SPE 116255

Numerical Inversion of Laplace Transforms in the Solution of Transient

Flow Problems With Discontinuities
N. Al-Ajmi, SPE, Colorado School of Mines; M. Ahmadi, SPE, Norwest Questa Engineering; and E. Ozkan, SPE,
and H. Kazemi, SPE, Colorado School of Mines

Copyright 2008, Society of Petroleum Engineers

This paper was prepared for presentation at the 2008 SPE Annual Technical Conference and Exhibition held in Denver, Colorado, USA, 21–24 September 2008.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents of the paper have not been
reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect any position of the Society of Petroleum Engineers, its
officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written consent of the Society of Petroleum Engineers is prohibited. Permission to
reproduce in print is restricted to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
Laplace transformation provides advantages in the solution of many pressure-transient analysis problems. Usually, these
applications lead to a solution that needs to be inverted numerically to the real-time domain. The algorithm presented by
Stehfest in 1970 is the most common tool in petroleum engineering for the numerical inversion of Laplace transforms. This
algorithm, however, is only applicable to continuous functions and this limitation precludes its use for a wide variety of
problems of practical interest. Other algorithms have also been used, but with limited success or popularity. A recent
algorithm presented by Iseger in 2006 removes the restriction of continuity and provides opportunities for many practical
applications. This paper exploits the useful features of the Iseger’s algorithm in the inversion of continuous as well as
singular and discontinuous functions that arise in the solution of pressure-transient analysis problems. The most remarkable
applications are in the problems that require the use of piecewise continuous and piecewise differentiable functions, such as
the use of tabulated data in the Laplace transform domain, deconvolution algorithms, and solutions that include step-rate
changes as in the mini-DST tests.

Introduction
Since the seminal work of van Everdingen and Hurst (1953), Laplace transformation has been a standard tool for transient
fluid-flow problems in porous media. Laplace transformation is mostly used for the solution of initial-boundary-value
problems and in the evaluation of integral equations of convolution type. For some of these applications, an exact (analytical)
inversion is not possible and numerical inversion is the only resort. For some others, numerical inversion is chosen also
because of convenience.
Kryzhniy (2006) defines three general cases encountered in the inversion of Laplace transforms:
1. The solution is obtained in closed form analytically in the Laplace domain: The inverse transform can be found
analytically or with the help of tables. Numerical inversion may also be used if the solution is too complicated.
2. The Laplace transform of the solution is computable in the complex half-plane of convergence: In this case, the
inverse Laplace transform may be obtained by the evaluation of the complex integral of inverse transformation.
Numerical inversion based on the calculation of the Bromwich integral (when possible) yields good results.
3. The Laplace transform of the solution is computable or measurable on the real and positive axis only: In this case,
there is no exact inversion formula. All numerical inversion methods encounter difficulties because the inversion
problem is severely ill-posed.
Kryzhniy (2004) points out that a large group of exponential analysis problems encountered in transient problems, which
involve real-valued Laplace transforms, are subject to inversion problems because no exact inversion formula is available for
real-valued functions.
The focus of this paper is the numerical inversion of Laplace transforms in the solution of transient-flow problems with
piecewise-differentiable data. These problems usually arise due to piecewise approximations to tabulated rates and pressures,
interruption of production by shut-ins, and step-rate sequences as in mini-DST applications. In some of these cases, the
function to be inverted may be continuous but not continuously differentiable (e.g., piecewise approximations to sampled
data). In others, the function may be only piecewise continuous and piecewise differentiable (e.g., a sequence of drawdown
and buildup periods). Laplace transformation of piecewise continuous functions include exponential contributions at the
boundaries of its sections, which cause problems in numerical inversion. If the function is discontinuous at section
boundaries, then most numerical inversion algorithms require special treatments or fail altogether. These piecewise
2 SPE 116255

differentiable or discontinuous functions commonly appear in deconvolution applications where the inherent numerical
problems due to ill-posed nature of deconvolution (Lamm, 2000) are further exacerbated by the errors in numerical Laplace
inversion.
The objective of this paper is not to present a new deconvolution algorithm, but we demonstrate that some desirable or
undesirable features of the Laplace inversion algorithm used in the deconvolution process may be falsely attributed to the
deconvolution algorithm. In other words, it is possible to improve the success of a given deconvolution algorithms by
improving the numerical inversion of Laplace transforms of discontinuous or piecewise differentiable functions. Along these
lines, Ilk et al. (2005) have noted the importance of the numerical inversion algorithm used to invert the Laplace transform of
the sampled rate function in their deconvolution approach using B-splines. Our renewed interest in the problem was driven
by a new algorithm presented by Iseger (2006) that claims to handle singularities and discontinuities. In petroleum
engineering, the most commonly used numerical inversion algorithm has been the one proposed by Stehfest (1970), but this
algorithm is only applicable to continuous functions, which limits its use for problems involving step-rate changes and shut-
ins. Other algorithms, such as the ones proposed by Bellman et al. (1966), Crump (1976), and Talbot (1979), have been
applied in these cases, but not gained as much popularity due to their complexity and inflexibility in standard applications.
Persistent problems and increased interest in deconvolution algorithms led to the continuous search for new Laplace
inversion algorithms, such as the Gaver-Wynn-Rho algorithm presented by Abate and Valkó, 2004 and Valkó and Abate,
2004.
The algorithm presented by Iseger (2006) removes the restriction of continuity and provides opportunities for many
practical applications. We have tested the Iseger algorithm for several common conditions requiring the use of piecewise-
differentiable and discontinuous functions, such as the use of tabulated data in the Laplace transform domain, problems
involving step-rate changes, build-up tests following a drawdown period, and mini-DST tests with a sequence of drawdown
and buildup periods. All of these cases are known to have inversion problems with the commonly used Laplace inversion
algorithms. We have compared the results of the Iseger algorithm with the the Stehfest algorithm. The objective of this
communication is to share our findings about the application and success of the Iseger algorithm in transient fluid-flow
problems.

