Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Geophys. J. Int. (2021) 225, 68–84 doi: 10.

1093/gji/ggaa579
Advance Access publication 2020 December 22
GJI Seismology

Frequency domain full-waveform inversion in a fluid-saturated


poroelastic medium
1
Qingjie Yang and Alison Malcolm2
1 Department of Earth Science, Khalifa University, Abu Dhabi, 127788, UAE. E-mail: qingjie.yang@ku.ac.ae
2 Department of Earth Science, Memorial University of Newfoundland, St. John’s, NL, A1B3X9, Canada

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
Accepted 2020 December 3. Received 2020 November 27; in original form 2020 July 8

SUMMARY
Determining subsurface properties is of fundamental importance in exploration seismic imag-
ing. Poroelasticity theory provides an opportunity to extract quantitative fluid- and attenuation-
dependent properties from seismic data. Following Pratt’s frequency-domain full-waveform-
inversion (FWI) procedure and extending the basic FWI equations from the elastic case to
the poroelastic case, we implement poroelastic FWI (PFWI) of fluid-saturated porous media.
By analysing the sensitivity kernels of poroelastic parameters, we explain the reason why
some parameters are more difficult to recover than others. We also show the analytical and
numerical radiation patterns based on which we predict the trade-offs among parameters. In
numerical experiments, we invert two models to demonstrate the feasibility and effectiveness
of the proposed PFWI and to verify our predictions about trade-offs for two-parameter PFWI.
Finally, we discuss the various factors that influence the inversion results.
Key words: Inversion; Attenuation; Fluids in rocks; Sensitivity; Permeability; Porosity.

1 I N T RO D U C T I O N
Using full-waveform inversion (FWI) to recover subsurface properties from seismograms is a challenging problem in both academia and
industry (Virieux & Operto 2009; Virieux et al. 2017). Because it makes use of all the available information in the observed data, this inversion
technique is expected to provide as much information as possible about fluid-filled rocks, such as porosity, solid frame moduli, fluid phase
properties and saturation. These rock physics coefficients are linked to the recorded seismic data via Biot theory (Biot 1956a, 1956b, 1962)
and its extensions (Johnson et al. 1987; Carcione 1996; Pride et al. 2004; Pride 2005; Liu et al. 2018) provide a complete framework. These
poroelastic theories successfully predict the existence of the slow compressional wave observed in the laboratory (Plona 1980). In spite of
inaccuracies in predicting both amplitude attenuation and velocity dispersion (Carcione et al. 2010), Biot theory is still the most popular and
the most important theory to describe the propagation of elastic waves in fluid-saturated porous media.
Compared to elastic or viscoelastic theories, poroelastic theory not only considers the properties of the solid rock, but also takes into
account the properties of the fluid and fluid–solid interactions. In real data, wave-induced fluid flow plays a non-negligible role in seismic
attenuation. For example, Chotiros (2002) finds that the predictions of a viscoelastic model are inconsistent with measured parameters of
water-saturated sand in the laboratory, but the poroelastic Biot model does better. Morency et al. (2011) demonstrate that there is a poroelastic
signature in time-lapse migration data collected for a synthetic CO2 sequestration monitoring project, and that this signature cannot be
captured using acoustic or elastic models. These results indicate that poroelastic theory is necessary to recover all of the relevant physical
properties of reservoir rocks. Although, the visco-acoustic or visco-elastic FWI can invert for both velocities and quality factors, these must
be further processed to discriminate between fluids or to infer information about solid and fluid properties. Generally, these inferences rely
upon empirical relations to link the velocities and quality factors with various properties of a porous medium. However, these empirical
relationships are limited and depend on the rock type (Dupuy et al. 2016b). An extensive compilation of velocity-porosity empirical relations
is presented by Mavko et al. (2009, chapter 7). But in these relations, only porosity affects the P-wave velocity, whereas in real media, other
rock properties will also affect velocities. Consequently, to avoid the limitations associated with the use of empirical relations, we study the
possibility of directly inverting poroelastic parameters by FWI from the seismograms. That said, the estimation of the properties of fluid-filled
reservoir rocks is challenging (Morency et al. 2009; De Barros et al. 2010). The difficulties of the traditional FWI, of course, are still present
in poroelastic FWI (PFWI). Further the additional parameters in PFWI and the more complex mechanisms in poroelasticity result in stronger
non-linearity of the inversion problem.


C The Author(s) 2020. Published by Oxford University Press on behalf of The Royal Astronomical Society. All rights reserved. For
68 permissions, please e-mail: journals.permissions@oup.com
Poroelastic FWI in frequency-domain 69

Recently, FWI has been successfully applied using (visco-) acoustic and (visco-) elastic, isotropic and anisotropic rheologies, in 2-D and
3-D geometries, both on synthetic and field data (Warner et al. 2008; Brossier et al. 2009; Zhu et al. 2012; Gholami et al. 2013a, b; Operto
et al. 2015; Bozdağ et al. 2016; Sun et al. 2017; Keating & Innanen 2019). These promising results encourage more development of FWI in
more complicated media, like poroelastic media. For example, De Barros & Dietrich (2008) compute the Fré chet derivatives of poroelastic
waves in the plane-wave domain using the Born approximation. Morency et al. (2009) derive finite-frequency sensitivity kernels associated
with wave propagation in porous media based on adjoint methods. Dupuy et al. (2016a, b) use a two-step inverse approach, time-lapse FWI
plus rock physics inversion, to estimate rock physics properties. Yang et al. (2019) present an analytical derivation of radiation patterns for
scattered waves due to different parameter perturbations in porous media. The work most closely related to that proposed here is the estimation
of the poroelastic properties of reservoir rocks developed by De Barros et al. (2010), who apply a frequency–wavenumber domain inversion
algorithm to estimate the poroelastic parameters of horizontally layered media. However, as they discuss in their paper, their assumption of
plane-layered media is too simple to correctly describe the structural features of realistic geologic media. The PFWI we present can invert

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
for arbitrarily heterogeneous models and is not limited by the empirical relationships between the rock properties and the elastic parameters
used in two-step inversion methods.
The objective of this paper is to extend frequency-domain FWI from acoustic and elastic media to fluid-saturated poroelastic media. In
the frequency domain, it is easy to execute a multiscale inversion that is routinely used to mitigate the non-linearity of FWI. So we implement
PFWI in the frequency domain by making small modifications to the algorithm proposed by Pratt et al. (1998). A second advantage of the
frequency domain is that we can easily implement the convolution that describes the attenuation mechanisms in the generalized Darcy’s law
(Carcione et al. 2010; Dupuy et al. 2011).
Before carrying out PFWI, we calculate the sensitivity kernels independently for nine parameters included in the poroelastic model and
compare them with their corresponding radiation patterns derived by Yang et al. (2019). By analysing the sensitivity kernels and radiation
patterns, following similar analyses done in Pan et al. (2018), we find that they are internally consistent. The radiation patterns present the
waveform variations with opening angle of the scattered wavefields generated by perturbing one parameter at a single point. The directions in
which the scattered fields overlap result in trade-offs between parameters. The study of radiation patterns in poroelastic media is important,
because one can visually assess to what extent more than one parameter can be inverted simultaneously.
The paper is organized as follows. In the next section, we outline the dynamic poroelastic model and introduce the constitutive parameters.
In Section 3, we extend the frequency-domain FWI workflow proposed by Pratt et al. (1998) to poroelastic media and review the optimization
algorithm; in Section 4, we examine the sensitivity kernels for each parameter independently to assess the possibility of recovering each
parameter, and compare them with their corresponding radiation patterns; in Section 5, we present the results of inverting for each parameter
independently in two models; in Section 6, we implement two-parameter inversion for six pairs of parameters in two models, and discuss the
factors that influence the accuracy of these results and in the final section, we discuss all results to conclude the paper.

