Current conceptual models underpinning parameterizations of the interaction between convection and the environment have relied on an assumption that slowly varying large-scale environment is in statistical equilibrium with a large number of small and short-lived clouds. Thus they fail to capture non-equilibrium transitions such as the diurnal cycle and formation of meso-scale convective systems. Informed by analysis of radar observations, cloud-permitting model simulations and theory, this work presents a probabilistic transition matrix model of cloud-cloud interactions and evolutions of the size distribution of convective cells and stratiform area. A machine learning algorithm is applied to cloud populations obtained from 12 years of C-band radar observations and equivalent model simulations at Darwin, Australia to construct the transition matrices. The model predicts the size distribution of cloud sizes and its diurnal cycle well. Implementation of the model in large-scale models as a non-equilibrium, stochastic alternative to traditional mass flux cumulus parameterization schemes will be discussed.