A system of forecasts that delegates the entire processing process to the AI, without moving from supercomputer, could trigger a revolution.
The future of the weather forecast could look more to a single researcher who in a few seconds develops accurate scenarios with a simple PC, than to a teamwork of hours, complex and based on the use of supercomputer. A new system of Machine Learning promises to use artificial intelligence To replace each phase, and not just a fewof the entire process required by traditional weather forecast. Thus arriving at faster, accurate and customizable elaborations.
The model, called Aardvark Weather and elaborated by the scientists of the University of Cambridge, of thelan Turing Institute, Microsoft Research and the European Center for Medium-Range Weather Forecasts (ECMWF), is described in an article in the magazine Nature.
Tailor -made. For Richard Turner, professor of Machine Learning of the University of Cambridge and the first author of the study, the new approach could serve to provide quick and adaptable forecasts to specific needs (and places): for example, to predict the temperatures of the next few days To optimize crops in Africa, or to understand which they will be The speeds of the winds To efficient the work of a renewable energy company in Europe. According to its creators, the model should be able to deliver accurate weather forecasts up to 8 days later Instead of up to 5, as the current ones do.
Where is the news? Modern weather forecasts are based on numerical prediction of time (Numerical Weather Prediction, NWP), which tries to obtain a vision of the future state of an atmospheric system starting from the data on its initial state. These calculations are extremely complex and require Supercomputer energy and performance In order to be done.
In recent years, the IA has gradually seen expanding its role in the weather forecast by replacing some phases of traditional models. But his contribution had so far limited himself to the forecast part: he had never concerned a passage called initializationin which the data collected by satellites, weather balloons and menterological stations scattered all over the world are collected, filtered and made interacting in an organized grid on which the forecast is based. All this first part, explains Turner, consume half of the computational resources required.
One more step. For the first time Aardvark Weather uses Machine Learning also for the initialization phase. It starts from the raw data collected by meteorological stations, satellites, balloons, ships and airplanes and, using only 10% of the input data used by existing systems, reaches results comparable to the most advanced traditional systems, over time of a second and using a calculation power thousands of times lower, accessible to a normal PC.
Perplexity. For the authors of the study, the discovery lays the foundations for the most refined forecasts of natural disasters such as hurricanes, fires and tornadoes, for more accurate forecasts of air quality, extension of sea ice and ocean currents. However, other scientists point out that the grille of the earth’s surface used by the US cell model of 1.5 degrees per side, wider than those of the most advanced traditional forecasting systems such as Era5, of the UCMWF (which are 0.3 degrees). The fear of some is, therefore, that the model can be too coarse to capture the arrival of complex and unexpected weather phenomena.
