minerals
Article
Three-Dimensional Magnetotelluric Inversion for Triaxial
Anisotropic Medium in Data Space
Jingtao Xie 1 , Hongzhu Cai 1,2, * , Xiangyun Hu 1,2 , Shixin Han 1 and Minghong Liu 1
Citation: Xie, J.; Cai, H.; Hu, X.;
Han, S.; Liu, M. Three-Dimensional
Magnetotelluric Inversion for Triaxial
Anisotropic Medium in Data Space.
Minerals 2022, 12, 734. https://
doi.org/10.3390/min12060734
Academic Editors: Binzhong Zhou,
Changchun Yin, Zhengyong Ren,
Xuben Wang and Amin
Beiranvand Pour
Received: 29 March 2022
Accepted: 6 June 2022
Published: 8 June 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Minerals 2022, 12, 734. https://doi.org/10.3390/min12060734
1 Institute of Geophysics and Geomatics, China University of Geosciences, Wuhan 430074, China;
jtxie@cug.edu.cn (J.X.); xyhu@cug.edu.cn (X.H.); hsx@cug.edu.cn (S.H.); liuminghong@cug.edu.cn (M.L.)
2 State Key Laboratory of Geological Processes and Mineral Resources, China University of Geosciences,
Wuhan 430074, China
* Correspondence: caihongzhu@hotmail.com or caihz@cug.edu.cn
Abstract: The interpretation of three-dimensional (3-D) magnetotelluric (MT) data is usually based
on the isotropic assumption of the subsurface structures, and this assumption could lead to erroneous
interpretation in the area with considerable electrical anisotropy. Although arbitrary anisotropy is
much closer to the ground truth, it is generally more challenging to recover full anisotropy parameters
from 3-D inversion. In this paper, we present a 3-D triaxial anisotropic inversion framework using
the edge-based finite element method with a tetrahedral mesh. The 3-D inverse problem is solved by
the Gauss-Newton (GN) method which shows fast convergence behavior. The computation cost of
the data-space method depends on the size of data, which is usually smaller than the size of model;
therefore, we transform the inversion algorithm from the model space to the data space for memory
efficiency. We validate the effectiveness and applicability of the developed algorithm using several
synthetic model studies.
Keywords: magnetotelluric; triaxial anisotropic; data-space; 3-D inversion; finite element
1. Introduction
With the fast development of computer hardware and numerical algorithm, 3-D
magnetotelluric (MT) inversion has been widely used in many applications, such as mineral
exploration [1,2], deep earth structure imaging [3–5], and geothermal exploration [6,7].
At present, the interpretation of magnetotelluric data is generally based on the isotropy
assumption of the Earth’s geoelectric structure, but numerous studies have revealed the
prevalence of electrical anisotropy in the Earth [8–10]. This electric anisotropy can affect
the magnetotelluric response [11,12]. If the isotropic inversion technique was adopted to
invert the data affected by electrical anisotropy, it could produce artifacts and even lead to
erroneous interpretation [13,14]. Although the arbitrary anisotropy assumption is much
closer to the ground truth, it needs to consider more model parameters compared to the
isotropic inversion. Moreover, it is more challenging to solve the arbitrary anisotropy
inversion due to the strong non-uniqueness [15].
The current research of anisotropic MT inversion mainly focuses on two-dimension
(2-D) inversion and 3-D triaxial inversion in the model space. Li et al. (2003) implemented
2-D anisotropic inversion using the Gauss-Newton method [16]. Pek et al. (2011) used the
nonlinear conjugate gradient (NLCG) method to solve 2-D arbitrary anisotropic inversion,
and they found that it is challenging to reconstruct the anisotropic dipping angle without
additional a priori information [17]. Yu et al. (2022) reported 2-D arbitrary anisotropic MT
inversion using the limited-memory quasi-Newton (Q-N) method [18]. For 3-D anisotropic
inversion, Cao et al. (2018) and Wang et al. (2021) realized 3-D magnetotelluric inversion
with triaxial anisotropy using the limited-memory Broyden–Fletcher–Goldfarb–Shanno
(L-BFGS) method [15,19]. Kong et al. (2022) reported 3-D arbitrary anisotropic MT inversion
using the NLCG method [20]. These published results are based on the model-space
https://www.mdpi.com/journal/minerals
Minerals 2022, 12, 734 2 of 24

