DeepMind says its AI weather models can give earlier cyclone warnings

In short: Google DeepMind says its A.I. weather models can sometimes deliver accurate cyclone forecasts about a day earlier than leading traditional systems.

What's going on

Google DeepMind has been testing a series of A.I. weather models that try to predict the weather by learning from past data, like a student who has studied years of old forecasts and outcomes. The company says these systems can match or beat some of the best “conventional” forecasting tools, which are physics-based simulations that use supercomputers to calculate how the atmosphere should move.

One earlier model, called GraphCast, was described in a Science paper and in DeepMind’s summary. In those tests, GraphCast produced more accurate 10-day forecasts than the European Centre for Medium-Range Weather Forecasts’ high-resolution system on more than 90% of 1,380 test variable and lead-time combinations. DeepMind also says it can run much faster than traditional numerical weather prediction (computer simulations of the atmosphere).

DeepMind later reported results for GenCast, saying it outperformed ECMWF’s ensemble system (a method that runs many slightly different forecasts to show uncertainty) on 97.2% of test targets overall. It reported especially strong results beyond 36 hours, and said skill extended up to 15 days.

The newest model, WeatherNext, is focused on cyclones. DeepMind says it gives forecasters an average extra day of useful accuracy, meaning its 3-day cyclone forecasts match what earlier systems only reached at 2 days.

What to watch

The evidence suggests “a day or more earlier” is most solid for cyclones and some medium-range forecasts, not every kind of weather in every place. Traditional models still matter, especially for very local, fast-changing events like short-range rain.

Source: NYTimes