Background
To set the stage for discussion, we will first summarize some general concepts of Laplace and inverse Laplace transforms,
Laplace transformation of sampled functions, and convolution and deconvolution applications encountered in the solution of
transient-flow problems that involve piecewise differentiable and discontinuous functions.

Laplace Transformation and Numerical Laplace Inversion

Laplace transformation is a linear operation that assigns a functional, f (s ) , to a function, f (t ) , defined for all t > 0 , by the
rule
∞
L{ f (t )} = f (s ) = ∫e
− st
f (t ) dt . (1)
0+

A set of conditions on the function, f (t ) , and the Laplace transform parameter, s, accompany this definition to guarantee the
convergence of the integral and thus the existence of the Laplace transform of the function, f (t ) (Churchill, 1958). As an
integral transformation, Laplace transformation reduces a partial differential equation in n independent variables to one in n-
1, which is expected to possess a simpler or known solution. In physics and engineering, Laplace transformation is usually
applied to time variable to reconstruct the problem in the Laplace domain (known as the frequency domain in physics) where
time-dependency does not appear explicitly but is preserved parametrically in the Laplace transform parameter, s. Another
advantage of Laplace transformation is the simplification of convolution operation into an algebraic product of functions,
which finds important applications in the solution of integral equations. In transient fluid flow in porous media, this property
is useful in dealing with variable-rate production problems, which is the focus of this paper.
To be useful, every Laplace transform must possess a unique inverse transform; that is, for f (s ) to be a Laplace
transform, there should be a unique function f (t ) as the inverse Laplace transform of f (s ) . A set of rules guarantees the
existence of the inverse Laplace transforms for certain classes of functions (Churchill, 1958). Although the complex
inversion formula,
γ + i∞
1
f (t ) = ∫e
st
f (s ) ds , (2)
2πi
γ−i∞

is well-known, the real-inversion formula

SPE 116255 3

{ }
L-1 f (s ) = f (t ) . (3)

is not available (Kryzhniy, 2004). Therefore, if the Laplace transform, f (s ) , is known in the complex half-plane of
convergence, then its inverse, f (t ) , may be found by the evaluation of complex integral in Eq. 2. In some problems of
practical interest to us, the Laplace transform is known only for real and positive values of s. In these cases, because no
inversion formula is available in the form of Eq. 3, the ill-posed problem of Volterra integral equation of the first kind given
in Eq. 1 needs to be evaluated to find the original function, f (t ) .
Since the advent of computers, interest in numerical methods to find accurate inversion of Laplace transforms has grown.
Cohen (2007) categorizes the numerical Laplace inversion methods as follows:
1. Series expansion methods
2. Quadrature methods (including Fourier series based methods)
3. Rational approximation methods
4. The method of Talbot
5. Methods based on the Post-Widder inversion formula
6. Regularization methods
Numerical Laplace inversion algorithms commonly used in petroleum engineering fall mostly into Categories 2, 4, and 5.
For example, the well-known Stehfest (1970) algorithm is based on the Post-Widder inversion formula. The algorithm
proposed by Bellman et al. (1966) use Gauss-Legendre quadrature rule and the Crump (1976) algorithm is a quadrature
method using Fourier series. The new algorithm by Iseger (2006), which is used in this work, is also as a Fourier series based
algorithm.
The algorithms mentioned above have some common features of the categories they belong to. As noted by Onur and
Reynolds (1998), for example, the Stehfest (1970) algorithm displays some smoothing characteristics. This is due to the basis
of the Stehfest algorithm in stochastic data analysis, which originates from Gaver’s (1966) probabilistic interpretation of the
Laplace transforms (Ozkan and Raghavan, 1997). Although this property is useful when smoother representation of the
inverted function is favored, such as derivative calculations (Onur and Reynolds, 1998), it causes some loss of accuracy and
character if the inverted function includes contributions from step-changes and discontinuities. Fourier series based
algorithms, such as Crump (1976) and Iseger (2006), on the other hand, provide better accuracy and preserve the character of
the function in dealing with step functions and functions with singularities and discontinuities, but they may lead to more
oscillatory inversion results. As we mentioned in the Introduction, the inversion of analytical or sampled responses that
include contributions from discontinuous forcing functions, such as a sequence of drawdown and buildup periods, is an
important problem and will be our subject in this paper.

Convolution and Deconvolution

An important property of Laplace transformation is expressed by the convolution or Duhamel’s theorem (Duhamel, 1833),
and is given by
t
{ }
L−1 f1 (s ) f 2 (s ) = ( f1 ∗ f 2 )(t ) = ∫ f (τ ) f
1 2 (t − τ ) dτ . (4)
0

This property of the Laplace transformation converts the convolution integral of two functions into an algebraic product of
their Laplace transforms and has important applications in transient flow problems in porous media associated with variable-
rate production. A particular application of convolution in obtaining the solution of pressure drop, Δp(M , t ) , due to variable
production rate, q(t ) , is of interest in this paper and we restrict the following discussion of convolution and deconvolution to
this particular application.
If Δp c′ (M , t ) is the time derivative of the pressure drop at a point M and time t for constant unit rate, then the pressure
drop, Δp(M , t ) , due to variable production rate, q(t ) , is given by the convolution of q(t ) and Δp c′ (M , t ) as follows:
t
Δp(M , t ) = q(τ )Δp c′ (M , t − τ ) dτ .
∫ (5)
0