algorithm, and the memory consumption depends on the size of model. Considering the
number of model parameters is relatively large for 3-D inversion, the iterative solver is
usually used to solve the model update in each inversion iteration.
We develop a 3-D triaxial anisotropic inversion scheme using the data space method,
and the Gauss–Newton method is used to solve this inversion problem. To accurately
simulate the rugged topography and complex geoelectric structures, the unstructured
tetrahedral mesh is used to discretize the computational domain. For simplicity, we
now only consider three principal resistivities. The anisotropic strike can be determined
by the phase tensor technique and induction arrows in some cases [21,22], and also by
considering the known geological information. At present, it is still challenging to resolve
the anisotropic dip and slant using the current inversion strategy. The anisotropic inversion
algorithm is transformed from the model space to the data space to reduce the memory
requirement [23–26]. Moreover, the direct solver is used to solve the normal equation in
each inversion iteration. Compared to the iterative solver, the direct solver can improve the
stability of the inversion process.
In this paper, we first introduce the anisotropic forward modeling with a secondary
field formulation. Following this, we describe the triaxial anisotropic inversion based on the
Gauss–Newton optimization in the data space. Finally, we demonstrate the effectiveness
and applicability of the developed algorithm using several synthetic models.

2. Basic Theory
2.1. 3-D Magnetotelluric Forward Modeling with a Secondary Field Formulation
We start from Maxwell’s equations and assume the time dependence term of eiωt . By
ignoring the displacement current, we can obtain the diffusion equation as follows:
 
^ ^
∇ × ∇ × Es + iωµ0 σEs = −iωµ0 σ − σp E p , (1)

where Es and E p are the secondary field and the primary field, respectively; ω is the angular
frequency; µ0 is the free space magnetic permeability; σp is the background conductivity,
^
and σ is the anisotropic conductivity tensor, described as follows [27]:
 
σx
^
σ = Rz (−αS )Rx (−α D )Rz (−α L ) σy Rz ( α S )R x ( α D )Rz ( α L ), (2)
σz

where σx , σy , and σz are three principal conductivities in the X, Y, and Z directions, respec-
tively; αS , α D , and α L are the anisotropic strike, dip, and slant, respectively; and Rz (αS ),
Rx (α D ), and Rz (α L ) represent the rotations around the Z axis and X axis through the angle
αS , α D and α L , respectively.
The large sparse system of linear equations can be obtained by applying finite element
analysis:
AEs = b. (3)
The Dirichlet boundary condition is adopted for solving the above equation by assum-
ing the secondary electrical field vanishes on the boundary of the modeling domain. We use
the tetrahedral mesh to discretize the computational domain. Compared to the total field
formulation, one has to select proper background conductivity when simulating models
with complex topography using the secondary field formulation. We solve Equation (3)
using the parallel direct solver MKL Paradiso [28]. Considering the independence of fre-
quency, we further parallelize the forward modeling algorithm over the frequencies to
speed up the computation.
Minerals 2022, 12, 734 3 of 24

2.2. Data-Space Inversion Theory for Triaxial Anisotropic Medium

For the triaxial anisotropic inversion, three anisotropic rotation angles are constant,
and the objective functional can be written as:


ϕ (m ) = ϕ d (m ) + β x ϕ x (m x ) + β y ϕ y my + β z ϕ z (mz ). (4)


where ϕd (m) is the data misfit; ϕ x (mx ), ϕy my , and ϕz (mz ) are the model regularization

