Representing cloud sub-gridscale variability in large-scale models

Abstract

Clouds are variable across an enormous range of spatial and temporal scales. Large-scale models, such as those used to predict weather and climate, almost universally truncate this variability at the model time step and grid size. When processes depend non-linearly on cloud properties, this unresolved variability leads to systematic biases. One of the best-known is the plane-parallel albedo bias, which was first highlighted in a series of papers almost ten years ago. Those papers note that models must be "tuned" in physically unrealistic ways in order to achieve radiation balance.

In this talk I'll describe how we have removed the plane-parallel albedo bias and its attendant tuning from GFDL's climate model AM2. We represent subgrid-scale structure by randomly generating sub-columns in each large-scale grid column. Each sub-column is homogeneous, but taken in aggregate they reconstruct the grid-mean profiles of cloud fraction and liquid water content. We use these columns to apply general assumptions about the structure of clouds in the vertical, and to estimate the internal variability of cloud water and particle size in a physically realistic way. A new algorithm for computing domain-averaged fluxes (the Monte Carlo Independent Column Aproximation, or McICA) make calculations with these many sub-columns affordable.