2 WAV E P R O PA G AT I O N I N F L U I D - S AT U R AT E D P O R O U S M E D I A

2.1 Poroelastic wave equations


Since Biot (1956a, b, 1962) first rigorously established the fundamental equations of wave propagation in a fluid-saturated porous medium,
many researchers (Levy 1979; Burridge & Keller 1981; Pride et al. 1992) have demonstrated that Biot theory is an accurate general model
for wave propagation in porous media. Assuming an e−iwt time dependence, based upon the absolute solid and fluid displacements u and U,
Biot’s equations of poroelasticity in the frequency-domain (Pride 2005; Dupuy et al. 2011; Yang & Mao 2017; Yang et al. 2019) are


⎪ −ω2 (ρ1 u + ρ2 U) = ∇ · τ,
⎨ τ = (λ − φC)∇ · uI + G ∇u + (∇u)T  + φC∇ · UI,

u  (1)

⎪ −ω2 (ρ f − φ ρ )u + φ
ρ U = −∇ P,

⎩ P = (C − φ M)∇ · u + φ M∇ · U,

where the stress tensor and the fluid pressure are denoted by τ and P, ω is the angular frequency, I is the identity matrix, φ is the porosity
of the material, λu is the undrained Lamé modulus, and G is the shear modulus (same for both undrained and drained conditions). The other
two poroelastic constants used here are Biot’s coupling modulus C, and the modulus M (also called the fluid storage coefficient, because it
represents how much fluid can accumulate in a sample of constant size when the fluid-pressure changes). For fluid-saturated porous materials,
these moduli or constants are exactly expressed in terms of the undrained bulk modulus Ku , the mineral modulus of the grains Ks , the fluid
modulus Kf , and the drained bulk modulus Kd through the Biot–Gassmann relations (Gassmann 1951; Biot & Willis 1957):
2 2
λu = K u − G = K d + α 2 M − G, (2)
3 3
C = α M, (3)

K f Ks
M= , (4)
(α − φ)K f + φ K s
70 Q. Yang and A. Malcolm

Table 1. Constitutive rock parameters of a poroelastic medium.


Parameter Value (Unit) Physical interpretation
Ks 30 (GPa) Grain bulk modulus
Kd 6.21 (GPa) Drained frame bulk modulus
G 4.55 (GPa) Frame shear modulus
ρs 2650 (kg m–3 ) Grain density
Kf 2.0 (GPa) Fuild bulk modulus
ρf 1000 (kg m–3 ) Fuild density
η 0.1 (Pa·s) Fuild viscosity
κ0 10−9 (m2 ) Static permeability
φ 0.2 (-) Porosity

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
where α is the so-called “Biot-Willis constant”, named for Biot & Willis (1957), defined as α = 1 − Kd /Ks (independent of fluid type). For
isotropic media, the drained bulk modulus can also be written as Kd = λ + 2μ/3, so that eq. (2) becomes λu = λ + α 2 M.
In eq. (1), the terms related to density are

ρ1 = (1 − φ)ρs , ρ2 = φρ f , (5)

where ρ s , ρ f are the density of the solid grains and the fluid, respectively. The third expression in eq. (1) is a variation of the frequency-domain
version of Darcy’s law, which not only allows for flow due to induced pressure gradients, but also accounts for flow created by the acceleration
of the framework of grains. The reader is referred to Masson & Pride (2010) for the time-domain version of Darcy’s law. The relative-flow
resistance ρ and the dynamic permeability κ(ω) in Darcy’s law (Johnson et al. 1987) are given by


ρ= , (6)
ω · κ(ω)
−1
ω ω
κ(ω) = κ0 1−i −i . (7)
2ω B ωB
The relative-flow resistance  ρ describes the dynamic loss of energy due to the fluid flow with explicit frequency dependence. It is
responsible for the intrinsic scattering of waves in the Biot poroelasticity theory (Biot 1956a, 1956b). Eqs (6–7) describe the attenuation
behavior of poroelastic waves, where η is the viscosity of the fluid, κ 0 is the intrinsic permeability, ωB is the characteristic or relaxation angular
frequency defined as ω B = ρηφ f κ0
. The dynamic permeability κ(ω) proposed by Johnson et al. (1987) connects the two frequency ranges, which
are the low frequency range (smaller than ωB ) where viscous friction effects are dominant, and the high frequency range (larger than ωB )
where inertial coupling effects prevail. Using this dynamic permeability we can use one unified system to describe the seismic wave behavior
over the entire frequency band (Pride 2005; Dupuy et al. 2011; Yang & Mao 2017; Liu et al. 2019).
Returning to eq. (2), Pride (2005) shows another mechanism to compute the drained moduli, also called the effective moduli, as
cs +1
K d = K s 1+c
1−φ

, and G = G s 1+1−φ3 c φ by including a consolidation coefficient cs , so the Biot–Willis constant α = 1 +c becomes independent
2 s φ s
of the properties of the solid grains. The parametrization including the consolidation coefficient cs is used by De Barros & Dietrich (2008)
and Morency et al. (2009). Yang et al. (2019) derive the radiation patterns for four of these parametrizations and suggest that we choose the
parametrization with the smallest overlap of radiation patterns between parameters during multiparameter inversion. Morency et al. (2009)
also consider two alternative parametrizations to present their finite-frequency kernels in porous media. All of the parameters we use in this
paper are measurable in the lab and are listed in Table 1. They are consistent with the (Kd , G) parametrization of Yang et al. (2019). In our
examples, we choose parameter values that generate all of the necessary wave phenomena for us to use in the inversion.

2.2 Forward modelling


Seismic data are non-linearly related to rock parameters through the wave equation given in eq. (1). As mentioned above, the introduction of
a complex dynamic permeability (eqs 5–6) makes it more natural to solve the forward problem in the frequency domain (Dupuy et al. 2011).
Therefore, we use the frequency-domain weighted-averaging finite-difference method proposed by Min et al. (2000) and extended to solve
the poroelastic wave equations by Yang & Mao (2017). This approach uses a 25-point stencil to accurately calculate the spatial derivatives of
displacements and models the full seismic wavefield across all frequencies. In this paper, we focus on the low frequency regime where the
slow-P (Biot) wave is strongly diffusive and attenuated so that the seismic acquisition system does not record it.
Fluid-saturated, isotropic, poroelastic media are described by the nine independent intrinsic parameters shown in Table 1. In the
frequency-domain, the poroelastic wave eq. (1) reduces to a system of linear equations. This system can be written compactly as

Aw = f, (8)
Poroelastic FWI in frequency-domain 71

where A is the impedance matrix, w is the poroelastic wavefield vector, and f denotes the source term. Explicitly, they are
⎛ 1,u x 1,u z 1,Uz
⎞ ⎛ ⎞
A(m,ω) A(m,ω) A1,U x
(m,ω) A(m,ω) ux
⎜A2,u x A2,u z A2,Ux A2,Uz ⎟ ⎜u ⎟
⎜ (m,ω) (m,ω) (m,ω) (m,ω) ⎟ ⎜ z⎟
A = ⎜ 3,u x 3,Uz ⎟ , w = ⎜ ⎟, (9)
⎝A(m,ω) A3,u(m,ω) A(m,ω) A(m,ω) ⎠
z 3,Ux
⎝Ux ⎠
4,u z 4,Uz
A4,u x 4,Ux
(m,ω) A(m,ω) A(m,ω) A(m,ω)
Uz
l,w
where A(m,ω)k
indicates the coefficient matrix of the kth wavefield component (w1 = ux , w2 = uz , w3 = Ux , w4 = Uz ) in the lth equation of eq.
(1). All blocks are sparse, diagonally dominant, complex-valued and frequency-dependent, but not symmetric due to the absorbing boundary
condition (Min et al. 2000). The structure of A is based on a 25-point stencil illustrated in fig. 4 of Yang & Mao (2017). To highlight which
l,wk
parameters each block A(m,ω) depends on, we write out their functional dependencies in Appendix A, which is an extension of eq. (9). For 2-D
forward modelling, eq. (8) can be solved by LU factorization which is efficient for multiple sources (Hustedt et al. 2004; Virieux & Operto

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
2009). The effectiveness and accuracy of the modelling method are demonstrated by Yang & Mao (2017).

