Ensemble Kalman Filter Data Assimilation for the MPAS system

Report
The 6th EnKF Workshop, May 18-22, 2014
Ensemble Kalman filter data assimilation for the MPAS system
So-Young Ha
National Center for Atmospheric Research
Collaborators
Chris Snyder, Bill Skamarock, Michael Duda, Laura Fowler in MMM/NCAR
Jeffrey Anderson, Nancy Collins, Tim Hoar in IMAGe/UCAR
Overview
System overview: MPAS, DART and the interface
Wind data assimilation strategy thru OSSE
Real data assimilation in MPAS/DART cycling
1.
2.
3.
1.
2.
3.
4.
5.
Sensitivity test on the quasi-uniform mesh
Comparison to CAM/DART on the uniform mesh
Comparison to WRF/DART on the variable mesh
Extended forecast verification
Summary and future plans
Based on unstructured centroidal Voronoi
(hexagonal) meshes using C-grid staggering and
selective grid refinement.
Jointly developed, primarily by NCAR and LANL/DOE
MPAS infrastructure - NCAR, LANL, others.
MPAS - Atmosphere (NCAR)
MPAS - Ocean (LANL)
MPAS - Ice, etc. (LANL and others)
MPAS-A development team: Bill Skamarock, Joe Klemp,
Michael Duda, Laura Fowler, Sang-Hun Park
A community facility for ensemble data assimilation
developed and maintained by the Data Assimilation
Research Section (DAReS) at NCAR
DART development team:
Jeff Anderson, Nancy Collins, Tim Hoar (IMAGe/UCAR)
MPAS-DART interface:
So-Young Ha and Chris Snyder (MMM/NCAR)
MPAS-Atmosphere
Unstructured spherical Centroidal Voronoi meshes
•
•
•
•
•
•
Mostly hexagons, some pentagons and 7-sided cells.
Cell centers are at cell center-of-mass.
Lines connecting cell centers intersect cell edges at right angles.
Lines connecting cell centers are bisected by cell edge.
Mesh generation uses a density function.
Uniform resolution – traditional icosahedral mesh.
C-grid staggering
Solve for normal velocities on cell edges.
Solvers
Fully compressible nonhydrostatic equations
Current Physics
•
•
•
•
•
Noah LSM, Monin-Obukhov surface layer
YSU PBL
WSM6 microphysics
Kain-Fritsch and Tiedtke cumulus parameterization
RRTMG and CAM longwave and shortwave radiation
Motivation: Global Mesh and Integration Options
Global Uniform Mesh
Global Variable Resolution Mesh
Voronoi meshes allows us to cleanly incorporate
both downscaling and upscaling effects (avoiding
the problems in traditional grid nesting) and to
assess the accuracy of the traditional downscaling
approaches used in regional climate and NWP
applications.
Regional Mesh - driven by
(1) previous global MPAS run
(no spatial interpolation needed!)
(2) other global model run
(3) analyses
MPAS/DART: Overview

Ensemble Kalman filter for MPAS


Broadly similar to WRF/DART






Implemented through the Data Assimilation Research Testbed (DART)
Interfaces with model, control scripting
Forward operators for conventional observations, AMVs, GPS RO data
Vertical localization in various vertical coordinates
Rejecting observations or reducing the weight of obs in the upper-level due to
the strong damping near the top in the model
Largely independent of model physics; facilitates testing with different
schemes during ongoing model development
New features in MPAS/DART



Use of unstructured grid meshes in the forward operator
Multiple options for updating winds
Interfaces to both MPAS-A and MPAS-O are available
MPAS/DART: Grid and Variables

Dual mesh of a Voronoi tessellation




All scalar fields and reconstructed winds
are defined at “cell” center locations (red
circles)
Normal wind speed (u) is defined at
“edge” locations (blue squares)
“Vertex” locations (green triangles) are
used in the searching algorithm for an
arbitrary observation point in the
observation operator
State vector in DART: Scalar
variables, plus horizontal velocity either reconstructed winds at cell
centers ( V ) or normal component on
the edges (u)
MPAS/DART-Atmosphere: Observation operators

