ECMWF Unveils AI-Driven Weather Forecasting System
The European Centre for Medium-Range Weather Forecasts (ECMWF) has launched the Artificial Intelligence Forecasting System (AIFS), an AI-powered weather prediction model. According to the ECMWF, the new system outperforms existing physics-based models, achieving up to 20% greater accuracy. This marks a substantial leap in forecasting capabilities, with AIFS proving faster and more energy-efficient than its predecessors.
Medium-range forecasts, ranging from three to fifteen days out, are crucial for various sectors. These forecasts aid in everything from disaster preparedness to everyday planning for individuals. Traditional models rely on solving complex physics equations, which, while effective, are inherently approximations of atmospheric dynamics. AI models offer an alternative by learning directly from historical weather data. This method allows them to identify complex patterns and relationships that might be missed by traditional equation-based models.
AIFS offers significant improvements in how quickly forecasts are generated and how much energy is consumed compared to traditional models. “At the moment, the resolution of the AIFS is less than that of our model (IFS), which achieves 9 km [5.6-mile] resolution using a physics-based approach,” said Florian Pappenberger, Director of Forecasts and Services at ECMWF.
The new system comes after Google DeepMind’s GenCast model, the next iteration of its weather prediction software. GenCast exceeded the performance of ECMWF’s existing model (ENS) in the majority of different weather variables and lead times. However, the ECMWF’s new system introduces a complementary approach with the intent to keep innovating to provide the most accurate product to their user community. “We see the AIFS and IFS as complementary, and part of providing a range of products to our user community, who decide what best suits their needs,” Pappenberger added.
The ECMWF team plans to explore hybrid models that blend data-driven and physics-based approaches, with a goal of achieving greater precision in weather forecasting. “Physics-based models are key to the current data-assimilation process,” said Matthew Chantry, Strategic Lead for Machine Learning at ECMWF and Head of the Innovation Platform. They use this same data-assimilation process to initialize machine learning models.
Chantry also co-authored a data-driven system, named GraphDOP, which doesn’t rely on physics-based reanalysis. The system uses observable quantities from polar orbiters to predict weather parameters up to five days ahead. The integration of artificial intelligence with physics-driven modeling represents a promising avenue for more precise weather forecasting. Though initial tests have been promising, the full potential of these models will be realized when they’re used outside of reanalysis data.
