examples of forward and inverse...
TRANSCRIPT
![Page 1: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/1.jpg)
Examples of forward and inverse simulations
M.Sofiev
Finnish Meteorological Institute
![Page 2: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/2.jpg)
Content
• Introduction
Forward and inverse problems of atmospheric dispersion
Data assimilation in atmospheric modelling
• Basic equations
Forward and adjoint diffusion equations
An optimization problem of data assimilation
• Materials and methods
Overview of SILAM dispersion model
Outline of the European Tracer Experiment ETEX
• Inverse problem solution for ETEX
Generation of the first guess with adjoint dispersion simulation
Iterative minimization of the cost function
Post-processing of emission fields
• Summary
![Page 3: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/3.jpg)
Forward dispersion problem
Chemical & physical reactionsTransportEmission
Diffusion in deep layersDiffusion in deep layers
LOADLOAD
![Page 4: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/4.jpg)
Inverse dispersion problem
Where’s it come from ??
![Page 5: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/5.jpg)
Forward diffusion equation
0( 0) ( ); ( ) ( , )b
C t C x C x C x t
( / )( ) ( )
.
i ii
i i i
C CLC U C K R C E
t x x x
advect diffusion sink source
![Page 6: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/6.jpg)
Forward diffusion equation: an exposure (load) functional
p is a weight (sensitivity) function of the load:
Population exposure
Ecosystem damage
Model-measurement comparison for K stations
T
dxghdthgspaceHilbert0
),(,
)(),( cityxxtxp
)()(),( ecosystemAtptxp
K
k
kk xxtxp1
)(),(
0
( , ) load functional
T
M dt C p dx C p
![Page 7: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/7.jpg)
Adjoint diffusion equation
* * * *
* *
( , ) ( , ) ( , ) ( , )
adjoint diffusion equation
C f C LC LC C p C
LC p
*( , ) ( , )M p C C f Load functional:
, * 0
*( ) 0
x
C C
C t T
* ( )i ii
i i i
L U K Rt x x x
![Page 8: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/8.jpg)
Forward problem:
Adjoint problem
Dual features of atmospheric dispersion problem
(1/ )( ) ; ; ( , )
i ii
i i i
L U K R LC E M p Ct x x x
* * * *( ) ; ; ( , )i ii
i i i
L U K R LC p M E Ct x x x
![Page 9: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/9.jpg)
Milestones
• Introduction
Forward and inverse problems of atmospheric dispersion
Data assimilation in atmospheric modelling
• Basic equations
Forward and adjoint diffusion equations
An optimization problem of data assimilation
• Materials and methods
Overview of SILAM dispersion model
Outline of the European Tracer Experiment ETEX
• Inverse problem solution for ETEX
Generation of the first guess with adjoint dispersion simulation
Iterative minimization of the cost function
Post-processing of emission fields
• Summary
![Page 10: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/10.jpg)
European Tracer Experiment ETEX
• Inert tracer release: 23-24.10.1994, north-western France
• ~160 observation stations over Europe
~90 hours sampling period
3-hour mean data
• ~30 numerical models, 6 sets of meteo data
• Over 20 statistical measures of model quality
![Page 11: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/11.jpg)
Plume evolution (SILAM)
![Page 12: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/12.jpg)
Concentrations at T+1 day
![Page 13: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/13.jpg)
Concentrations at T+2 days
![Page 14: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/14.jpg)
Milestones
• Introduction
Forward and inverse problems of atmospheric dispersion
Data assimilation in atmospheric modelling
• Basic equations
Forward and adjoint diffusion equations
An optimization problem of data assimilation
• Materials and methods
Overview of SILAM dispersion model
Outline of the European Tracer Experiment ETEX
• Inverse problem solution for ETEX
Generation of the first guess with adjoint dispersion simulation
Iterative minimization of the cost function
Post-processing of emission fields
• Summary
![Page 15: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/15.jpg)
Input and output of adjoint runs
• First-guess generation
Zero-source first guess modelled concentrations 0
Two kinds of measurements as input for adjoint runs
– Non-zero value => sensitivity distribution covers possible source area
– Zero value => sensitivity distribution covers “no-source” area
• Iterations
Three kinds of model-measurement relations in input to adjoint runs
– Model – measurement > 0 – over-estimation
