Introduction to AI-Powered Weather Forecasting
The European Center for Medium-Range Weather Forecasts (ECMWF) has recently launched an artificial intelligence (AI)-powered forecasting model, which outperforms state-of-the-art physics-based models by up to 20%. This new model, dubbed the Artificial Intelligence Forecasting System (AIFS), operates at faster speeds and requires significantly less energy to make forecasts, approximately 1,000 times less than traditional models.

Background on ECMWF and Weather Forecasting
The ECMWF, now in its 50th year of operation, has produced one of the world’s leading medium-range weather prediction models, known as ENS. Medium-range forecasting involves predicting weather patterns between three days and 15 days in advance, although ECMWF also provides forecasts up to a year ahead. These models are crucial for governments and local authorities to prepare for extreme weather events and for everyday needs, such as planning vacations.

Limitations of Traditional Weather Prediction Models
Traditional weather prediction models rely on solving physics equations, which are approximations of atmospheric dynamics. A significant limitation of these models is that they may not capture complex relationships and dynamics in weather patterns. In contrast, AI-driven models can learn these relationships directly from data, rather than relying solely on known equations.

Recent Developments in AI-Powered Weather Forecasting
The ECMWF’s announcement follows Google DeepMind’s introduction of the GenCast model, which outperformed the ECMWF’s leading weather prediction model, ENS, on 97.2% of targets across different weather variables. With lead times greater than 36 hours, GenCast was more accurate than ENS on 99.8% of targets. The European Center is also innovating, with the launch of AIFS-single being the first operational version of the system.

Future Directions and Potential
According to Florian Pappenberger, Director of Forecasts and Services at ECMWF, the AIFS and IFS models are complementary and will provide a range of products to users. The team plans to explore hybridizing data-driven and physics-based modeling to improve weather prediction precision. Matthew Chantry, Strategic Lead for Machine Learning at ECMWF, notes that physics-based models are essential to the current data-assimilation process, which is also vital for initializing machine learning models. The next frontier for machine learning weather forecasting is the data-assimilation step, which could enable a fully machine learning-based weather forecasting chain.

Emerging Research and Technologies
Chantry is a co-author of a study describing a data-driven, end-to-end forecast system called GraphDOP, which uses observable quantities to form a coherent latent representation of Earth System state dynamics and physical processes. This system can produce skilled predictions of relevant weather parameters up to five days into the future. Integrating artificial intelligence methods with physics-driven weather prediction modeling holds promise for more precise forecasting. While testing indicates that AI-powered forecasting can outperform historical models, further research is needed to determine the technology’s forecasting abilities when forced to go off-script.