Scribd
Upload
103 views21 pages

A 3D Ray Tracing Approach

1. The document presents research on using a projected spectral gradient (PSG) method to solve ray tracing and inverse problems. 2. Numerical experiments were conducted applying the PSG method to synthetic models, demonstrating its ability to efficiently find minimum travel times for ray tracing and optimize velocity models for tomography inversion. 3. The PSG method was shown to outperform other optimization techniques like conjugate gradient methods, requiring less storage and computation per iteration while achieving good precision quickly for large problems like full waveform inversion.

Uploaded by

Debora Cores Carrera
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
103 views21 pages

A 3D Ray Tracing Approach

1. The document presents research on using a projected spectral gradient (PSG) method to solve ray tracing and inverse problems. 2. Numerical experiments were conducted applying the PSG method to synthetic models, demonstrating its ability to efficiently find minimum travel times for ray tracing and optimize velocity models for tomography inversion. 3. The PSG method was shown to outperform other optimization techniques like conjugate gradient methods, requiring less storage and computation per iteration while achieving good precision quickly for large problems like full waveform inversion.

Uploaded by

Debora Cores Carrera
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
You are on page 1/ 21

Universidad Simn Bolvar

A 3-D Ray Tracing and Inverse


Problem Approach
Debora Cores Carrera
SOVG 2004
Noviembre 14-17, 2004

Universidad Simn Bolvar


OUTLINE
The Ray tracing problem (RT)
The Inverse problem approach (IP)
Brief historical overview
The optimization Solver
Adavantages of the solver
Numerical Results for RT and IP
Full waveform inversion
Conclusions

Universidad Simn Bolvar


The Ray Tracing Problem

Minimize TXX r
s

Xr

2 n 1
li
dl

v( x, y, z ) i 2 vi
Xs

v ( x , y, z ) is the group velocity and

along the ray


The number of
layers is given
by n

l2

dl is the differential

Universidad Simn Bolvar


Tomography Inverse problem

1
Minimize || Tobs T (v ) ||22
l vu 2

l 2j

l5j

l3j

l 4j

T (v ) (T1 (v ),...Tnr (v ))T

2 n 1

l ij
T j (v )
i 2 v i

Universidad Simn Bolvar


Brief Historical Overview
Ray Tracing Approaches
Solving Differential Equations
Especially in the 70s

Solving Optimization Problems


More recently

P.L. Jacson (1970)

Um and Thurber (1987)

H. Jacob (1970)

Prothero et al. (1988)

R.L. Wesson (1970-1971)

Mao and Stuard (1997)

Julian and Gubbins (1970-1971)

Cores et al. (2000)

Pereyra et al. (1980)

Universidad Simn Bolvar


Brief Historical Overview
Inverse tomography Approaches
Reconstruction Techniques
Dines and Lytle
(1979)

Damped Gauss Newton


Bishop et al. (1985)
Chiu et al. (1986)

Ivansson (1983)

Zhu and Brown (1987)

Lines and Treitel


(1984)

Farra and Madariaga


(1988)

Conjugate Gradient type methods


Pica et al. (1990)
Michelena et al. (1993)

Universidad Simn Bolvar


The Optimization Approach used for solving both Problems
The Projected Spectral Gradient (PSG)
Method (Raydan et al. (2000))
Considered a low cost
and storage technique as
any of the extensions of
conjugate gradient
methods (Polak-Ribiere,
Hestenes-Stiefel) for a
nonlinear optimization
problem.

Local Storage requirements


Few floating point operations
per iteration
Fast Local Convergence
Do not require to solve a
linear system of equation per
iteration

Universidad Simn Bolvar

1. Given
2. If || P( x

x0 n

, 0

g k ) xk || 0

3. Compute

, stop
and set

f ( xk j ) g kT d k

0 j min{k , M 1}

, then

k , xk 1 xk k d k , yk g k 1 g k , sk xk 1 xk

5.

l xu

and M 0

d k P ( xk k g k ) xk

4. If f ( xk 1 ) max

f (x)

Min
s. t.

Projected Spectral Gradient (PSG)


Method

go to step 5

skT sk
k 1 T
sk y k

n
Where: P is the projection on {x / l x u} and

g k f ( xk )

Universidad Simn Bolvar


Advantages of the Optimization Approach
1.

The projection over is simple and has low computational


cost

2.

The objective function is not monotonicaly decreasing because


of step lenght and the non monotone line search (step 4).
Implying less function evaluations to converge from any initial
point (Global convergence).

3.

The step size k is not the classical choice for the steepest
descent method. It speeds up the convergence of the PSG
method.

