Professional Documents
Culture Documents
A fast ridge regression inversion technique has algorithms. Although these algorithms are con-
been devised for the interpretation of simple two- ceptually quite different, the resulting matrix equa-
dimensional resistivity and induced-polarization tions are very similar. The efficiency of either method
data. The program will determine the rectangular depends mainly on the scheme chosen for solving the
source under a single layer of overburden which best resultant large system of linear equations.
fits the observed data. Once the data bank has been created, it is possible
Several advantages are derived from using the to obtain inverse solutions for less cost than the
ridge regression method; they include convergence computation of one finite elemenl or transmission-
from very poor initial guesses, stability in the pres- surfaceforward problem. Tests on theoreticaldata and
ence of high-frequency geologic noise, readily ob- field data show the inversion technique to be reason-
tained estimates of parameter statistics, and the ably accurate, stable, and fast.
ability for simultaneous inversion of multiple data The statistics estimated by the inversion program
sets. Unfortunately, each ridge regression inversion provide additional useful information on the uncer-
requires a great many forward problem evaluations; tainty in the parameters of the derived model and on
thus in order to achieve speed and reasonablecost, it high correlations between parameters. The most
is essential to reduce the calculation time for the highly correlated parameters arc. as might be
forward problem to an absolute minimum. One anticipated, the resistivity and the width of thin
method of achieving this is to store in the computer conductive bodies. Two practical methods for carry-
a data bank containing solutions for the entire range ing out inversion in spite of highly correlated param-
of expected parameter combinations. The forward eters are, preferably, to add extra data sets which
problem then reduces to numerical interpolation provide more information on some of the parameters
between these precalculated data sets. or, alternatively, to fix some of the parameters at
For compilation of the data bank of forward solu- geologically reasonable values and invert to a more
tions, two main numerical methods were investi- restricted model.
gated: the finite element and transmission-surface
-a+
geometry. The purpose of this paper is to outline one
method for 2-D inversion which circumvents the ax
= -pJ,,
above two difficulties. No claim is made that it is an
-a+
Downloaded 05/03/13 to 132.203.235.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
av
Transmission surface algorithm - = -ZI,, (10)
dX
z, = $T’
Z
I
(18)
Downloaded 05/03/13 to 132.203.235.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
z was
problem
first applied
by Coggon
to the two-dimcn~ional
( I Y7 I ). In recent years enpincers
resistivity
L I I
FIG. 3. Plots showing typical behavior of the trans- direction (strike direction) to ohtain
formed potential as a function of the Fourier transform
variable X. Since the function is very smooth, only
seven calculated values are required to predict ac-
curately the behavior of the function over its entire A
I I I I
+ &1.<6(.~)6(;) dA = 0. (24)
1
- x’
computer programs arose when we used the finite
element routine to solve for the secondary potential.
-iY
Ia4 If we use the subscriptsr, p, and s to denote total,
primary, and secondary potential, respectively, we
may write
,,’ 01
(23)
or MODEL PARAMETERS
K4~s = -K,,+iu. (28)
CENTER
Downloaded 05/03/13 to 132.203.235.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
K,, are everywhere zero except for those associated DEPTH OVERBURDEN
4
with nodes in the inhomogeneity. Thus, in order to .R. ,BODY
solve for the secondary field we need only calculate THICK R HOST
.‘.:,@.., BQRY
4 h/Jwhere the inhomogeneity exists. +-WIDTH--( IP HOST
As a consequence.of solving for the secondary
potential, the 4hs shown in Figure 4 are much smaller FIG. 6. Diagram showing the nine parameters used
in magnitude and different in form from the I#Qshown in the inversion.
in Figure 3. Since the functions are no longer
monotonically decreasing, it is slightly more difficult
to obtain a highly accurate interpolation from a The fast forward routine
limited number of calculated points. As a result of the good agreement between the
There are perhaps two main numerical advantages two pseudosectionsshown in Figure 5, we were rea-
to solving for the secondary potential. The most sonably confident that no major programming errors
important of these is that a fine mesh is not required had been made and that the two routines produced
near each source, since the actual point sources are results of acceptable accuracy for our inversion re-
replaced by fictitious sources, quirements. The next step was to compile a data bank
S, = -Ko4hp, (29) of dipole-dipole pseudosectionsin order to implement
the fast forward routine desired for ridge regression.
