Weather forecasting may be one of the clearest examples of AI improving an old, computationally expensive science rather than merely automating office work. A study highlighted today reports that Google’s WeatherNext can provide a three-day cyclone forecast with accuracy comparable to a two-day forecast from conventional methods.
Traditional numerical weather prediction solves enormous sets of physics equations on supercomputers. AI forecasting instead learns atmospheric patterns from historical and simulated data, allowing forecasts to be generated much faster once the model is trained.
The important word is complement. Meteorologists are not simply throwing away physics models. AI forecasts can be run quickly in large ensembles, helping forecasters estimate uncertainty and compare possible storm tracks. An extra useful day of warning can materially improve evacuation and emergency planning.
A Nature study reports that Google’s WeatherNext AI forecasting system improves tropical-cyclone prediction enough that its three-day forecast can match the accuracy of conventional two-day forecasts. The bigger promise is faster, cheaper forecasting that complements physics-based supercomputer models.