It only takes a minute for an artificial intelligence developed by Google researchers to generate 10-day weather forecasts for every region of the world.
The “robot” is therefore (much) faster and cheaper than the computer models currently used.
But are his predictions more accurate? This is what the researchers say in their study published on Tuesday in the journal Science: This “model” called GraphCast, developed by the British company DeepMind from Google, would have been 10% fairer than the traditional “European model” at more than 90%. % of “meteorological variables” evaluated – which includes both daily forecasts and forecasts of “extreme events” such as hurricanes or severe heat or cold waves.
One of the two models commonly used in meteorology (the other is the “American model”) is called the “European model”. Both make their forecasts based on complex mathematical equations, which in turn are based on the basic principles of physics (temperatures, atmospheric and oceanic currents, etc.). The scope of the calculations requires a lot of computing power.
In comparison, new models that have emerged in the last two years amid rapid advances in AI, such as Microsoft’s ClimaX or Nvidia’s FourCastNet, are based on trends they discover through updated weather data. The work requires less computer power and takes less time: a minute or less, rather than an hour or more. It is guaranteed that AI can also learn when new data arises.
The 18 signatories of the study, all of whom work at Google DeepMind in London or Google Research in California, even affirm that their offspring are better than their main competitor among young AIs of the same type, namely Pangu-Weather from the Chinese company Huawei.
But they also recognize its Achilles heel: the lack of precision. For example, in extreme weather events: Because these are rare, the AI lacks data that it can rely on to accurately predict their strength or trajectory. A particularly tricky question when it comes to hurricanes…
Due to the same logic – the lack of data – AI models like GraphCast or Pangu-Weather have their limitations: they cannot make such precise predictions for such a large number of parameters (precipitation, wind speed and direction, temperature, pressure, humidity). etc.) than the European or American models. This makes AI, summarizes the Washington Post meteorologist, “less useful for predicting small-scale phenomena, like a storm or flood.”
As observers noted last summer, climate change may eventually increase the difficulty if AI is too trained to make predictions based on data from the last 40 years.
Meanwhile, meteorologists themselves face a learning curve: One of their jobs has always been to interpret (and properly disseminate) the information they receive based on what they have learned over decades about the strengths. and weaknesses of traditional models. However, currently few people other than the developers of the new models know how they work, or more precisely, “why an AI model makes the predictions it does.” “These models are still in their infancy,” data visualization researcher Jacob Radford of Colorado State University explains in the post, “and trust still needs to be built, both in the research community and among the people making the forecasts.”