distributed throughout the inhomogeneity. The sec- Our first task was to identify all parameters which
ond numerical advantage is realized when r#~*,~is are truly important in characterizing a resistivity and
small relative to 4hf. Any relative inaccuracy in the IP section. As shown in Figure 6, we arrived at nine
calculation of $J*,~will be diminished by the addition model parameters. Notably absent from these is the
of the analytically known (bhp to form 4*,. Of
course, when 4hp and 4i,q are of nearly the same
magnitude with opposite sign, the addition processis
Id_ ,
disadvantageousand a small percentage error in 4hr I I I
B /ago,
2
a logo,
(Seigel, 1959) where p,, is the apparent resistivity
c‘ d ’ I I I I I I
and p2 is the true resistivity of the body, is all that is
5 6 7 8 9 IO II
required in order to calculate the apparentpolarizabil-
CENTER ity anomaly caused by a polarizable body. To handle
FIG. 8. Typical variation in apparent resistivity for accurately host rock and overburden IP effects, this
changes in the horizontal location of the body. Each process should have been repeated for the host rock
curve relates to one pseudosectiondata point. resistivity and the overburden resistivity. We were
running into rather large storage requirements, how-
ever, and the host rock and overburden polarizability
dip of the target. It turns out that dip is not at all well usually provide rather small additional contributions
resolved in most dipole-dipole pseudosection data. to the total apparent polarizability. As a result we
However, even with this omission, we still ended calculated and stored only those dilution factors per-
up with enough different parameter combinations in taining to the target, and simulated the effect of a
compiling the data bank to keep our Univac 1108 “background polarizability” by the simple expedient
computer busy for 13 straight hours! of adding a small constantto the apparentpolarizabil-
Once we had stored 576 apparent resistivity ity pseudosection.
pseudosectionson file, we immediately converted all With this last modification the fast forward routine
values to logs. This conversion aided in two ways: was essentially complete. When nine model param-
(1) it became very easy to accommodate changes in eters are specified, the routine immediately locates
background resistivity by simple addition of a con- the nearest stored pseudosection, then by spline
stant, and (2) the interpolation in logarithmic space interpolation it determines the necessary corrections
became smooth and accurate. As illustrated by to both the dilution factor and log apparent resistivity
Figure 7, three typical pseudosection data points, in order to compensate for the slight difference be-
corresponding to the three curves, are well-behaved tween the input model parameters and those of the
functions of depth. Consequently, pseudosections nearest stored pseudosection. The corrected dilution
calculated for only four different values of depth are factor is then multiplied by the polarizability of the
all that are required in order to ensure that the cubic source and added to the background polarizability to
spline interpolating function accurately approxi- give the apparent polarizability. The log apparent
mates the true apparent resistivity versus depth resistivity is formulated first in terms of contrasts,
dependence. then the backgroundlog apparentresistivity is simply
The model parameter which gave the most diffi- added. The resulting forward problem is more than
culty in accurate interpolation was the one deter- 100 times faster than the transmission surface
mining horizontal position along the line. For the algorithm which was used to compile the data bank
particular case of a large resistivity contrast, shallow and, in addition, readily provides all the derivatives
depth, and no conductive overburden, a slight shift required by ridge regression inversion.
Resistivity and Induced Polarization 795
to”
7
Downloaded 05/03/13 to 132.203.235.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
IO2
7
5
3
2
IO’
7
5
R BODY
FIG. 9. Typical variation in apparentresistivity for changes in the resistivity of the body. Each curve relates to
one pseudosectiondata point.
