Marine boundary-layer clouds play a critical role for the Earth’s energy balance, as they have a substantial cooling effect. Aerosols change the properties of these clouds, and thus changes in the composition and loading of atmospheric aerosols can modify this cooling potential. However, the aerosol-induced changes in marine boundary-layer clouds have proven difficult to constrain with observations, as many processes occur at the same time, and cloud systems can be buffered against changes in aerosol concentration via other atmospheric parameters. Global climate models cannot resolve these cloud processes in full detail and rely on uncertain parameterizations. Due to the deficiencies in observation analysis as well as models, uncertainties associated with aerosol-cloud interactions remain among the largest within climate science, hampering estimates of climate sensitivity.
The proposed research project addresses these challenges and will quantify the effects of aerosols on marine boundary-layer clouds using global satellite observation data, and use this as a basis to evaluate global climate model parameterizations. In recent studies, machine learning models have helped understand the aerosol-cloud-meteorology system, due to their capability of isolating aerosol effects from meteorological confounders. Within the proposed research project novel machine learning techniques capable of predicting multiple targets (parameters) at once will be applied to predict all relevant cloud properties at the same time, thereby specifically accounting for and quantifying buffers within the cloud system. Within the framework of the statistical models, process relationships will be analyzed using both global satellite-based data sets as well as global climate model output. In this way, a process-oriented evaluation of global climate model parameterizations concerning marine boundary-layer clouds will be accomplished that goes far beyond the comparison of observed and modeled climatological mean cloud fields. This will advance observation-based constraints of model parameterizations, and help quantify the effective radiative forcing due to aerosol-cloud interactions.

• In our latest paper (ACP, 2024), we use machine learning to quantify the sensitivity of marine low cloud cover (CLF) to droplet number concentration (Nd) as a proxy for aerosols (see figure). We find that CLF is particularly sensitive to Nd in the transition regions of stratocumulus to cumulus clouds. Using explainable machine learning techniques, we quantify the influence of large-scale environmental cloud controls on the Nd-CLF sensitivity, and find that in these transition regions, the Nd–CLF sensitivity is amplified by higher SSTs, potentially pointing to increases in Nd delaying this transition in these conditions.

• In ongoing work we apply the machine learning framework of our prior publication to outputs from the ICOsahedral Non-hydrostatic-Hamburg Aerosol Module (ICON–HAM) global atmospheric-aerosol model, 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.