“The writing’s on the wall that this is going to transform things.” That was Richard Turner, a professor of machine learning at the University of Cambridge, on a new AI weather system called Aardvark Weather. In his words, it is “going to be the new way of doing forecasting.”
The research was published in Nature, by teams from Cambridge, the Alan Turing Institute, Microsoft Research and the European Centre for Medium-Range Weather Forecasts (ECMWF). The Guardian reported that a single researcher with a desktop computer could use it to produce accurate forecasts. It runs tens of times faster than conventional systems and needs thousands of times less computing power.
One model instead of a chain
Today’s forecasts come out of a series of stages. Each takes several hours on purpose-built supercomputers, and large teams of experts have to build and maintain the whole chain.
Aardvark skips most of that. It trains an AI directly on raw observations from weather stations, satellites, balloons, ships and aircraft, and it makes predictions from them. Earlier work by Huawei, Google and Microsoft had shown that one step, the numerical solver that calculates how weather evolves over time, could be swapped for AI. The ECMWF already uses that approach.
Using only 10% of the input data existing systems need, the researchers said, Aardvark outperformed the US national GFS forecasting system in certain respects. It also held its own against forecasts from the United States Weather Service.
Forecasts on demand
Turner said the approach could deliver tailored forecasts quickly, such as temperatures for African agriculture or wind speeds for a renewable energy firm in Europe. A custom system built the traditional way takes years of work by large research teams. He expects the model eventually to give accurate eight-day forecasts, against five days now, along with hyper-local predictions.
Dr Scott Hosking of the Alan Turing Institute said the work could “democratise forecasting” by putting powerful tools in the hands of developing nations. It could also help policymakers, emergency planners and weather-dependent industries.
Dr Anna Allen, the paper’s lead author at Cambridge, said the findings open the way to better forecasts of hurricanes, wildfires and tornadoes, and of air quality, ocean dynamics and sea ice.
