Using a Machine Learning Approach for Isolating Aerosol Effects on Cloud Droplet Number Concentration in Marine Stratocumulus Clouds

Abstract Understanding the complex relationship between aerosols and clouds is crucial for accurately estimating anthropogenic aerosol radiative forcing and its impact on weather and climate. However, quantifying the relationship between cloud droplet number concentration (CDNC) and cloud condensation nuclei (CCN) concentration remains challenging due to extensive heterogeneities in cloud properties, processes occurring during cloud lifecycle, and observational uncertainties. This study integrates satellite observations with a global climate model and employs advanced statistical techniques to improve estimates of cloud droplet susceptibility to aerosol perturbations, particularly focusing on the CCN‐CDNC slope. A key challenge in determining this relationship is that CDNC is influenced by factors beyond CCN alone, such as atmospheric dynamics and cloud microphysics. These additional factors introduce variability that complicates direct correlation between CCN and CDNC, making it difficult to ascertain their true relationship. In addition, traditional methods like Ordinary Least Squares regression can produce biased slope estimates due to uncertainties in both CCN and CDNC measurements. When applied to CCN‐CDNC data, advanced curve fitting methods, such as bivariate least squares, often yield slopes exceeding 1, deviating from expected physical behavior. Our machine learning analysis identifies updraft velocity, among other key predictors, as a major factor leading to this larger‐than‐one slope between CCN and CDNC. To address this, we utilized Elastic Net Regression to isolate the effect of changes in CCN concentration on CDNC. This method refines the slope estimates by accounting for factors affecting CDNC, resulting in a slope that better captures the susceptibility of CDNC to CCN changes.