Efficient Spectral IP-modelling in The Time Domain
Efficient Spectral IP-modelling in The Time Domain
www.elsevier.com/locate/jappgeo
Abstract
We show that the modelling of time-domain spectral induced polarization (SIP) data can be carried out using a DC resistivity
code. The intrinsic resistivity at each point in space is calculated by Fourier transforming the complex, frequency-dependent
resistivity, and a separate DC problem is solved for each time channel. This procedure constitutes an approximation, because the
exact formulation involves a convolution of the time-dependent conductivities with the electric field. The approximation
dramatically simplifies numerical modelling of time-domain SIP data. Implicitly, it is already in use in inversion routines where
data are inverted independently for separate time channels. However, the accuracy has not been investigated in detail.
We first formulate the underlying equations in terms of a maximum chargeability M. We expand the exact and the approximate
solutions with respect to the maximum chargeability and show that the differences are at least second order in M. Next, we
investigate the reflection factor for a simple two-layer model, where the second layer resistivity is assumed to exhibit Cole–Cole
behaviour. We calculate exact and approximate solutions analytically and show that the early- and late-time limits are identical. In
the intermediate time range the error of the approximation is limited to a few percent. Finally, we calculate approximate and exact
apparent resistivities for dipole–dipole configurations over a layered and a full 3-D model and show that the differences are far
below the typical measurement noise level. The conclusion is that the approximation is excellent and will work in all practical
situations.
D 2005 Elsevier B.V. All rights reserved.
Cole–Cole parameters from the decay curves (Tombs, caused by the complex, frequency-dependent conduc-
1981; Johnson, 1984), and the interpretation of time- tivity r(x,r) (e.g. Weller et al., 1996)
domain SIP profiles was frequently based on pseudo-
sections of apparent Cole–Cole parameters (Johnson, jd r x;rP P E x;rP ¼ Id3 P r P r V H ðxÞ ð1Þ
1984; Vanhala and Peltoniemi, 1992; Seigel et al., Here, I is the transmitter current amplitude, d is the
1997). More recently, a concept for a 2-D inversion Dirac delta function, E the electric field, and r denotes
of time-domain SIP data was presented by Yuval and the receiver location. The source term on the right hand
Oldenburg (1997). They first carry out a separate 2-D side describes a single current electrode at location r .V
inversion of voltages for each time channel to obtain The frequency dependence of the transmitter current is
intrinsic time-dependent chargeabilities at each point in denoted H(x), in analogy to the Heavyside function
space. In the next step, they calculate Cole–Cole para- normally used to describe the transmitter waveform in
meters from the time-dependent chargeability and ob- the time domain. Without loss of generality we will
tain a 2-D distribution of intrinsic Cole–Cole restrict ourselves to the case where we use Eq. (1) to
parameters. calculate the time-domainpstep
ffiffiffiffiffiffiffiffiffi responses, in which case
However, treating the time channels as independent H(x) = 1/ix, where i ¼ 1. For the following deri-
of each other involves an approximation which has not vations, it will be useful to express the equation in
been discussed in the literature. The calculation of the terms of the complex resistivity. With
IP decay curve due to an inhomogeneous distribution of
frequency-dependent conductivities may be done by 1
q x;rP ¼ ð2Þ
Fourier transforming the results from the frequency r x;rP
domain into the time domain. Alternatively, the under-
lying differential equation may be transformed into the Eq. (1) becomes:
time domain. The multiplication of the frequency-de-
pendent conductivity with the electric field becomes a E x;r
P P
jd ¼ Id Pr Pr V H ðxÞ ð3Þ
convolution of a time-dependent conductivity with the q x;rP
electric field at each point in space. The resulting
voltage measured at any time will thus depend on all In the time domain, the induced polarisation re-
the previous time channels, and an independent inver- sponse can be calculated by solving Eq. (1) and trans-
sion of time channels is strictly not possible. Conse- forming the results via a Fourier transform. For a direct
quently, Soininen (1984a) suggested to transform the solution in the time domain, Eq. (1) needs to be Fourier
time domain data and to carry out the entire analysis in transformed as:
frequency domain.