3 FWI METHODOLOGY
The theory of FWI is presented by many authors (Tarantola 1984, 1986; Mora 1987; Koren et al. 1991; Pratt et al. 1996; Sambridge &
Mosegaard 2002; Virieux & Operto 2009; Sourbier et al. 2009). We follow the compact matrix formalism proposed by Pratt et al. (1998),
extending it to poroelastic media by making some small modifications. As in acoustic and elastic media, the PFWI problem is solved by
iteratively minimizing the misfit function E in a least-squares form
1 t ∗
E(m) = δw δw , (10)
2
where m refers to the model parameters in a given parametrization. The superscript t denotes the matrix transpose; and ∗ denotes the complex
conjugate (refer to eq. 8 in Pratt et al. 1998). But, the key difference between the poroelastic case and the acoustic or elastic case is that,
the poroelastic case should include both solid and fluid components in the residual vector, that is, δw = [δu, δU]t , the difference between
the synthetic data wsyn and the observed data wobs . As far as we know, the fluid residual δU is unavailable in the observed data wobs as no
instruments are capable of measuring fluid displacements within the pore-space. So, we just consider solid residual in the misfit function, that
is, δw = δu.
Using a least-squares algorithm, the model is updated along the gradient direction via

m(k+1) = m(k) + β (k) δm, (11)

where β is the step length in the kth iteration, and δm is the model perturbation or update direction. The role of the step length is to scale the
(k)

units of the model perturbation vector to model dimensions. We apply the parabolic fitting method (Vigh et al. 2009) to get the optimal step
length. The details of this calculation can be found in Liu et al. (2013). Eq. (11) describes iterative gradient descent, which is the same for the
acoustic, elastic or poroelastic cases. We solve for the model perturbation using the l-BFGS method (Nocedal & Wright 1999, pp. 177–180;
Brossier et al. 2009). According to Nash (2000), l-BFGS becomes more effective as the optimization problem becomes more nonlinear,
making it ideally suited to our problem.
As in the acoustic and elastic cases, the model perturbation δm is along the gradient direction ∇ m E evaluated by taking partial derivatives
of the misfit equation with respect to the model parameters m, yielding
∂E
δm = ∇m E = = {Jt δu∗ }, (12)
∂m
where R{} indicates the real part, and J denotes the Jacobian matrix or the sensitivity matrix. Explicitly, the gradient in the poroelastic case
can be expanded to
⎧ ⎡ ⎤⎫

⎪ δu∗x ⎪

⎪
⎨ ⎪
 ⎢δu∗ ⎥⎬
∂ux ∂uz ∂Ux ∂Uz ⎢
t t t t
z⎥
(∇m E)i =  ⎢ ⎥ , (13)


∂m i ∂m i ∂m i ∂m i ⎣ 0 ⎦⎪


⎩ ⎪

0
which corresponds to eq. (14) in Pratt et al. (1998), the difference is the two ‘0’ vectors where there should be fluid-phase residuals. We
assume that there are N gridpoints and n receivers in our data set, each component of the residual then has dimension N × 1 and n nonzero
values at the receiver positions. This formulation does highlight that the inversion problem for the poroelastic case is quite similar to that for
the acoustic or elastic case, with the key differences being the size of the wavefield vector (here we have both the solid and fluid residual) and
additional parameters in the model. Obviously, calculating the gradient using eq. (13) is not cost-efficient. Therefore, we calculate gradients
with the adjoint-state method as done by eq. (25) in Pratt et al. (1998). In the poroelastic case, the gradient is
  t
∂A
(∇m E)i =  wt v , (14)
∂m i
72 Q. Yang and A. Malcolm

where

v = A−1 δu∗ (15)

is the adjoint wavefield or the ‘backpropagated wavefield’, and the column of A−1 corresponds to the Green’s functions for unit im-
pulse sources located at the receiver positions where the data residuals δu are recorded. ∂A/∂m i is the derivative of the impedance
matrix A with respect to mi (one parameter of m at the ith gridpoint) and is extremely sparse because mi is supported only at, or
l,wk
near, the diagonal of some block matrices A(m,ω) . Explicitly, each element of the gradient in the poroelastic case can be expressed
by
⎧ ⎡ ⎤⎫

⎪ ǔx ⎪ ⎪

⎨   ⎪
 t t t t  ∂At ⎢ ⎥⎬

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
⎢ ǔz ⎥
(∇m E)i =  ux , uz , Ux , Uz ⎢ ⎥ , (16)

⎪ ∂m i ⎣Ǔx ⎦⎪ ⎪

⎩ ⎪

Ǔz

where ǔx ǔz Ǔx Ǔz are the components of the adjoint wavefield v which is driven by the solid-phase residual only (In other words, δu from
eq. (15) is [δux , δuz , 0, 0]).
The matrix ∂A/∂m i controls the radiation pattern related to the parameter mi resulting from a point perturbation in this parameter located
at position i (Pratt et al. 1998, Virieux & Operto 2009). Radiation patterns are well studied for isotropic acoustic and elastic waves (Wu &
Aki 1985; Tarantola 1986; Forgues & Lambaré 1997), viscoelastic waves (Ribodetti & Virieux 1996), anisotropic elastic waves (Pan et al.
2016; He & Plessix 2017) and poroelastic waves (Yang et al. 2019).
In summary, the procedure of PFWI we follow is the same as that of Pratt et al. (1998) except for the modelled wavefields and the data
residuals. These differences are responsible for the significant computational cost from the modelling of poroelastic theory. The combination
of two phases of data and many parameters leads us to expect PFWI to be more nonlinear and more likely to fall into local minima than elastic
FWI is, because with more wavefields there are more chances of cycle skipping and with more parameters there are more models that may fit
the data equally well.
In this paper, we consider the nine parameters shown in Table 1 and try to recover them individually and in pairs. Before dis-
cussing the inversion, we calculate the sensitivity kernels of these nine parameters to assess the expected capabilities and resolution of our
PFWI.

4 SENSITIVITY KERNELS
From the two gradient expressions eqs (12) and (14), the element of the sensitivity kernel or the Jacobian matrix with respect to parameter m
is,
 t
∂A
Jit = wt Gr , i = 1, 2, . . . N , (17)
∂m i

where Gr = A−1 δr indicates the Green’s function generated by an unit impulsive source δ r located at receiver position r. These sensitivity
kernels describe the area where the source wavefield (w) and the Green’s function from a receiver (Gr ) interfere (Operto et al. 2006; Liu et al.
2009; Sun et al. 2017). The central zone of the sensitivity kernel is the first Fresnel zone which illustrates the large wavelength structure, the
outer fringes are isochrones that describe the short wavelength structure, the width of which give some insight into the vertical resolution of
FWI (Virieux & Opterto 2009).
Using eq. (17), we compute the sensitivity kernels in a 500 m × 500 m homogeneous fluid-saturated porous medium with grid
intervals of 5 m, with constant parameters given in Table 1. As listed in Table 1, the characteristic or relaxation frequency fB is 3183 Hz,
which is much higher than the seismic band (100 to 102 Hz). Our computation frequency is 25 Hz, so the slow-P wave vanishes due to
strong dissipation and dispersion. We consider one source and one receiver located at (250 m, 100 m) and (250 m, 400 m), respectively.
These two simulations (w and Gr ) have the same impedance matrix A and are solved in the same way, but with different sources, the
forward simulation uses an explosive source, the back-propagated adjoint simulation uses a point force source in the solid phase, in
order to be consistent with the inversion procedure and the expected receiver model. Fig. 1 shows the real parts of the frequency domain
sensitivity kernels of the nine parameters given in Table 1, after normalizing each plot by its individual maximum absolute value. The
amplitudes of the sensitivity kernels for different parameters are inversely proportional to the values of the parameters. So, the magnitudes
of the kernels for the moduli are much smaller than those of other parameters before normalizing. This can be seen from eq. (17), for
different parameters, the only difference in kernel expressions is ∂At /∂m i which is approximatively inversely proportional to mi . During the
inversion, the step length scales the gradients, so we compare the kernels after normalizing to understand the update directions, ignoring this
scaling.
Poroelastic FWI in frequency-domain 73

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
Figure 1. The frequency domain sensitivity kernels of the nine parameters listed in Table 1. (a) the solid-phase, (b) the fluid-phase. (1) Kd , (2) G, (3) ρ s , (4)
ρ f , (5) Ks , (6) Kf , (7) φ, (8) η, (9) κ 0 . They are normalized and displayed on a scale of [–1, 1].