For many applications of pressure-transient analysis, it is convenient to remove the effect of rate-transients on the measured
wellbore pressures, Δp w (t ) = Δp(M w , t ) . This requires deconvolution of Eq. 5 to recover the underlying constant-rate
response, Δp ′wc (t ) = Δp c′ (M w , t ) , from the known (measured) data of Δp w (t ) and q(t ) . Deconvolution, however, corresponds
to the solution of the convolution integral given in Eq. 5, which is a Volterra integral equation of the first kind. Because of
the smoothing effect of convolution (Lamm, 2000), different kernels, Δp′wc (t ) , yield the same convolution response, Δp w (t ) ,
for a given rate function, q(t ) (Ilk et al., 2005). In deconvolution, this causes small errors in data, especially in q(t ) , to yield
4 SPE 116255

large variations (errors) in the deconvolved constant-rate response. To alleviate the problem, some deconvolution algorithms
condition the deconvolved constant-rate response by using its expected behavior, such as strict monotonicity (Coats et al.,
1964), non-positive second derivative (Kuchuk et. al., 1990), and non-negativity (Schroeter et al., 2004, and Levitan, 2003).
Another approach is regularization that smoothes the response by removing non-standard or irregular features (Ilk et al.,
2005).
Performing deconvolution in the Laplace domain is usually a favored approach because the application of the convolution
theorem (Eq. 4) to Eq. 5 yields the following algebraic deconvolution expression in the Laplace domain:
Δp w (s )
Δp wc (s ) = . (6)
sq (s )
Despite the simplicity of Eq. 6, Laplace transformation does not remove the inherent ill-posed nature of deconvolution. More
importantly, when the rates include step changes and discontinuities, Laplace inversion becomes severely ill-posed itself
(Kryzhniy, 2006). Therefore, the potential of deconvolution is not realized if an appropriate Laplace inverter that can handle
singularities and discontinuities is not used. The objective of this paper is to contribute to the solution of this problem.

Laplace Transformation of Sampled Functions

In field applications of deconvolution, the input data, Δp w (t ) and q(t ) , are available as sampled (tabulated) functions at a
finite number of time points. When deconvolution is performed in Laplace domain by using Eq. 6, the sampled input data are
expressed in Laplace domain by using techniques, which are variants of the suggestions of LePage (1961). The general idea
is connecting the sampled data points by line segments or by using polynomial fits, and then taking the Laplace
transformation through analytical techniques. Laplace transformation, however, requires the knowledge of the function over
the entire semi-infinite domain while the sampled functions are normally available over a finite interval. Therefore, some
form of extrapolation to infinity is required to compute the Laplace transform of the sampled data. For most applications, the
initial value of the function is known (e.g., pressure or sandface rate with wellbore storage effect), but the sampling frequency
requires interpolation between the initial value and the first sampling point. If the behavior of the data at the two endpoints of
the Laplace integral is not known (e.g., constant wellbore storage, radial flow, pseudosteady state, etc.), then insufficient
accuracy of the interpolation and extrapolation leads to errors known as tail effects.
To deal with the different functional approximations required at the two endpoints, the Laplace integral in Eq. 1 can be
written as follows:
t1 tN ∞
f (s ) = ∫ e − st
f (t )dt + e −st f (t )dt + e −st f (t )dt .
∫ ∫ (7)
0+ t1 tN

Roumboutsos and Stewart (1988) have proposed a simple, piecewise-linear interpolation to connect the consecutive data
points of the sampled function for all three integrals. For the first integral in Eq. 7, this approach leads to a reasonably good
approximation of f (s ) if the early-time response includes wellbore-storage effects. To evaluate the third integral, they
proposed to extend the linear slope of the last interval of the data ( t N −1 ≤ t ≤ t N ) to infinity. The Roumboutsos-Stewart
algorithm for the Laplace transform of a sampled function is given by

f 0 f N − st N f ′ 1 − e − st1( ) N −1
(
f i′ e − sti − e − sti +1 )+ f ′ e
f (s ) =
s
−
s
e + 0 2
s
+ ∑
i =1 s 2
s
N
2
− st N
, (8)

where f i are the values of the function at points ti and f i′ is the linear slope of the data in the interval ti ≤ t ≤ ti +1 .
Onur and Reynolds (1998) investigated different approaches to represent the sampled function, f (t ) , in each integral in
the right-hand side of Eq. 7. For the second integral, they suggested the use of the piecewise-linear interpolation (same as
Roumboutsos and Stewart) or piecewise quadratics. For most applications, simple, piecewise-linear approximation of f (t ) in
the second integral of Eq. 7 yields reasonably accurate results. For the first and the third integrals in Eq. 7, Onur and
Reynolds suggest the use of a log-linear (power law) approximation for f (t ) . This approach yields the following Onur-
Reynolds algorithm for the Laplace transformation of a sampled function:

α0 f1e − st1 f e − st N
N −1
(
f i′ e − sti − e − sti +1 )+ α
f (s ) =
s β +1
0
[Γ(β 0 + 1) − Γ(β 0 + 1, st1 )] +
s
− N
s
+ ∑i =1 s 2
s β +1
[Γ(β + 1, st N )] . (9)

Onur and Reynolds suggest that the constants, α 0 and β 0 , be determined by fitting a least-square straight line of the form
ln f (t ) = β 0 ln t + ln α 0 through the sampled data in the interval t1 ≤ t ≤ t1∗ where ln t1∗ t1 ≥ Ll with 0.1 ≤ Ll ≤ 1 . The ( )
SPE 116255 5

constants, α and β , should be determined by a similar procedure for the interval t N∗ ≤ t ≤ t N that includes minimum five
( )
data points and satisfies ln t N t N∗ ≥ Lr with 0.1 ≤ Lr ≤ 1 .
For our purposes in this paper, we use both the Roumboutsos-Stewart and Onur-Reynolds algorithms. Although the
particular form of the Onur-Reynolds algorithm given by Eq. 9 yields more accurate results than the Roumboutsos-Stewart
algorithm given by Eq. 8, it cannot be used for cases including shut-in periods (because the log-linear approximations used
for the first and the third integrals require the logarithms of the function). Therefore, we will use Roumboutsos-Stewart
algorithm for the examples that include shut-in periods in this paper. It is possible to modify both Roumboutsos-Stewart and
Onur-Reynolds algorithms or develop other algorithms to better represent the Laplace transform of a specific sampled
function. A detailed discussion for most common cases encountered in transient fluid flow is provided by Onur and Reynods
(1998) and will not be repeated here.