T
term in the X, Y, and Z directions, respectively; m = mx my mz is the model parame-
ter and mx = lnσx , my = lnσy , and mz = lnσz are the model parameters in three principal
directions; and β x , β y , and β z are the regularization parameters which are used to balance
the data fitting and model constraints.
The data misfit term is described as:
2
ϕd (m) =k Wd (d pre − dobs ) k , (5)

where d pre = F (m) is the predicted data, F is the forward modeling operator, and dobs rep-
resents the observed data. Wd is the diagonal data weighting matrix, which is constructed
based on the reciprocal of the standard error of the observed data, and the readers can refer
to Appendix A for more details.
The model regularization term along each principal direction can be written as:

re f 2
ϕi =k Li (mi − mi ) k , i = x, y, z, (6)

re f
where mi is reference model parameter; and Lx , Ly , and Lz are the model roughness
matrices for the conductivities in the X, Y, and Z directions, respectively. We adopted a
method similar to Cai et al., 2021, to calculate the roughness matrix. We can control the
smoothness of the inverted model by choosing the number of adjacent elements [29,30]. The
roughness matrix can stabilize the inversion by providing a measure of model variations
and avoiding spurious structures [31].
The Gauss–Newton method is used to minimize the objective functional, and we solve
the normal equation in each Gauss–Newton iteration to obtain the model update:

[ Re{(Wd J) H Wd J} + LT L]δm
(7)
= −[ Re{(Wd J) H Wd (d pre − dobs )} + LT L(m − mre f )],

J = [J x , Jy , Jz ], (8)
p 
β x Lx p
L= β y Ly p
, (9)
β z Lz
re f T
h i
re f re f
mre f = m x , my , mz , (10)

where Jx , Jy , and Jz are the sensitivity matrices for anisotropic inversion; H denotes the
Hermitian transpose; and T denotes the transpose.
We transform the system of normal equations from the model space to the data space
using the Sherman–Morrison–Woodbury formula [32], and the inverse of the reduced
Hessian matrix of Equation (7) can be written as:

[ Re{(Wd J) H Wd J} + LT L]−1 = U−1 − U−1 DΓ−1 D H U−1 , (11)

D = [ Re{(Wd J) H }, Im{(Wd J) H } ], (12)

U = LT L + e·diag{LT L}, (13)
Γ = I + D H U−1 D. (14)
Minerals 2022, 12, 734 4 of 24

where I is an identity matrix, and e is a small positive number, and it is added to Equation (13)
to enable the positive definiteness of matrix. Numerical tests show that its influence on the
inversion results is negligible [25,26].
We multiply both sides of Equation (7) by the inverse of the reduced Hessian matrix,
and substituting Equation (11) into Equation (7) to obtain the system of normal equations
in the data space:
δm = U−1 R − U−1 DΓ−1 D H U−1 R, (15)
R = −[ Re{(Wd J) H Wd (d pre − dobs )} + LT L(m − mre f )]. (16)
For simplicity, Equation (15) is written as:

δm = X1 − X2 X3 , (17)

X1 = U−1 R, (18)
X2 = U−1 D, (19)
X3 = Γ−1 D H X1 . (20)
It is computationally expensive to directly calculate the inverse of matrix U. We obtain
X1 and X2 by solving the linear equation with U as the coefficient matrix. U is a sparse
matrix with the size of Nm × Nm (Nm is the size of model), and it is efficient to obtain X1
and X2 using the direct solver. We notice that the size of matrix D is 2Nd × Nm (Nd is the
size of the data set), and we can solve the matrix X2 by hybrid programming with MPI and
OpenMP [33]. The matrix X3 is solved using the direct solver for dense matrix.
The model-space method requires dealing with the sensitivity matrix J ∈ C Nd × Nm and
the dense matrix of (Wd J) H Wd J ∈ C Nm × Nm . The calculation and storage of the latter term
are considerably expensive, and it is generally impractical for the anisotropic inversion.
The data-space method deals with matrix J ∈ C Nd × Nm and D ∈ R2Nd × Nm . Fortunately,
these matrices can be stored on the hard disk and some of them can be read into RAM. The
matrix of Γ ∈ R2Nd ×2Nd also needs to be stored, but Nd is usually much smaller than Nm .
Moreover, for the model-space method based on the direct solver, the construction of the
coefficient matrix and the right-hand sides of this equation are unsuitable for parallelization.
In contrast, most of the solutions of the data space method can be efficiently parallelized.
The regularization parameter is important to obtain geophysical meaningful solutions.
We choose the regularization parameters in a similar way as Cai et al., 2021 [30]:

