Extensions and Nonlinear Inverse Scattering - Rice Universitytrip.rice.edu/downloads/iprpi.pdf ·...

Extensions and Nonlinear InverseScattering

William W. Symes

IPRPI, April 2004

Outline

1. The acoustic model of reflection seismology

2. Linearization

3. Why Least Squares doesn’t work

4. Extensions

5. Annihilators

6. Etc.

7. References

1

1. The Acoustic Model of Reflection

Seismology

2

Marine Acquisition90%+ of all datacollected worldwide

3

Data parameters: timet, source locationxs, and receiver locationxr, (vector)halfoffseth = xr−xs

2, scalar half offseth = |h|. Experiment =shot, single experiment

data =shot record.

4

Typical Marine Record

0

1

2

3

4

5

time

(s)

-4 -3 -2 -1offset (km)

Shot record, Gulf of Mexico (thanks: Exxon)

5

Mechanical Characteristics of Sedimentary Rock

1000 1200 1400 1600 1800 2000 2200 2400 2600 2800 3000500

1000

1500

2000

2500

3000

3500

4000

4500

depth (m)

Well logs from North Sea borehole. Top curve: compressionalwave velocity (m/s);middle curve: density (kg/m3); bottom curve: shear wave velocity (m/s). (thanks:Mobil R&D, Viking Graben)

6

Constant Density Acoustic Model

acoustic potentialu(x, t) related to pressurep and particle velocityv by

p =∂u

∂t, v =

1

ρ∇u

Second order wave equation for potential(

1

c(x)2∂2

∂t2−∇2

)

u(x, t) = w(t)δ(x − xs)

plus initial, boundary conditions. RHS models localized energy source, “no lowfrequencies” -many wavelengthsbetween source and target.Useful idealization:w(t) = δ(t).

Forward map:F [c] ≡ p|Y , whereY = {(t,xr,xs) : 0 ≤ t ≤ T, ...} is acquisitionmanifold.

7

2. Linearization

8

Nonlinear inverse scattering

Inverse problem: givend ∈ L2(Y ) find c ∈ C s. t.F [c] ' d.

Many difficulties:

• What is C?

• What is'?

• If ' means “close inL2”, could pose asleast squaresproblem: findc ∈ C tominimize‖F [c] − d‖2.

• Results of numerical experimentation mixed.

• Theoretical foundation inadequate - few results re relevant properties ofF .

9

(Partly) linearized inverse scattering

Formally,F [v(1 + r)] ' F [v] +F [v]r whereF [·] is linearized forward mapdefinedby

(

1

v(x)2∂2

∂t2−∇2

)

δu(x, t) = 2r(x)

v2(x)

∂2u

∂t2(x, t)

F [v]r = δp|Y

• basis of most practical data processing procedures.

• Beylkin (1985) and many others: good understanding of thelinear mapr 7→

F [v]r and the associatedlinear inverse problem forr givenv;

• v is no more known thanr, inverse problem for[v, r] still nonlinear!

10

Linearization error

Critical question: If there is any justiceF [v]r = directional derivativeDF [v][vr]

of F - but in what sense? Physical intuition, numerical simulation, and not nearlyenough mathematics: linearization error

F [v(1 + r)] − (F [v] + F [v]r)

• smallwhenv smooth,r rough or oscillatory on wavelength scale - well-separatedscales

• largewhenv not smooth and/orr not oscillatory - poorly separated scales

2D finite difference simulation: shot gathers with typical marine seismic geometry.Smooth (linear)v(x, z), oscillatory (random)r(x, z) depending only onz(“layeredmedium”). Source waveletw(t) = bandpass filter.

