Quantification of aerosol effects on marine boundary layer clouds with machine learning

Abstract

<p>Satellite observations are used in regional machine learning models to quantify sensitivities of marine boundary-layer clouds (MBLC) to aerosol changes.</p><p>MBLCs make up a large part of the global cloud coverage as they are persistently present over more than 20% of the Earth’s oceans in the annual mean.They play an important role in Earth’s energy budget by reflecting solar radiation and interacting with thermal radiation from the surface, leading to a net cooling effect. Cloud properties and their radiative characteristics such as cloud albedo, horizontal and vertical extent, lifetime and precipitation susceptibility are dependent on environmental conditions. Aerosols in their role as condensation nuclei affect these cloud radiative properties through changes in the cloud droplet number concentration and subsequent cloud adjustments to this perturbation. However, the magnitude and sign of these effects remain among the largest uncertainties in future climate predictions.</p><p>In an effort to help improve these predictions a machine learning approach in combination with observational data is pursued:</p><p>Satellite observations from the collocated A-Train dataset (C3M) for 2006-2011 are used in combination with ECMWF atmospheric reanalysis data (ERA5) to train regional Gradient Boosting Regression Tree (GBRT) models to predict changes in key physical and radiative properties of MBLCs. The cloud droplet number concentration (N<sub>d</sub>) and the liquid water path (LWP) are simulated for the eastern subtropical oceans, which are characterised by a high annual coverage of MBLC due to the occurrence of semi-permanent stratocumulus sheets. Relative humidity above cloud, cloud top height and temperature below the cloud base and at the surface are identified as important predictors for both N<sub>d</sub> and LWP.  The impact of each predictor variable on the GBRT model's output is analysed using Shapley values as a method of explainable machine learning, providing novel sensitivity estimates that will improve process understanding and help constrain the parameterization of MBLC processes in Global Climate Models.</p>