Machine-Learning Based Analysis of Liquid Water Path Adjustments to Aerosol Perturbations in Marine Boundary Layer Clouds Using Satellite Observations

Changes in marine boundary layer cloud (MBLC) radiative properties in response to aerosol perturbations are largely responsible for uncertainties in future climate predictions. In particular, the relationship between the cloud droplet number concentration (N-d, a proxy for aerosol) and the cloud liquid water path (LWP) remains challenging to quantify from observations. In this study, satellite observations from multiple polar-orbiting platforms for 2006-2011 are used in combination with atmospheric reanalysis data in a regional machine learning model to predict changes in LWP in MBLCs in the Southeast Atlantic. The impact of predictor variables on the model output is analysed using Shapley values as a technique of explainable machine learning. Within the machine learning model, precipitation fraction, cloud top height, and N-d are identified as important cloud state predictors for LWP, with dynamical proxies and sea surface temperature (SST) being the most important environmental predictors. A positive nonlinear relationship between LWP and N-d is found, with a weaker sensitivity at high cloud droplet concentrations. This relationship is found to be dependent on other predictors in the model: N-d-LWP sensitivity is higher in precipitating clouds and decreases with increasing SSTs.

Automated summary

In this study, satellite observations and atmospheric reanalysis data are used to predict changes in cloud liquid water path (LWP) in the marine boundary layer clouds (MBLCs) in the Southeast Atlantic using a regional machine learning model. The study identifies precipitation fraction, cloud top height, and cloud droplet number concentration (N-d) as important cloud state predictors for LWP, while dynamical proxies and sea surface temperature (SST) are found to be the most important environmental predictors. The study also finds a positive nonlinear relationship between LWP and N-d, with a weaker sensitivity at high cloud droplet concentrations, and this relationship is dependent on other predictors in the model, such as precipitation and SST.