11

0

0.2

0.4

0.6

0.8

1.0

z (k

m)

0 0.5 1.0 1.5 2.0x (km)

1.6

1.8

2.0

2.2

2.4

,

0

0.2

0.4

0.6

0.8

t (s)

0 0.2 0.4 0.6 0.8 1.0x_r (km)

Left: Total velocityc = v(1 + r) with smooth (linear) backgroundv(x, z), oscilla-tory (random)r(x, z). Std dev ofr = 5%.Right: Simulated seismic response (F [v(1 + r)]), wavelet = bandpass filter 4-10-30-45 Hz. Simulator is (2,4) finite difference scheme.

12

0

0.2

0.4

0.6

0.8

1.0

z (k

m)

0 0.5 1.0 1.5 2.0x (km)

1.6

1.8

2.0

2.2

2.4

,

0

0.2

0.4

0.6

0.8

1.0z

(km

)

0 0.5 1.0 1.5 2.0x (km)

-0.10

-0.05

0

0.05

0.10

Model in previous slide as smooth background (left,v(x, z)) plus rough perturba-tion (right,r(x, z)).

13

0

0.2

0.4

0.6

0.8

t (s)

0 0.2 0.4 0.6 0.8 1.0x_r (km)

.

0

0.2

0.4

0.6

0.8

t (s)

0 0.2 0.4 0.6 0.8 1.0x_r (km)

Left: Simulated seismic response of smooth model (F [v]),Right: Simulated linearized response, rough perturbationof smooth model (F [v]r)

14

0

0.2

0.4

0.6

0.8

1.0

z (k

m)

0 0.5 1.0 1.5 2.0x (km)

1.4

1.6

1.8

2.0

2.2

2.4

,

Model in previous slide as rough background (left,v(x, z)) plus smooth 5% pertur-bation (r(x, z)).

15

0

0.2

0.4

0.6

0.8

t (s)

0 0.2 0.4 0.6 0.8 1.0x_r (km)

.

0

0.2

0.4

0.6

0.8

t (s)

0 0.2 0.4 0.6 0.8 1.0x_r (km)

Left: Simulated seismic response of rough model (F [v]),Right: Simulated linearized response, smooth perturbation of rough model (F [v]r)

16

0

0.2

0.4

0.6

0.8

t (s)

0 0.2 0.4 0.6 0.8 1.0x_r (km)

,

0

0.2

0.4

0.6

0.8

t (s)

0 0.2 0.4 0.6 0.8 1.0x_r (km)

Left: linearization error (F [v(1+ r)]−F [v]−F [v]r), rough perturbation of smoothbackgroundRight: linearization error, smooth perturbation of rough background (plotted withsame grey scale).

17

Implications

• Some geologies have well-separated scales - cf. sonic logs -linearization-basedmethods work well there. Other geologies do not - expect trouble!

• v smooth,r oscillatory⇒ F [v]r approximatesprimary reflection = result ofwave interacting with material heterogeneity only once (single scattering); errorconsists ofmultiple reflections, which are “not too large” ifr is “not too big”,and sometimes can be suppressed.

• v nonsmooth,r smooth⇒ error consists oftime shiftsin waves which are verylarge perturbations as waves are oscillatory.

No mathematical results are known which justify/explain these observations in anyrigorous way, except in 1D.

18

3. Why Least Squares doesn’t work

19

minc ‖F [c] − d‖2

• Problems are so large that iterative methods (variants of Newton) are only option(3D: millions of unknowns, billions of equations)⇒ can only find stationarypoints;

• For any choice of norm in domain,DF has very poor condition - very large,very small singular values (cf. examples);

• Poor approximation ofF by linearization⇒ poor approximation of least squaresfunction by quadratic;

• Observed behaviour isnonconvex⇒ many stationary points exist with largeresiduals.

• Same remarks apply (and are a bit easier to justify) forpartially linearized leastsquaresminv,r ‖F [v]r − (d − F [v])‖2.

20

The good news...

