The objective of this research is to better understand global low-cloud feedbacks and aerosol indirect effects by using innovative methods in global climate simulation to explicitly resolve the small (250-m) scale turbulent eddies that form boundary-layer clouds. This has been a decades-long parameterization challenge for global climate models, leading to a large spread of simulated low cloud feedbacks and climate sensitivity that are a primary source of uncertainty in climate projections for the 21st century. Even the most advanced global cloud resolving models do not exceed 1-10 km horizontal grid resolution. This does not resolve boundary layer eddies, so their key physics must still be approximated, producing comparable modeling uncertainties to conventional global climate models.
We plan to sidestep this problem using innovative software engineering tailored to modern coprocessor-accelerated petascale computing systems in order to explicitly resolve the important turbulent eddy scales (250 m in the horizontal and 25 m in the vertical). This can appear computationally intractable for a global model, due to the extremely high space-time resolution needed, especially around the sharp inversions that form at the top of many marine stratocumulus cloud layers. We envision two related approaches to break this computational barrier. Both use a global climate model with embedded domains of explicit convection (i.e. cloud superparameterization) but add novel computational strategies.
Our first step will be to radically enhance the computational capabilities of the SuperParameterized Community Atmosphere Model v.5 (SPCAM5) by implementing a new parallel scaffold to liberate it from a current scale limit of 2500 cores to instead bring over fifty thousand conventional computing processors to bear. Our next step adds a layer of software engineering to exploit new coprocessor technology. Graphical processor unit co-processors will be used to provide incremental acceleration of the computationally intensive double moment microphysics and scalar advection calculations in each embedded cloud resolving model. In sum this allows 25x more computing power to be brought to bear.
Our primary strategy, “ultraparameterization” is to then resolve both boundary-layer cloud processes and deep cumulus convection in a statistically satisfactory manner in each of SPCAM5’s > 10k embedded cloud resolving arrays. The computational overhead due to the necessary increase in interior resolution appears to be 640x but can be reduced to the resources available at GPU+petascale (25x) through a series of algorithm improvements including interior domain size minimization, efficiency gains in cloud resolving model radiative transfer, advection and subgrid turbulence improvements, and a novel time-acceleration scheme at the interior resolved scale. We will also explore a secondary strategy, “mini-LES” that is less computationally intensive - embedding two interior scale regimes in SPCAM5 – one resolving boundary layer cloud processes in an especially tiny array, alongside the typical large domain used to represent deep convection, using a switch function to blend both solutions.
The payoff is a new pair of modeling systems with a fundamentally more robust treatment of the interaction of low cloud processes with climate and aerosols. We will optimize their representation of low clouds against observations and apply them to make new estimates of global low cloud-climate feedbacks and aerosol indirect effects on climate change – for the first time with a global model that makes minimal assumptions about the relevant scales of turbulence and therefore may make more robust climate projections.
