replace.

  2. climatemusings

    I went to a talk at the American Geophysical Union last year, which pointed out that pattern matching was the exact approach used to predict the weather for D-Day - except that manual labor was used to compare historical weather maps to present conditions. Of course, once there were computer models, those easily beat the labor-intensive human comparison process.

    But with AI, pattern matching becomes easy. And pattern matching works at the resolution of the real world, whereas computer weather models generally run at a resolution of a few km. So it isn't that surprising that AI can beat weather models. (but this kind of pattern matching may not work for future climate projections, as the whole point of climate change is that we are moving away from historic patterns)

    1. FrankM

      But presumably the AI could just as easily be trained to recognize the changes over time and account for them it its predictions. Over the short time scales we're talking about the extrapolation would be small. That removes another degree of freedom, though, which would slightly degrade the accuracy of the prediction.

  3. jdubs

    Wow, the weather in 2019?! Amazing! Let's shovel more money that away!

    Predicting earthquakes 5 years after the fact will mean even more funding!

    1. aldoushickman

      The point of "predicting" the past is to be able to assess how well the model performs. It's a pretty common approach to demonstrating whether a model works.

      1. emjayay

        It seems that jdubs doesn't understand the concept. I expected a more familiar screen name when I checked who made that comment.

  4. Doctor Jay

    "Plus GenCast eliminates the need for huge, expensive supercomputers"

    This is true, and also misleading. Current state of the art is that there is a complex model that is a lot like, if not solely, a PDE solver (PDE stands for partial differential equation). Traditionally, this is the meat and potatoes for a supercomputer, and it doesn't surprise me at all that some of them maybe still run on supercomputers.

    AND, these days a "supercomputer" might be a rack of 80 high performance intel machines (running Linux) with a high speed network connection between them all.

    So, it seems likely that GenCast can run on a few machines that sport maybe 4 high-end Nvidia boards. By the way, by the standards of say, 2001, each one of those boards is a supercomputer.

    Another interesting tidbit. Nvidia chips are only single precision. All supercomputers were double precision. Because those PDEs were sometimes kind of unstable numerically, and the extra precision was really valuable.

    However, the extra precision is not valuable to a neural network. The extra information it needs is carried in some other way, extracted earlier and held in the state space differently.

    Sooo, it's not wrong. Nobody needs to buy a Cray anymore. This is the truth. Of course, Cray hasn't existed for probably 20 years, so there's that.

    1. TheMelancholyDonkey

      Cray still exists. It's a division of Hewlett-Packard. Cray still has a manufactory in Chippewa Falls, WI.

      1. Doctor Jay

        I dunno, seems like the entity you reference might be better called "Zombie Cray" than "Cray".

        Cray was a failing business in the 90's and to my mind ceased to exist when it got bought out by Silicon Graphics in the 90's.

  5. Justin

    I really never look at the weather 15 days out. For the record, the weather channel says it will be some kind of partial cloudy ⛅️ with high in the upper 30s and low in low 30s. There is a 24% chance of rain that day which seems not particularly useful. So what would make that inaccurate? No rain? No clouds? High of 50?

  6. msobel

    Obviously this programs required careful regulatory oversight. Otherwise they could end up promulgating the climate change hoax.

  7. NeilWilson

    Deep Mind has been amazing since it developed Alpha Go and Alpha Zero almost a decade ago.

    It created a system that learned from itself.

    The thing that will carry over to other uses of AI is the fact that Alpha Zero was so much better than Alpha Go. Alpha Go studied the best humans and learned from them. Alpha Zero started from scratch and learned everything on its own.
    Human knowledge is often a bad thing when AI is learning how to complete a task.

    I am not sure what the ramifications of that is.

  8. emjayay

    Meanwhile (not based on any specific information) Elonvivek will no doubt recommend eliminating NOAA/National Weather Service since the private sector and AI are doing the job. They might not be as good at it with no data though.

    1. cld

      They'll just make it illegal to access the information without a license because that would be Big Government walking all over the little guy.

  9. tango

    How long before the profession of meteorology collapses? Hope none of you or your loved ones is in that field, or whatever AI will come for next?

  10. pjcamp1905

    You might imagine that but, as the disclaimer goes, past results do not predict future performance. I can imagine that, with climate change in play, predictions will get harder, not easier, as weather patterns that haven't existed for millions of years come back into play. My imagination is as legitimate as yours.

    And the thing is this will help you plan a vacation but it is not science. If you don't understand why the software made the prediction it made then you haven't learned a damn thing.

  11. HalfAlu

    The AI predicts the weather will be similar to previous years when conditions were the same. It will work as long as local weather patterns do not change, good thing the climate isn't changing then. And it will fail to predict (notice) if unusual conditions today are heading for some unusual weather, which the physics based predictions will find.

    1. TheMelancholyDonkey

      They'll work so long as weather patterns don't change too quickly. You can train the AI to have a recency bias, weighting the last few years more heavily than those farther in the past.

      I also don't think you understand how it works. The system doesn't make predictions that require similar conditions to previous years. It makes predictions based upon current conditions, and how the weather developed in very similar conditions in previous years. That there will be more very hot days isn't a problem. It only becomes a problem if the same set of conditions as happened in the past lead to different outcomes in the present.

  12. D_Ohrk_E1

    A plot of the ensemble against GenCast is not as useful as seeing it compared to each outlet's cyclone paths.

    You know, I'd like AI to predict El Nino. That would be a fabulous tool for long-range planning.

  13. Altoid

    Does this flavor have the same capacity to hallucinate as the LLM-based ones? It would be a caution flag, at the very least, if it did

  14. bouncing_b

    Even if AI forecasts more accurately than the physical-equation models (as seems very possible), it won't save that much money because it still needs training and initializing data. Better and ever more complete data.

    Satellite data. Radar data. Soundings. Buoys.

    We're still going to need NOAA for that.

  15. jeffreycmcmahon

    You can tell Mr. Drum doesn't know anything about geology by his assumption that earthquakes can be "predicted" like weather, but they can't, and the reason is that the data that we would need to be able to do so is buried miles deep in the Earth's crust and is largely inaccessible.

Comments are closed.