Expanding eq. (17), gives


∂u x ∂u z ∂Ux ∂Uz
Jit = + + +
∂m i ∂m i ∂m i ∂m i
! "# $ ! "# $
solid− phase f luid− phase

%4 % 4 l,wk
∂A(m,ω)
= ( wk )G rwl
l=1 k=1
∂m i
1,u 1,U
∂A1,u x
(m,ω) ∂A(m,ω)
z
∂A1,U x
(m,ω) ∂A(m,ω)
z

= (u x + uz + Ux + Uz )G ru x
∂m i ∂m i ∂m i ∂m i
2,u 2,U
∂A2,u x
(m,ω) ∂A(m,ω)
z
∂A2,U x
(m,ω) ∂A(m,ω)
z

+ (u x + uz + Ux + Uz )G ru z
∂m i ∂m i ∂m i ∂m i
3,u 3,U
∂A3,u x
(m,ω) ∂A(m,ω)
z
∂A3,U x
(m,ω) ∂A(m,ω)
z

+ (u x + uz + Ux + Uz )G rUx
∂m i ∂m i ∂m i ∂m i
4,u 4,U
∂A4,u x
(m,ω) ∂A(m,ω)
z
∂A4,U x
(m,ω) ∂A(m,ω)
z

+ (u x + uz + Ux + Uz )G rUz , (18)
∂m i ∂m i ∂m i ∂m i
where G rwl represents G ru x , G r z , G rUx and G r z , the components of Gr . For eq. (18), we observe that the kernel can be separated into the solid-
u U

(the first two terms) and fluid- (the last two terms) phase parts as annotated. So we show solid- and fluid-phase kernels separately in Fig. 1.
From Fig. 1 we note several interesting features of the sensitivity kernels:
1.The Kd , Ks and Kf (Figs 1-a1, b1, a5, b5, a6, b6), kernels are ellipse-shaped with foci located at the source and receiver positions. The other
six parameters have more complex kernels especially in their first Fresnel zones which have less regular shapes with quite low amplitude
along the direct wave path.
2.For G and ρ s , their fluid-phase kernels (Figs 1-b2, b3) are zero.
3.For η and κ 0 , their solid-phase kernels (Figs 1-a8, a9) are zero.
4.For Kd and φ (Figs 1-a1, b1, a7, b7), the solid and fluid kernels have opposite vibration directions.
5.For φ, η and κ 0 (Figs 1-a7, b7, b8, b9), the kernels have much weaker amplitudes than those for other parameters, this is particularly
pronounced for η and κ 0 .
74 Q. Yang and A. Malcolm

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
Figure 2. The analytical (1) and numerical (2) radiation patterns of nine parameters listed in Table 1 and the max values of numerical scattered wavefields.
Some parameters have the same radiation patterns; we show only those that are unique. (a) Kd , Ks , and Kf ; (b) G; (c) ρ s , ρ f , η and κ 0 ; (d) φ.

6.The shapes of the kernels for porosity φ (Figs 1-a7, b7) are more elliptical with a little bit of a saw-tooth shape. And the shape of the kernel
of G (Fig. 1-a2) is unique with a stranger saw-tooth shape.

From eq. (18), as mentioned above, the sensitivity kernels and the radiation patterns are closely related because both are controlled by
the derivative matrix ∂A/∂m i . The radiation patterns of poroelastic media are presented by Yang et al. (2019). We recalculate those analytical
and numerical radiation patterns using the parameters listed in Table 1 and show them in Fig. 2. Some parameters have the same scattering
wavefronts but different amplitudes of radiation patterns, we mark these parameters in Fig. 2. In Fig. 2(a), we observe that perturbations in Kd ,
Ks and Kf only generate scattered P waves, while perturbations in the other six parameters generate scattered P and S waves simultaneously.
By comparing each parameter’s radiation pattern with its sensitivity kernel, we infer that in the first Fresnel zones the elliptical shape is
caused by P waves and the saw-tooth shape comes from the S waves. In Fig. 2(b), we observe that the S-wave scattered by perturbations
in G is stronger than the scattered P wave, both P and S waves have four sidelobes which is unique and consistent with the unique spatial
distribution of the kernel for G shown in Fig. 1(b). In Fig. 2(c), the scattered S-wave dominates in the radiation patterns of ρ s , ρ f , η and κ 0 , so
the saw-tooth shape is clearly present in their kernels (see Figs 1c, d, h and i). But in Fig. 2(d), the opposite is true for φ where the elliptical
shape is dominant. So we find that the sensitivity kernels of the poroelastic parameters agree perfectly with their corresponding radiation
patterns. It is worth noting that the energy of the scattered wave due to perturbations in η and κ 0 is at least three orders of magnitude smaller
than the others, which means the observed wavefields are not sensitive to η and κ 0 perturbations.
3,wl 4,wl
The four empty kernels can be explained if we look carefully at eq. (A1), where we note that A(m,ω) , and A(m,ω) are not functions of G
∂Uz ∂Uz
do not depend on η and κ 0 , which implies that ∂U ∂Ux
0, ∂u = ∂u
1,wl 2,wl
and ρ s , and similarly A(m,ω) and A(m,ω) ∂G
x
= ∂G
= 0, ∂ρs
= ∂ρs
= ∂η
x
∂η
z
= 0 and
∂u x ∂u z
∂κ0
= ∂κ0
= 0.
Poroelastic FWI in frequency-domain 75

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
Figure 3. Simple box model and acquisition geometry. The stars represent sources, the dash-line represents the receiver distribution surrounding the box to
provide complete seismic illumination.

5 O N E - PA R A M E T E R I N V E R S I O N

To check the feasibility of our PFWI algorithm and to determine the accuracy of the inversion procedure described above, we test our
procedure on two numerical experiments: a simple box model, and a modified Marmousi model, for each parameter independently (i.e. we
fix all other parameters to their true values and invert for each parameter on its own).

5.1 Simple box model

We first use a simple box model shown in Fig. 3. The model is 500 m × 500 m with a grid interval of 10 m. We place an array of 10 sources
on the top, at 30 m depth and 200 receivers distributed across each boundary of the model to represent an ideal acquisition geometry. The
parameter value inside the box is 10 percent higher than its background value for all parameters. When estimating each parameter, we use the
model without the box as the initial model, and assume that the other eight parameters are known perfectly and equal to the constants shown
in Table 1.
In our PFWI procedure, we use the l-BFGS method to compute the model update δm, in which we store 4 previous gradients of the
previous iterations. We invert six frequencies (5, 15, 25 35, 45, 55 Hz) in sequence from low to high.
The reconstructed results for each of the nine parameters are displayed in Fig. 4. We extract the vertical profiles from the middle position
to compare our results with the true models and initial models. The final results for Kd , G, ρ s , ρ f , Ks and Kf (Figs 4a–f) are very accurate
with clear boundaries, correct amplitudes, and only small artefacts at the bottom of the models because of the inhomogeneous distribution of
sources. But the inversion for η, and κ 0 (Figs 4g, h) fail even with this ideal acquisition geometry. The porosity φ (Fig. 4k) is not accurately
recovered. Although there is a clear outline of the box, there is also strong noise around the box. Checking the procedure of inverting for
porosity φ we find the data are not sensitive to this parameter at low frequencies, the first updating happens at 25 Hz. So we speculate that φ
could be recovered better if we use an initial model with more information at large wavelengths. In summary, in this ideal case we find that we
can recover 6 of the 9 parameters of the poroelastic model. We can recover a seventh parameter, φ, with less accuracy and we cannot recover
η and κ 0 . We cannot recover these last two parameters because perturbations in these parameters generate only very weak scattered energy
as shown in the radiation patterns (Fig. 2) and also as predicted by De Barros & Dietrich (2008) and Morency et al. (2009). De Barros et al.
(2010) also do not recover these two parameters using the frequency–wavenumber domain FWI for the same reason.

