Upload
others
View
6
Download
0
Embed Size (px)
Citation preview
Evolution of CESM Cloud Microphysics:Parameterization of Unified Microphysics
Across Scales (PUMAS)
A. Gettelman, H. Morrison (NCAR/MMM), K. Thayer-Calder (NCAR/CGD), T. Eidhammer, G. Thompson (NCAR/RAL), D. Barahona
(NASA/GSFC), C. Chen, C. McCluskey (NCAR/CGD), D. J. Gagne, J. Dennis (NCAR/CISL), C. Bardeen (NCAR/ACOM), R. Forbes (ECMWF)
Cloud Microphysics Kills!
• Clouds are responsible for most severe weather• Tornadoes, Thunderstorms, Hail,
Tropical Cyclones
• Critical processes depend on cloud microphysics (Thunderstorms, Hail, even Tornadoes)
Outline
• What is cloud microphysics• CAM Microphysics Evolution: PUMAS• Some current work• Unified Ice• CAM6 Perturbed Parameter Ensemble (PPE)• MG3 in IFS (ECMWF: Starting with OpenIFS SCM)
• Summary
What is Cloud Microphysics?
Dynamics
Unified Turbulence
Radiation
Aerosols
Clouds (Al), Condensate (qv, ql)
Mass, Number Conc
A, qv, ql, qi, qr qs,rei, rel, rer, reg
Surface Fluxes
Precipitation
Cloud fraction & Condensate: T, Adeep, Ash
A = cloud fraction, q=H2O, re=effective radius (size), T=temperature (i)ce, (l)iquid, (v)apor, (r)ain, (s)now
Finite Volume
4-ModeLiu, Ghan et al
2 MomentMorrison & GettelmanIce supersaturationPrognostic 2-moment Precip
Crystal/DropActivation
CLUBB
Sub-StepMicrophysicsZhang-McFarlane
Deep Convection
Community Atmosphere Model (CAM6)
CCN,IN
Still need Microphysics Regardless of scale…
Dynamics
Unified Turbulence
Radiation
Aerosols
Clouds (Al), Condensate (qc, qi, qg, qr)
A, qc, qi, qg, qr, qvrei, rel, rer, reg
Surface Fluxes
Precipitation
Clouds & Condensate: T, PDF(Cld)
A = cloud fraction, q=H2O, re=effective radius (size), T=temperature (i)ce, (l)iquid, (v)apor , (g)raupel (rimed ice)
Non-Hydrostatic
CLUBB
Microphysics
What if you ran CAM at 7km? 2km? 1km?
Mass, Number Conc
CCN,IN
Types of Microphysical Schemes
‘Explicit’ or BinRepresent the number of particles in each size ‘bin’Number and/or mass by bin for each ’class’ of hydrometeor. Computationally expensive, but ‘direct’
Predict the total mass (Usually) represent the size distribution with a function (fixed shape, size)Mass for different ‘Classes’ (Liquid, Ice, Precip)Computationally efficient
Predict mass and number giving more flexibilityRepresent the size distribution with a function (fixed or variable shape)Distribution function for different ‘Classes’ (Liquid, Ice, Precip)
Two Moment (Bulk)
One Moment (Bulk)
Lagrangian or“Superdroplet’
Follow Lagrangian trajectories of “super-droplets” Represent individual drop interactions with each other
CAM Microphysics Evolution
• CAM4: 1 moment (Rasch & Krisjansson 1998)• CAM5: 2 moment (MG: Morrison & Gettelman 2008)• CAM6: 2 moment+ Prognostic Precip (MG2: Gettelman & Morrison 2015)
• Convective microphysics available (Song and Zhang 2011)• CAM6.2: 2 moment + Rimed Ice (MG3: Gettelman et al 2019) [On CAM Trunk]• CAM6.2: Pulling microphysics from a separate repository (PUMAS)• Other efforts
• MG2 + Unified Ice (Eidhammer et al 2016)• Optimization and GPU directives added (CISL)
• MG2 or MG3 currently being used/available in versions of:• CAM5/6 derivative modes (e.g. E3SM & NorESM, others), GISS (Ackerman), GEOS (Barahona)• Ported to CCPP as part of NOAA testing• Also ported to ECMWF-IFS for comparisons (current work)• Made available to Climate Modeling Alliance (CliMA: Caltech effort)
PUMASPUMAS = Parameterization of Unified Microphysics Across Scales• State of the art bulk 2 moment microphysics for all scales of modeling
• LES to Mesoscale/Regional to GCM for Weather/Climate• Scalable and efficient for different models• Served from an external github repository: more modern software• Include unit tests• Interfaces for different models• Enable broader contributions• Incorporate performance improvements and GPU directives
• Based on MG3. • Not just M & G anymore!
