Next time you roll your eyes at a weather forecast that got it wrong, just remember that predicting the weather is one of the most complex problems in science. Now, Google has put AI to work as a weatherman, and shown that in just one minute on a single machine, it can make accurate predictions up to 10 days in advance, a task that normally takes a room full of supercomputers hours to achieve.
The famous butterfly effect posits that whether or not a storm brews could be influenced by something as small as a butterfly flapping its wings in another part of the world. It’s the job of weather forecasting to wrangle all these proverbial butterflies into accurate models that tell you if you should go ahead with planning that picnic next Saturday.
Doing so involves what’s called Numerical Weather Prediction (NWP), which uses current weather observations around the world as input data and runs it through complex physics equations run on supercomputers. But now, Google has unveiled an AI system called GraphCast that can crunch the numbers much faster, on less powerful hardware.
This AI was trained on 40 years’ worth of weather reanalysis data, gathered by satellite images, radar and weather stations. GraphCast takes the state of the weather six hours ago and the current state, then uses its treasure trove of data to predict the weather state six hours from now. From this, it can project forward in six-hour increments to build a forecast up to 10 days out.
GraphCast does this across more than a million grid points around the Earth’s surface, each measuring 0.25 degrees in longitude and latitude. At each of these points, the model accounts for five variables – such as temperature, pressure, humidity and wind speed and direction – at the surface and six in the atmosphere at 37 different altitudes.
In tests, GraphCast running on a single Google TPU v4 machine was compared to the current gold-standard for weather prediction – a simulation system called the High Resolution Forecast (HRES), running on supercomputers. GraphCast was able to make 10-day forecasts in under a minute, and was more accurate than HRES on 90% of the test variables and forecast lead times. When the models were focused on the troposphere – the lowest layer of the atmosphere, where accurate predictions are most useful and applicable to everyday life – GraphCast outperformed HRES 99.7% of the time.
Even more impressive, GraphCast demonstrated an ability to identify severe weather events earlier than HRES – even though it hadn’t been specifically trained to do so. In one real-world example, the AI accurately predicted where a hurricane would make landfall nine days in advance, while traditional forecasts could only confirm it six days ahead.
Google says that GraphCast’s code is open source, allowing scientists around the world to experiment with it and incorporate it into everyday weather forecasts. This kind of number-crunching feels like the perfect job for AI, so they can leave the art and writing to us humans.
The research was published in the journal Science.