Data and Code for "Enhancing Deep Learning Cloud Parameterization to Improve Hybrid GCM Simulation"

This repository contains two components: 1. NN-cloud (in NN-cloud.tar.gz): A compact residual convolutional neural network that predicts cloud fraction, ice crystal number concentration, and cloud droplet number concentration from temperature, cloud water and ice contents, and surface pressure. The archive includes training and testing scripts, the trained model checkpoint, offline evaluation results, and plotting notebooks for reproducing Figures 1, 2, and S1 in the paper. The preprocessed SPCAM training dataset is hosted at https://portal.nersc.gov/cfs/m4035/spcam_cloud_full/ 2. HyCAM evaluation (in plot_HyCAM.tar.gz): Plotting notebook and global-mean total energy and water time series for SPCAM, HyCAM, NCAM, and CAM5. The multi-year climate simulation output (monthly-mean NetCDF files) for HyCAM, SPCAM, and NCAM, along with observational datasets (CERES-EBAF, CloudSat/CALIPSO, CMAP, ERA5, ERBE, IMERG, ISCCP, NVAP) used in Table 1, are hosted at https://portal.nersc.gov/cfs/m4035/HyCAM/ NN-cloud operates alongside the ResCu neural network (available at https://doi.org/10.5281/zenodo.17127771) to form HyCAM, a hybrid climate model in which all moist physics parameterizations in CAM5 are replaced by neural networks. HyCAM achieves a stable, non-drifting decade-long simulation with substantially improved cloud radiative forcing, atmospheric thermodynamic states, and precipitation compared to the previous hybrid model (NCAM) that used ResCu with conventional cloud parameterizations.