The Iseger (2006) Algorithm

As noted in the introduction, our renewed interest in the numerical inversion of singular and discontinuous functions was
driven by a new algorithm presented by Iseger (2006). The Iseger algorithm removes the restriction of continuity required by
some of the common inversion algorithms, such as the Stehfest (1970) algorithm, and provides opportunities for many
practical applications. The algorithm proposed by Iseger (2006) is a Fourier series method, which is based on Poisson’s
summation formula. In this algorithm, Poisson’s summation relates an infinite sum of Laplace transform values to z-
transform of the function values. The infinite sum is approximated by a finite sum based on the Gaussian quadrature rule and
the time-domain values of the function are computed by a Fourier transform algorithm. Iseger (2006) shows that the results of
the algorithm are in near machine precision.
The appeal of the Iseger algorithm is in its ability to compute inverse Laplace transforms of functions with all kinds of
discontinuities, singularities, and local nonsmoothness, even if the points of discontinuity and singularity are not known a
priori. The implementation of the algorithm may be slightly more involved than the Stehfest (1970) algorithm, but it is
comparable to the implementation of the other common algorithms. We refer the reader to the original paper by Iseger (2006)
for the details of the implementation of the algorithm. Here, we will comment on a few features and parameters of the Iseger
algorithm that we have noted during our applications to fluid flow problems in porous media.
A critical parameter used in the Iseger algorithm is Δ computed from the following relation:
T
Δ= , (10)
M
where T is the period for which the inversions are computed and M is the number of points at which the inverse Laplace
transforms are computed. To obtain accurate results, we have found that (1 Δ ) ∈ N and the following condition must be
satisfied:
M ≥ 10ΔM ; that is, M ≥ 10T . (11)
To demonstrate the use of these conditions, we consider the inversion of the unit step function for a period of 20 hr as shown
in Fig. 1. By Eq. 11, the minimum number of inversion points required for this case is M ≥ 200 . If 220 points are used, the
condition in Eq. 11 is satisfied and (1 Δ ) = 11 ∈ N . Fig. 1 shows that we obtain accurate inversion of the unit step function
with 220 data points. If we use 103 data points, for example, the condition in Eq. 11 is not satisfied and if we use 223 data
points, the condition in Eq. 11 is satisfied but (1 Δ ) ∉ N . For both cases, inversions are not sufficiently accurate. We also
note from our experience that if the characteristics of the function change in short intervals (e.g., rate and pressure variations
due to short drawdown and buildup periods in a mini-DST test), a large M may be required for accurate inversions.

2
220 data
Unit Step Function (One Step Change)

1.5 103 data

233 data

0.5

-0.5

-1
0 5 10 15 20
Time, hr

Figure 1 – Numerical Inversion of Unit Step Function by the Iseger algorithm; Effect of the number of inversion points.
6 SPE 116255

The Iseger algorithm uses M 2 = nrp × M oversampling points to compute inversions at M data points. It was
recommended by Iseger (2006) to use nrp = 8 especially for well behaved functions. This particular choice of nrp was
dictated by the Fast Fourier Transform originally used by Iseger, which required M and nrp be a power of 2. For our
applications, however, we did not obtain stable inversions with nrp = 8 . We have found that using nrp = 3 with Discrete
Fourier Transform generated more stable inversions for our applications. Fig. 2 shows the effect of nrp on numerical
inversion of a typical pressure function encountered in fluid flow problems by using the Iseger algorithm.

1.4

1.2

1
Function

0.8

0.6

0.4
nrp=3
0.2 nrp=7

0
0 5 10 15 20
Time, hr
Figure 2 – Effect of nrp in the numerical Inversion of a function by the Iseger algorithm.

Verification Examples
In this section, we present examples to verify the success of the Iseger algorithm in the inversion of functions that are
piecewise differentiable or discontinuous. We compare the results to the Stehfest algorithm to delineate the differences from
standard algorithms.

Discontinuous and Piecewise Differentiable Functions

Fig. 3 shows the inversion of a unit step function (Heaviside function) with the pole at t = 10 hr using the Iseger and Stehfest
algorithms. For the Iseger algorithm, inversion was performed with M = 220 and nrp = 3 . Three values of the parameter,
N = 6, 8, and 12, were used for the Stehfest inversions ( N controls the number of functional evaluations in the Stehfest
algorithm and theoretically, higher N values yield better inversions). Fig. 3 shows that the Iseger algorithm recovers the true
step-wise character whereas the Stehfest algorithm smears the function around the point of discontinuity ( t = 10 hr).

2.0

1.5
Original Function

Iseger
Unit Step Function

1.0 Stehfest, N=6

Stehfest, N=8

0.5 Stehfest, N=12

0.0

-0.5
0 5 10 15 20
Time, hr

Figure 3 – Numerical Inversion of a unit step function; comparison of Iseger and Stehfest algorithms.

In Fig. 4, we consider a function with multiple step changes. This function corresponds to the rate sequence used in a
synthetic deconvolution example below. For the results in Fig. 4, we used M = 512 and nrp = 3 in the Iseger algorithm and
N = 6, 8, and 12 in the Stehfest algorithm. The smearing of the function at the points of discontinuity with the Stehfest
algorithm completely destroys the local and global characteristics of the original function. The success of the Iseger
algorithm in this example is remarkable.
SPE 116255 7

Original Rate
300 Iseger
Stehfest, N=6
Stehfest, N=8
Stehfest, N=12

200

q, rbbl/d
100

0
0 4 8 12 16
Time, hr

Figure 4 – Numerical Inversion of a function with multiple step changes; comparison of Iseger and Stehfest algorithms.

