Search for a command to run...
Abstract Atmospheric blocking is a primary driver of high-impact weather extremes but remains challenging to forecast. Recent advances in machine learning (ML) have led to the development of data-driven weather prediction models. However, their capability in forecasting blocking has not been evaluated. This study evaluates the skill of four state-of-the-art ML models (Pangu-Weather, FuXi, GraphCast, and NeuralGCM) against the ECMWF High-Resolution Forecast System (HRES) in predicting winter blocking over the Northern Hemisphere (NH) during 2018–2022. All models are assessed under consistent experimental setups using ERA5 as the reference. FuXi yields the lowest root-mean-square error (RMSE) in 500-hPa geopotential height (Z500), but substantially underestimates blocking frequency, particularly in the Pacific. In contrast, NeuralGCM demonstrates the highest overall skill in blocking prediction, achieving 73.68% and 39.59% accuracy at 5- and 9-day lead times, respectively. It also outperforms other models in forecasting sector-specific and persistent blocking episodes, likely due to its physics-informed hybrid architecture. HRES remains competitive in Pacific blocking prediction, outperforming fully data-driven models (e.g. FuXi, Pangu-Weather) in that region. Case studies further reveal rapid degradation in forecast skill beyond day 5, particularly with earlier initializations. While ML models perform well in short-range forecasts, their long-range forecast skill varies markedly with architecture and region. These findings highlight the promise of physics-informed ML models for improving NH blocking prediction and highlight the need to incorporate key boundary conditions and enhance model interpretability in future developments.