Developing cloud parameterizations for large-scale models that incorporate sub-grid scale variability: Guidance from cloud resolving models and techniques for radiative transfer

Abstract

A new generation of cloud parameterizations for large-scale models is under development. These schemes account for the resolution-dependent sub-grid scale variability in cloud water and ice, which we hypothesize is a significant factor driving ad hoc model tuning.

The development and evaluation of these schemes requires detailed knowledge of the four-dimensional structure of clouds and precipitation. This information is not available from observations alone, so cloud resolving model output is particularly attractive. We use simulations by small-scale, high resolution (512 km domain, 2 km grid size) cloud models to inform the development of a new statistical cloud scheme. We examine a one-month simulation of summertime deep convection at the ARM SGP site to determine how the distribution of water depends on the domain size, and to explore the vertical correlation of condensate. Our results suggest some of the ways the parameterization should behave: variance in total water should decrease with spatial resolution, and skewness is non-negligable at scales larger than about 64 km. The simulations can can also provide new approaches to parameterization: it appears that the correlation of cloud water and ice between two model levels decreases exponentially with layer separation, but that the length scale depends on the amount of convection.

If large-scale models produce complicated clouds in each grid cell they will require fast, flexible techniques for computing radiative fluxes. Existing techniques won't work, since they weave assumptions about cloud structure right into the fabric of the radiative transfer solver. We think that it's better to solve the right problem approximately than the wrong problem exactly, and offer a fast but approximate technique for computing fluxes in variably cloudy skies. The method, essentially a Monte Carlo integration of the Independent Column Approximation, introduces substantial random errors into each calculation, but this error does not affect forecast skill at any time scale.