To show that special algorithms, such as the Iseger algorithm, may be required even when the function is continuous but
only piecewise differentiable, we present the results in Fig. 5. To generate the function in this example, we used the function
in Fig. 4 as the rate sequence and generated the corresponding pressure changes. The Laplace transform of this tabulated
pressure-change function was obtained by the Roumboutsos-Stewart algorithm (the Onur-Reynolds algorithm generated the
same results) and inverted back by the Iseger and Stehfest algorithms. As shown in Fig. 5, while the Stehfest algorithm
returns a smooth, continuously differentiable function, the Iseger algorithm accomplishes an accurate inversion that preserves
all the characteristics of the piecewise differentiable function.
300

Original Function
Iseger
Stehfest

200
Δp, psi

100

0
0 4 8 12 16
Time, hr

Figure 5 – Numerical Inversion of a piecewise differentiable function; comparison of Iseger and Stehfest Algorithms.

Wellbore Pressure Solution for a Combined Drawdown and Buildup

This is a classical example where the Stehfest algorithm do not yield good results. For the particular example here, we
assume that the well is produced at a constant rate until time t p = 8 hr and then shut in. This rate sequence introduces a
discontinuity into the problem. Pertinent data for the example is given in Table 1.

TABLE 1 – DATA FOR DRAWDOWN FOLLOWED BY BUILDUP

q B kh C rw S η
rbbl/d rbbl/stb md.ft bbl/psi ft md-psi/cp
736 1.00 1000.0 0.01 0.30 0.0 29300

Correa and Ramey (1986) provided the following solution in the Laplace domain for the wellbore pressure change for this
problem:

Δp wf =
141.2qBμ ( )
K 0 s η rw + S s η rw K1 s η rw 1 − e p ( )( − st
) . (12)
kh ⎧
s ⎨ s η rw K1 s η rw +
⎩
(
141.2Cμ
kh
)
s K 0 s η rw + S s η rw K1 [ ( ) ( ⎫
s η rw ⎬
⎭
)]
( )
The term 1 − e − stp in Eq. 12 introduces the effect of the step-rate-change at t p = 8 hr into the solution.
8 SPE 116255

Fig. 6 shows the results of the numerical inversion of the solution in Eq. 12 for the data in Table 1. As expected, the
Stehfest algorithm cannot accurately compute the behavior of the solution at the point of rate discontinuity ( t p = 8 hr). Chen
and Raghavan (1996) suggested a practical approach to handle this problem by using Stehfest algorithm. Their approach is
based on the superposition solution of the pressure buildup problem and requires the inversion of two drawdown solutions
with a time shift. This approach, however, is not convenient for all applications that involve step-rate changes. As shown in
Fig. 6, inversion by the Iseger algorithm yields accurate results and does not require any special treatment of the solution (we
used M = 1024 and nrp = 3 for this example).

900

800 Iseger
Stehfest
700

600
Δp, psi

500

400

300

200

100

0
0 2 4 6 8 10 12 14 16 18
Time, hr

Figure 6 – Numerical Inversion of wellbore pressure change for a combined drawdown and buildup sequence.

Deconvolution of Pressure Responses for a Sequence of Step-Rate Changes

To demonstrate the use of the Iseger algorithm in deconvolution applications, we consider the step-rate sequence and
corresponding pressure change in Figs. 4 and 5, respectively. The pertinent data for the example are given in Table 2. For this
variable-rate (step-rate changes) example, we used the Laplace domain deconvolution given in Eq. 6 to generate the constant
rate responses (with wellbore storage) shown in Fig. 7. The Laplace transform of the tabulated pressure responses, Δp w (t ) ,
were generated by the Roumboutsos-Stewart algorithm. The derivatives of the pressure responses were computed in the
Laplace domain for the Stehfest algorithm and numerically in the time domain for the Iseger algorithm.

TABLE 2 – DATA FOR DECONVOLUTION EXAMPLE

rw pi μ φ ct h kh S C B
ft psi cp fraction psi-1 ft md.ft bbl/psi rbbl/stb
0.3 5000 1.0 0.1 3.0E-6 30 1000 0.0 0.01 1.0

10.0000

1.0000
Δp and dΔp/dln t, psi

0.1000

0.0100
Isege, pressure
Iseger, pressure derivative
0.0010 Stehfest, pressure
Stehfest, pressure derivative

0.0001
0.0100 0.1000 1.0000 10.0000 100.0000
Time, hr
Figure 7 – Deconvolution of pressure responses for a step-rate sequence; comparison of Iseger and Stehfest algorithms.
SPE 116255 9

Fig. 7 shows the success of the Iseger algorithm in computing the inverse Laplace transform of the deconvolved pressures
(analysis of the deconvolved responses yields very accurate results with respect to the input data). Considering the poor
performance of the Stehfest algorithm in representing the Laplace transforms of the sampled rate and pressure functions
shown in Figs. 4 and 5, respectively, relatively smooth deconvolution results of the Stehfest algorithm may be unexpected.
This, however, is a result of the smoothing effect of the Stehfest algorithm mentioned earlier and is not necessarily an
indication of success. As shown by the derivative responses for the Stehfest algorithm, some important characteristics of the
data may be lost in the inversion. For example, the late-time behavior of the deconvolved pressures indicates pressure support
whereas the inversion results by the Iseger algorithm show the true nature of the infinite-acting flow regime.

