An artificial-intelligence (AI) model developed by Microsoft can accurately forecast weather and air pollution for the whole world — and it does it in less than a minute.
The model, called Aurora, is one of a slew of AI weather forecasters being developed by tech giants, including Google DeepMind’s GraphCast and Nvidia’s FourCastNet. But Aurora’s ability to quickly predict air pollution globally is pioneering, say researchers.
DeepMind AI accurately forecasts weather — on a desktop computer
“This, for me, is the first big step in a journey of atmospheric chemistry and machine learning,” says machine-learning researcher Matthew Chantry at the European Centre for Medium Range Weather Forecasts (ECMWF) in Reading, UK.
Conventional weather forecasting uses mathematical models of physical processes in the atmosphere, land and sea. To predict air-pollution levels, researchers have previously used machine learning along with conventional mathematical models, says Chantry. Aurora seems to be the first entirely AI model to predict air pollution globally — a much more complex task than weather forecasting, says Chantry.
“That was the thing where I went: wow, that’s a really cool result,” he says. The benefit of AI models is that they often require less computational power to make predictions than conventional models, says Chantry.
AI researcher Paris Perdikaris at the Microsoft Research AI4Science in Philadelphia, Pennsylvania, and his colleagues found that Aurora could predict the levels of six major air pollutants worldwide: carbon monoxide, nitrogen oxide, nitrogen dioxide, sulphur dioxide, ozone and particulate matter — for five days in under a less than a minute. It can do it “at orders of magnitude smaller computational cost” than a conventional model used by the Copernicus Atmosphere Monitoring Service at ECMWF, which predicts global air pollution, the team wrote in a preprint published on arXiv on 20 May.
How AI is improving climate forecasts
The quality of Aurora’s preditions were similar to those of the conventional model. Such predictions are used by policymakers to track and protect against the health harms of air pollution, which has been linked to an increased risk of asthma, heart disease and dementia.
The researchers trained Aurora on more than a million hours of data output from six weather and climate models. After training, the team tweaked the model to predict pollution and weather globally. The model generates a 10-day global weather forecast alongside the air-pollution prediction.
Although the team says that Aurora could outperform other AI weather-forecasting models, such as Graphcast — which can outperform conventional models and make global weather predictions in minutes — at some tasks, it is too early to make a definitive comparison, says Chantry. “You’d have to spend a lot of time and probably have access to the models themselves, to be able to really go into detail and say with some certainty that model A is better than Model B,” he says.
Further research will reveal whether ‘foundational’ AI models trained on diverse data sets, such as Aurora, perform better than those trained on a single data set, such as Graphcast. “There’s lots of cool science to be done,” he says.