Comparison between model and observational cloud fraction adjustment using explainable machine learning

This ongoing study uses machine learning to quantify and compare observation- and global climate model-based sensitivities of cloud fraction (CF) for marine boundary layer clouds (MBLCs) to atmospheric aerosols. In addition, differences in the meteorological influence on these sensitivities between the model and observation are examined. Aerosol-cloud interactions in MBLCs remain one of the most substantial sources of uncertainties in climate simulations. Recent studies have reported that climate forcing from an increase in low-level liquid cloud fraction due to aerosol perturbations may be dominant. However, the impact of ambient meteorological conditions on the aerosol influence on CF continues to pose challenges as their covariability and interactions obscure the quantification of the aerosol–CF relationship. We established a data-driven framework based on cloud droplet number concentration (Nd, as a proxy for aerosol) and CF retrieved from Moderate Resolution Imaging Spectroradiometer (MODIS) and meteorological parameters from the European Centre for Medium-Range Weather Forecasts Reanalysis v.5 (ERA5). The eXtreme Gradient Boosting (XGBoost) machine learning is applied to the daily collocated MODIS-ERA5 data (2011-2019) from 60°N to 60°S to predict CF with Nd and meteorological predictors. The Nd–CF sensitivity and its dependence on meteorological factors are quantified by SHapley Additive exPlanation (SHAP) values and SHAP interaction values. We found that both CF sensitivities and their interactions with meteorology derived from the SHAP approach exhibit distinct and coherent regional characteristics. The ongoing work is intended to implement an identical XGBoost-SHAP setup on outputs from the ICOsahedral Non-hydrostatic-Hamburg Aerosol Module (ICON–HAM) global atmospheric-aerosol model, and to compare the magnitudes and geographical patterns of the sensitivities and interactive effects derived from observations with those from ICON-HAM. Discrepancies may point to the physics parameterization schemes in ICON-HAM which may need further evaluation of their representativity with respect to relevant processes. This novel explainable machine learning framework can potentially provide insights into parameterization tuning and enhance our knowledge of the complex aerosol-cloud-climate system.