Vous êtes sur la page 1sur 1

S53A-4489

Chao Gao and Ved Lekic


Department of Geology, University of Maryland, College Park

TRADITIONAL INVERSION VS TBI WITH SURFACE WAVE DISPERSION


Transdimensional inversion with random starting model

PVelocities
Depth=4km
1200

10

800

600

600

400

400

200

200

1.5

Discontinuity
Probability Density

PVelocities
PVelocities

HZ Ratio

10

PPDF of Vs
True value of Vs

1000

800

SVelocities
SVelocities

1200
PPDF of Vs
True value of Vs

1000

10

Depth=4km

0.9

10

10

10

3.1

3.2

3.3

3.4

3.5

3.1

Vs (km/s)
Depth=10km

1200

1000

1000

800

800

600

600

400

400
200
3.15

3.2

3.25

3.3

3.35

3.4

3.45

0
3.1

3.5

3.15

3.2

3.25

Depth=60km

300
250

200

200

150

150

100

100

70

70

Vs (km/s)

Vp (km/s)

3.4

3.45

3.5

40

4.2

4.4

4.6

2000

2500

4.8

30

30

0.6
30

30

0.4

0.5

40

4.2

4.4

4.6

4.8

40

40

50

50

60

60

60

H
60

60

Horizontal Component

500

1000

1500

2000

2500

70

70

Vs (km/s)
Vs (km/s)

500

1000

1500

2000

2500

570

7
6

Vp (km/s)
Vp (km/s)

Frequency(Hz)

70
00

5000

10000

2
1000

15000

Correlation Coefficient

15 15

20 20

0.4 0.4

0.2 0.2

0.20.2

0.40.4

0.60.6

0.80.8
25 25
5 5

10 10

15 15

20 20

1 1

25 25

1200

1300

1400

2
1

0.2

0.8

0.8

0.3

0.3

0.6

0.6

0.4

0.4

0.4

0.4

0.2

0.2

0.5

0.6

0.5

0.6

0.7

0.7

0.8

0.8

0.9

0.9

0.2 0.2

0.4 0.4

0.6 0.6

0.8 0.8

0.2

0.2 0.3

0.3 0.4

0.4 0.5

Death'
Randomly'choose'one'
Voronoi'nucleus'and'
delete''

SVelocities

PVelocities

10

10

10

20

20

20

Depth (km)

30

PVelocities

30

30

10

40

40

50

50

60

60

70

40

70

50

30

60

Vs (km/s)

40
70
2500

3000

Vp (km/s)

3500

Density(kg/m3)

Proposed'model'
G'

Calculate'the'
likelihood'func4on'
p(dobs|G)'''

5000

60

Accepted'
model'1'

30

40

40

40

3000

3500

4000

4500

50

50

50

60

60

60

0.5

0.5

0.5 0.6

0.6 0.7

0.7 0.8

0.8 0.9

0.9

70

70

We compute the correlation matrix by


adding different realizations of noise
into synthetic waveforms and then
using our HZ ratio measurement
technique. The top middle panel is an
example of processing the noisy
waveform. After choosing the waveform
centered around the maximum
amplitude, we measure the HZ ratio by
inverting for the best-fit ellipse. As with
SWD, we find that the errors of HZ ratio
measurements are correlated.

(top left) Reference model in our


synthetic test (blue line). The model
is parameterized with a variable
Transdimensional joint inversion of SWD, HZ Ratio and Ps Receiver Function
number of Voronoi nuclei (red
squares) which define the seismic
PVelocities
SVelocities
Density
Discontinuity
SVelocities
Discontinuity
structure. The
vertical location of Density
nuclei define the geometry of layers
while Vs, Vp and density values are
represented by horizontal
positioning of nuclei. (bottom left)
Covariance matrix generated from
pre-event noise at station ANMO of
the GSN. (bottom) Random sample
of noise generated from our
covariance matrix; note the
dominance of long period energy,
similar to the microseism.
Randomly-generated noise
samples are used to determine the
data error covariance matrix
0