Assimilation of scalar variables (x)
 finds a triangle with the closest cell center ( ) to a given
observation point ( )
 barycentric interpolation in the triangle
x2
A3
b
yobs
= (A1 x1 + A2 x2 + A3 x3 ) (A1 + A2 + A3 )
x3
A1
observation
A2
x1
The dua
a Delau
corners
points o
Fields in
•
“cell”
genera
•
“edge
points
interse
•
“verte
MPAS/DART-Atmosphere: Observation operators
Radial Basis
:
Function
Edge_wind
Cell_wind
Wind DA strategy: OSSE in MPAS/DART

Truth: 15-km quasi-uniform global mesh

Observation network:1,200 evenly distributed locations on the globe

Simulated observation types: sounding temperature, u- and v-wind, geopotential
height at 11 mandatory pressure levels and surface pressure

Observation errors: 1 K for T, 2 m/s for wind, 10 mb for Psfc, 25 – 450 m for Z

Ensemble filter data assimilation design: localization (H/V), adaptive inflation in
prior states, 6-hrly cycling for one month of August 2008.

Model configuration: 80-member ensemble at ~1-degree quasi-uniform mesh, 41
vertical levels w/ the model top at 30-km

WRF-Physics: WSM6 microphysics, YSU
PBL, NOAH LSM, Tiedtke cumulus,
RRTMG SW/LW radiation schemes
Two different wind DA options:
Cell_wind vs. Edge_wind

Assimilation of horizontal winds:
Sounding verification of 6-hr forecast (RAOB_U)
RMSE RADIOSONDE_U_WIND
Southern Hemisphere
RADIOSONDE_U_WIND
Northern Hemisphere
# of obs (o=poss, +=used)
# of obs (o=poss, +=used)
# of obs (o=poss, +=used)
30000
40000
13333.3
20000
14500
33333.3
22000
Cell_wind pr=2.99
Edge_wind pr=3.06
29500
# of obs (o=poss, +=used)
37000
14285.7
18571.4
22857.1
27142.9
31428.6
35714.3 40000
Cell_wind pr=2.25
Edge_wind pr=2.27
Cell_wind pr=1.87
Edge_wind pr=1.88
100
150
200
250
300
100
150
200
250
300
400
400
400
400
500
500
500
500
700
700
700
700
850
925
1000
2.5
850
850
925
1000
850
hPa
100
150
200
250
300
hPa
100
150
200
250
300
925
1000
3
3.5
4
2.6
RMSE
Tropics
2000022500250002750030000325003500037500
2.8
3
RMSE
Globe
3.2
51428.6
62857.1
74285.7
85714.3
97142.9
108571.4120000
Cell_wind pr=2.85
Edge_wind pr=2.95
925
1000
1.5
22000
Cell_wind pr=2.99
Edge_wind pr=3.06
2
2.5
SPREAD
Tropics
27000
32000
3
37000
1.5
40000
2
SPREAD
Globe
62222.2
Cell_wind pr=1.85
Edge_wind pr=2.00
84444.4
2.5
106666.7
Cell_wind pr=2.00
Edge_wind pr=2.06
100
150
200
250
300
100
150
200
250
300
100
150
200
250
300
400
400
400
400
500
500
500
500
700
700
700
700
850
925
1000
850
850
925
1000
850
925
1000
2.5
3
RMSE
3.5
2.5
3
RMSE
3.5
hPa
100
150
200
250
300
hPa
hPa
Cell_wind pr=3.14
Edge_wind pr=3.16
26666.7
hPa
20000
Southern Hemisphere
925
1000
1
1.5
2
SPREAD
2.5
1
1.5
2
SPREAD
Cell_wind shows slightly better fits to the sounding observations everywhere.
=> Cell_wind is default.
2.5
hPa
10000
hPa
Spread RADIOSONDE_U_WIND
RADIOSONDE_U_WIND
Northern Hemisphere
Assimilation of real observations in MPAS/DART

Model configuration: 80-member ensemble at ~2-degree uniform mesh, 41
vertical levels w/ the model top at 30-km

Conventional observations (NCEP PrepBUFR) + GPS RO

Ensemble filter data assimilation design: localization (1200H/4V), adaptive
inflation in prior state, 6-hrly cycling for one month of August 2008.