5.2 Modified Marmousi model


Using a more complicated model to test our PFWI algorithm also gives some insight into the factors governing the resolution of PFWI. Due
to large computational cost, we try to recover rock properties using a modified Marmousi model shown in Fig. 5(a) which has a size of 720
× 1800 m with a 10 m spatial sampling interval in both x and z. We use a surface acquisition geometry: 35 sources with a spacing of 50 m
and 180 receivers with a spacing of 10 m on the top of the model. The initial model is made by applying a Gaussian filter to the true model,
shown in Fig. 5(b). Owing to the limited offsets in the surface acquisition geometry of this experiment, the received data do not include many
diving or refracted waves, which are responsible for the recovery of the long wavelength structures of the model. To mitigate this lack of
long wavelength information, we start the inversion from a lower frequency at 3 Hz. The inverted frequencies are 3, 5, 7, 10, 14, 20, 28, 40,
76 Q. Yang and A. Malcolm

Figure 4. Mono-parameter PFWI synthetic example of simple box models. The results of recovered nine parameters are shown individually, the right-hand Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
panels are vertical profiles through the centre of the box showing the true model (solid line), the initial model (dashed line) and the final model (dotted line) of
(a) Kd , (b) G, (c) ρ s , (d) ρ f , (e) Ks , (f) Kf , (g) η, (h) κ 0 and (k) φ.

56 and 80 Hz from low to high following the multiscale strategy and a linear depth-weighting function from 0 to 1.0 is used. We use twenty
iterations per frequency. As we are inverting for one parameter, the other parameters are assumed to be perfectly known. Fig. 6 shows all the
final results of PFWI for this model.
In the inversion results presented in Fig. 6, we observe that the results of three moduli, Kd , Ks and Kf (Figs 6a, e and f), are worse in
the circled regions than the recovery of the densities ρ s and ρ f in the same region (Figs 6c and d), this is particularly notable for the inclined
body marked by arrows on the left side, which is smoothed out. This is likely because the radiation patterns of densities not only include the
scattered P wave but also the scattered S wave and the latter is responsible for short wavelength recovery. On the other hand, the scattered S
wave provides the information from wider apertures (see the numerical radiation patterns in Figs 2). The inclined body is located near the
boundary of the model so not many receivers record the signals reflected from it and transmitted through it, which also leads to a poorer
inversion result. The shear modulus G (Fig. 6b) is recovered well for the same reason as for densities, but the final model is not as smooth
Poroelastic FWI in frequency-domain 77

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
Figure 5. The modified Marmousi used in our final inversion tests. When inverting for each parameter, the true model is like (a), and the initial model is like
(b) with different magnitudes and units for each parameter.

Figure 6. One-parameter PFWI final results for the modified Marmousi. (a) Kd , (b) G, (c) ρ s , (d) ρ f , (e) Ks , (f) Kf , (g) η, (h) κ 0 , (k) φ.
78 Q. Yang and A. Malcolm

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
Figure 7. Seismograms corresponding to the inversion for Kd of part of the Marmousi model. The data residuals (Diff-) are displayed with a scale range that
is two orders-of-magnitude smaller than the synthetic data (-Data) to highlight the details. TrueData are the synthetic data with the true model, InitialData are
associated with the initial model, FinalData are the synthetic data computed in the final recovered model. Diff-T-I are the data residuals between the TrueData
and InitialData, Diff-T-F are the data residuals between the TrueData and FinalData.

as the recovered density models. For viscosity η, and permeability κ 0 (Figs 6g and h), their reconstructions fail. The result of porosity φ
(Fig. 6k) is without clear structures, just updating at the last five frequencies (20, 28, 40, 56, 80 Hz) and stopping after only a few iterations,
emphasizing that φ is a difficult parameter to invert for. Better recovery of φ, using additional offsets and other strategies is left for future
work. In general for other parameters, the main structures of the modified Marmousi model are accurately recovered, but we need more
iterations if we want more accurate small details.
Taking Kd as an example, we show the seismograms corresponding to the true, initial models and final models and compute the residuals
among them in Fig. 7. From Fig. 7 we find the synthetic data look similar because of the strong direct waves, but the final data residual
Diff-T-F is clearly reduced from the initial data residual Diff-T-I.
The computational cost of FWI is always of concern. FWI in the frequency domain proceeds from low frequency to high frequency
one by one. Model updating requires several iterations at each frequency. Here, we compare the average time consumed by one iteration of
one frequency of the FWIs in the elastic and poroelastic cases for the above two models, taking the solid density ρ s (ρ in elastic case) as
our example. Our computational platform is Matlab R2014b in Windows 10 system with Intel(R) Core(TM) i7-7700HQ CPU @2.80 GHz.
Table 2 shows the model size, the impedance matrix size, the computational time and the memory requirement of the two FWIs. All data
shown in Table 2 are obtained at 25 Hz. From Table 2, we observe that it is much more expensive to do poroelastic FWI than elastic FWI.
In our algorithm, each iteration of FWI needs three simulations at least (one for forward modelling, one for backpropagation modelling, and
Poroelastic FWI in frequency-domain 79

Table 2. Comparison between the elastic and poroelastic single-parameter FWI of the simple box model, and the
Marmousi model.
Simple box model Marmousi model

EFWI PFWI EFWI PFWI


Model size with PML 60×60 60×60 92×200 92×200
Impedance matrix size 7200×7200 14400×14400 38800×38800 77600×77600
Computational time 10.93 s 35.43 s 77.66 s 347.25 s
Memory requirement 0.71 Mb 3.06 Mb 7.12 Mb 32.33 Mb

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
Figure 8. Gaussian-anomaly models for two-parameter inversion.

one for the step length), or more if the step length requires refinement, which is the most time-consuming. So, the computational cost of FWI
is basically proportional to the time of forward modelling. However, the time required for forward modelling increases significantly with the
increasing size of the impedance matrix, which is why PFWI is so much more expensive.

6 T W O - PA R A M E T E R I N V E R S I O N

The coupling effect among parameters in a specific parametrization can be assessed with the radiation patterns. Generally, overlap between
radiation patterns of two inverted parameters is the source of trade-off, because at overlapping orientations FWI cannot distinguish which
parameter generates the signal recorded by the receivers. Fig. 2 shows the radiation patterns and the strength of scattered waves perturbed
by the parameters of poroelastic media, based on which we can analyse the trade-off among all parameters. We do not consider η and κ 0 in
multiparameter inversion because we could not recover them in single-parameter inversion.
First, we know that parameters having identical radiation patterns can not be recovered together. So we check trade-offs by type of
radiation patterns. In Fig. 2, we show four types of radiation patterns. Looking at the scattered P waves (outer waveform) in all panels, it
is hard to distinguish different parameters only based on P waves recorded at near offsets (330◦ –30◦ ) in a surface survey, but because the
scattered S waves (inner waveform) exist in Figs 2(b), (c) and (d), one of Kd , Ks and Kf (Fig. 2a) has a chance to be distinguished from any
one of the others. To do this we will need S-wave information which means that we need enough long offsets in a surface survey or a vertical
seismic profile (VSP) configuration. To recover G (Fig. 2b) with one of densities ρ s and ρ f (Fig. 2c), we have to use a VSP data set because
differences of their scattered P and S waves are only around 0◦ and 180◦ . For G and φ (Figs 2b and 2d), the situation is somewhat similar to
that of G and densities, because of the similarity of radiation patterns of densities and φ, even generating the same shape for the scattered S
waves. For the same reason, therefore, recovering one of ρ s and ρ f with φ simultaneously becomes very difficult for both a surface survey or
VSP data set. We note also the different strength of the scattered waves. In Figs 2(a) and (d), the P waves dominate, conversely the S waves
dominate in Figs 2(b) and (c). We infer that parameters that generate a strong amplitude in the data residual might be recovered better than
those generating a relatively weak amplitude.
For the (Kd , G) parametrization, which has four kinds of radiation patterns, we only consider six sets of two-parameter pairs to verify
our above analysis, that is, selecting one from each type of radiation patterns. In this section, we apply our PFWI on two models, one is a
Gaussian-anomaly model and the other one is a complex modified Marmousi model.
In Fig. 8, the Gaussian-anomaly models for two parameters are constructed consisting of 50 × 50 gridpoints with an interval of 10 m.
Each of the parameter 1 and 2 models have a Gaussian anomaly, located symmetrically on each side of the model. Our acquisition geometry
includes 9 sources distributed evenly on the top of model and 196 receivers at four edges around model which means an ideal geometry for
our inversion and thus the inverted results are only affected by inter-parameter trade-offs. The initial models are homogeneous with constants
listed in Table 1. We invert six frequencies (5, 15, 25 35, 45, 55 Hz) in sequence from low to high and 10 iterations per frequency.
Fig. 9 shows the inversion results for six sets of two-parameters. They are (a) Kd and G, (b) Kd and ρ s , (c) Kd and φ, (d) G and ρ s , (e) G
and φ, (f) ρ s and φ. Looking at Fig. 9(a), Kd and G are recovered correctly, in particular the G-anomaly is well-recovered with a clear shape
and just a little bit of an imprint of a Kd -anomaly on the left-hand side. Compared with G, the amplitude of Kd is not recovered as accurately.
The better recovered G is because of the strong scattered S waves. In Fig. 9(b), the results of Kd and ρ s are very similar with those of Kd and
80 Q. Yang and A. Malcolm

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
Figure 9. Results of the Gaussian-anomaly models for two-parameter inversion. (a) Kd and G, (b) Kd and ρ s , (c) Kd and φ, (d) G and ρ s , (e) G and φ, (f) ρ s
and φ.