Here, we investigate the approximation in the time jd r t;rP 4E P t;r
P ¼ Id Pr Pr V H ðt Þ ð4Þ
domain, which consists of replacing the convolution
where r(t) is the Fourier transform of r(x). The
with conductivity by a multiplication with the recip-
multiplication of conductivity and electric field
rocal of the time-dependent resistivity. We discuss the
becomes a convolution, denoted by the * symbol.
approximation using theoretical considerations and
H(t) is the Heavyside function describing a current
forward modelling and show that it will be sufficiently
step-on at t = 0. Omitting the spatial dependence for
accurate in most practical situations. As a conse-
convenience, we now define a time-dependent resis-
quence, spectral IP-modelling and inversion in the
tivity by:
time domain can be carried out with a DC resistivity
code. Z l
1
qðt Þ ¼ qðxÞH ðxÞeixt dx: ð5Þ
2p l
2. Theoretical background
With this definition, the time-dependent resistivity
We start with the basic equation in the frequency and conductivity may be considered as inverse to each
domain to express the behaviour of electric fields. It can other, because it is easily shown that
be derived from Maxwell’s equations in the frequency
qðt Þ4rðt Þ ¼ H ðt Þ ð6Þ
domain, assuming that the frequencies are sufficiently
low so that displacement currents and induction effects i.e. the convolution of resistivity and conductivity
can be neglected. The physics is described as a DC equals the step-function. It also follows that q(t) must
problem, where the only frequency dependence is be causal, i.e. q(t) = 0;t b 0.
154 A. Hördt et al. / Journal of Applied Geophysics 59 (2006) 152–161
The approximation to the exact Eq. (4) is now least second order in M. Since it is known from
defined by: laboratory and field observations that the chargeabil-
ity is usually a small number, and theoretically cannot
Ê t;r
jd P P
¼ Id Pr P
r V H ðt Þ ð7Þ exceed 1, errors in the approximation are expected to
q t;rP be small.
First, we expand both the solution in time and
where the ^ denotes the approximate solution. The
frequency domain in a Taylor series with respect to M:
equation is only defined for t z 0. Here, the convolution
is ignored and all electric fields in time are considered
BE
independent of each other. Yuval and Oldenburg (1997) E x;rP ;M ¼ P E x;rP þ P
x;rP dM
P
M ¼0 BM M ¼0
implicitly used this approximation by inverting data for
each time channel separately to reconstruct a time- 1 B2PE
þ x;r d M2 þ O M3
dependent chargeability m(t). 2 BM 2 P
M ¼0
The approximate Eq. (7) gives the exact solution in 1
¼ P0
E x; P r þE x;rP d M þ P2 E x;rP M 2
the case of a homogeneous fullspace or halfspace, P1 2
where conductivity gradients vanish (q(t,r ) = q(t)). þ O M3 ð11Þ
This can be shown by rewriting Eq. (7):
E 0, E 1, and E 2 denote the zero, first and second order
jd PÊ t;rP ¼ Id P r V H ðt Þqðt Þ ¼ Id P
r P r V qðt Þ
r P terms of the expansion, respectively.
ð8Þ In the time domain, we obtain analogously for the
approximate electric fields:
where the causality of q(t) was used. By convolution of
Eq. (8) with r(t), making use of Eq. (6), we obtain the
Ê t; P Ê t;rP þ P1
r ; M ¼ P0 Ê t; P
r M
exact Eq. (4). If there are nonzero conductivity gradi- P
where f(x, r ) is a function in space that expresses the The procedure is carried out in detail in Appendix
frequency dependence of q. In the following, we will A. The conclusion is that the differences between ap-
show that the error between the exact solution in proximate and exact solution are at least second order
the time domain and the approximate solution is at in M.