In the case of simultaneousinversion of several data Since we always have more data than parameters,
sets, it is necessaryto obtain first a weighted deriva- the matrix A is overdetermined and cannot be inverted
tive matrix, readily in its present form; thus, we first multiply
both sides of (33) by AT, where T denotestranspose,
A=WB,
to obtain
and a weighted data difference vector, ATAAp = ATAy. (34)
Ay = WAg, (32) The matrix AT A is now square, symmetric, and posi-
796 Pelton et al
FIG. IO. Theoretical data obtained by the trans- where n is the number of data points and m is the
mission-surface algorithm were contaminated by the number of parameters. If we have assumed relative
addition of 22 percent random noise (top pseudo-
section). Although the initial guess differs greatly
from the theoretical model, the inversion converged
O
‘ l’
stably to yield the best least-squaresfitting data shown
in the bottom pseudosection.
Ap = (ATA)-‘ATAy, (35)
could produce a parameter change which may in-
c‘reuse instead of decrease the least squares re-
sidual A yTA y.
41
In order to stabilize ATA and prevent divergence,
CONVERGENCE
we resort to a technique described by Levenberg
(1944). Foster (1961), and Marquardt (1963). We
add a small positive constant k to the diagonal
terms of ATA before inversion, so that our algorithm lG3- I I I I
becomes 0 2 4 6 8 IO
From the reduced chi-square, we may now obtain 414 367 263 3w 343 306 242 276 333 29
the parameter covariance matrix NR 436 304 333 236 386 308 267 326 268 304
cov P,,
car Pij = (39)
(cov Pii . cov P,)“”
Table 1. Parameters, percentage parameter standard deviations, and parameter correlation coefficients obtained
from the inversion of theoretical data contaminated by 22 percent random noise.
Percent
Theoretical Inverted standard Correlation
Parameters model model deviation coefficients
NR
ness is 0.94, it suggests that only the resistivity-
thickness quotient or conductivity-thickness product
is accurately known. All the other parameter corre-
lation coefficients are relatively small, indicating that
for this particular data set the remaining model param-
eters may be individually resolved.
apparent resistivity data produced a model which 568 310 \660/ 350 324 980 w 820
roughly corresponds to the location of the known y 63 “Ofl@ 36&4 565 530
427 550 437 296 390 462 380 329 431 NR
sulfide mineralization. Although the apparent re-
0
sistivity contours of the calculated data do not
Downloaded 05/03/13 to 132.203.235.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/
oRE_~ OVERBURDEN
exactly correspond to those of the field data, the
basic pattern of an apparentresistivity low of approxi- LIMESTONE
I
28 percent, respectively. The thicknessof the deposit.
however. was essentially unresolved.
Using the body geometry and resistivity contrast
obtained from the apparent resistivity data, we then FIG. 16. Observed and calculated apparent resistivity
inverted the apparent polarizability data to obtain data (p/2n in R-ft, x = 200 ft) from inversion of
an average polarizability of 3.4 percent FE for the field results (Hallof, 1972) over strata-bound lead-
zinc mineralization near Pine Point, N. W. T
body and 2.0 percent FE for the host rock. Both the
apparent resistivity and apparent polarizability inver-
sions were stable and converged in four iterations
for a total cost of $.20 on the University of Utah
Univac 1I08 computer.
Mendoza, Argentina
A more difficult data set to interpret was obtained
6 4 IO 6 II 1.0 IO IO by Hallof (1972) from a porphyry copper prospect
near Mendoza, Argentina. Since the mineralization
is buried deeply beneath alluvial valley sediments,
it is difficult to obtain an accurate estimate of the
0
i ” ” c : 1’
true resistivity and polarizability of the deposit. In
addition to the complications of depth, it was ncces-
sary first to contend with a second anomaly appear-
ing on the left-hand side of the pseudosection.Since
our fast forward .problem was limited to calculating
effects due to a single causativebody, we effectively)
IP MODEL
ignored the second anomaly by setting equal to zero
12 12 13 14 13 14 I3 I2
II II I4 17 17 17 17 14 II
the elements of the weight matrix corresponding to
IO IO 14 the shaded portion of the pseudosection shown in
9 9 dsko9
15 Figures I4 and 15. We next constrainedthe horizontal
position of the deposit and its depth extent, since
these parameters were very poorly resolved. and
simultaneously inverted the apparent resistivity and
apparent polarizability data to obtain the sections
shown in the figures Since there is information on
both depth and resistivity contrast in the apparent
FIG. 15. Observed and calculated apparent polari- polarizability data, the simultaneous inversion of the
zability data (PFE 0.31 to 5.0 Hz, x = 200 m) from
data sets helped to resolve the model more accurately
inversion of field results (Hallof, 1972) over the
Yalguaraz porphyry copper prospect near Mendoza, than either data set alone. Computer costs for the
Argentina. combined inversion were again less than S I .OO.