4.

The PSG method is related to the Quasi Newton methods. It can


be view as a two point method.

5.

The PSG method is competitive and many times out performs


the extensions of CG methods (CONMIN and PR+)

6.

The method converge to the global minimun if we have an


stratified and dipped model with constant velocity between
layers

Universidad Simn Bolvar


Numerical Results for Ray Tracing

5 layer synthetic model where P-S converted waves velocities are considered

Universidad Simn Bolvar


Numerical Results for Ray Tracing
1.

157 recievers and 3


sources randomly
genereted at the
surface.

2.

The average CPU


time for 1 shot is 3 s
(from different initial
rays).

3.

Convergence to the
global minimum is
obtained.

5 layer synthetic model where P-S converted wave velocities are considered

Universidad Simn Bolvar


Numerical Results for Ray Tracing
1.

157 recievers and 5 sources


randomly generated at the
surface.

2.

Lateral heterogeneous
model :
v( x, y ) ax by c,
a (0,1.7,1.5,1.3,0.8,0.8,1.3,1.5,1.7)T ,
b (0,1,1,1,1,1,1,1,1)T ,
c (0,800,700,500,150,150,500,700,800)T

3.

We can not guarantee


convergence to the global
minumum.

4.

The average CPU time for


the first shot was 50 s (from
different initial rays).

4 layer synthetic lateral heterogeneous model of complex stratigraphy

Universidad Simn Bolvar


Numerical Results for Ray Tracing
We consider a 5 layer ellipsoidal anisotropic medium,where the velocities are
given by the formula:
(xi' ) 2
(yi' ) 2
1 1 ( zi' ) 2

,
vi li ((v z , j ) i ) 2 ((v[ x , z ], j ) i ) 2 ((v[ x , y ], j ) i ) 2
xi' xi cos( i ) cos(i ) yi sin(i ) cos( i ) zi sin(i ),
yi' xi sin( i ) yi cos(i ),
zi' xi sin( i ) cos(i ) yi sin(i ) sin(i ) zi cos(i ),

Where i

and i denote the polar and azimuthal rotation angles in the

layer i, and j=P,SV,SH, i=1,2,...,2n+1


If the medium is an stratified or dipped model, the approach converges to a
global minimum

Universidad Simn Bolvar


Numerical Results for Ray Tracing
157 receivers at the surface and 1
source in the origen.

(v z , p )i 1500 100 * i m / s,
(v[ x , z ], p ) i 1200 50 * i m / s,
(v[ x , y ], p ) i 1350 80 * i m / s,
(v z , s ) i 1400 100 * (n 3 i ) m / s,
(v[ x , z ], s ) i 1000 50 * (n 3 i ) m / s,
(v[ y , z ], s ) i 1150 80 * (n 3 i ) m / s
for i=2,...,n+1

5 layer synthetic ellipsoidal anisotropic medium

Universidad Simn Bolvar


Numerical Results for the tomography inversion

Universidad Simn Bolvar


Numerical Results for the tomography inversion

Universidad Simn Bolvar


Numerical Results for the tomography inversion

We fixed CPU time and the


grid size (500x500) to observe
the reduction in the gradient
and the residual during that
period of time

Universidad Simn Bolvar


Numerical Results for the tomography inversion

We used a (20x20)
grid size to measure

Real velocities

Initial velocities

The initial velocities have an error of 50% from the real velocities

the precision of PR+


and PSG

Final velocities (PSG)

Final velocities (PR+)

The quality of the solution by the 2 methods are almost the same

Universidad Simn Bolvar


Numerical Results for the tomography inversion
1.

SIRT has low computational cost per iteration but requires too many
iterations and therefore consumes more CPU time.

2.

PSG, PR+ and CONMIN reach quickly a good precision (10e-03) when
compared to SIRT and Gauss Newton methods.

3.

Gauss Newton is fast, in CPU time, for very small size of the grid.

4.

The PSG and PR+ methods outperform CONMIN for very large
problems.

5.

The PSG method is always slightly faster , in CPU time, than PR+.

Universidad Simn Bolvar


Conclusions

1. The PSG method is a simple, global and fast method for


large scale problems (Example: inversion and ray tracing).
2. The PSG method reachs quickly to a good precision (For
example 10e-02 or 10e-03).
3. The PSG method only requires firts order information.
4. The PSG method does not require exhastive line search
which implies less function evaluations per iteration.
5. We also used the method for Full waveform inversion,
obtaining very good results.

Universidad Simn Bolvar


Full waveform inversion (for Modified Marmousi model)

You might also like