NOAA is rolling out a new generation of AI-driven weather models

Three operational systems, including a hybrid physical-AI system, producing 16-day forecasts in 40 minutes using only a fraction of the usual computing resources.

The US National Oceanic and Atmospheric Administration has launched a suite of global weather forecasting models based on artificial intelligence. The main advantage is not just accuracy but cost: a 16-day forecast produced by the AIGFS system uses 0.3 per cent of the computing resources of the traditional GFS model and is completed in around 40 minutes, making the data available to forecasters much earlier.

The suite comprises three applications. AIGFS improves the forecasting of large-scale structures and significantly reduces track errors for tropical cyclones at long lead times, although version 1.0 does, however, result in a deterioration in intensity forecasts. AIGEFS, a 31-member ensemble, achieves a quality comparable to that of the conventional operational ensemble whilst using just 9 per cent of its resources. Finally, HGEFS combines the 31 physical members and the 31 AI members into a ‘large ensemble’ comprising 62 members.

It is this latter ensemble that represents a true first: to the best of NOAA’s knowledge, no operational meteorological centre had previously deployed a hybrid ensemble combining physical modelling and AI modelling. It consistently outperforms each of the two systems taken in isolation on most validation metrics. Developed as part of the EAGLE project, this suite is based on Google DeepMind’s GraphCast model as a starting point, which is then refined using analyses from NOAA’s data assimilation system.