Professional Documents
Culture Documents
'
Depth and Velocity Estimation in a Lateral
Heterogeneous Medium
by
Debora Cores
Bolvar
Universidad Simon
&
'
OUTLINE
Resumen
The Tomography Problem in Lateral Heterogeneous Medium (LHM)
Historical Overview
Discretized Problem
Numerical Approach
Preliminary Numerical Results
Conclusions
&
'
Resumen
In this work we present the depth and velocity estimation problem as a set of
non-linear constrained optimization problems, where few depth and velocity
parameters are needed to get the approximated estructure of the model. On
the other hand, to solve this optimization problem we use the recently
developed Spectral Projected Gradient Method (SPG), that allow us to handle
box constraints, which correspond to velocity bounds. This optimization
technique is a low optimization technique that only requieres first order
&
'
Minimizef (S ) = 12 kT T (S )k22
T : IR6 +6 ! IR travel time function,
T = (T1 (S ); T2 (S ); : : : ; T (S )) where,
r
T (S ) =
ns nr n
ns nr nn
Rayi
IR
V (x; y; z )
ns nr n
dl
real travel
time vector.
IR
&
6n3n
parameters in
LHM.
'
Historical Overview
In a 2D Medium
Gauss Newton Approach:
Levenberg and Marquardt method
using Gauss Seidel with Successive
Overrelaxation: T. Bishop et al, 1985
In a 3D Medium
T.
et al., 1985
&
Madariaga, 1988.
Low Storage Opt. Techniques:
Spectral Gradient Method: Castillo,
Cores and Raydan , 2000.
'
DISCRETIZED PROBLEM: Travel time function
The travel time function ray j reflecting in the layer k , can be represented as
the sum on i of the straight segments of the ray.
i;j;k (
S) =
P =2+1
k
li;j;k
vi;j;k
P2=2+1+2
n
li;j;k
k vi;j;k
where
i;j;k
&
i;j;k
px
(
a (x
i
i;j;k
i;j;k
x 1
+x 1
2
2 + (yi;j;k
;j;k )
;j;k )
b (y
i
i;j;k
y 1 )2 + (f
+y 1
)
+c
i
;j;k
;j;k
i;j;k
f 1
i
k = 1; : : : ; n, i = 2; : : : ; k + 1; 2n + 2 k; : : : ; 2n + 1 (the straight
segments of the ray) and j = 1; : : : ; ns nr n.
for
;j;k )
'
We only consider dip interfaces in the problem. Assume that for any interface
fi;j;k =
+ i1
where,
&
'
velocity and three parameters for the depth. So, if the problem has n layers
the number of model parametrs is 9n.
&
a = (a2 ; a3 ; : : : ; a2n+1 );
b = (b2 ; b3 ; : : : ; b2n+1 );
c = (c2 ; c3 ; : : : ; c2n+1 );
k = (k11 ; k12 ; k13 ; k21 ; k22 ; k23 ; : : : ; kn1 +1 ; kn2 +1 ; kn3 +1 ):
8
'
Min 12 kT T (S )k22 =
s:t: L S U
r
P =1 P =1 P =1 T
ns
nr
ri;j;k
S ))2
i;j;k (
where,
T : < ! <9 ,
S = (a; b; c; k) 2 <9 ,
a = (a2 ; a3 ; : : : ; a2 +1 );
b = (b2 ; b3 ; : : : ; b2 +1 );
c = (c2 ; c3 ; : : : ; c2 +1 );
k = (k11 ; k12 ; k13 ; k21 ; k22 ; k23 ; : : : ; k1 +1 ; k2 +1 ; k3 +1 ):
L = (l1 ; l2 ; : : : ; l9 );
U = (u1 ; u2 ; : : : ; u9 );
The Jacobian matrix, rT (S ), is sparse.
ns nr n
&
'
NUMERICAL APPROACH
MinimizekTr
T (S )k2
rf (S ) = rT (S )(T (S ) Tr )
Global convergence
Fast local convergence
Low computational cost and storage
Box constraints
&
10
'
NUMERICAL APROACH
For tracing rays in this medium we use the algorithm in A fast and global
two point low storage optimization technique for tracing rays in 2D and 3D
isotropic Media, Cores, Fung and Michelena, Journal of Apllied
Geophysics, vol. 45, 2000.
Minimize T (x; y; z )
&
Minimize
s:t:
k
Tr T (S )k22
2
LSU
11
'
NUMERICAL APPROACH
Spectral Projected Gradient Method (SPG)
Step 1: If
kP (S
rf (S
k ))
=
Step 2.2: Set S+ = P (S
rf (S ))
Step 2.3: If f (S+ ) max0
1 f (S ) +
(S+ S )
= , S +1 = S+ , W = S +1 S , y = rf (S +1 )
Step 2.1: Set
k;M
k
rf (S ) then
rf (S ), go
k
to Step 3.
&
else,
T
k
max
T
k
max
12
; maxf
min
ak
bk
gg
'
We consider a synthetic model with five dip layers and converted wave. We first fix the
real velocities and we estimate the depth parameters. Second, we fix the depth
parameters and estimate the P wave velocities and S wave velocities .
&
Synthetic Model
13
'
PRELIMINARY NUMERICAL RESULTS
Depth parameter estimation: Initial Model Vs. Final Model
&
14
'
PRELIMINARY NUMERICAL RESULTS
Velocity estimatimation: Initial Model vs. Final Model
&
15
'
CONCLUSIONS
The SPG method estimates depth parameters and velocities with good
prescision, few iterations and few function evaluations since it utilized a
nonmonotone line search.
The ray tracing takes most CPU time required for the inversion. However,
&
16