Constraints on cloud fraction adjustment to aerosols using explainable machine learning

This study applies explainable machine learning to near-global daily satellite and reanalysis data. We quantify and analyse the sensitivities of cloud fraction (CLF) to aerosol changes and their dependence on meteorological parameters.

Aerosol-cloud interactions have prolonged influences on the Earth’s radiation budget but remain one of the most substantial uncertainties in the climate system. Marine boundary layer clouds (MBLCs) are particularly important since they cover a large portion of the Earth’s surface. One of the biggest challenges in quantifying aerosol-cloud interactions from observations lies in isolating the CLF adjustments due to aerosol perturbations from the covariability of local meteorology and quantifying the influences of meteorology on the aerosol-CLF relationship. In this study, 10 years (2011-2020) of near-global daily satellite cloud products are used in combination with reanalysis data of meteorological confounders. Using cloud droplet number concentration (CDNC) as a proxy for aerosol, MBLC CLF is predicted by region-specific gradient boosting machine learning models. By means of SHapley Additive exPlanation (SHAP) regression values, the predictions are explained by quantifying the sensitivities of CLF to the predictors. Furthermore, the meteorological influences on the CLF adjustments are analysed with SHAP interaction values to define an interaction index (IAI). Globally, the regional ML models are able to capture on average 32% and up to around 71% of the variability of CLF. Global patterns of CLF sensitivities show that CLF is positively associated with CDNC and lower tropospheric stability (LTS), strongest in low-cloud regions. Increased sea surface temperature (SST) on the other hand will lead to reduced CLF, probably by increasing the vertical moisture gradient. The CDNC-CLF sensitivities are especially strong in stratocumulus-to-cumulus transition regions. Negative CLF sensitivities to the u wind component at 700 hPa are found for most regions which may indicate an influence of facilitated turbulence at the cloud top. In terms of the interactive effects of meteorological parameters, a significant dependence of the CDNC-CLF relationship on LTS and SST is found in low-cloud regions, and the patterns coincide with sensitivities. Positive IAIs are shown globally for SST and LTS, indicating that the CDCN-CLF sensitivity is stronger with high SST/LTS values.

The ongoing work shows that CDNC-CLF sensitivity is positive globally after accounting for meteorological covariations. Globally, SST and LTS can influence the positive CDNC-CLF relationship significantly, which is especially the case in stratocumulus regions. Detailed investigations will be carried out for not only SST/LTS but also other predictors to dig out the physics and causality behind the statistics.