A. Hördt et al. / Journal of Applied Geophysics 59 (2006) 152–161 155
4. Reflection factor for a two-halfspace model with expressions can be found for c = 0.5 and c = 1. Here,
Cole–Cole model behaviour we treat the case of c = 0.5. We re-write Eq. (9) as:
The considerations above were fully general, with- q2 ðxÞ ¼ q0 ð1 mf ðxÞÞ ð17Þ
out any restrictions to the complexity of the model or
where
the frequency dependence of the resistivity. We now
consider a simple model, a 2-layer fullspace, that can be 1 ðixsÞc
treated analytically (Fig. 1). The first layer is not f ðx Þ ¼ 1 c ¼ ð18Þ
1 þ ðixsÞ 1 þ ðixsÞc
polarisable (i.e. the resistivity q 1 is real and indepen-
dent of frequency), the second layer is polarisable (i.e. After some algebra we obtain from Eqs. (16)–(18):
q 2 = q 2(x)). Both the current electrode and the potential
electrode are in the first layer. b1 2bmf ðxÞ
k ðxÞ ¼ þ ð19Þ
From the theory of electrical image sources (e.g. b þ 1 ðb þ 1Þðb þ 1 mf ðxÞÞ
Telford et al., 1990), we find that the potential at
point P (Fig. 1) is given by: where
IH ðxÞq1 q1
IH ðxÞq1 b¼ ð20Þ
/Pr ¼ k ðx Þ ; ð15Þ q2 ðx ¼ 0Þ
4pr 4pr V
where k is the reflection factor: is the DC resistivity contrast. In Appendix B, we re-
write Eq. (19) in terms of the Laplace variable s = ix
q1 q2 ðxÞ and consider s as a scaling factor. We then substitute
k ðxÞ ¼ ð16Þ c = 0.5 and bring Eq. (19) in a form that can be found in
q1 þ q 2 ð x Þ
a lookup table (Abramowitz and Stegun, 1964) to show
The frequency dependence is fully characterised by that the step response of the reflection factor is given
the reflection factor. Factor H(x) in Eq. (15) is required by:
to obtain the step response when transforming to the
rffiffiffi!
time domain, and thus it does not constitute a frequency b1 2mb a2 st t
dependence of the solution. Therefore, for a comparison k ðt Þ ¼ þ e erfc a
b þ 1 ðb þ 1Þðb þ 1 mÞ s
of the approximate and exact solutions, it is sufficient to
consider the reflection factor instead of the full potential ð21Þ
in the subsequent derivations. The reflection factor can
where
also be used to describe more complicated models, such
as a layered halfspace, a sphere, or a vertical dike bþ1
(Telford et al., 1990; Militzer and Weber, 1985), and a¼ ð22Þ
bþ1m
it thus represents a wide class of models.
We now assume Cole–Cole behaviour for the second and erfc is the complementary error function defined
layer resistivity q 2(x) (Eq. (2)). The aim is to bring Eq. by:
(16) in a form that can easily be transformed into the Z x
time domain via a Laplace transform. Analytical 2 2
erfcð xÞ ¼ 1 erf ð xÞ ¼ 1 pffiffiffi ey dy ð23Þ
p 0
b1 2mbf ðt Þ
k̂k ðt Þ ¼ þ ð25Þ
b þ 1 ðb þ 1Þðb þ 1 mf ðt ÞÞ
The late-time limit is equal to the DC reflection factor, hardly visible (Fig. 3a). Differences are largest when
which, as expected, does not depend on chargeability. the lower layer is more resistive. For the parameters
The early-time limit (Eq. (27)) of the reflection factor chosen here, the reflection factor changes sign and the
depends on the chargeability. An interesting feature is calculation of a relative error is not meaningful. Fig. 3b
that the reflection factor may change sign with time, zooms into the critical range, showing that the maxi-
which is the case if q 1 N q 2 and m N 1 b. We also show mum absolute error is about 0.03. However, the reflec-
in Appendix C that even the late-time asymptotic be- tion factor is only a measure of the secondary response
haviour is identical
pffi for approximate and exact solution, caused by the resistivity contrast. In practice, this will
namely a 1= t - convergence in both cases. be added to a primary component as indicated in Eq.
