AI based forecasting of Air Quality in Europe

This proposal develops an AI-driven, multi-pollutant model for operational regional air-quality forecasting over Europe, implemented within ECMWF’s Anemoi framework. The model employs a graph-neural-network architecture to emulate transport and chemistry while integrating externally supplied meteorology and sectoral emissions. Training uses 22 years of hindcast data and 11 years of validated analysis with assimilated observations. The operational target is a 0.1° (~10 km) limited-area model delivering calibrated 96-hour probabilistic forecasts for NO2, O3, PM2.5, PM10 and SO2. The project transfers recent advances in AI weather forecasting to air-quality modeling, enabling inference in seconds rather than hours and preserving assimilated observational information to reduce forecast drift. Anticipated impacts include improved timing and detection of regulatory exceedances to support decision-making by individuals and public health authorities. The implementation builds on an existing prototype that is containerised and based on PyTorch/Torch Lightning with ROCm support with scaling tests on LUMI-G demonstrating >92% efficiency up to sixteen nodes. Large-scale GPU resources are requested to complete full-scale training, hyperparameter exploration, and preparation for operational deployment.