G, just Kd is recovered somewhat less accurately in this case. This is due to the stronger scattered S waves of ρ s than those of G which limits
the recovery of Kd . The results of these two pairs are consistent with our prediction. But in Fig. 9(c), we observe some artefacts, particularly
in the φ result below the source positions. The imprint of Kd is also obvious in φ. The trade-off comes from their similar scattered P waves,
the strong artefacts, like in the one-parameter inversion result of φ, is due to its weak sensitivity kernel. In Fig. 9(d), we find that G and ρ s
are accurately recovered. The data at all orientations of non-overlapping radiation patterns of G and ρ s are recorded, which means the ideal
geometry suppresses the trade-off between them. ρ s is recovered better than G because of the stronger scattered waves for ρ s . In Fig. 9(e),
here G and φ are recovered clearly, however, there are many strong artefacts appearing in φ below the source positions. In Fig. 9(f), ρ s and φ
are poorly recovered due to trade-offs. In the result for φ, there are strong artefacts, the imprint of ρ s is even stronger than φ. This is expected
because ρ s and φ have very similar P wave radiation patterns, and the same shape of S wave radiation patterns (Figs 2 c and d). We conclude
that our inversion results are largely in agreement with our predictions based on the radiation patterns and sensitivity kernels. With this ideal
acquisition geometry, we are able to recover all parameters except φ accurately in pairs, with the caveat that recovering ρ s and Kd together
reduces the accuracy of the recovered Kd model.
Our next experiment is to apply our PFWI algorithm to invert for two parameters in a more complicated model, the modified Marmousi
model, in which the data sets not only include simple reflected and transmitted waves, but also diving waves and refracted waves. For
computational efficiency, this model is small enough that we can run PFWI without resorting to larger computing resources. Each inverted
parameter has an initial model as shown in Fig. 7. However, to alleviate the trade-off in the complex inversion experiments, we assume there
are two wells on the right- and left-hand sides of the model which means that in addition to the surface survey we also have the VSP survey.
We invert the same six pairs of two parameters as in the Gaussian-anomaly model. The inverted frequencies are 3, 5, 7, 10, 14, 20, 28, 40, 56
and 80 Hz from low to high and 20 iterations per frequency. A linear depth-weighting function from 0 to 1.0 is used.
Fig. 10 illustrates the inversion results of the modified Marmousi models for six parameter pairs. We observe that Kd and G (Fig. 10a), Kd
and ρ s (Fig. 10b) are recovered correctly. However, Kd is not inverted accurately due to trade-off and the differences in the strengths of their
radiation patterns, just like in the previous Gaussian-anomaly model. In Fig. 10(c), the results of Kd and φ are special, Kd changes very little
from the initial model but φ is recovered quite well. G and ρ s (Fig. 10d) are recovered very well simultaneously. Because of the receivers in
the wells their results are even better than their corresponding results in single parameter PFWI (in Figs 6b and c). In Fig. 10(e), G and φ are
Poroelastic FWI in frequency-domain 81

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
Figure 10. Results of the modified Marmousi models for two-parameter inversion. (a) Kd and G, (b) Kd and ρ s , (c) Kd and φ, (d) G and ρ s , (e) G and φ, (f) ρ s
and φ.

recovered poorly, the deep parts of model are blurry and fuzzy. In Fig. 10(f), the results are disappointing, but the outline of ρ s is recovered,
however the result of φ fails completely.

7 DISCUSSION
Our first example (box model) of single parameter inversion shows that with an ideal acquisition geometry, all parameters except η and κ 0
can be correctly reconstructed individually by our proposed PFWI. The failure of the inversion for viscosity η and permeability κ 0 is due to
their much weaker sensitivity kernels than others’ which is also pointed by De Barros & Dietrich (2008) and De Barros et al. (2010). For the
other parameters, our results represent a first-step toward recovering and using these parameters in a real case, but leaves a number of open
questions and issues that we discuss here.
From the inversion results of the modified Marmousi model, we find that a wider acquisition aperture allows us to better recover densities
ρ s , ρ f and shear modulus G. In other words, the parameters with a radiation pattern including both P and S waves are better recovered with
82 Q. Yang and A. Malcolm

wider aperture data because S waves are responsible for shorter wavelength structures. Basically, for the recoveries of moduli and densities,
the characteristics of regular FWI (Wu & Aki 1985; Tarantola 1986; Sun et al. 2017) are applicable in PFWI. From this example, we believe
that our PFWI can recover arbitrarily heterogeneous models.
For two-parameter PFWI, the inversion results conform to the predictions from radiation patterns in both the simple Gaussian-anomaly
and the complicated Marmousi model. Particularly, G and ρ s have the best inversion results when we include the VSP data sets. However,
for all of the inversions that include φ, the recovered φ models have strong artefacts and trade-offs with other parameter. So, joint inversion
including φ remains an open question. It is worth noting that in rock physics inversions P- and S-wave velocities have strong sensitivity to
porosity (Dupuy et al. 2016a), but the P- and S-wave velocities mentioned here refer to two composite quantities which depend on almost all
constitutive parameters in our parametrization (Kd , G) and so are not explicitly sensitive to only porosity as Vp /Vs is in elasticity theory. From
this perspective, maybe the two-step inversion method is more suitable to be applied to inversion of porosity.
The proposed PFWI is a non-linear problem with a big model space and data space. In this paper, we do not use any regularization

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
techniques to mitigate this non-linearity. Adding regularization terms to the objective function and using a strategy such as Tikhonov
regularization could improve the resolution of our PFWI. We also do not consider noisy-data to test PFWI yet, but we believe when the noise
is strong, the inversion will fail especially for the parameters with weak sensitivity kernels. In elastic or in poroelastic media, using noisy-data
for FWI is a big challenge.
There are also many inversion strategies for multiparameter FWI, such as choosing a proper parametrization and acquisition geometry
to avoid parameter trade-offs, or using the more accurate Gauss–Newton method or the truncated-Newton method to suppress parameter
trade-off (Sun et al. 2017; Keating & Innanen 2017; Pan et al. 2018, 2019). If one can pay the computational cost, these techniques should
be effective for multiparameter PFWI in theory, but work remains to be done to verify this. Specific to PFWI, De Barros et al. (2010)
explore two strategies to get around the poroelastic parameter coupling. Their first strategy is to use as much external information as possible
and to combine parameters that are physically interdependent into new composite parameters. The second one is to consider only inverting
subsurface properties that change over time in underground fluid-filled reservoirs. These strategies are useful and helpful in some situations.
After all, some of the parameters, such as the fluid and grain parameters, can be obtained easily, at least at limited spatial locations from log
data or geologic knowledge.

8 C O N C LU S I O N S
We extended the traditional frequency domain FWI proposed by Pratt et al. (1998) to the poroelastic case. Our PFWI can directly take
advantage of all of the information in the full wavefield to estimate the properties of porous media. In this paper we focus on inverting for
one-parameter and two-parameters in isotropic fluid-saturated porous media. Through studying the sensitivity kernels of each parameter and
their radiation patterns, we predict the feasibility of inverting for and the trade-off between six pairs of two parameters. We show that our PFWI
algorithm is valid in the standard seismic frequency range. We successfully implemented the inversions of seven parameters in poroelasticity
although we could not recover the viscosity η and the permeability κ 0 . Our two parameter results validate our predictions of the trade-offs
among different parameters based on their radiation patterns. We recommend that long offsets are required to recover multiple parameters,
such as a VSP survey, but even with such ideal data one needs to avoid inverting porosity together with any other parameter. In summary, we
set the stage and test the possibility of doing poroelastic multiparameter inversion. We put the PFWI into the same frequency-continuation
framework commonly used for elastic or acoustic FWI, allowing them to be compared side-by-side.