k Re{(Wd Ji ) H (Wd Ji x)} k2

γi = , i = x, y, z, (21)
k LiT (Li x) k2

max γx , γy , γz
β i = qi c , i = x, y, z, (22)
niter
where 0 < qi < 1 is an empirical scaling parameter, niter is the iteration number, and c is
a positive integer that is usually less than 3, based on our numerical tests [34]. We have
tested the influence of qi and c on the inversion results as shown in Appendix B.
Once the model update direction δm is solved using the data-space method [23], a
new model is obtained as:
mn+1 = mn + αδmn . (23)
where α is the step length, which ensures the decreasing of the objective functional. An
appropriate step length can be obtained by a simple line search strategy [35].
The definition of RMS in this study is given as follows [36]:
s
(Wd (d pre − dobs ))T (Wd (d pre − dobs ))
RMS = . (24)
Nd
 where 𝛼 is the step length, which ensures the decreasing of the objective functional. An
appropriate step length can be obtained by a simple line search strategy [35].
The definition of RMS in this study is given as follows [36]:
𝑇
(𝐖𝑑 (𝐝𝑝𝑟𝑒 − 𝐝𝑜𝑏𝑠 )) (𝐖𝑑 (𝐝𝑝𝑟𝑒 − 𝐝𝑜𝑏𝑠 )) (24)
Minerals 2022, 12, 734 RMS = √ . 5 of 24
𝑁𝑑

The workflow for 3-D triaxial anisotropic inversion in data space is shown in Figure
1. The workflow for 3-D triaxial anisotropic inversion in data space is shown in Figure 1.

Figure 1.
Figure 1. Workflow
Workflowfor
for3-D
3-Dmagnetotelluric
magnetotelluricinversion
inversionwith
withtriaxial
triaxialanisotropy.
anisotropy.

3.
3. Synthetic
Synthetic Model
Model Studies
Studies
The
The developed code
developed code is
is run
run on
on the
the high-performance
high-performance cluster
cluster at
at China
China University
University of
of
Geosciences,
Geosciences, Wuhan, China. Each node has 2 Intel Xeon 2.5 GHz CPUs with 20
Wuhan, China. Each node has 2 Intel Xeon 2.5 GHz CPUs with 20 cores
cores per
per
CPU
CPU and
and384
384GB
GBRAM.
RAM.We We use two
use cluster
two nodes
cluster in this
nodes study.
in this For each
study. node,node,
For each we set
we2 set
MPI2
processes and 40 OpenMP threads. The model-space inversion algorithm is solved by the
direct solver for dense matrix.
Firstly, we use an example to compare the performance of the model-space and data-
space algorithm based on the direct solver. The total number of synthetic observed data
is 16,224 (including the real and imaginary parts). We use different discretization to test
the performances. The size of models for these different discretization are 20,196, 30,945,
41,280, 61,455, and 82,512, respectively. The memory requirements and CPU time for these
models are shown in Figure 2 for the model-space and data-space methods. Although the
model-space method is more efficient than the data-space method when the model size is
small, the memory consumption and calculation time increase approximately quadratically
 is 16,224 (including the real and imaginary parts). We use different discretization to test
space algorithm based on the direct solver. The total number of synthetic observed data
the performances. The size of models for these different discretization are 20,196, 30,945,
is 16,224 (including the real and imaginary parts). We use different discretization to test
41,280, 61,455, and 82,512, respectively. The memory requirements and CPU time for these
the performances. The size of models for these different discretization are 20,196, 30,945,
models are shown in Figure 2 for the model-space and data-space methods. Although the
41,280, 61,455, and 82,512, respectively. The memory requirements and CPU time for these
model-space method is more efficient than the data-space method when the model size is
Minerals 2022, 12, 734 models are shown in Figure 2 for the model-space and data-space methods. Although 6 ofthe
24
small, the memory consumption and calculation time increase approximately quadrati-
model-space method is more efficient than the data-space method when the model size is
cally with the increase of model size for the model-space method. For the data-space
small, the memory consumption and calculation time increase approximately quadrati-
method, the memory consumption and computation time increase linearly with the in-
cally with the increase of model size for the model-space method. For the data-space
crease
with theof modelofsize.
increase modelThesize
model-space method easily
for the model-space method.becomes
For the impractical becausethe
data-space method, the
method, the memory consumption and computation time increase linearly with the in-
memory consumption
model parameters canand computation
be extremely time
large. increase linearly
Therefore, with themethod
the data-space increase
is of model
more fea-
crease of model size. The model-space method easily becomes impractical because the
size.
sibleThe model-space
for the large-scalemethod easilyinversion
anisotropic becomesin impractical because the model parameters
certain scenarios.
model parameters can be extremely large. Therefore, the data-space method is more fea-
can be extremely large. Therefore, the data-space method is more feasible for the large-scale
sible for theinversion
anisotropic large-scale
in anisotropic inversion in certain scenarios.
certain scenarios.