We have now shown that both the early- and late- (15), and the effect of the approximation errors will be
time limits of the approximate solution are exact, and even smaller.
differences occur only in between. These, however, are
expected to be small, and, as was shown in the section 5. Numerical modelling
above, at least second order in chargeability. Fig. 2a
shows comparisons of the exact and approximate solu- In this section we discuss some numerical modelling
tion for a relatively high chargeability of m = 0.5. There results for dipole–dipole configurations for both 1-D
is no visible difference for any of the three selected and 3-D models, in order to show that the approxima-
resistivity contrasts. The relative differences (Fig. 2b) tion will be valid in real situations. First, we calculated
show that the error depends on the resistivity contrast IP responses for the 3-layer model and the configura-
and that the maximum error for this case is below 3%. tion shown in Fig. 4. The calculation of the exact
Even for the extreme case of m = 0.9, the differences are response was done using a 1-D DC resistivity code,
A. Hördt et al. / Journal of Applied Geophysics 59 (2006) 152–161 157
Fig. 4. Dipole–dipole configuration and layered model to test the Fig. 6. 3-D model used to calculate exact and approximate responses.
approximation. Time constant s and frequency exponent c are 1 s and A vertical dike is embedded in an otherwise homogeneous halfspace.
0.5 for the third and last layer. Fig. 5 shows results for a = 3 m and The dipole–dipole configuration is centered over the dike (after
n = 8. Soininen, 1984b).
158 A. Hördt et al. / Journal of Applied Geophysics 59 (2006) 152–161
of polarisability may be considered as a reduction of made significant contributions to the project that moti-
conductivity by an amount proportional to the charge- vated this work. Roland Blaschek carried out the 3-D
ability. He subsequently shows that apparent charge- simulations. The comments of Niels B. Christensen and
abilities may be calculated solely from the apparent an anonymous reviewer helped to improve the manu-
resistivity, and thus DC theory is sufficient. His deriva- script. The project was partly sponsored by the German
tions, which correspond to the early-time limit imme- Science foundation (DFG-SFB 419: Teilprojekt A1).
diately after the transmitter shut-off, were later
modified by Roy and Poddar (1981) who state in Appendix A. Derivation of the first order terms
their conclusions that da complete dynamic theory in-
cluding the time decay has not been worked outT. We Here, we calculate the derivatives of Eqs. (3) and (7)
believe that with our considerations of the reflection with respect to M in order to compare the terms of
factor we made a step towards a complete dynamic different order in M of the approximate and exact
theory, but at the same time we have shown that such solutions. The derivative of Eq. (3) with respect to M is:
a theory is not necessary for the practical application. !
It may be surprising that such an apparently crude B E x;r B
jd P P
¼ Id3 P r Pr V H ðx Þ
approximation, i.e. replacing a convolution by a multi- BM q x;rP BM
plication, works so well. To understand this, it may be ðA1:1Þ
useful to remember that we replace the convolution
with r(t) by multiplication with 1 / q(t). As shown in The right hand side vanishes, because it expresses
the theoretical background section, q(t) is defined such the current of a single point electrode at r̄ ,V which does
that it already contains the properties of the convolution not depend on M. The left hand side yields:
with r(t) if there are no conductivity gradients. The 0 BE 1
P
x;rP
approximation only deals with the charge transfer E x;r Bq C
jd @ BM
B P P2
through layer boundaries in the mid-time or mid-fre- x;rP A ¼ 0
q x;rP q x;rP BM
quency range.