Smoothing and Noisy Data

We have noted the smoothing effect of the Stehfest algorithm in the step-function examples above. To comment on the
smoothing effect of the Iseger algorithms, we reconsider the deconvolution example in Fig. 7 and add some noise to the input
data. In Figs. 8 and 9, we demonstrate the effects of 5% and 10% noise, respectively, on the inversion of deconvolution
results by the Iseger algorithm. These results indicate that, despite its accuracy, the Iseger algorithm does not smooth the
inverted function and may need some form of conditioning or regularization to work with noisy field data. It is, however,
interesting to note from Figs. 10 and 11 that the inversion by the Stehfest algorithm is less sensitive to noise. This shows that
some mild regularization combined with the smoothing effect of the Stehfest algorithm may display a stable deconvolution
result; but, as demonstrated in Fig. 7, accuracy may be questionable. Regularization, however, is outside the scope of this
paper and we have presented this discussion to highlight the perils of the most modern deconvolution algorithms which rely
heavily on regularization/conditioning techniques to generate stable results.

1.E+01

1.E+00
Δp and dΔp/dlnt, psi

1.E-01

Pressure, 0% noise

1.E-02 Pressure derivative, 0 % noise

Pressure, 5 % noise

Pressure derivative, 5% noise

1.E-03
1.E-02 1.E-01 1.E+00 1.E+01 1.E+02
Time, hr
Figure 8 – Deconvolution with 5% noise in the input; inversion by the Iseger algorithm.

1.E+01

1.E+00
Δp and dΔp/dlnt, psi

1.E-01

Pressure, 0 % noise

1.E-02 Pressure derivative, 0% noise

Pressure, 10% noise

Pressure derivative, 10% noise

1.E-03
1.E-02 1.E-01 1.E+00 1.E+01 1.E+02
Time, hr

Figure 9 – Deconvolution with 10% noise in the input; inversion by the Iseger algorithm.
10 SPE 116255

1.E+01

1.E+00

Δp and dΔp/dln t, psi

1.E-01

1.E-02
Pressure, 0% noise
Pressure derivative, 0% noise
1.E-03
Pressure, 5% noise
Pressure derivative, 5% noise
1.E-04
1.E-04 1.E-03 1.E-02 1.E-01 1.E+00 1.E+01 1.E+02

Time, hr
Figure 10 – Deconvolution with 5% noise in the input; inversion by the Stehfest algorithm.

1.E+01

1.E+00
Δp and dΔp/dln t, psi

1.E-01

1.E-02

1.E-03 Pressure, 0% noise

Pressure derivative, 0% noise

1.E-04 Pressure, 10% noise

Pressure derivative, 10% noise

1.E-05
1.E-04 1.E-03 1.E-02 1.E-01 1.E+00 1.E+01 1.E+02

Time, hr
Figure 11 – Deconvolution with 10% noise in the input; inversion by the Stehfest algorithm.

Field Examples
Here we present three field examples of deconvolution by using the Iseger algorithm. The first two examples, taken from
Meunier et al. (1985) and Fetkovich and Vienot (1984), are classical for the discussion of variable sandface flow rates due to
wellbore storage effects. The third example is a mini-DST application which represents a severe test for any deconvolution
and numerical Laplace inversion algorithm.

Sandface Rate Deconvolution – Muenuier et al., (1985) Example

In this example, we discuss the deconvolution of an eight-hour buildup test with afterflow effects considered by Muenuier et
al. (1985). This is a standard example to demonstrate that deconvolution can reveal the underlying dual-porosity behavior
masked by the wellbore storage effect. Fig. 12 shows the original and deconvolved data. We followed the suggestions of
Onur and Reynolds (1998) for the deconvolution of these data by using the Stehfest algorithm. At first sight, the Stehfest
algorithm seems to be more successful especially in generating smoother derivative responses (this is important to estimate
the storativity ratio from the depth of the valley on derivative responses). The results from the Iseger algorithm indicate a
deeper valley compared to the results from the Stehfest algorithm but may not permit an estimate of the storativity ratio.
The observations from Fig. 12 may seem to be in favor of the Stehfest algorithm but may be easily challenged by
considering the numerical inversions of the sandface rates and wellbore pressures shown in Figs. 13 and 14, respectively. For
these results, we obtained the Laplace transforms of the rates and pressures by using the Onur-Reynolds algorithm and then
inverted into the time domain by using the Stehfest and Iseger algorithms. The Iseger algorithm generates very accurate
inversions except at the very end. Although the magnitude of the errors may look relatively small for the Stehfest algorithm,
the changes in the characteristics of the functions are alarming because small changes in the rate function may cause
completely different deconvolution results (Lamm, 2000, and Ilk et. al. 2005). Therefore, considering the poor performace of
the Stehfest algorithm in recovering the original data, the smooth derivatives of the deconvolved responses in Fig. 12 may be
questioned as an artifact of the smoothing effect of the Stehfest algorithm.
SPE 116255 11

1000

100

Δp and dΔp/dln t, psi

10

0.1 Original Pressure Original pressure derivative

Iseger pressure Iseger pressure derivative
Stehfest pressure Stehfest pressure derivative

0.01
0.001 0.010 0.100 1.000 10.000
Time, hr
Figure 12 – Sandface-rate deconvolution to remove wellbore storage effect; Muenuier et al. (1985) example.

10000.0
qsf, rbbl/d

Original rate
Iseger
Stehfest
1000.0
0.1 1 10
Time, hr

Figure 13 – Numerical inversion of sandface rates from tabulated data; Muenuier et al. (1985) example.

1000

Original pressure
Iseger
Stehfest
Δp, psi

100
0.1 1 10
Time, hr
Figure 14 – Numerical inversion of pressure drop from tabulated data; Muenuier et al. (1985) example.
12 SPE 116255

Sandface Rate Deconvolution – Fetkovich and Vienot (1984) Example

In the previous example, the Iseger algorithm generated more oscillatory deconvolution responses compared to those from
the Stehfest algorithm. We attributed this result to the smoothing effect of the Stehfest algorithm. To demonstrate that the
Iseger algorithm does not always generate more oscillatory results, here we consider a buildup test with afterflow effects. The
well in this example is Phillips’ Oil Well Nr. 1 and is hydraulically fractured. Relevant information is given in the paper by
Fetkovich and Vienot (1984). Fig. 15 shows that both Stehfest and Iseger algorithms generate similar deconvolution results.
The results from the Iseger algorithm are slightly better in general and especially at early times.