– Model – measurement < 0 – under-estimation (equal to above non-zero case)
– Model = measurement = 0 – all-zero case, (equal to above zero case)
• A way to combine in the adjoint runs output:
make two / three adjoint SILAM runs, one for each case;
subtract C* for zeroes and C* for over-estimation cases from C* for under-estimation cases
![Page 16: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/16.jpg)
Iterative cost function minimization
Source
f
SILAM
forward
SILAM
adjoint
Model-
measurements
comparison
knkn ~
sf fJf
J
1iter ~
SILAM
adjoint SILAM
adjoint fJ ~*
![Page 17: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/17.jpg)
True source:
(20 W, 48.050 N)
Release time:
23.10.1994 16:00 ->
24.10.1994 3:50
(duration ~12 hours)
Evolution of sensitivity distribution (1st guess)
![Page 18: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/18.jpg)
Emission from assimilation cycle 7 (before post-processing)
Map
Time evolution
2 3 4 5 6 7
0
200
400
600
800
1000
1200
1400
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
iss
ion
[m
g s
ec
-1] True src/10
iter 2
0
200
400
600
800
1000
1200
1400
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
issio
n [
mg
sec-1
]
True src/10
iter 2
iter 3
0
200
400
600
800
1000
1200
1400
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
issio
n [
mg
sec-1
]
True src/10
iter 2
iter 3
iter 4
0
200
400
600
800
1000
1200
1400
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
iss
ion
[m
g s
ec
-1]
True src/10
iter 2
iter 3
iter 4
iter 5
0
200
400
600
800
1000
1200
1400
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
iss
ion
[m
g s
ec
-1]
True src/10
iter 2
iter 3
iter 4
iter 5
iter 6
0
200
400
600
800
1000
1200
1400
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
iss
ion
[m
g s
ec
-1]
True src/10
iter 2
iter 3
iter 4
iter 5
iter 6
iter 7
![Page 19: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/19.jpg)
Comparison with observations. Cost function
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
1 2 3 4 5 6 7
iterations
Mean meas.
Mean model
Correl.coef
FMT
Mean RMSE
![Page 20: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/20.jpg)
Emission from assimilation cycle 7 (before post-processing)
Post-processing the emission
map (no feedback to
assimilation iterations!):
• low-pass filtering of high-
frequency noise
• cutting out the background
level of emission
![Page 21: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/21.jpg)
Final emission distribution
Map iteration:
Time evolution
2 3 4 5 6 7
0
100
200
300
400
500
600
700
800
900
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
iss
ion
[m
g s
ec
-1] True src/10
iter 2
0
100
200
300
400
500
600
700
800
900
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
iss
ion
[m
g s
ec
-1] True src/10
iter 2
iter 3
0
100
200
300
400
500
600
700
800
900
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
iss
ion
[m
g s
ec
-1]
True src/10
iter 2
iter 3
iter 4
0
100
200
300
400
500
600
700
800
900
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
iss
ion
[m
g s
ec
-1]
True src/10
iter 2
iter 3
iter 4
iter 5
0
100
200
300
400
500
600
700
800
900
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
iss
ion
[m
g s
ec
-1]
True src/10
iter 2
iter 3
iter 4
iter 5
iter 6
0
100
200
300
400
500
600
700
800
900
23.1
0.9
4
24.1
0.9
4
25.1
0.9
4
26.1
0.9
4
27.1
0.9
4
27.1
0.9
4
9 12 18 0 6 12 18 0 6 12 18 0 6 12 18 0 6 9
em
iss
ion
[m
g s
ec
-1]
True src/10
iter 2
iter 3
iter 4
iter 5
iter 6
iter 7
![Page 22: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/22.jpg)
Summary
• Adjoint formalism and data assimilation can be used for solving inverse dispersion problem
The constraint and resolving differential operator are linear
There exist convenient observation operator and cost function
Measurements are comparably easy and plentiful
• Probabilistic interpretation of footprint allows for inclusion of zero-reporting measurements into the simulations
• ETEX application confirmed the method strengths and highlighted necessity for a proper assessment of observation quality and representativeness
• A problem of the absolute emission rate is to be addressed
![Page 23: Examples of forward and inverse simulationssilam.fmi.fi/open_source/SILAM_school/day1_forward_inverse_20130… · Content Introduction ¾ Forward and inverse problems of atmospheric](https://reader034.vdocuments.mx/reader034/viewer/2022052009/601e0b43e8f0791520176d9e/html5/thumbnails/23.jpg)
Acknowledgements
• SILAM programming team
M.Salonoja (FMI, 1995-1999)
M.Ilvonen (VTT)
P.Siljamo (FMI)
I.Valkama (FMI)