Listen to the audio version of the article
With the climate crisis, the instability of weather conditions and the need for increasingly accurate forecasts grows, to save human lives and minimize the economic damage of the most serious events. Hence the rapid growth of the sector, in search of a future in which every farmer is able to manage sowing and harvesting safely, in which it is easy to plan energy consumption based on wind and solar production times, in where it is possible to plan the construction times of buildings or know in advance the arrival of a flood.
Why is it worth investing?
Every euro invested in accurately forecasting floods and droughts generates a return of as much as 400 euros, according to estimates by the European Center for Medium-range Weather Forecasts, an intergovernmental organization headquartered in Reading, supported by 23 EU members (plus 12 other countries), whose computer system contains the largest archive of weather data in the world, in one of the largest supercomputer complexes in Europe, installed a year ago in the Tecnopolo of Bologna.
The recent experiments with artificial intelligence applied to weather forecasts were based precisely on this data, from which the extraordinary results obtained by Google DeepMind’s GraphCast emerged. The GraphCast model “marks a turning point in predictions,” its developers claim in an article published in mid-November in the journal Science. An in-depth evaluation has shown that GraphCast is more accurate than the conventional system operated by Ecmwf of Reading, currently the world leader in forecasts with a lead of three to ten days. In 90% of the 1,380 atmospheric variables examined, the Google DeepMind model was more accurate than the forecasts produced by the European Center, which included temperature, pressure, humidity, wind speed and direction at different levels of the atmosphere. Matthew Chantry, head of machine learning at the European Center, certified these results.
Accurate and rapid technology
The European Center has run real-time tests with AI models from Huawei and Nvidia, as well as DeepMind, along with its own built-in prediction system, and Chantry has endorsed DeepMind’s claim that its system is the most accurate. “GraphCast is consistently proven to be better than other machine learning models, Huawei’s Pangu-Weather and Nvidia’s FourCastNet, and is also more accurate than our prediction system on many points,” says Chantry. And that’s not the only advantage guaranteed by this technology, which in addition to being accurate is also very fast.
The standard meteorological simulations, used by the Ecmwf of Reading or by the National Oceanic and Atmospheric Association of the United States, are based on the so-called “numerical weather forecast”, which uses mathematical and physical models of the Earth’s atmosphere starting from the weather condition known at the time of prediction. Over the last few decades the accuracy of this method has improved drastically, but it remains very complex and requires hours of work from several supercomputers, given the enormous amount of data to be analyzed. The model developed by DeepMind researchers, led by Rémi Lam, can be used on a laptop and provides accurate results in less than a minute.