WRF-Physics: WSM6 microphysics, YSU PBL, NOAH LSM, Tiedtke
cumulus parameterization, CAM SW/LW radiation schemes
Sensitivity test

Filter design




Adaptive inflation: on and off
Localization radius: horizontal and vertical
Ensemble size
Model design


Grid resolutions: {1- vs. 2-degree} and {uniform vs. variable} mesh
Different physics parameterizations
Uniform
240km X1.10242
Variable
240-60km
X4.40962
60-15km
x4.535554
120km X1.40962
60km
X1.163842
15km
X1.2621442
Sensitivity test: Adaptive inflation (on and off)
Common
OBS
RMSE
-
SPRD
-
Adaptive covariance inflation (Anderson
2009) improved the 6-h forecast skills by up
to 15% throughout the period.
Without inflation, ensemble spread was
quickly reduced and remained small
rejecting more observations, which led to a
bad performance.
Adaptive inflation (cont’d)
Sensitivity test: Horizontal localization
Sensitivity test: Vertical localization
Larger localization, larger spread and
larger error.
Sensitivity test: Grid resolutions
1-deg vs. 2-deg
Uniform vs. variable mesh
• In quasi-uniform meshes, double the resolution increased the 6-h
forecast skill by ~5% (in a verification against common observations).
• A variable mesh with a 1:4 ratio reduces the grid resolution from 180km in the globe down to 45-km resolution over the CONUS.
• In the variable mesh, the fine-mesh area showed the better fits to the
observations.
Comparison w/ CAM/DART






CAM/DART run by Kevin Reader (IMAGe/NCAR)
CCSM4.0 on ~2-degree resolution w/ the model top at 3 mb
Assimilating same observations
Very similar filter configuration
Verification for the same month of August 2008 in observation
space
CAM/DART cycling continuously for 10 yrs starting in 2000;
MPAS/DART begins 1August 2008
Sounding verification: Comparison w/ CAM/DART

MPAS/DART poor initially (by construction), then improves over 2-3 d.

CAM/DART and MPAS/DART are broadly comparable and reliable.
GPSRO_REFRACTIVITY verification: 6-h forecast
RMSE
BIAS
MPAS showed
slightly larger
errors in the lower
atmosphere but
greatly improved
the bias in the
entire atmosphere.
Summary and future plans for MPAS/DART

The MPAS/DART interface is available with the full capability now, released
in the latest version of DART.

The analysis/forecast cycling was successfully tested assimilating real
observations for one month of August 2008.

MPAS/DART on the quasi-uniform mesh seems to be reliable and broadly
comparable to CAM/DART.

MPAS/DART on the variable mesh is promising compared to the quasiuniform mesh, showing a positive impact of higher resolution grids.

The performance skill of MPAS can be further improved by more physics
options such as GFS or CAM physics.

MPAS/DART will be available in CESM for coupled models soon.

More tests will be done for a longer period on the higher resolution meshes
focusing on the direct comparison of quasi-uniform and variable meshes on
the simulation of regional-scale features, eventually compared to WRF/DART
on the mesoscales.
Current status


MPAS Version 2.0 was released on 15 Nov 2013 (for both
MPAS-Atmosphere and MPAS-Ocean core)
http://mpas-dev.github.io/
The latest official release (e.g., the “Lanai” version) of DART
includes the MPAS-A and MPAS-O interfaces.
http://www.image.ucar.edu/DAReS/DART/Lanai_release.html
or contact [email protected] or [email protected]

similar documents