In this study, we develop an approach to correct biases in the atmospheric component of the Community Earth System Model (CESM) using convolutional neural networks (CNNs) to create a corrective model parameterization for online bias reduction. By learning to predict systematic nudging increments derived from a linear relaxation towards the ERA5 reanalysis, our method dynamically adjusts the model state, significantly outperforming traditional corrections based on climatological increments alone. Our results demonstrate substantial improvements in the root mean square error (RMSE) across all state variables, with precipitation biases over land reduced by 25-35\%, depending on the season. Beyond reducing climate biases, our approach enhances the representation of major modes of variability, including the North Atlantic Oscillation (NAO) and other key aspects of boreal winter variability. A particularly notable improvement is observed in the Madden-Julian Oscillation (MJO), where the CNN-corrected model successfully propagates the MJO across the maritime continent, a challenge for many current climate models. This advancement underscores the potential of using CNNs for real-time model correction, providing a robust framework for improving climate simulations. Our findings highlight the efficacy of integrating machine learning techniques with traditional dynamical models to enhance climate prediction accuracy and reliability. This hybrid approach offers a promising direction for future research and operational climate forecasting, bridging the gap between observed and simulated climate dynamics.

Traditional post-processing methods have relied on point-based applications that are unable to capture complex spatial precipitation error patterns. With novel ML methods using convolution to more effectively identify and reduce spatial biases, we propose a modified U-Net convolutional neural network (CNN) to post-process daily accumulated precipitation over the US west coast. For training, we leverage 34 years of deterministic Western Weather Research and Forecasting (West-WRF) reforecasts. On an unseen 4-year data set, the trained CNN yields a 12.9-15.9% reduction in root mean-square error (RMSE) over West-WRF for lead times of 1-4 days. Compared to an adapted Model Output Statistics baseline, the CNN reduced RMSE by 7.4-8.9% for all events. Effectively, the CNN adds more than a day of predictive skill when compared to West-WRF. The CNN outperforms the other methods also for the prediction of extreme events, highlighting a promising path forward for improving precipitation forecasts.