Slash Boxes

SoylentNews is people

SoylentNews is powered by your submissions, so send in your scoop. Only 19 submissions in the queue.
posted by hubie on Sunday November 19, @11:06PM   Printer-friendly
from the AI-overlords dept.

On Tuesday, the peer-reviewed journal Science published a study that shows how an AI meteorology model from Google DeepMind called GraphCast has significantly outperformed conventional weather forecasting methods in predicting global weather conditions up to 10 days in advance. The achievement suggests that future weather forecasting may become far more accurate, reports The Washington Post and Financial Times.

In the study, GraphCast demonstrated superior performance over the world's leading conventional system, operated by the European Centre for Medium-range Weather Forecasts (ECMWF). In a comprehensive evaluation, GraphCast outperformed ECMWF's system in 90 percent of 1,380 metrics, including temperature, pressure, wind speed and direction, and humidity at various atmospheric levels.

Original Submission

This discussion was created by hubie (1068) for logged-in users only. Log in and try again!
Display Options Threshold/Breakthrough Mark All as Read Mark All as Unread
The Fine Print: The following comments are owned by whoever posted them. We are not responsible for them in any way.
  • (Score: 5, Interesting) by crafoo on Sunday November 19, @11:44PM (2 children)

    by crafoo (6639) on Sunday November 19, @11:44PM (#1333558)

    I think this is a particular area that people believe AI should be very good at. non-linear, highly-coupled dynamic systems. Chaotic dynamical systems, or "ordered chaos". very sensitive to initial conditions but which confine themselves to a set domain over time/iterations. Maybe we will even see design tools for turbulent flow soon?

    Oh, looks like people are all over it: [] []

    • (Score: 4, Interesting) by PiMuNu on Monday November 20, @09:12AM (1 child)

      by PiMuNu (3823) on Monday November 20, @09:12AM (#1333588)

      > AI should be very good at. non-linear, highly-coupled dynamic systems

      Indeed, for weather forecasting they have a massive training data set; and it is a very large system with many initial parameters that needs a high-dimensional approach.

      Does anyone know how 10 days compares with, say, the rotational frequency of one of these big weather cells (e.g. hadley cell or ferell cell)?

      By the way here is the arxiv for your nature paper: []

      • (Score: 2) by crafoo on Thursday November 23, @06:01AM

        by crafoo (6639) on Thursday November 23, @06:01AM (#1333927)

        indeed. thank you for the arxiv link. I cannot answer your weather cell question but it directly points to the heart of the matter.

  • (Score: 3, Touché) by EJ on Monday November 20, @01:06PM (1 child)

    by EJ (2452) on Monday November 20, @01:06PM (#1333598)

    So, AI does better than a guy flipping a coin.

    • (Score: 0) by Anonymous Coward on Monday November 20, @10:27PM

      by Anonymous Coward on Monday November 20, @10:27PM (#1333660)

      Down wind of Lake Erie (western NY State) the weather is highly changeable. Long ago I heard a local say, "Don't like the weather? Just wait 20 minutes!" and it often works that way. Once, driving for 45 minutes across the Buffalo metro area I went through three bands of near-whiteout snow, with clear bright sun in between.

      With that said, in the last decade or so, the local forecasts have gotten much better. What might have been fairly accurate looking ahead one or two days is now pretty reliable 3-4 days out. This is, afaik, almost completely due to improvements in weather modeling. In the last few years they have even added experimental predictions of local high snow accumulations, which are quite good looking 24 hours ahead. It will be interesting to see if adding AI to the toolbox can stretch this out another day or two, but I don't have my hopes up.