We actually know something aboutF [v], besides its representation whenw(t) =δ(t):

F [v]r(t,xr,xs) =∂2

∂t2

∫

dx

∫

dτ G(x,xr, t − τ )G(x,xs, τ )2r(x)

v2(x)

• F [v] : E ′(X) → D′(Y ) (X = Earth) is aFourier Integral Operatorassociated toacanonical relation(Lagrangian submanifold ofT ∗(X × Y )) (Rakesh, 1988);

• when canonical relation is graph, representation asGeneralized Radon Trans-form (Beylkin, 1985)⇒ many practical computations;

• when canonical relation is a graph (Beylkin 1985, Rakesh 1988) and sometimeseven when it isn’t (Smit, Verdel, tenKroode 1998, Nolan 1997, Stolk 2000),F [v]∗F [v] is pseudodifferential operator⇒ construction ofleft parametrixorapproximate microlocal inverse.

21

minr ‖F [v]r − (d −F [v])‖2, given v

0

500

1000

1500

2000

2500

Dep

th in

Met

ers

0 200 400 600 800 1000 1200 1400CDP

Approximate linear least squares solution apres Beylkin (“GRT inversion”), Mis-sissippi Canyon, Gulf of Mexico, 2D survey (750 MB, 500 shots). Thanks: Exxon.

22

4. Extensions

23

Extended models

Extensionof F [v] (akaextended model): manifoldX and mapsχ : E ′(X) → E ′(X),F [v] : E ′(X) → D′(Y ) so that

F [v]E ′(X) → D′(Y )

χ ↑ ↑ idE ′(X) → D′(Y )

F [v]

commutes, i.e.

F [v]χr = F [v]r

Extension is “invertible” iffF [v] has aright parametrixG[v], i.e. I − F [v]G[v]issmoothing, or more generally ifF [v]G[v] is pseudodifferential (“inverse except forwrong amplitudes”). Also require existence of a left inverseη for χ: ηχ = id.

NB: The trivial extension -X = X, F = F - is virtually never invertible.

24

Grand Example

The Standard Extended Model:

• X = X × H, H = offset range.

• χr(x,h) = r(x) (so r ∈ range ofχ ⇔ plots of r(·, ·, z,h) (“image gathers”)appearflat)

•

F [v]r(xr,xs, t) =∂2

∂t2

∫

dx

∫

dτ G(x,xr, t − τ )G(x,xs, τ )2r(x,h)

v2(x)

(recallh = (xr − xs)/2)

NB: F is “block diagonal” - family of operators (FIOs) parametrized byh.

25

Reformulation of inverse problem

Givend, find v so thatG[v]d ∈ the range ofχ.

Claim: if v is so chosen, then[v, r] solves partially linearized inverse problem withr = ηG[v]d.

Proof: Hypothesis means

G[v]d = χr

for somer (whence necessarilyr = ηG[v]d), so

d ' F [v]G[v]d = F [v]χr = F [v]r

Q. E. D.

26

Application: Migration Velocity Analysis

Membership in range ofχ is visually evident

⇒ industrial practice: adjust parameters ofv by hand(!) until visual characteristicsof R(χ) satisfied - “flatten the image gathers”.

For the Standard Extended Model, this means: untilG[v]d is independent ofh.

Practically: insist only thatF [v]G[v] be pseudodifferential, so adjustv until G[v]d

is “smooth” inh.

27

0.6

0.8

1.0

1.2

1.4

1.6

1.8

time

(s)

10 20 30 40 50channel

,

0

0.5

1.0

1.5

2.0

time

(s)

0.5 1.0 1.5 2.0 2.5offset (km)

Left: shot record (d) from North Sea survey (thanks: Shell Research), lightly pre-processed.Right: restriction ofG[v]dobs to x, y = const (function of depth, offset): shows rel.sm’ness inh (offset) for properly chosenv.

28

5. Annihilators

29

Automating the reformulation

SupposeW : E ′(X) → D′(Z) annihilates range ofχ:

χ WE ′(X) → E ′(X) → D′(Z) → 0

and moreoverW is bounded onL2(X). Then

J [v; d] =1

2‖WG[v]d‖2

minimizedwhen[v, ηG[v]d] solves partially linearized inverse problem.

