The European Center for Medium-Range Weather Forecasts (ECMWF) has introduced an AI-driven forecasting model that reportedly surpasses state-of-the-art physics-based models by up to 20%. Named the Artificial Intelligence Forecasting System (AIFS), the model operates at higher speeds than traditional physics-based models and requires approximately 1,000 times less energy for generating forecasts, as per ECMWF’s announcement.
Celebrating its 50th year in operation, ECMWF is renowned for producing ENS, a leading medium-range weather prediction model. Medium-range forecasting involves predictions made between three to 15 days ahead, but ECMWF also extends its forecasts up to a year in advance. Weather prediction models are critical for governmental bodies and local authorities to prepare for extreme weather conditions, as well as for routine needs such as planning vacations.
Traditional forecasting models depend on solving physics equations, which provide approximations of atmospheric dynamics. In contrast, AI-driven models hold the potential to learn complex relationships and dynamics in weather patterns directly from data, without relying solely on pre-existing equations.
ECMWF’s announcement follows the release of Google DeepMind’s GenCast model for AI-driven weather prediction, which incorporates NeuralGCM and GraphCast. GenCast has demonstrated superior performance against ECMWF’s ENS on 97.2% of targets across various weather variables, achieving greater accuracy than ENS on 99.8% of targets when the lead time exceeded 36 hours.
ECMWF is also advancing in innovation with the launch of AIFS-single, marking the first operational version of the system. Florian Pappenberger, Director of Forecasts and Services at ECMWF, emphasized the significant effort required to ensure stability and reliability in model operations. Currently, AIFS offers a lower resolution than ECMWF’s IFS model, which achieves a 9 km resolution using a physics-based approach.
Pappenberger views AIFS and IFS as complementary elements within a broader array of products offered to ECMWF’s user community, allowing them to choose what best suits their needs. ECMWF plans to explore a hybrid approach combining data-driven and physics-based modeling to enhance weather prediction accuracy.
Matthew Chantry, Strategic Lead for Machine Learning at ECMWF, highlighted the crucial role of physics-based models in the current data-assimilation process, which initializes machine learning models for daily forecasts. He acknowledged that advancing data-assimilation represents a frontier for machine learning weather forecasting, potentially enabling a fully machine learning-based forecasting chain.
Chantry co-authored a study under peer review that introduces an end-to-end forecast system called GraphDOP. This system leverages observable data, such as brightness temperatures from polar orbiters, to create a coherent latent representation of Earth’s system state dynamics and physical processes. It is capable of making skilled predictions of relevant weather parameters up to five days in advance.
The integration of AI methods with physics-driven weather models holds promise for more precise forecasting. Initial tests suggest AI-powered models may outperform legacy models, though they have relied on reanalysis data thus far. Field observations have been crucial for training these models, and it remains to be seen how effective the AI technology will be when operating independently of such data.