• There are 5 NCAR labs + outside collaborators on the title slide
Current PUMAS Status
• GitHub repository established (https://github.com/ESCOMP/PUMAS)• Pulled into CAM as an external (micro_mg3_0.F90, micro_mg_utils.F90)
• Developing unit tests for code• New workflow facilitates development across model systems• Will work to have sample interfaces, CCPP port, etc. • Average CAM user doesn’t really have to care about this • Science efforts
• Unified Ice• Machine Learning• Efficiency (Vectorization & GPU directives)• Independent aerosol formulations (separate lightweight aerosols)• Ice Nucleation• Convert a version to CCPP interface
Unified Ice
• Similar to Predicted Particle Properties (P3: Morrison and Milbrant) for combining ice and snow. Single category of ice. Not rimed yet.• Based in Eidhammer et al (2016)• Putting it together with a separate rimed ice variable (in MG3)• Have it running globally. Trying to understand how to balance the
climate with a new ice/snow class combined. (Non-trivial)
Unified Ice results
• Unified ice code has smaller sizes than MG2 snow. Less ice mass, and lower net Cloud Radiative Forcing
• Still working on if this can be adjusted (‘tuned’) to match observations
• New code eliminates many uncertain parameters, e.g. ice autoconversion
• This has not been used in a global model except by Eidhammer et al 2016, with similar issues in CAM5
• Embarking on a Perturbed Parameter Ensemble (PPE) experiment
MG2 ControlPUMAS-Unified Ice
SWCRE
LWCRE
CAM6 Perturbed Parameter Ensemble (PPE)
• Goal: understand sensitivity of CAM6 to parameter changes. Also help tune new parameterizations• Method: develop a list of parameters and ranges, do a Latin
hypercube sampling of the parameters (3-5x as many sets as parameters). Run experiments and then develop an ‘emulator’ from the output. • Have some CPU time to do this. Starting now with SCAM tests• Will make ALL output available to the community.• Intent is to modify CAM6 physics: microphysics, aerosols, deep
convection, CLUBB
Parameters
• Have a list, let me know if interested• Is your favorite parameter here?• Hope to start SCAM testing in a
month or so• Simulations will be 1˚ standard
CAM6, AMIP and/or nudged• Will also look at forcing (PI
Aerosols) and feebacks (SST+4K)
MG3 in the ECMWF IFS (SCM)MICRE: 10 of 90 days over Macquarie Island (56˚S): Jan-Mar 2017
Cloud Fraction
With Richard Forbes, ECMWF
MG3 in the ECMWF IFS (SCM)MICRE: 10 of 90 days over Macquarie Island (56˚S): Jan-Mar 2017
Liquid (Supercooled)
MG3 in the ECMWF IFS (SCM)MICRE: 10 of 90 days over Macquarie Island (56˚S): Jan-Mar 2017
Total Ice = Ice + Snow
Other work (I know about)
• S. Ocean Cloud Studies (supercooled liquid) • Machine learning the warm rain process• Optimization and GPU Port of MG2 microphysics (CISL, Dennis Group)• CCPP version of MG3 from NOAA• Ice nucleation, especially in mixed phase (McCluskey)• CARMA sectional aerosols/cloud modeling with CAM
Summary
• MG2 and 3 (graupel) are currently available in CAM6• Microphysics is now pulled from an external repository• That means need to manage_externals to see all the code.
• Goal is developing flexible options in the microphysics• Use in different models• Microphysics across scales (LES or certainly mesoscale à ESM)• Different complexity available• Explore different science goals/approaches
• CAM6 Perturbed Parameter Ensemble (PPE) experiment starting now
Southern Ocean Cloud Studies (SOCRATES)
Gettelman et al 2020 (Submitted to JGR)
CAM6 simulations along flight tracks in the S. Ocean (Jan-Feb 2018)• Model size distributions compared to observations.• CAM6 does very well (this is a 100km global GCM v. in-situ aircraft)• Not enough supercooled liquid water (distribution not peaked enough)• Too much warm rain• Using this to play with the functional form of size distributions• Work ‘recommends’ we switch Autoconversion to Seifert & Behang (2001)
Machine LearningCan we do the warm rain process better?Replace autoconversion, accretion and self collection Use a stochastic collection kernel:
1. Break distributions of cloud and rain into bins2. Run stochastic collection kernel from a bin microphysics code3. Use altered distributions to estimate AUTO+ACCRE tendency
Results:Similar climate, but different process rate tendenciesRain formation process looks more like a bin model…different precip structureSame ACI as control modelDifferent S. Ocean cloud feedbacks!
So try to emulate it with a neural network. Mostly works: with Guardrails
Control Bin (Emulated)
Control =significant rain for all ReBin = Little rain for Re<15um (all LWP)
Machine Learning: Forcing and Feedback
ACI Is unchanged with Stochastic CollectionConclusion: Autoconversion/Accretion may not be the issue!
Cloud feedback reduced with Stochastic CollectionReduced DCldFraction and +FICE in present day
MG3 in the ECMWF IFS (SCM): 90 day summaryMICRE: 10 of 90 days over Macquarie Island (56˚S): Jan-Mar 2017
Total Ice = Ice + Snow
MG3 in the ECMWF IFS (SCM): 90 day summaryMICRE: 10 of 90 days over Macquarie Island (56˚S): Jan-Mar 2017
Liquid
SFC SW SFC LW
TOA SW TOA LW
(wm/2) MG3 refIFS
SFC SW Down 178 157
SFC LW Down 288 296
TOA SW 168 148
TOA LW -223 -204
MG3 in the ECMWF IFS (SCM): 90 day summaryMICRE: 10 of 90 days over Macquarie Island (56˚S): Jan-Mar 2017
Now Proceeding to 3D simulations in IFS
Radiative Fluxes