Investigating the Performance of CMIP6 Seasonal Precipitation Predictions and a Grid Based Model Heterogeneity Oriented Deep Learning Bias Correction Framework

Abstract Climate change is expected to alter the magnitude and spatiotemporal patterns of hydro‐climate variables such as precipitation, which has significant impacts on the ecosystem, human societies and water security. Global Climate Models are the major tools to simulate historical as well as future precipitation. However, due to imperfect model structures, parameters and boundary conditions, direct model outputs are subject to large uncertainty, which needs serious evaluation and bias correction before usage. In this study, seasonal precipitation predictions from 30 Coupled Model Inter‐comparison Project Phase 6 (CMIP6) models and Climate Research Unit observations are used to evaluate historical precipitation climatology in global continents during 1901–2014. A grid based model heterogeneity oriented Convolutional Neural Network (CNN) is proposed to correct the ensemble mean precipitation bias ratio. Besides, regression based Linear Scaling (LS), distribution based Quantile Mapping (QM) and spatial correlation CNN bias correction approaches are employed for comparison. Results of model performance evaluation indicate that generally precipitation prediction is more reliable in JJA than DJF on the global scale. Most models tend to have larger bias ratio for extreme precipitation. In addition, current CMIP6 models still have certain issues in accurate simulation of precipitation in mountainous regions and the regions affected by complex climate systems. Moreover, the proposed grid based model heterogeneity oriented CNN has better performance in ensemble mean bias correction than LS, QM, and spatial correlation CNN, which could consider the relative model performance and capture the features similar to actual climate dynamics.