0 0

10

10

1010

10

10

10

20

20

2020

20

20

20

3030

30

30

30

4040

40

40

40

0.1

30

30

40

40

50

50

50
50

50

50

50

60

60

60
60

60

60

60

20

30
4700

LagTime(s)
Time(s)

40

50

4600

70
40

70

Vp (km/s)

70
2500

Vs (km/s)

3000

3500

Density(kg/m3)

70
70
4

6
7
100

Vp (km/s)

60

80

8
200

70

2000

4000

6000

8000 10000

Joint inversion of surface wave


dispersion and HZ ratio data
shows several improvements in
retrieving the 1D velocity structure
compared to inversion with SWD
only. The inclusion of HZ data
reduces the variance of the
posterior Vs distributions (top left
Vs structure), and increases their
Gaussianity (left). Due to its
sensitivity to Vp at shallow depths
(<20km), HZ ratio also enables
more accurate retrieval of Vp.

4900
20

Vp (km/s)

10

4800

A schematic overview of the reversible-jump Markov chain Monte Carlo algorithm used for
inversion. Every 100th accepted model is included in the ensemble solution to increase the
independence of sampled models.

30

0.05

50

Generate'synthe4c'
data'from'forward'
code''

30

FUTURE WORK

Density

20

Accepted'
model'NJ3'

4500

Gaussianity analysis of Vs at 2km

Frequency(Hz)
Frequency(Hz)

Depth (km)

Birth'
Randomly'choose'a'depth'
along'the'1D'model'and'
add'one'Voronoi'nucleus'

4000

Vs (km/s)

Correlation matrix of HZ ratio data error

Probability Density

Move'
Randomly'choose'one'
Voronoi'nucleus'and'
change'its'depth'along'
the'1D'model'

20

1100

Frequency(mHz)
Frequency(mHz)

Depth (km)

Change'

3500

Time(s)

We compute the correlation matrix


by measuring surface wave
dispersion from a synthetic
waveform contaminated by many
different realizations of noise. The
non-diagonal terms of the
correlation matrix (left) indicate that
data error of SWD measurements is
not uncorrelated as is typically
assumed and that the full
covariance matrix should be used in
SWD inversion.

Randomly'choose'one'
Voronoi'nucleus'and'
change'the'associated'
parameter'

20

Particle Motion of Selected Waveform

TRANSDIMENSIONAL BAYESIAN INVERSION WITH rjMCMC

Accepted'
model'NJ2'

20

Time(s)
Selected Waveform After Normalization

Particle motion of Rayleigh wave

Frequency(Hz)

Frequency(mHz)
Frequency(mHz)

The root-mean-squaredifference of misfit functions


suggest that the Markov chains
from different starting models
are indistinguishable,
suggesting that inversion
results are independent from
the starting model for surface
wave dispersion.

Accepted'
model'NJ1'

10

0.1

Correlation matrix of surface wave dispersion data error

10 10

Calculate'the'acceptance'
probability'by'
'a(G|G)=min[1,prior'
ra4o*likelihood'
ra4o*proposal'ra4o*|J|]'

3000

Vertical Component
Horizontal Component

70

0.6 0.6

Change'G'to'G'by'
randomly'selec4ng'
1'of'the'following'4'
possibili4es':'

10

Vertical Component

0.3

Measure HZ ratio
at each frequency

0.3
50

50

0.8 0.8

Accepted'
model'N'

10

0.2

0.2

If'rejected,'
iterate'with'
model'G''

4500

0.4

0.1

Comparison between traditional inversion (left


panels) and TBI inversion (right panels) of
posterior probability density function for Vs at
different depths (denoted with red bars)

Current'model'G'
with'likelihood'
func4on'p(dobs|G)'''

4000

0.2
4

Convergence test of surface wave dispersion from different starting models

If'accepted,'
save'G'into'
MC'and'update'
G'with'G'

3500

Time(s)
Envelope Function of Waveform

Direction of
propagation

40

Vs (km/s)

Markov chain

3000

0.4

50

Vs (km/s)

