Weather: A Google Deepmind Ai Proves To Be More Accurate Than All Current Predictive Models.
Artificial intelligence continues to revolutionize methods and increase efficiency across numerous sectors. If you find that medium-term weather forecasts are not at all reliable, it's worth noting that artificial intelligence could soon make a significant difference in this field. An AI developed by Google DeepMind has proven to be more accurate than all current forecasting models. Here's how.
A breakthrough in the field of weather forecasting.
Weather forecasts are important for both individuals and professionals. They allow them to adjust their activities according to inclement weather and to plan their schedules. However, they are not always reliable, especially in the medium term...
In this field, as in many others, artificial intelligence could help forecasters make a real leap forward, according to a recent article published in the prestigious journal Science.
A team from Google Deepmind, the group's subsidiary specializing in artificial intelligence, has developed a 10-day weather forecasting program called GraphCast and has achieved impressive results.
Accurate forecasts in record time.
The AI developed by Google Deepmind has managed to surpass almost all existing forecasting tools for medium-term weather predictions. But its effectiveness is not its only quality. To achieve this excellent result, this artificial intelligence used only a fraction of the computing power required by HRES, an acronym for High Resolution Forecast which refers to the prediction system of the European Centre for Medium-Range Weather Forecasts.
The performance of GraphCast, revealed in the journal Science, is impressive. According to the authors of this article, this algorithm achieved better results than those of HRES on more than 99% of the meteorological variables and in 90% of the 1300 regions tested!
This artificial intelligence has also proven capable of predicting disasters with great accuracy. It anticipated the exact impact point of Hurricane Lee, which hit the Canadian province of Nova Scotia in September, 9 days in advance. In contrast, traditional forecasting systems only managed to do so 6 days before the hurricane's arrival and with less geographical precision.
But the most revolutionary aspect is the speed at which GraphCast delivers its results. While the HRES system needs several hours to provide a 10-day weather forecast, Google Deepmind's algorithm provides a result in less than a minute!
A radically different method.
GraphCast stands out from current weather forecasting systems by using artificial intelligence and also by the data used to achieve a result.
Today, the best weather prediction systems like HRES all work in the same way: they use thermodynamic equations and fluid mechanics to calculate parameters such as temperature, humidity, and atmospheric pressure.
Since these equations are extremely complex, it takes a lot of time and great expertise to define them and translate them into algorithms. Weather forecasting also requires the use of cutting-edge supercomputers, offering phenomenal computing power, to apply these calculations on a large scale.
Google Deepmind engineers have chosen a completely different approach. Instead of calculating atmospheric variations from equations, they preferred to base their predictions on real-world data. They trained their artificial intelligence with a huge database of meteorological data collected over several decades.
This training allowed GraphCast to learn the cause-and-effect relationships that govern weather evolution. Google's system uses the rules it has learned to quickly determine the weather for the coming 10 days, without having to go through very complex equations.
A tool that is not infallible.
Given its impressive performance, one might think that GraphCast will soon replace all other predictive systems. However, this will not happen in the near future, as this method of weather forecasting has a major flaw.
With this type of artificial intelligence, one can only observe the results and possibly commend their accuracy. But it is not possible to precisely determine how the algorithm achieved them. The magic happens in a kind of "black box" that remains opaque to our understanding… If the algorithm were to start offering results disconnected from reality, it would be almost impossible to realize it and identify the exact source of the problem.
Conversely, traditional predictive models are much more robust. Moreover, the accuracy of weather forecasts is constantly improving, not only because the supercomputers used are increasingly powerful, but also because forecasters continue to refine their equations over time by comparing forecasts with actual weather.
Even researchers at Google Deepmind acknowledge that their invention is not intended to replace modern predictive systems like HRES. They rather believe that the two tools are complementary. GraphCast is a new tool whose experimentation will continue. It interests meteorological researchers because it can be very useful, provided that one does not rely entirely on it. At least, for now…
