Quantifying the oscillatory evolution of simulated boundary-layer cloud fields using Gaussian process regression

Abstract. Average properties of the cloud field, such as cloud size distribution and cloud fraction, have previously been observed to evolve periodically. Identifying this behaviour, however, remains difficult due to the intrinsic variability within the boundary-layer cloud field. We apply a Gaussian process (GP) machine-learning model to the regression of the oscillatory behaviour in the statistical distributions of individual cloud properties. Individual cloud samples are retrieved from high-resolution large-eddy simulation, and the cloud size distribution is modelled based on a power-law fit. We construct the time series for the slope of the cloud size distribution b, a slope that is consistent with satellite observations of marine boundary-layer clouds, by observing the changes in the slope of the modelled cloud size distribution. Then, we build a GP model based on prior assumptions about the cloud field following observational studies: a boundary-layer cloud field goes through a phase of relatively strong convection where large clouds dominate, followed by a phase of relatively weak convection where precipitation leads to formation of cold pools and suppression of convective growth. The GP model successfully identifies the oscillatory behaviour from the noisy time series, within a period of 95±3.2 min. Furthermore, we examine the time series of cloud fraction fc and average vertical mass flux M‾, whose periods were 93±2.5 and 93±3.7 min, respectively. The oscillations reveal the role of precipitation in governing convective activities through recharge–discharge cycles.