Google Deepmind researchers have developed a weather forecast program enhanced with artificial intelligence and will improve forecasts beyond three days.
A study published Wednesday, November 15 in the journal Science concludes that 10-day weather forecasts could become much more reliable thanks to artificial intelligence. Researchers at Google Deepmind have developed a new weather forecasting program called GraphCast. For several months, it has been put to the test against what is best today in the field. And the result is clear: in more than 90% of cases, its forecasts are both more reliable and more precise.
During Hurricane Lee which brushed the Antilles at the beginning of September, the program predicted its speed and trajectory nine days before, whereas with traditional tools, we had not seen anything coming for six days. It could therefore become very valuable for anticipating extreme weather phenomena such as storms, heatwaves or torrential rains.
Introduction of artificial intelligence to augment forecast calculations
All weather forecasters rely on computer models for their medium- to long-term forecasts. These are models that need enormous computing power since they must simulate all the movements of air masses and all their interactions in the atmosphere.
The introduction of artificial intelligence makes it possible to extrapolate from a 40-year history of weather data, without losing speed. Thanks to it, ten-day forecasts are calculated in less than a minute and on a single machine, whereas the best models today need several hours on a network of supercomputers. So, not only is it more reliable, but it’s also faster and ultimately less energy-intensive.
However, artificial intelligence will not replace current forecasting models. It should rather be seen as a complement. This is a technology that appears to be very effective for large-scale global phenomena. But much less so for more granular, more local forecasts.
In fact, the two methods should work in parallel, feeding off each other, to perhaps finally give us truly reliable forecasts beyond three days. Today, with more than three days left, we are talking about “trends”.