3.35

30

50

1500

0.2

PPDF of Vs
True value of Vs

350

250

20

Depth=60km

300

60

20

20
0.7

400
PPDF of Vs
True value of Vs

50

60

3.3

Vs (km/s)

400
350

1000

0.2

Depth (km)

1400

1200

Vs (km/s)

50

20

PPDF of Vs
True value of Vs

1600

1400

0
3.1

50

20

Depth (km)

PPDF of Vs
True value of Vs

40

500

0.4

3.5

Depth=10km

200

40

3.4

1800

1600

30

3.3

Vs (km/s)

1800

30

3.2

Time(s)
Narrow Band Filtered Waveform

Correlation Coefficient

Discontinuity

1
1.5

HZ Ratio

20

PVelocities

0.5

0.8

20

Horizontal Component

Rayleigh)Wave)

10

10

SVelocities

Vertical Component

0.5

Depth (km)

Transdimensional joint inversion with surface wave


dispersion and HZ Ratio

Synthethic Waveform contaminated with noise

SWD

SVelocities
0

Depth (km)

Due to their different and complementary


sensitivities to subsurface structure,
multiple seismic observables are
increasingly often combined to image the
Earths deep interior. In this project, we
use a reversible jump Markov chain Monte
Carlo (rjMCMC) algorithm to incorporate
complementary seismic observables
including surface wave dispersion, particle
motion ellipticity (HZ ratio), and receiver
functions into transdimensional, Bayesian
inversion for the profiles of shear velocity
(Vs), compressional velocity (Vp), and
density beneath a seismic station.
Compared to traditional approaches, a
Bayesian approach yields an ensemble of
models, making it possible to fully quantify
uncertainty and trade-offs between model
parameters to. Furthermore, to properly
account for the noise characteristics of the
different seismic observables, we compute
covariance matrices of data errors for each
data type by measuring the seismic
observables of interest from synthetic
waveforms contaminated by many different
realizations of noise. The non-diagonal
terms of the covariance matrices we obtain
for different seismic data types indicate
that typical assumptions of uncorrelated
data errors are not always justified.

Traditional inversion with fixed layers

SYNTHETIC TEST OF TBI WITH HZ RATIO

Frequency(Hz)

ABSTRACT

SWD+HZ Ratio

EAR-1352214

70
2
300

Vs (km/s)

70
2500

3000

Density(kg/m3)

3500

70

100

200

300

As can be seen from the inversion


results presented above, both SWD
and HZ Ratio provide a relatively
weak constraint on sharp velocity
changes in the 1D structure.
Therefore, a complementary
seismic observable is needed to
better constrain the 1D model. We
combine Ps receiver functions in
the joint inversion due to their
sensitivity to seismic discontinuities.
We define the misfit function using
three terms: 1. The amplitude
difference between the observed
and proposed RF; 2. The time-shift
needed to maximize the correlation
between the proposed and
observed RFs; 3. One minus the
maximum correlation coefficient
between the proposed and
observed RFs. Preliminary joint
inversion results suggest improved
recovery of seismic discontinuities.
Further work will be aimed at
improving the low acceptance rate
and resulting slow convergence of
the rjMCMC algorithm.

60
4500100

REFERENCES

70
4400120

80

90

100

4300140

10

20

30

40

50

Time(s)

60

Lag Time(s)

70

80

90

100

160

10

20

30

40

50

Time(s)

60

70

80

90

100

Sambridge, M., Bodin, T., Gallagher, K., & Tkali, H. (2013). Transdimensional inference in the geosciences. Philosophical Transactions of the Royal Society A:
Mathematical, Physical and Engineering Sciences, 371(1984), 20110547.
Bodin, T., & Sambridge, M. (2009). Seismic tomography with the reversible jump algorithm. Geophysical Journal International, 178(3), 1411-1436.
Tanimoto, T., & Rivera, L. (2008). The ZH ratio method for long-period seismic data: sensitivity kernels and observational techniques. Geophysical Journal
International, 172(1), 187-198.

Vous aimerez peut-être aussi