Construction ofannihilatorof R(F [v]) (Guillemin, 1985 - cf deHoop’s talks):

d ∈ R(F [v]) ⇔ G[v]d ∈ R(χ) ⇔ WG[v]d = 0

30

Annihilators, annihilators everywhere...

For Standard Extended Model, several popular choices:

•

W = (I − ∆)−1

2∇h

(“differential semblance” - WWS, 1986)

•

W = I −1

|H|

∫

dh

(“stack power” - Toldi, 1985)

•

W = I − χF [v]†F [v]

⇒ minimizingJ [v, d] equivalent to least squares.

31

But not many are good for much...

Sinceproblem is huge, only W giving rise to differentiablev 7→ J [v, d] are useful -must be able to use Newton!!! Once again, idealizew(t) = δ(t).

Theorem (Stolk & WWS, 2003):v 7→ J [v, d] smooth⇔ W pseudodifferential.

i.e. onlydifferential semblancegives rise to smooth optimization problem, regard-less of source bandwidth.

NB: Least squares embedded in larger family of optimizationformulations, some(others) of which are tractable.

Numerical evidence using synthetic and field data: WWS et al., Chauris & Noble2001, Mulder & tenKroode 2002. deHoop et al. 2004.

32

6. Etc.

33

Invertible Extensions

• Beylkin (1985), Rakesh (1988): if‖∇2v‖C0 “not too big” (no caustics appear),then the Standard Extended Model is invertible.

• Nolan & WWS 1997, Stolk & WWS 2004: if‖∇2v‖C0 is too big (caustics,multipathing), Standard Extended Model isnot invertible! Not in any version -common offset, common source, common scattering angle,...

• Stolk & deHoop 2001:Claerbout extensionis invertible under much weakercondition (absence of turning rays).

• WWS, Stolk, Biondi 2003: generalized Claerbout extension to accommodateturning rays.

34

Beyond Born

Nonlinear effects not included in linearized model:multiple reflections. Conven-tional approach: treat ascoherent noise, attempt to eliminate - active area of re-search going back 40+ years, with recent important developments.

Why not model this “noise”?

Proposal:nonlinear extensionswith F [v]r replaced byF [c]. Create annihilators insame way (now also nonlinear), optimize differential semblance.

Nonlinear analog of Standard Extended Model appears to beinvertible - in factextended nonlinear inverse problem isunderdetermined.

Open problems: no theory. Also must determinew(t) (Delprat & Lailly 2003).

35

And so on...

• Elasticity: theory of Born inversion at smooth background in good shape (Beylkin& Burridge 1988, deHoop & Bleistein 1997). Theory of extensions, annihilators,differential semblance partially complete (Brandsberg-Dahl et al 2003).

• Anisotropy - see deHoop’s talk, this meeting.

• Anelasticity - in the sedimentary section,Q = 100 − 1000, lower in gassy sedi-ments and near surface. No results.

• Source determination - actually always an issue. Some success in casting as aninverse problem - Minkoff & WWS 1997, Routh et al SEG 2003.

• ...

36

Conclusion

• Least error formulation of (waveform) reflection seismic inverse problemin-tractable- very irregular with large residual stationary points⇒ no influence onpractice.

• Linearizedextended modelsprovide framework for both (industry standard) in-terpretive velocity analysis and automated techniques based on construction ofrange annihilators.

• Only (pseudo)differential annihilatorsyield smooth objective functions.

• Not all extensions suitable for use in “complex structure” (strong refraction).

• May be able to account for more nonlinearity (multiple reflections) via nonlinearextensions.

37

Thanks to...

National Science FoundationDepartment of Energy

Sponsors of The Rice Inversion ProjectVeritas, Schlumberger, and Seismic Consultants for graphics

Exxon, Mobil, and Shell for dataThe Organizers, for inviting me.

http://www.trip.caam.rice.edu

38

7. Selected References

39

Cohen, J. K. and Bleistein, N.: An inverse method for determining small variationsin propagation speed,SIAM J. Appl. Math. 32, 1977, pp. 784-799.

