Artificial intelligence first surpassed exactly the traditional models of weather forecasts

Scientists consider this a “turning point” in the prediction of hurricanes, cyclones and extreme heat

The Graphcast weather prediction model based on artificial intelligence first surpassed exactly existing forecast systems. Graphcast was developed by a group of scientists from Google Deepmind. Unlike traditional methods, the GraphCast model works on the basis of machine learning using historical data in which it can find non -obvious patterns. Thanks to this, her forecasts are more precise. In his study, published in the journal Science, a group of scientists calls the result of the comparison of Graphcast with the traditional models of the “turning point” in predicting hurricanes, cyclones and extreme temperatures.

One of the most accurate weather systems in the world is the High Resolution Forecast (Hres) system, which is used by the European Center for Mediciper Weather Forecasts (ECSPP).

This system, like all other traditional models of weather predictions in the world, uses the so -called “numerical forecast”. In fact, it is a solution to many equations containing various meteorological data using supercomputers. At the same time, as scientists note, over the years the accuracy of forecasts has increased to such an extent that the trajectory of the hurricane can be predicted in many days, which was unthinkable several decades ago.

However, this approach has its drawbacks: firstly, it requires very large computational resources. Secondly, it is limited by the laws that meteorologists study and consciously add to the model in the form of additional equations and their clarifications. Such a model itself cannot study on historical data - for this she needs a “translator” in the form of a researcher.

In recent years, an alternative to this approach has been machine learning, which uses historical data and independently builds a forecast on the basis of many years of statistics, without taking into account the “physics” of atmosphere behavior.

The development of this approach was a model developed by scientists called Graphcast. It creates an accurate 10-day forecast in less than a minute on one Google Cloud Tpu V4 device-that is, it spends several orders of magnitude less computing data than existing supercomputer systems. As the initial data for the forecast, the model takes two states of the weather on Earth - at the current time and six hours ago, collected by ECSPP based on global meteorological observations.

Scientists compared the results of Graphcast with the forecasts of the HRes system, which is used by the ECSP, in a number of indicators, including temperature, pressure, speed and direction of wind, as well as humidity at different levels of the atmosphere. Graphcast surpassed HRes with 90.3% of 1380 indicators.

In addition, scientists tested the abilities of Graphcast in the prediction of phenomena, the forecasting of which it was not specially trained - tropical cyclones, atmospheric rivers (these are narrow zones of high concentration of water vapor in the atmosphere) and extreme temperatures.

As an example of a successful forecast, scientists are given by a hurricane “Lee”, who came to Canada in September 2023. According to the host of the author of the article by Remya Lama, Graphcast was able to predict that “Lee” would go ashore in New Scotland nine days before this happened - compared to six days for traditional approaches. “This gave people for another three days to prepare for his arrival,” Lama told the newspaper Financial Times to reporters.

However, as FT notes, with the prediction of the strengthening of the hurricane Otis off the Pacific coast of Mexico at the end of October, artificial intelligence coped no better than traditional models.

Another advantage of models based on artificial intelligence scientists call the cost of their improvements - it is possible to increase the effectiveness of their work through the use of modern deep learning equipment (this is one of the types of machine learning), not supercomputers.

According to the coordinator of the machine learning of the ECSPP Matthew Cantri, after training, Graphcast will be "extremely cheap in operation." “We can say that it will be about a thousand times cheaper from the point of view of energy consumption,” Cantrie said.

At the same time, the models using machine learning are critically dependent on the quality and volume of data, the source of which in this case is the archive of traditional “numerical” forecasts. Therefore, as scientists note, Graphcast should not be considered as a replacement of traditional methods, but as evidence that models based on machine learning can solve the problems of real forecasting.

Scientists themselves call the results of their research “turning point” in the weather forecasting.