Obviously, our results have a direct impact on the ðA1:2Þ
inversion. When inverting time-domain SIP data, the Substituting M = 0 and using
time channels may be inverted separately, and the
convolution can be ignored. This procedure, which Bq
x;rP ¼ q0 P
r f x;rP ðA1:3Þ
was already applied by Yuval and Oldenburg (1997) BM M ¼0
is now perfectly justified by our considerations. Hönig
we obtain:
(2002) and Hönig and Tezkan (submitted for publica- !
tion) present further examples of 2-D Cole–Cole inver- E x;r E x;rP
P1 P þ f x;rP
sion both with synthetic and field data. More generally, jd P0
¼0 ðA1:4Þ
q0 P
r q0 Pr
any procedure that uses apparent parameters as the
basis for the interpretation implicitly makes use of the where E 0(x,r ) and E 1(x,r ) are the zero and first order
approximation. In particular, this applies to apparent terms of the Taylor series expansion of the solution, as
resistivity pseudosections of Cole–Cole parameters explained in the main section. We denote the DC
(e.g. Johnson, 1984; Vanhala and Peltoniemi, 1992; solution E 0(r ), such that the frequency dependence of
Seigel et al., 1997). The relationship between intrinsic E 0(x,r ) is given by E 0(x,r ) = E 0(r )H(x). With
and apparent time-dependent resistivities in the inter-
mediate time range was not resolved until now, and the f t;rP ¼ FT f x;rP H ðxÞ ðA1:5Þ
meaning of apparent Cole–Cole parameters calculated
from transient decay curves was not clear. Our results the Fourier transform of Eq. (A1.4) becomes:
show that the relationship is essentially the same as that !
E t;r E r
for DC resistivities and depends only on geometry and jd P1 P
þ f t;rP
P0 P
¼0 ðA1:6Þ
the spatial distribution of resistivities. q0 P
r q0 Pr
where s = ix. We omit the algebra that is left to show that this is
In Eq. (18), the frequency is always multiplied with equivalent to Eq. (21).
decay time s, which may be considered as a frequency
scaling factor. Therefore, we note that: Appendix C. Late-time limit of the reflection factors
t
1 1 1 1 Here, we derive the late-time limits of the reflection
L f ðssÞ ¼ sL f ðssÞ ¼f ðA2:2Þ factors for approximate and exact solution (Eqs. (21)
s ss s
and (25)). First, we use the Taylor series expansion of
With these considerations, and using c = 0.5, we now the error function for large arguments (Abramowitz and
can use a lookup table to show that the step response Stegun, 1964) and truncate the series after the first term
corresponding to Eq. (18) is (Abramowitz and Stegun, to show that
1964, 29.3.43): pffiffiffiffiffiffi 1 t
et=s erfc t=s c pffiffiffipffiffiffiffiffiffi ; NN1 ðA3:1Þ
pffiffiffiffi p t=s s
1 1 ss 1 1
f ðt Þ ¼ L pffiffiffiffi ¼ sL pffiffiffiffi pffiffiffiffi
s 1 þ ss ssð1 þ ssÞ and
rffiffiffi !
t t rffiffiffi!
¼ e s erfc ðA2:3Þ a2 st t 1 t
s e erfc a c pffiffiffipffiffiffiffiffiffi ; NN1 ðA3:2Þ
s a p t=s s
which is equal to Eq. (26).
With Eq. (A3.2), we can describe the late-time be-
Next, we show that Eq. (21) is the result of inverse
haviour of the exact solution (Eq. (21)) as:
Laplace transforming Eq. (19). First, it can easily be
demonstrated that
b1 2mb 1
k ðt Þc þ pffiffiffipffiffiffiffiffiffi
1 b1 ðb 1 þ mÞðixsÞc b þ 1 ðb þ 1Þðb þ 1 mÞ a p t=s
k ðxÞ ¼ þ
b þ 1 m a þ ðixsÞc a þ ðixsÞc b1 2mb 1
¼ þ 2 pffiffiffipffiffiffiffiffiffi ðA3:3Þ
ðA2:4Þ b þ 1 ð b þ 1Þ p t=s
A. Hördt et al. / Journal of Applied Geophysics 59 (2006) 152–161 161
where the definition of a (Eq. (19)) was used for the Militzer, H., Weber, F., 1985. Angewandte Geophysik. Band, vol. 2.
second step. Springer.
Pelton, W.H., Ward, S.H., Hallof, P.G., Sill, W.R., Nelson, P.H., 1978.