10000
Original pressure Original pressure derivative
Iseger, pressure Iseger, pressure derivative
Stehfest, pressure Stehfest, pressure derivative
1000
Δp and dΔp/dln t, psi

100

10

1
0.010 0.100 1.000 10.000 100.000
Time, hr
Figure 15 – Numerical inversion of pressure drop from tabulated data; Fetkovich and Vienot (1984) example.

Mini-DST Example
This final example is a semi-synthetic mini-DST. We regenerated the rate and pressure measurements shown in Fig. 16 from
a field test by cleaning the inconsistencies, rescaling the pressures, and recalculating the rates from cumulative volumes.
Table 3 shows the input data for this mini-DST. This example tests the Iseger algorithm with a large set of data including a
large number of discontinuities.

4650 16

4600
12
Pressure, psi

Rate, stb/day

4550

4500

4
4450
Pressure
Rate
4400 0
0 1 2 3 4 5 6
Time, hr

Figure 16 – Rates and pressures recorded during a mini-DST test.

TABLE 3 – INPUT DATA AND PERMEABILITY ESTIMATES FOR THE MINI-DST TEST
rw pi μ φ ct h zw kh S C B kz/k
ft psi cp fraction psi-1 ft ft md,ft bbl/psi rbbl/stb
0.354167 4638.8 0.45 0.2 8.0E-6 4 2 27.2 0.0 4.41E-4 1.0 51.22
SPE 116255 13

Fig. 17 shows the deconvolution of the mini-DST data by using the Iseger algorithm. We have analyzed the deconvolved
constant-rate responses to obtain the estimates of k and kz/k shown in Table 3. These estimates are consistent with the
estimates from the analysis of the last buildup period (Fig. 16).

1.E+04
Pressure
1.E+03 Pressure derivative

1.E+02
Δp, dΔp/dln t, psi

1.E+01

1.E+00

1.E-01

1.E-02

1.E-03

1.E-04
1.E-04 1.E-03 1.E-02 1.E-01 1.E+00 1.E+01
Time, hr
Figure 17 – Deconvolution of mini-DST responses; 15000 inversion points.

To complete our discussion on Fig. 17, we need to comment on the scattering of the derivative responses at late times.
This behavior is a combined effect of the noise on the input data, accuracy of the Laplace transformation of sampled data, and
accuracy of the numerical inversion of the deconvolved results. We used 15000 inversion points in the Iseger algorithm to
generate the results in Fig. 17. We have found that using 40000 inversion points reduced the scattering of the derivatives at
late times but did not completely eliminate it (using 60000 and 80000 inversion points, however, only yielded marginal and
diminishing improvement).

Concluding Remarks
Laplace transformation is an important tool for the solution of transient fluid flow problems in porous media and the
existence of accurate numerical Laplace inversion algorithms greatly enhances its utility. The shortcomings of the common
numerical inversion algorithms in handling some practical conditions, however, have long been known. In this paper, we
have explored the potential of a new algorithm presented by Iseger in the solution of transient fluid flow problems. Our
interest in the Iseger algorithm was because of its claim to accurately invert functions with discontinuities. We have tested the
Iseger algorithm for problems involving the use of tabulated data in the Laplace transform domain, step-rate changes, build-
up tests following a drawdown period, and mini-DST tests with a sequence of drawdown and buildup periods. All of these
cases are known to have inversion problems with the commonly used Laplace inversion algorithms. The success of the Iseger
algorithm was very good in most cases and promising in the others. In standard applications, the results of the Iseger
algorithm were as good as or better than those of the Stehfest algorithm. In one of the deconvolution examples, we argued
that the more oscillatory results obtained with the Iseger algorithm could be more accurate than the seemingly more stable
results obtained with the Stehfest algorithm. We have, however, concluded that because of the high level of accuracy in
handling sharp changes of the function, the Iseger algorithm retaines the effect of noise in the inversion. Therefore, some
smoothing may be required for the successful application of the Iseger algorithm in the deconvolution of noisy field data.

Acknowledgment
This work has been conducted under Marathon Center of Excellence for Reservoir Studies (MCERS) at Colorado School of
mines. Support and consulting provided by Mr. Peter Den Iseger in converting his algorithm to FORTRAN and resolving
application problems have been greatly appreciated. The authors would like to thank Mr. Dilhan Ilk for his help to analyze
some of data using deconvolution. Thanks also go to Norwest Questa Engineering for the time dedicated to this work. Parts
of this work will appear in the PhD dissertations of Nab Al-Ajmi and Mahmood Ahmadi. Financial support for Nab Aljmi’s
graduate studies has been provided by the Kuwait Embassy.
14 SPE 116255

Nomenclature
C = wellbore storage coefficient, bbl/psi
ct = total compressibility, psi-1
f i = value of f (t ) at time ti
f (t ) = function
f (s ) = laplace transform of f (t )
f i′ = linear slope of the data in the interval ti ≤ t ≤ ti +1
h = thickness, ft
k = horizontal permeability, md
kz = vertical permeability, md
Ll = region of least square fitting for left-hand endpoint
Lr = region of least square fitting for right-hand endpoint
N = parameter used in the Stehfest algorithm
nrp = parameter for the Iseger algorithm
L = Laplace transform
L-1 = inverse Laplace transform
pi = initial reservoir pressure, psi
Δp(M , t ) = pressure drop at a point M and time t, psi
Δpc′ (M , t ) = time derivative of pressure drop at point M and time t for constant unit rate, psi
Δp w = wellbore pressure drop, psi
Δp w = Laplace transform of wellbore pressure drop, psi
q = production rate, rbbl/d, stb/d
q = production rate in laplace space, rbbl/d, stb/d
rw = wellbore radius, ft
s = Laplace-transform parameter
S = skin factor, dimensionless
t = time, hr
T = time period, hr
tp = producing time, hr
zw = distance from the center of perforated interval to the bottom of formation, ft

