Enhancing surface wind speed and temperature prediction using surface-layer emulator and transfer learning

Abstract Accurate prediction of surface wind speed and temperature is crucial for many sectors. Physical schemes in numerical weather prediction (NWP) and data-driven correction approaches have limitations due to uncertainties in parameterization and lack of robustness, respectively. This study introduces a physics-informed data model, PhyCorNet, which combines a deep-learning-based physics emulator (PhyNet) and a subsequent correction network (CorNet). PhyNet imitates the Revised Surface-Layer parameterization scheme (default option of the Weather Research and Forecast (WRF) model). CorNet refines predictions by mitigating the difference between PhyNet and observations. PhyCorNet enables gridded forecasts despite the training data being point-based. Therefore, PhyCorNet can be regarded as a re-diagnostic scheme for surface wind speed at 10 meters (WS 10 ) and temperature at 2 meters (T 2 ). Compared to WRF, it reduced diagnostic root-mean-square errors in the 24-hour forecasts for WS 10 and T 2 by 40% and 36%, respectively, across China, with unbiased forecasts at almost all sites. PhyCorNet addresses the over-smoothing prediction issue of other deep-learning models by providing the ability to represent fine-scale features and perform well in statistically extreme samples. In grid cells without observations for training, PhyCorNet performed much better than WRF, demonstrating the zero-shot learning capability. This study implies that the emulator plus bias correction provided by PhyCorNet could be used as a simple but effective approach to improve the performance of other diagnostic quantities in NWP.