For the approximate solution, we note that f(t) (Eq. Mineral discrimination and removal of inductive coupling with
(26)) is small for large t and thus: multi-frequency IP. Geophysics 43, 588 – 609.
Roy, A., Poddar, M., 1981. A simple derivation of Seigel’s time
b1 2mbf ðt Þ b1 domain induced polarization formula. Geophys. Prospect. 29,
k̂k ðt Þc þ c
b þ 1 ðb þ 1Þðb þ 1 mf ðt ÞÞ b þ 1 432 – 437.
2mbf ðt Þ b 1 2mb 1 t Seigel, H.O., 1959. A theory for induced polarization effects (for step-
þ c þ pffiffiffipffiffiffiffiffiffi ; NN1 function excitation). In: Wait, J.R. (Ed.), Overvoltage Research
ðb þ 1Þ2 b þ 1 ðb þ 1Þ2 p t=s s and Geophysical Applications. Pergamon Press, London.
Seigel, H.O., Vanhala, H., Sheard, S.N., 1997. Some case histories of
ðA3:4Þ
source discrimination using time-domain spectral IP. Geophysics
Comparison of Eqs. (A3.3) and (A3.4) shows that 62, 1394 – 1408.
Soininen, H., 1984a. Inapplicability of pulse train time-domain
approximate and exact solution not only converge to
measurements to spectral induced polarization. Geophysics 49,
the same constant at late times, but also have the same 826 – 827.
asymptotic behaviour with time. Soininen, H., 1984b. The behavior of the apparent resistivity phase
spectrum in the case of a polarizable prism in an unpolarizable
References half-space. Geophysics 49, 1534 – 1540.
Spitzer, K., 1995. A 3-D finite-difference algorithm for DC resistivity
modelling using conjugate gradient methods. Geophys. J. Int. 123,
Abramowitz, M., Stegun, I.A., 1964. Handbook of Mathematical
903 – 914.
Functions. Dover Publ., New York.
Telford, W.M., Geldart, L.P., Sherriff, R.E., 1990. Applied Geophys-
Christensen, N.B., 1990. Optimized fast Hankel transform filters.
ics. Cambridge University Press, Cambridge.
Geophys. Prospect. 38, 545 – 568.
Tombs, J.M.C., 1981. The feasibility of making spectral IP measure-
Cole, K.S., Cole, R.H., 1941. Dispersion and absorption in dielectrics:
ments in the time domain. Geoexploration 19, 91 – 102.
1. Alternating current fields. J. Chem. Phys. 9, 341.
Vanhala, H., Peltoniemi, M., 1992. Spectral IP studies of Finnish ore
Hönig, M., 2002. Entwicklung und Anwendung von Auswertealgor-
prospects. Geophysics 57, 1545 – 1555.
ithmen für die induzierte Polarisation im Zeitbereich in bis zu drei
Vanhala, H., Soininen, H., Kukkonen, I., 1992. Detecting organic
Dimensionen. Ph.D. Dissertation, Universität zu Köln.
chemical contaminants by spectral-induced polarization method
Hönig, M, Tezkan, B., submitted for publication. 1-D and 2-D Cole–
in glacial till environment. Geophysics 57, 1014 – 1017.
Cole inversion of time domain induced polarization data, Geo-
Wait, J., 1959. Overvoltage Research and Geophysical Applications.
phys. Prospect.
Pergamon Press, London.
Johnson, I.M., 1984. Spectral induced polarization parameters as
Weller, A., Seichter, M., Kampke, A., 1996. Induced-polarization
determined through time-domain measurements. Geophysics 49,
modelling using complex electrical conductivities. Geophys. J.
1993 – 2003.
Int. 127, 387 – 398.
Kemna, A., Binley, A., Slater, L., 2004. Crosshole IP imaging
Yuval, Oldenburg, D.W., 1997. Computation of Cole–Cole para-
for engineering and environmental applications. Geophysics 69,
meters from IP data. Geophysics 62, 436 – 448.
97 – 107.