AC K N OW L E D G E M E N T S
We thank Christina Morency, Rene-Edouard Plessix, Jean Virieux and an anonymous reviewer for providing insightful comments, which
helped improve the manuscript. We thank Bastien Dupuy for a very helpful discussion. We also thank Dr Liu Xu who provided the analytical
solution of poroelastic media. This work is supported by Chevron and with grants from the Natural Sciences and Engineering Research
Council of Canada Industrial Research Chair Program (Grant No. IRCJ491051, 2016-492958) and InnovateNL (Grant No. 5405-1085-104),
by the Khalifa University of Science and Technology under Award No. CIRA-2018-48 and by the National Natural Science Foundation of
China (Grant No. 41704144).

REFERENCES Biot, M.A. & Willis, D.G., 1957. The elastic coefficients of the theory of
Biot, M.A., 1956a. Theory of propagation of elastic waves in a fluid satu- consolidation, J. Appl. Mech., 24, 594–601.
rated porous solid, Part I: low frequency range, J. acoust. Soc. Am., 28(2), Bozdağ, E., Peter, D., Lefebvre, M., Komatitsch, D., Tromp, J., Hill,
168–178. J., Podhorszki, N. & Pugmire, D., 2016. Global adjoint tomogra-
Biot, M.A., 1956b. Theory of propagation of elastic waves in a fluid sat- phy: first-generation model, Geophys. J. Int., 207(3), 1739–1766,
urated porous solid, Part II: high frequency range, J. acoust. Soc. Am., doi.org/10.1093/gji/ggw356.
28(2), 179–191. Brossier, R., Operto, S. & Virieux, J., 2009. Seismic imaging of complex on-
Biot, M.A., 1962. Mechanics of deformation and acoustic prop- shore structures by 2D elastic frequency-domain full-waveform inversion,
agation in porous media, J. Appl. Phys., 33(4), 1482–1498, Geophysics, 74(6), WCC105–WCC118.
doi.org/10.1063/1.1728759.
Poroelastic FWI in frequency-domain 83

Burridge, R. & Keller, J.B., 1981. Poroelasticity equations de- Liu, X., Greenhalgh, S., Zhou, B. & Greenhalgh, M., 2019. Frequency-
rived from microstructure, J. Acoust. Soc. Am., 70, 1140–1146, domain seismic wave modelling in heterogeneous porous media using the
doi.org/10.1121/1.386945. mixed-grid finite-difference method, Geophys. J. Int., 216, 34–54.
Carcione, J.M., 1996. Wave propagation in anisotropic, saturated porous Liu, Y., Dong, L., Wang, Y., Zhu, J. & Ma, Z., 2009. Sensitivity kernels for
media: plane-wave theory and numerical simulation, J. acoust. Soc. Am., seismic Fresnel volume tomography, Geophysics, 74, U35–U46.
99(5), 2655–2666. Masson, Y.J. & Pride, S.R., 2010. Finite-difference modeling of Biot’s poroe-
Carcione, J.M., Morency, C. & Santos, J.E., 2010. Computational lastic equations across all frequencies, Geophysics, 75(2), N33–N41.
poroelasticity—a review, Geophysics, 75(5), 75A229–75A243. Mavko, G., Mukerji, T. & Dvorkin, J., 2009. The Rocks Physics Handbooks:
Chotiros, N.P., 2002. An inversion for Biot parameters in water-saturated Tools for Seismic Analysis in Porous Media, 2nd edn, Cambridge Univ.
sand, J. acoust. Soc. Am., 112(5), 1853–1868. Press.
De Barros, L. & Dietrich, M., 2008. Perturbations of the seismic reflectivity Min, D.J., Shin, C., Kwon, B.D. & Chung, S., 2000. Improved frequency
of a fluid-saturated depth-dependent poroelastic medium, J. acoust. Soc. domain elastic wave modeling using weighted-averaging difference op-
Am., 123(3), 1409–1420. erators, Geophysics, 65(3), 884–895.

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
De Barros, L., Dietrich, M. & Valette, B., 2010. Full waveform inversion of Mora, P., 1987. Nonlinear two-dimensional elastic inversion of multioffset
seismic waves reflected in a stratified porous medium, Geophys. J. Int., seismic data, Geophysics, 52, 1211–1228.
182, 1543–1556. Morency, C., Luo, Y. & Tromp, J., 2009. Finite-frequency kernels for wave
Dupuy, B., Barros, L.D., Garambois, S. & Virieux, J., 2011. Wave propa- propagation in porous media based upon adjoint methods, Geophys. J.
gation in heterogeneous porous media formulated in the frequency space Int., 179(2), 1148–1168.
domain using a discontinuous Galerkin method, Geophysics, 76(4), N13– Morency, C., Luo, Y. & Tromp, J., 2011. Acoustic, elastic and poroelastic
N28. simulations of CO2 sequestration crosswell monitoring based on spectral-
Dupuy, B., Garambois, S., Asnaashari, A., Balhareth, H.M., Landrø, M., Sto- element and adjoint methods, Geophys. J. Int., 185, 955–966.
vas, A., et al., 2016b. Estimation of rock physics properties from seismic Nash, S.G., 2000. A survey of truncated-Newton methods, J. Comput. Appl.
attributes. Part 2: applications, Geophysics, 81(4), M55–M69. Math., 124, 45–59.
Dupuy, B., Garambois, S. & Virieux, J., 2016a. Estimation of rock physics Nocedal, J. & Wright, S.J., 1999, Numerical Optimization, Springer.
properties from seismic attributes. Part 1: strategy and sensitivity analysis, Operto, S., Miniussi, A., Brossier, R., Combe, L., Metivier, L., Monteiller,
Geophysics, 81(3), M35–M53. V., Riodetti, A. & Virieux, J., 2015. Efficient 3-D frequency-domain
Forgues, E. & Lambaré, G., 1997. Parameterization study for acoustic and mono-parameter full-waveform inversion of ocean-bottom cable data:
elastic ray+Born inversion, J. Seism. Explor., 6, 253–278. application to Valhall in the visco-acoustic vertical transverse isotropic
Gassmann, F., 1951. Úber die elastizitát poróser medien, Vierteljahrsschr. approximation, Geophys. J. Int., 202, 1362–1391.
Nat. Ges. Zurich, 96, 1–23. Operto, S., Virieux, J., Dessa, J. & Pascal, G., 2006. Crustal seismic imag-
Gholami, Y., Brossier, R., Operto, S., Prieux, V., Ribodetti, A. & Virieux, J., ing from multifold ocean bottom seismometer data by frequency domain
2013b. Which parameterization is suitable for acoustic vertical transverse full waveform tomography: application to the eastern Nankai trough, J.
isotropic full waveform inversion? Part 2: synthetic and real data case geophys. Res., 111, B09306, doi:10.1029/2005JB003835.
studies from Valhall, Geophysics, 78(2), R107–R124. Pan, W., Innanen, K.A. & Geng, Y., 2018. Elastic full-waveform inversion
Gholami, Y., Brossier, R., Operto, S., Ribodetti, A. & Virieux, J., 2013a. and parameterization analysis applied to walk-away vertical seismic pro-
Which parameterization is suitable for acoustic vertical transverse file data for unconvertional (heavy oil) reservoir characterization, Geo-
isotropic full waveform inversion? Part 1: sensitivity and trade-off analy- phys. J. Int., 213, 1934–1968.
sis, Geophysics, 78(2), R81–R105. Pan, W., Innanen, K.A., Geng, Y. & Li, J., 2019. Interparameter trade-off
He, W. & Plessix, R.É., 2017. Analysis of different parameterisations of quantification for isotropic-elastic full-waveform inversion with various
waveform inversion of compressional body waves in an elastic transverse model parameterizations, Geophysics, 84(2), R185–R206.
isotropic Earth with a vertical axis of symmetry, Geophys. Prospect., Pan, W., Innanen, K.A., Margrave, G.F., Fehler, M.C., Fang, X. & Li, J., 2016.
65(4), 1004–1024. Estimation of elastic constants for HTI media using Gauss-Newton and
Hustedt, B., Operto, S. & Virieux, J., 2004. Mixed-grid and staggered-grid full-Newton multiparameter full-waveform inversion, Geophysics, 81(5),
finite difference methods for frequency domain acoustic wave modelling, R275–R291.
Geophys. J. Int., 157, 1269–1296. Plona, T., 1980. Observation of a second bulk compressional wave in a
Johnson, D., Koplik, J. & Dashen, R., 1987. Theory of dynamic permeabil- porous medium at ultrasonic frequencies, Appl. Phys. Lett., 36(4), 259–
ity and tortuosity in fluid-saturated porous media, J. Fluid Mech., 176, 261.
379–402. Pratt, R.G., Shin, C. & Hicks, G.J., 1998. Gauss-Newton and full Newton
Keating, S. & Innanen, K.A., 2017. Cross-talk and frequency bands in trun- methods in frequency space seismic waveform inversion, Geophys. J. Int.,
cated Newton an-acoustic full waveform inversion, in Proceedings of the 133, 341–362.
SEG International Exposition and 87th Annual Meeting, 27 September, Pratt, R.G., Song, Z.M., Williamson, P.R. & Warner, M., 1996. Two-
pp. 1416–1421, Society of Exploration Geophysicists. dimensional velocity model from wide-angle seismic data by wavefield
Keating, S. & Innanen, K.A., 2019. Parameter crosstalk and modeling er- inversion, Geophys. J. Int., 124, 323–340.
rors in viscoacoustic seismic full-waveform inversion, Geophysics, 84(4), Pride, S.R., 2005. Relationships between seismic and hydrological proper-
R641–R653. ties, in Hydrogeophysics, pp. 253–291, eds Rubin, Y. & Hubbard, S. S.,
Koren, Z., Mosegaard, K., Landa, E., Thore, P. & Tarantola, A., 1991. Monte Springer.
Carlo estimation and resolution analysis of seismic background velocities, Pride, S.R., Berryman, J.G. & Harris, J.M., 2004. Seismic attenuation due
J. geophys. Res., 96, 20 289–20 299. to wave-induced flow, J. geophys. Res., 109(B01201), 1–19.
Levy, T., 1979. Propagation of waves in a fluid-saturated porous elastic solid, Pride, S.R., Gangi, A.F. & Morgan, F.D., 1992. Deriving the equations of
Int. J. Eng. Sci., 17, 1005–1014. motion for porous isotropic media, J. acoust. Soc. Am., 92, 3278–3290.
Liu, C., Gao, F., Feng, X., Liu, Y. & Liu, Y., 2013. Incorporating attenuation Ribodetti, A. & Virieux, J., 1996. Asymptotic theory for imaging the at-
effects into frequency-domain full waveform inversion from zero-offset tenuation factors Qp and Qs: Conference of Aixe-les-Bains. INRIA—
VSP data from the Stokes equation, J. Geophys. Eng., 10, 035004. French National Institute for Research in Computer Science and Control,
Liu, X., Greenhalgh, S., Zhou, B. & Greenhalgh, M., 2018. Effective Biot Expanded Abstracts, p. 334–353.
theory and its generalization to poroviscoelastic methods, Geophys. J. Sambridge, M. & Mosegaard, K., 2002. Monte Carlo methods in geophysical
Int., 212, 1255–1273. inverse problems, Rev. Geophys., 40, 1–29.
84 Q. Yang and A. Malcolm

