tions of cloud cover/depth/thickness, precipitation, and top-of-atmosphere radiative fluxes.
Using clouds as a testbed, some promising new approaches to improving parameterizations are being explored, including perturbed parameter ensembles to explore the range of simulated climates possible by changing parameters within an individual climate model, uncertainty quantification to systematically optimize uncertain parameters, and stochastic parameterization. Traditional parameterizations give a single best-guess estimate of the aggregate effect of a subgrid process such as turbulence or clouds averaged over a grid cell. Stochastic parameterization instead provides a random plausible realization of that aggregate effect, drawn from an appropriate probability distribution function. A conventional parameterization of subgrid fractional cloud cover might specify it in terms of the grid-mean relative humidity, while a stochastic parameterization will randomly choose a cloud cover scattered around that deterministic value. This can help maintain grid-scale variability that conventional parameterizations may artificially damp. Stochastic parameterization has been successfully demonstrated in numerical weather prediction (e.g., Buizza et al., 1999; Palmer et al., 2009; Shutts and Palmer, 2007) and monthly to seasonal prediction (Weisheimer et al., 2011).
A nonstochastic parameterization of a random subgrid process such as cumulus convection cannot produce statistically robust results unless there are many cumulus clouds in each grid cell. As the spatial and temporal resolution in climate models is refined, this “scale-separation” assumption breaks down well before a single cumulus cloud is well resolved by the model grid, creating a “grey zone” in which neither the parameterization nor an explicit simulation of the process is theoretically justified. Many global weather prediction models are approaching that resolution for cumulus convection, and climate models are likely to do so within the next 20 years. Designing parameterizations that can function through this range of resolutions is an important challenge for the next decade. Stochastic parameterization may be a particularly useful strategy in the grey zone.
While a revolution in computational approaches or capabilities is not impossible, in simulating clouds and in the broader challenges of climate modeling, incremental improvements are more likely. Improvements are possible by tapping into model capabilities that already exist in some cases, through strategic cooperation of the sometimes disparate global and regional modeling streams, as well as increased cooperation of global, regional, research-based, and operational modeling efforts. Such improvements will involve unified, scale-invariant physical treatments of key processes, conservative coupling schemes, and, in some cases, two-way coupling.