When you roll your eyes because weather forecasts are wrong, you need to know that predicting the weather is one of the most complex problems in science. Now, Google has let artificial intelligence act as a weather forecaster. The results show that after just one minute of running on a machine, artificial intelligence can make accurate predictions 10 days in advance, which usually takes a roomful of supercomputers several hours to complete.
The famous "butterfly effect" holds that the occurrence or non-occurrence of a storm may be affected by something as small as the flapping of a butterfly's wings in another part of the world. It's the weather forecast's job to turn these proverbial butterflies into accurate models that tell you whether you should go ahead and plan your picnic for next Saturday.
This is called "Numerical Weather Prediction" (NWP), and it uses current weather observations from around the world as input data and calculates them through complex physics equations run on supercomputers. But now, Google has released an artificial intelligence system called GraphCast that can calculate data faster on less powerful hardware.
The artificial intelligence was trained on 40 years of weather reanalysis data collected from satellite imagery, radar and weather stations. GraphCast extracts weather conditions six hours ago and current weather conditions, then uses its treasure trove of data to predict weather conditions six hours from now. On this basis, it can extrapolate forward in 6-hour increments to make weather forecasts up to 10 days.
GraphCast makes predictions at more than 1 million grid points on the Earth's surface, each with a latitude and longitude of 0.25 degrees. At each grid point, the model considers five variables: temperature, air pressure, humidity, wind speed and direction at the surface, as well as six variables at 37 different altitudes in the atmosphere.
In tests, GraphCast running on a Google TPUv4 machine was compared to the current gold standard for weather forecasting, a simulation system called High Resolution Forecasting (HRES) running on a supercomputer. GraphCast is able to produce 10-day forecasts in one minute and is more accurate than HRES on 90% of the test variables and forecast preparation time. When the model focused on the troposphere (the lowest layer of the atmosphere where accurate predictions are most useful for daily life), GraphCast outperformed HRES 99.7% of the time.
Even more impressively, GraphCast identified severe weather events earlier than HRES, even though it was not specifically trained to do so. In one real-life example, AI accurately predicted where a hurricane would make landfall nine days in advance, while traditional forecasts could only confirm it six days in advance.
Google said that the code of GraphCast is open source, and scientists around the world can experiment with it and apply it to daily weather forecasts. This kind of number crunching is the ideal job for artificial intelligence, so they can leave art and writing to us humans.
The research was published in the journal Science.
learn more:
https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/