000 Pelton et al
Iso-Copperfields, Quebec
To illustrate further the range of applications of
the program, we inverted resistivity data from the
Iso-Copperfields massiv,e sulfide deposit near
Noranda, Quebec (Fountain, 1973). Dipole-dipole
pseudosections for both 100 and 200 ft electrode
spacings were available and were simultaneously
inverted to obtain the common model shown in
IP MODEL Figures 20 and 21. Although the model is relatively
complex, all four parameters which were included
in the inversion were relatively well resolved. The
estimated standarddeviations for the depth and over-
burden resistivity were less than IO percent, whereas
the estimated standard deviations for the host rock
resistivity and target resistivity were less than 35
percent.
Since the calculateddata are derived from a vertical
dike model, the very close fit to the measured data
FIG. 17. Observed and calculated apparent polari-
zability data (PFE 0.31 to 5.0 Hz, x = 200 ft) from obtained over the prominently dipping ore body
inversion of field results (Hallof, 1972) over strata- tends to confirm that dip is indicated by only re-
bound lead-zinc mineralization near Pine Point, latively minor asymmetry in the anomaly pattern.
N.W.T. For most data sets this asymmetry is not recognizable
Resistivity and Induced Polarization 801
49 39 39
NR iR 4.9 36 30 NR NR NR 25 NR NR Posiedon Orebody, W. A.
Our last example of inversion involves very noisy
field data shown in Figures 22 and 23 from the
Posiedon nickel sulfide deposit in Western Australia
(Hallof, 1972). Two anomalies arc evident in the
observed data: one due to a wide band of iron forma-
tion and the other apparently due to the pyrrhotite
and nicke! sulfide ore deposit. We first weighted tn
zero the anomaly due to the iron formation and then
inverted simultaneously the remaining apparent
resistivity and polarizability data. In spite of the noisy
character of the pseudosections, caused by small-
scale geologic inhomogeneities in the weathered
layer, the inversion converged in thme iterations to
yield reasonable estimates for the true resistivity
and polarizability of the nickel deposit.
FIG. 19. Observed and calculated apparent polari- DISCUSSION AND CONCLUSIONS
zability data (PFE 0. I5 to I .2 Hz, x = 200 ft) from We have discussed a computer algorithm which
inversion of field results (Schwenk, 1976) over the
Flambeau massive sulfide deposit near Ladysmith, automatically interprets a certain class of 2-D dipole-
Wisconsin. dipole apparent resistivity and apparent polarizability
pseudosections. The present program is limited in
FIG. 20. Observed and calculated apparentresistivity FIG. 21. Observed and calculated apparentresistivity
data (p/2rr in a-ft, x = 200 ft) from inversion of data (p/27r in s2-ft, x = 100 ft) from inversion of
field results (Fountain, 1973) over the Iso-Copper- field results (Fountain, 1973) over the Iso-Copper-
fields massive sulfide deposit near Noranda, Quebec. fields massive sulfide deposit near Noranda, Quebec.
802 Pelton et al
inversion problems. Certainly the method is fast, Inman, Joseph Robert, 1975, Resistivity inversion with
ridge regression: Geophysics. v. 40, p. 798-817.
inexpensive, and viable. Levenberg, K., 1944, A method for the solution of certain
nonlinear problems in least squares:Quart. Appl. Math.,
v. 2. D. 164-168.
Downloaded 05/03/13 to 132.203.235.189. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/