Modeling multiscale interactions in the climate system Christopher S. Bretherton Atmospheric Sciences and Applied Mathematics University of Washington 08.09.2017 Aqua Worldview
Motivation Weather and climate are intrinsically multiscale Some climate phenomena are dominated by large scales (jet streams, atmospheric tides, ocean gyres, deserts) Others are not (cumulus convection, gravity waves, mesoscale ocean eddies, ice streams) This talk: 1. What are some strategies and issues for models covering a large range of spatial scales? 2. Frontiers for high-resolution climate models that explicitly simulate small-scale atmospheric and ocean eddies.
Multiscale cloud patterns
Variability in a GCM grid column
Grey zones for cloud-controlling circulations - Grid spacing comparable to dominant eddy size - With coarser grid, should parameterize - With finer grid, should explicitly simulate Deep cumulonimbus convection 5 km Shallow boundary-layer cumulus convection 500 m Marine stratocumulus 500 m 25 m
Simulating convection, turbulence, clouds and eddies 1. Conventional AGCM; Δx ~100 km, 20-100 levels: Parameterize subgrid variability of turbulence, cumulus convection, clouds (uncertain, ad-hoc, subjective, efficient) OGCM: parameterize mixing from mesoscale ocean eddies. 2. Cloud-resolving model (CRM); Δx < 5 km: Simulate cumulus convection, parameterize boundary-layer turbulence 3. LES; Δx < 250 m, Δz < 100 m : Explicitly simulate cumulus and large turbulent eddies important for shallow clouds. 4. Eddy-permitting ocean model; Δx < 20 km. Explicit meddies. Computational tradeoff: Grid resolution vs. domain size vs. simulation length Multiyear global LES is computationally infeasible; will remain so until compute energy/(grid cell x time step) reduced by 10-3.
A palette of models for climate processes Mix and match to suit purpose Trade-offs between resolution and run length IPCC 2013 Fig. 7.8
Some high-resolution models ECCO2 data-assimilating ocean model (~18 km grid)
NICAM global CRM 3.5-14 km: Tomita et al. 2005 0.9 km (for 24 hr): Miyamoto et al. 2013 Many applications: MJO, tropical storms cloud feedbacks etc.
NGAqua 4 km L34 tropical channel aquaplanet CRM 20480x10240 km (Bretherton and Khairoutdinov 2015; Narenpitak et al. 2017) 50 days = 105 core hours (on 103 processors) Used to study convective organization and cloud feedbacks
Regional hi-res weather/climate models COSMO 4.19 on GPUs = 2 km ICON LES = 150 m >105 cores Leutwyler et al. 2016 Heinze et al. 2017
Large-domain CRM/LES Self-aggregation Muller and Held 2012 Mesoscale ShCu Giga-LES Khairoutdinov et al. 2009 Pockets of open cells Wang et al. 2010 Seifert et al. 2015
Multiscale modeling frameworks Superparameterization: Small 2D CRM in each GCM grid column, typically Δx = 4 km, 32 levels, for deep Cu Ultraparameterization: Ultrafine 2D CRM in each GCM grid column (Δx = 250 m, Δz = 20 m for z=0.5-2 km) Khairoutdinov & Randall 2001 0 8 km Parishani et al. 2017
Numerical convergence issues In grey zone (too coarse grid), results are sensitive to grid spacing and advection scheme and may not be trustworthy. Under-resolved Convergence testing and intercomparison of high-resolution models is essential to their credibility. Resolved shallow Cu Parishani et al. 2017 Sensitive to alternate advection scheme Insensitive to alternate advection scheme Narenpitak et al. 2017 in prep
Some climate applications of high-resolution models Complement conventional IPCC-type GCMs for problems sensitive to subgrid parameterization assumptions SST and tropical weather/climate variability and biases Clouds, aerosol/cloud interaction, cloud feedbacks, Precipitation: trends, extremes, orographic effects Improve process fidelity and understanding Benchmark datasets to improve coarser-resolution GCMs? Local downscaling Potential for more skillful seasonal forecasts?