Greek
α = constant used in Eq. 9
α0 = constant used in Eq. 9
B = formation volume factor, rbbl/stb
β = constant used in Eq. 9
β0= constant used in Eq. 9
Δ = parameter used in the Iseger algorithm
φ = porosity of reservoir rock, fraction
μ = viscosity, cp
η = diffusivity constant, md-psi/cp
Г = gamma function

References
Abate, J. and Valkó, P. P. 2004. Multi-Precision Laplace Transform Inversion. Computers Math. Applic. 48: 629-636.
Bellman, R., Kalaba, R. E., and Lockett, J. A. 1966. Numerical Inversion of Laplace Transform: Applications to Biology, Economics,
Engineering, and Physics. American Elsevier Publishing Co. Inc., New York, NY.
Chen, C. C. and Raghavan, R.1996. An Approach to Handle Discontinuities by the Stehfest Algorithm. SPEJ (Dec.): 363.
Churchill, R. V. 1972. Operational Mathematics. 3rd Ed. McGraw-Hill, Inc. New York.
Coats, K. H., Rapoport, L. A., McCord, J. R., and Drews, W. P. 1964. Determination of Aquifer Influence Functions from Field Data. JPT
(Dec.): 1417-1424.
Cohen, A. M. 2007. Numerical Methods for Laplace Transform Inversion. Springer Science+Business Media LLC. New York.
Correa, A. C. and Ramey, H. J. Jr. 1986. Combined Effects of Shut-in and Production: Solution with a New Inner Boundary Condition.
Paper SPE 15579 presented at the SPE Annual Technical Conference and Exhibition, New Orleans, LA. 5-8 Oct.
Crump, K. S. 1976. Numerical Inversion of Laplace Transforms Using a Fourier Series Approximation. J. ACM 233: 89-96.
SPE 116255 15

Duhamel, J. M. C. 1833. Mémoire sur la méthode générale relative au mouvement de la chaleur dans les corps solides polongé dans les
milieux dont la température varie avec le temps. J. de Ec. Polyt. (Paris) 14: 20-77.
Fetkovich, M. J. and Vienot, M. E. 1984. Rate Normalization of Buildup Pressure by Using Afterflow Data. JPT (Dec.): 2211-2224.
Gaver, D. P. 1966. Observing Stochastic Process and Approximate Transform Inversions. Oper. Res. 14, 3: 459.
Ilk, D., Valkó, P. P., and Blasingame, T. A. 2005. Deconvolution of Variable-Rate Reservoir Performance Data Using B-Splines. Paper
SPE 95571 presented at the SPE Annual Technical Conference and Exhibition. Dallas, TX. 9-12 Oct.
Iseger, P. D. 2006. Numerical Transform Inversion Using Gaussian Quadrature. Probability in the Engineering and Informational Sciences,
20. 1-44.
Kryzhniy V.V. 2004. High-resolution exponential analysis via regularized numerical inversion of Laplace transforms. J. Comp. Phys.,
Vol.199/2, 618- 630
Kryzhniy, V. V. 2006. Numerical Inversion of the Laplace Transform: Analysis via Regularized Analytic Continuation.
http://laplacetransform.org/.
Kuchuk, F. J., Carter, R. G., and Ayesteran, L. 1990. Deconvolution of Wellbore Pressure and Flow Rate. SPEFE (March): 53-59.
Lamm, P. K. 2000. A Survey of Regularization Methods for First-Kind Volterra Equations. Surveys on Solution Methods for Inverse
Problems. ed. D. Colton, H. W. Engl, A. Louis, J. R. McLaughlin, W. Rundell. Springer, Vienna and New York: 53-82.
LePage, W. R. 1961. Complex Variables and the Laplace Transform for Engineers. McGraw-Hill Inc. New York.
Levitan, M. M. 2003. Practical Application of Pressure-Rate Deconvolution to Analysis of Real Well Tests. Paper SPE 84290 presented at
the SPE Annual Technical Conference and Exhibition, Denver, CO. 5-8 Oct.
Meunier, D., Wittmann, M. J., and Stewart, G. 1985. Interpretation of Pressure Buildup Test Using In-Situ Measurement of Afterflow. JPT
(Jan.): 143-152.
Onur, M. and Reynolds, A. C. 1998. Numerical Laplace Transformation of Sampled Data for Well-Test Analysis. SPEREE (June): 268-
277.
Ozkan, E. and Raghavan, R. 1997. Some Strategies to Apply the Stehfest Algorithm for a Tabulated Set of Numbers. SPE Journal. 2
(Sept.): 363-372.
Roumboutsos, A. and Stewart, G. 1988. A Direct Deconvolution or Convolution Algorithm for Well-Test Analysis. Paper SPE 18157
presented at the SPE Annual Technical Conference and Exhibition. Houston, TX. 2-5 Oct.
Stehfest, H. 1970. Numerical Inversion of Laplace Transforms. Communications, ACM 13 (1): 47–49.
Talbot, A. 1979. The Accurate Numerical Inversion of Laplace Transforms, J. Inst. Maths. Applics., 23: 97-120.
Valkó, P. P. and Abate, J. 2004. Comparison of Sequence Accelerators for the Gaver Method of Numerical Laplace Transforms Inversions.
International Journal for Numerical Methods in Engineering. 60: 979-993.
van Everdingen, A. F. and Hurst, W. 1953. The Application of the LaPlace Transformation to Flow Problems in Reservoirs, Trans. AIME
198: 171-176.
Von Schroeter, T., Hollaender, F., and Gringarten, A. C. 2004. Deconvolution of Well-Test Data as a Nonlinear Total Least-Squares
Problem. SPEJ (Sept.): 375-390.