The predictability of deep convection in cloud-resolving simulations over land

Cloud-resolving models (CRMs) can be used to provide subgrid information for use in improving the representation of convection in large-scale models. However, it is shown here that, for a 24-hour simulation of convection over land, the total daily rainfall and instant rain rates can often be hard to predict. An ensemble of 2D CRM runs, which differ only in the initial white-noise temperature perturbations added to the lowest 200 m, show a large spread of results. The size of the spread is shown to be dependent on the nature and strength of the forcing with the most significant differences between ensemble members when only a small number of convective cells exist in the domain. Further sensitivity experiments show that stronger forcing or increased domain size can reduce the spread of results due to an increased number of clouds. However, in all the simulations described in this paper, convection tended to organize into fewer but larger clouds during the day and, as this occurs, the ensemble members began to diverge even with a domain size of 1000 km.