Three‐Dimensional Full Variables Bias Correction of CMA‐GFS Analysis Field Using Spatial Convolutional Long Short‐Term Memory Network

Abstract Numerical weather prediction models are essential for modern meteorological forecasting, with accuracy critically hinging on the precision of initial analysis fields. Compared to the ERA5 global reanalysis data, the analysis fields produced by the China Meteorological Administration Global Forecast System (CMA‐GFS) exhibit notable biases in both atmospheric and surface variables, significantly impacting the forecast accuracy of CMA‐GFS. This paper introduces a deep learning model, termed spatial convolutional long short‐term memory network (SCL‐net), designed for correcting these biases in CMA‐GFS analysis fields. SCL‐net employs an image‐to‐image translation approach with a patch‐based processing pipeline, enabling an efficient correction of three‐dimensional biases in CMA‐GFS analysis fields while handling high‐resolution global data. Experimental results reveal that SCL‐net achieves a bias reduction of 20% or more for atmospheric variables such as geopotential, specific humidity, and temperature, as well as for the surface variable of sea level pressure. Notably, the improvements are particularly pronounced in the upper levels of the atmosphere for geopotential and specific humidity. Furthermore, the application of SCL‐net within the Pangu Weather Model has further validated the enhanced forecast skill following data correction, indicating its effectiveness in improving the medium‐range forecast capability of the analysis field. Specifically, for 500 hPa medium‐range forecasts, the average anomaly correlation coefficient (ACC) improved by 0.07, with the most significant improvement in geopotential, where the ACC increased by 0.1 on average, is twice that of the anomaly‐based numerical correction method, thus highlighting the substantial potential of the SCL‐net in bias correction research for analysis fields.