kumaresh singh department of computer science, virginia tech
DESCRIPTION
Assimilating Tropospheric Emission Spectrometer profiles in GEOS-Chem. Kumaresh Singh Department of Computer Science, Virginia Tech Mohammed Jardak, Paul Eller, Adrian Sandu, Daven Henze, Kevin Bowmann, Dylan Jones, Changsub Shim. Motivation. - PowerPoint PPT PresentationTRANSCRIPT
![Page 1: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/1.jpg)
Kumaresh SinghDepartment of Computer Science, Virginia Tech
Mohammed Jardak, Paul Eller, Adrian Sandu, Daven Henze, Kevin Bowmann, Dylan Jones,
Changsub Shim
Assimilating Tropospheric Emission Spectrometer profiles in GEOS-Chem
![Page 2: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/2.jpg)
Motivation
Kumaresh Singh
Adjoint models are powerful tools widely used in meteorology and oceanography for applications
data assimilation model tuning sensitivity analysis determination of singular vectors
04/08/2009
![Page 3: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/3.jpg)
GEOS-Chem Adjoint (GCv7_ADJ)
Created an adjoint model of geos-4 v7 of GEOS-Chem Tested each science process adjoint separately Consistency check after integrating all processes together Completely parallelized adjoint code
Added 4-D variational data assimilation and sensitivity analysis capabilities
Provided with choices of operations to choose from as per the need, plug-n-play functions for cost function calculations
Adjoint code quite similar to forward mode – same coding convention
Kumaresh Singh04/08/2009
![Page 4: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/4.jpg)
Forward and Adjoint Code Flow
Kumaresh Singh04/08/2009
![Page 5: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/5.jpg)
Definitions
Sensitivity Analysis Adjoint model is efficient in calculating sensitivities of a few
output variable or metrics with respect to a large number of (input) parameters
Data AssimilationVariational data assimilation allows the optimal combination of three sources of information: an a priori (background), estimate of the state of the atmosphere, knowledge about the physical and chemical processes
04/08/2009 Kumaresh Singh
![Page 6: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/6.jpg)
Sensitivity Analysis (emission species)
Sensitivity of the O3 column measured by TES with respect to total NOx emissions over Asia on April 1st, 2001
Kumaresh Singh04/08/2009
![Page 7: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/7.jpg)
Sensitivity Analysis (tracer concentrations)
Kumaresh Singh04/08/2009
![Page 8: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/8.jpg)
4-D Variational Data Assimilation Framework At iteration 0, x0 = cp
0
At each subsequent iteration i (i≥1),
xi+1 ← L-BFGS (xi, f, g)
cop0 ← xi+1
(f, g) ← Reverse Mode (cop0)
where, f is the cost function and g is the gradient of the cost function. In our test case, the cost function and its gradient are defined as:
04/08/2009 Kumaresh Singh
N
kkkk
B
N
kkkk
Tkk
BTB
yHxRxxBxg
yHxRyHxxxBxxxJ
1
100
10
1
100
1000
)()()(
)()(2
1)()(
2
1)(
![Page 9: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/9.jpg)
4-D Variational Data Assimilation
A 4-day animation of plots from difference between background trajectory and analysis trajectory through TES profile retrievals for 2006 summertime GEOS-Chem data
Kumaresh Singh04/08/2009
![Page 10: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/10.jpg)
3-D Variational Data Assimilation Framework At iteration 0, x0 = cp
0
At each subsequent iteration i (i≥1),
xi+1 ← L-BFGS (xi, f, g)
cop0 ← xi+1
(f, g) ← Observation Operator and its adjoint (cop0)
where, f is the cost function and g is the gradient of the cost function. In our test case, the cost function and its gradient are defined as:
04/08/2009 Kumaresh Singh
00
100
10
001
00001
000
)(
2
1
2
1)(
yHxRxxBxg
yHxRyHxxxBxxxJ
B
TBTB
![Page 11: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/11.jpg)
3-D Variational Data Assimilation
04/08/2009
A plot of difference between background trajectory and analysis trajectory through TES profile retrievals for 2006 summertime GEOS-Chem data through 3-D variational data assimilation for 2 months with diagonal background error covariance matrix.
Kumaresh Singh
![Page 12: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/12.jpg)
Non-Diagonal Background Covariance Matrix
04/08/2009 Kumaresh Singh
![Page 13: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/13.jpg)
3-D Variational Data Assimilation
04/08/2009
A 3-week animation of plots from difference between model predictions trajectory and assimilated trajectory through TES profile retrievals for 2006 summertime GEOS-Chem data for 3-D variational data assimilation with diagonal(left) and non-diagonal (right) background covariance.
Kumaresh Singh
![Page 14: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/14.jpg)
Time Distribution of GC processes
(a) 1 cpu (12 min 46 sec) (b) 8 cpus (2 min 7 sec)
GEOS-Chem simulation times on one and on eight cores. The majority of the computational time is spent in three processes: chemistry, transport, and convection.
04/08/2009 Kumaresh Singh
![Page 15: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/15.jpg)
KPP Chemistry using CUDA
Ironical - Working on improving air quality by running model codes on power hungry architectures
Successful implementation of double precision CUDA chemistry codeAchieved 2x speed up as compared to the serial CPU version
Kumaresh Singh04/08/2009
![Page 16: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/16.jpg)
Conclusion
Designed and developed parsers to automatically interface KPP with GEOS-Chem v7
Developed an adjoint model of GEOS-Chem v7 Parallelized adjoint code completely Added 3-D and 4-D Variational data assimilation, and
sensitivity analysis capabilities Implemented non-diagonal background error covariance
matrix Added TES satellite observation operator and its adjoint Developed CUDA KPP chemistry for GPGPUs
Kumaresh Singh04/08/2009
![Page 17: Kumaresh Singh Department of Computer Science, Virginia Tech](https://reader035.vdocuments.mx/reader035/viewer/2022081520/568151a9550346895dbfd752/html5/thumbnails/17.jpg)
Questions