Do hi-res models better simulate current climate? A mixed bag for large-scale non-orographic phenomena e. g. eddy-permitting ocean changes SST biases in CCSM3.5 coupled climate model but no clear overall improvement with 10 km ocean grid (HRC) vs. 100 km grid (LRC) Kirtman et al. 2012
Superparameterization improves diurnal cycle of rain but not mean tropical rainfall distribution
SP for tropical ensemble forecasting Ensemble of 10 day SP-IFS rainfall forecasts initialized 21 Oct. 2011 using different initial random noise in CRMs show reasonable spread Subramanian and Palmer 2017 JAMES
Hi-res model response to climate perturbations
SPCAM has much smaller aerosol indirect effects than CAM5 ΔSWCRE (PD-PI) -0.8 W m-2-1.8 W m-2 Larger ΔSWCRE in CAM5 due mainly to 3-fold larger dlwp/dccn
NICAM global cloud-resolving model cloud feedbacks (Tsushima et al. 2014 JAMES) Δx = 14 km for 90 days/δx = 7 km for 30 days; 40 vertical levels. With specified +2K SST increase, large tropical cirrus increase strong positive longwave cloud feedback, Results are sensitive to subgrid turbulence parameterization (MYNN), snow fall speed, Δx uncertainties in GCRM, too! 2 1
Cloud feedbacks: CGILS (10 day LES) Blossey et al. 2013
LES, CRM don t yet constrain GCM cloud feedback spread Bretherton et al. 2015
Extreme precipitation in NGAqua: +4K vs. CTL 30% Daily (20 km)2 block averages Extreme precipitation scales as Clausius-Clapeyron (7% K-1) Less extreme precipitation increases much less True in all latitude belts in NGAqua
Can hi-res simulations help subgrid parameterizations? High-resolution models have progressed faster than moist physics parameterizations in GCMs They are conceptually simpler than GCMs, because cloud properties and air velocity don t vary much within a grid cell, so a complicated model for their subgrid covariability is not needed. They must still parameterize smaller-scale processes (e. g. cloud droplets a few microns wide, complex ice crystals, turbulence, aerosols). Like other models, they are works in progress needing constant testing vs. observations. Still, high-resolution simulations provide realistic reference datasets for parameterizing subgrid cloud process variability
Original GCSS/GASS view of model hierarchy LES and CRM produce process understanding and benchmark statistics which model developers use in SCM framework to improve the fidelity of global weather and climate models. Randall et al. 2003 Dussen et al. 2013
The human bottleneck The flip side of GCSS/GASS has been that progress has been slow compared to the availability of new hi-res simulations and observations, because a few skilled humans concoct parameterizations and their brains are getting information overload.
How about machine learning? We have increasingly large and comprehensive training datasets from high-resolution simulations, if we trust them. A coarse-graining problem: With variables computed by the coarse grid model (e. g. temperature, moisture and wind profiles), use the fine-grid model to return needed quantities to the coarse-grid model (e. g. rainfall, vertical profiles of fractional cloud cover, turbulence, atmospheric heating and drying). Ideally, the needed quantities should include a stochastic part derived from the internal variability of the hi-res simulations conditioned on the coarse-grid variables. Could machine learning techniques help? Maybe - there are many challenges.
Some discussion questions 1. Should we put a lot more of our climate modeling effort (human and computational) into global and regional high-resolution models simulating a large range of length scales? 2. What aspects of microphysics, subgrid turbulence parameterization, and numerics most need improvement for use in such models? How do we know? 3. What kind of new observations or tests do we most need to strengthen LES and CRMs within the climate model palette? 4. Does machine learning have a role in climate modeling, e. g. in using high-resolution models to improve coarser-resolution models, or observationally-based model tuning?