Sourbier, F., Operto, S., Virieux, J., Amestoy, P. & L’Excellent, J.Y., Virieux, J. & Operto, S., 2009. An overview of full-waveform inversion in
2009. FWT2D: a massively parallel program for frequency-domain full- exploration geophysics, Geophysics, 74, WCC1–26.
waveform tomography of wide-aperture seismic data: part 1, algorithm, Warner, M., Stekl, I. & Umpleby, A., 2008. 3D wavefield tomography: syn-
Comp. Geosci., 35, 487–95. thetic and field data examples, in Proceedings of the 2008 SEG Annual
Sun, M., Yang, J., Dong, L., Liu, Y. & Huang, C., 2017. Density reconstruc- Meeting, SEG Technical Program Expanded Abstracts, 9–14 November,
tion in multiparameter elastic full-waveform inversion, J. Geophys. Eng., Las Vegas, Nevada, pp. 3330–3334.
14, 1445–1462. Wu, R.S. & Aki, K., 1985. Scattering characteristics of elastic waves by an
Tarantola, A., 1984. Inversion of seismic reflection data in the acoustic elastic heterogeneity, Geophysics, 50, 582–595.
approximation, Geophysics, 49, 1259–1266. Yang, Q., Malcolm, A., Rusmanugroho, R. & Mao, W., 2019. Analysis of
Tarantola, A., 1986. A strategy for nonlinear inversion of seismic reflection radiation patterns for optimized full waveform inversion in fluid-saturated
data, Geophysics, 51, 1893–1903. porous media, Geophys. J. Int., 216, 1919–1937.
Vigh, D., Starr, E.W. & Kapoor, J., 2009. Developing earth models with full Yang, Q. & Mao, W., 2017. Simulation of seismic wave propagation in 2-D
waveform inversion, Leading Edge, 28, 432–35. poroelastic media using weighted-averaging finite difference stencils in

Downloaded from https://academic.oup.com/gji/article/225/1/68/6044230 by USP- Reitoria-Sibi (inst. bio) user on 17 July 2024
Virieux, J., Asnaashari, A., Brossier, R., Métivier, L., Ribodetti, A. & Zhou, the frequency–space domain, Geophys. J. Int., 208, 148–161.
W., 2017. An introduction to full waveform inversion, in Encyclopedia Zhu, H., Bozdağ, E., Peter, D. & Tromp, J., 2012. Structure of the Euro-
of Exploration Geophysicspp. R1–1-R1-40, eds Grechka, V. & Wapenaar, pean upper mantle revealed by adjoint tomography, Nat. Geosci., 5(7),
K., Society of Exploration Geophysics. 493–498.

A P P E N D I X : T H E D E P E N D E N C I E S O F T H E C O E F F I C I E N T M AT R I X
l,w
k
To highlight which parameters each block A(m,ω) depends on, we list out their functional dependencies as
⎧ 1,u x 1,u

⎪ A(m,ω) = A11 ∂x x + A55 ∂zz + ω2 ρ1 ∂ A ; A(m,ω)z
= (A13 + A55 )∂x z ;




1,U x
= α Mφ∂x x + ω ρ2 ∂ A ;
2 1,U
A(m,ω) = α Mφ∂x z ;
z
⎪ A(m,ω)


⎪ A2,u x
= (A13 + A55 )∂x z ;
2,u z
A(m,ω) = A33 ∂zz + A55 ∂x x + ω2 ρ2 ∂ A ;

⎪ (m,ω)
⎨ A2,Ux = α Mφ∂ ; 2,Uz
A(m,ω) = α Mφ∂zz + ω2 ρ2 ∂ A ;
(m,ω) xz
(A1)
⎪ A(m,ω) = (α − φ)M∂x x + ω (ρ f − φ

3,u x 2 3,u z
ρ )∂ A ; A(m,ω) = (α − φ)M∂x z ;


⎪ = + ω φ ρ ∂
3,Uz
(m,ω) = Mφ∂x z ;
3,Ux
⎪ A
⎪ (m,ω) Mφ∂ x x
2
A ; A


⎪ A4,u

x
= (α − φ)M∂ ; A
4,u z
(m,ω) = (α − φ)M∂zz + ω (ρ f − φ
2
ρ )∂ A ;

⎩ (m,ω)
x z
4,Uz
A(m,ω) = Mφ∂x z ;
4,Ux
A(m,ω) = Mφ∂zz + ω φ 2
ρ∂A.
where A11 = A33 = Kd + 4/3G + α(α − φ)M, A13 = Kd − 2/3G + α(α − φ)M, and A55 = G; ∂ xx , ∂ zz and ∂ xz are the 25-point FD stencil of
partial derivatives, and ∂ A represents the acceleration term stencil. More details about the 25-point stencil can be found in Min et al. (2000)
or Yang & Mao (2017). Other quantities in eq. (A1) are defined in the main text.

You might also like