Resolving Clouds in Atmospheric Models

1
Resolving Clouds in Atmospheric Models
Bill Skamarock NCAR/MMM
2
(No Transcript)
3
Clouds in the Atmosphere
Weather Precipitation rain, snow, hail
Wind, radiation, visibility
Chemistry/Air-Quality Chemical processing
(acid rain) Ozone chemistry Transport of
pollutants Wet deposition
Climate Moisture redistribution and
precipitation hydrological cycle
Radiation
4
Representation of Clouds in Atmospheric Models
Large-scale models ?h gt 30 km

• The effects of the clouds are diagnosed
  (parameterized) from
• the predicted water vapor field
• precipitation
• vertical transport and redistribution of
  moisture and heat
• radiative effects
• turbulence

5
Representation of Clouds in Atmospheric Models
Meso-scale models 8 km lt ?x lt 30 km

• The effects of the clouds are partially prognosed
  from
• predicted fields water vapor, cloud water and
  ice, and frozen
• and liquid precipitation.
• Some portions of the cloud effects are still
  diagnosed (parameterized).
• some precipitation
• some vertical transport and redistribution of
  moisture and heat
• turbulence

6
Representation of Clouds in Atmospheric Models
Cloud-scale models 100 m lt ?x lt 8 km
The effects of the clouds are entirely prognosed
from predicted fields water vapor, cloud water
and ice, and frozen and liquid precipitation.
7
Problems with Modeled Clouds

• Large-scale models (clouds completely diagnosed)
• Poor diagnosis of cloud type, composition, and
  precipitation.
• Clouds and cloud-systems do not know about
  vertical wind shear.
• Implications
• (1) Large uncertainty in climate-model
  predictions
• (2) A key limiting factor for weather-forecast
  accuracy

8
Meso-/Cloud-Scale Model (WRF) Hurricane Katrina
Reflectivity at Landfall
29 Aug 2005 14 Z
Mobile AL Radar
4 km WRF, 62 h forecast
9
Realtime WRF 4 km BAMEX Forecast
12 h forecast Initialized 5/24/03 00Z
Composite NEXRAD Radar
Reflectivity Forecast
10
Realtime WRF 4 km BAMEX Forecast
12 h forecast Initialized 5/24/03 00Z
Composite NEXRAD Radar
Reflectivity Forecast
11
Vertical Velocity at z 5 km, t 5 h
Along-line cell spacing 6 to 8 Dh until Dh lt
500 m (cell diameter is 3 to 4 km in converged
solutions)
(Courtesy of G. Bryan, NCAR/MMM)
12
Simulations using ?x 4 km to ?x 250 m
?x 4000 m
?x 1000 m
Weak-shear case Vertical cross-section of
tracer concentration at 6 h (not a line-average).
?x 250 m
(Courtesy of G. Bryan, NCAR/MMM)
13
Surface rain rate, weak shear

• 250 m solution close to convergence
• 1, 2, 4 km solutions over-predict precipitation.

(Courtesy of G. Bryan, NCAR/MMM)
14
Problems with Cloud Models

• Solutions do not statistically converge until
• ?h lt O(100 m) - turbulence problem

When will our applications get there? (assume
comp. speed doubles every 18 months)
Climate - not in my lifetime Weather - global
(state-of-the-art ?h 25 km) 36 years
(maybe in my lifetime) Weather - regional
(state-of-the-art ?h 7 km)
19 years (hopefully in my lifetime but
will I be retired?)
15
Cloud Models

• Cloud models solve the 3D Euler equations and
  transport equations for water vapor and
  liquid/solid water species with subgrid models
  for turbulence and other models
  (parameterizations) for everything else (moisture
  phase changes, radiation, land-surface,
  ocean-surface, etc.)
• Generally speaking, there are 2 flavors
• (1) Semi-Implicit (implicit treatment of acoustic
  and gravity waves)
• usually found in global models on lat-long
  grids pole problem.
• (2) Explicit (explicit treatment of acoustic and
  gravity waves)
• some form of splitting is usually used to
  advance acoustic and
• gravity waves with a shorter timestep.

16
WRF-ARW

• Terrain-following hydrostatic pressure vertical
  coordinate
• Arakawa C-grid
• 3rd order Runge-Kutta split-explicit time
  integration
• Conserves mass, momentum, entropy, and scalars
  using flux form prognostic equations
• 5th order upwind or 6th order centered
  differencing for advection
• Limited area (not global)

(more info - http//www.mmm.ucar.edu/wrf/users/)
17
Why Explicit

• Explicit time integration with splitting is more
  efficient than implicit solvers (operations for a
  given level of accuracy).
• Solver needs little tuning for application at
  different grid resolutions and problem sizes.
• Easily parallelized for SM, DM and SM/DM
  architectures.

18
Time Integration in ARW
3rd Order Runge-Kutta time integration
advance
Amplification factor
19
Time-Split Runge-Kutta Integration Scheme
dt is the RK3 timestep
acoustic timestep (in this case dt/4)
20
Time-Split Runge-Kutta Integration Scheme
In DM applications A small amount of data is
communicated within each acoustic step.
21
Time-Split Runge-Kutta Integration Scheme
In DM applications A small amount of data is
communicated within each acoustic step. A larger
amount is data is communicated after each RK
substep.
22
Parallelism in WRF Multi-level Decomposition
Logical domain
1 Patch, divided into multiple tiles

• Single version of code for efficient execution
  on
• Distributed-memory
• Shared-memory
• Clusters of SMPs
• Vector and microprocessors

Inter-processor communication

• Model domains are decomposed for parallelism on
  two-levels
• Patch section of model domain allocated to a
  distributed memory node
• Tile section of a patch allocated to a
  shared-memory processor within a node this is
  also the scope of a model layer subroutine.
• Distributed memory parallelism is over patches
  shared memory parallelism is over tiles within
  patches

23
WRF Software Framework Overview

• Implementation of WRF Architecture
• Hierarchical organization
• Multiple dynamical cores
• Plug compatible physics
• Abstract interfaces (APIs) to external packages
• Performance-portable

24
Courtesy of J. Michalakes see http//box.mmm.ucar
.edu/wrf/WG2/bench/ for more info
25
Petascale Computing and Clouds

• Many effects of clouds on climate and weather are
  largely unknown/uncertain (observations lacking,
  models at coarse resolution have poor
  representation of clouds). Most important
  problem confronting dynamicists and modelers
  today.
• Cloud-resolving (Dh O(100 m)) simulations of
  cloud systems are needed to understand cloud
  dynamics and to improve parameterizations - a
  petascale computing challenge.

cloud- mixing eddies
cloud systems
planetary waves synoptic systems
clouds
meters to 100s meters
102 - 104 meters
105 - 106 meters
gt106 meters
26
Petascale Computing and Clouds

• Split-explicit cloud models are easiest to scale
  to peta-computing - no global data exchange or
  implicit solver needed, numerics are not scale
  dependent.
• We can scale our problems to bigger machines.
• Questions
• What will new machine architectures look like?
• Will we maintain efficiency with scaling and
  changes in machine architecture?
• What code architecture changes will be needed?
• Other problems load balancing, analysis, I/O.