Rakesh: A linearized inverse problem for the wave equation,Comm. PDE 13, 1988,pp. 573-601.

Beylkin, G.: Imaging of discontinuities in the inverse scattering problem by inver-sion of a causal generalized Radon transform,J. Math. Phys 26, 1985, pp. 99-108.

Ten Kroode, A. P. E., Smit, D. J., and Verdel, A. R.: A microlocal analysis ofmigrationWave Motion 28, 1998, pp. 149-172.

40

deHoop, M. and Bleistein, N.: Generalized Radon Transform inversions for reflec-tivity in anisotropic elastic media,Inverse Problems 16, 1998, pp. 669-690.

Brandsberg-Dahl, S., De Hoop, M., and Ursin, B.: Focussing in dip and AVAcompensation on scattering angle/azimuth common image gathers,Geophysics 68,2003, pp. 232-254.

Stolk, C. and Symes, W. W.: Smooth objective functionals forseismic velocityinversion,Inverse Problems 19, 2003, pp. 73-89.

Stolk, C. and Symes, W. W.: Kinematic artifacts in prestack depth migration,Geo-physics 68, May-June 2004.

Minkoff, S. E. and Symes, W. W.: Full waveform inversion of marine reflectiondata in the plane wave domain,Geophysics 62, 1997, pp. 540-553.

41

Grau, G. and Versteeg, R.:The Marmousi Experience: Practical Aspects of Inver-sion, IFP-Technip, Paris, 1991.

Lions, J. L.:Nonhomogeneous Boundary Value Problems and Applications, SpringerVerlag, Berlin, 1972.

Operto, S., Xu, S. and Lambare, G.: Can we image quantitatively complex modelswith rays?,Geophysics 65, 2001, pp. 1223-1238.

Nolan, C. J. and Symes, W. W.: Global solution of a linearizedinverse problem forthe wave equation,Comm. PDE 22, 1997, pp. 919-952.

Stolk, C.: Microlocal analysis of a seismic linearized inverse problem,Wave Motion32, 2000, pp. 267-290.

42

Stolk, C. and deHoop, M.: Seismic inverse scattering in the wave equation ap-proach, MSRI preprint, 2001.

Hagedoorn, F.: A process of seismic reflection interpretation,Geophysical Prospect-ing, 1954, pp. 85-127.

Duistermaat, J.:Fourier Integral Operators, Birkhauser, NY, 1996.

Taylor, M.: Pseudodifferential Operators, Princeton University Press, 1981.

Hormander, L.:The Analysis of Linear Partial Differential Operators, vols. I-IV,Springer Verlag, 1983.

43

Nirenberg, L.: Lectures on Partial Differential Equations, CMBS Lecture Notes,AMS, 1972.

Treves, F.:Introduction To Pseudodifferential Operators And FourierIntegral Op-erators, Plenum, 1980.

Claerbout, J.:Basic Earth Imaging(“BEI”) and Imaging the Earth’s Interior(“IEI”),available through SEP web site,sepwww.stanford.edu.

Miller, D., Oristaglio,, M. and Beylkin, G.: A new slant on seismic imaging,Geo-physics 52, pp. 943-964, 1987.

Bleistein, N., Cohen, J. and Stockwell, J.:The Mathematics of Seismic Imaging,Migration, and Inversion, Springer, 2000.

44

Dobrin, M. and Savit, C.:Introduction to Geophysical Prospecting

Yilmaz, O.: Seismic Data Processing, 2nd edtion, SEG, 1999.

Symes, W.:Mathematics of Reflection Seismology, Lecture Notes,www.trip.caam.rice.edu1995.

Symes, W.: 2000 MGSS Notes,www.trip.caam.rice.edu.

45