Figure 2. The performance of the model-space and the data-space inversion algorithms.

Figure2.2. The
The performance
performanceof
Figure
3.1. Algorithm Validation ofthe
themodel-space
model-spaceand
andthe
thedata-space
data-spaceinversion
inversionalgorithms.
algorithms.

3.1.
3.1. Algorithm
The SM3Validation
Algorithm model from Cao et al., 2018 [15] is used to demonstrate the correctness of
Validation
the developed
The algorithm, as shown in Figure 3. The anisotropic anomaly with the size of
The SM3
SM3 model
model from
from Cao
Cao et et al.,
al., 2018
2018 [15]
[15] is
is used
used to
to demonstrate
demonstrate the the correctness
correctness of of
3.6 km × 3.6 km
the × 1 km is shownburied inFigure
the homogeneous half-space withwith a resistivity of
the developed
developed algorithm,
algorithm,as as shownin in Figure3.3. The
The anisotropic
anisotropic anomaly
anomaly with the the size
size of
of
300
3.6 km Ω𝑚
× .3.6The
km ×three
1 kmprincipal
is buried in resistivities of thehalf-space
the homogeneous anisotropic
withblock are 𝜌of𝑥 /𝜌
a resistivity 𝑦 /𝜌
300 𝑧 =
Ωm.
3.6 km × 3.6 km × 1 km is buried in the homogeneous half-space with a resistivity of
10 Ωm/30
The Ωm/50resistivities
three principal Ωm, and the values
of the of threeblock
anisotropic rotation
are ρxangles
/ρy /ρzboth Ωm/30
= 10are 0. TheΩm/50
central Ωm,
loca-
300 Ω𝑚 . The three principal resistivities of the anisotropic block are 𝜌𝑥 /𝜌𝑦 /𝜌𝑧 =
tion of the anomalous body is (0, 0, 1) km. The resistivity of the air
and the values of three rotation angles both are 0. The central location of the anomalous layer is chosen as
10 9Ωm/30 Ωm/50 Ωm, and the values of three rotation angles both are 0. The central loca-
body is (0, 0, 1) km. The resistivity of the air layer is chosen as 10 Ωm. We simulate the full
10 Ω𝑚. We simulate the full impedance tensor data for 5 9
frequencies (0.1 Hz, 1 Hz, 5 Hz,
tion of the anomalous body is (0, 0, 1) km. The resistivity of the air layer is chosen as
10 Hz and tensor
impedance 100 Hz) at for
data 1695sites and contaminate
frequencies the synthetic
(0.1 Hz, 1 Hz, 5 Hz, 10 Hz data
andby100
2%Hz)Gauss random
at 169 sites
109 Ω𝑚. We simulate the full impedance tensor data for 5 frequencies (0.1 Hz, 1 Hz, 5 Hz,
noise.
and contaminate the synthetic data by 2% Gauss random noise.
10 Hz and 100 Hz) at 169 sites and contaminate the synthetic data by 2% Gauss random
(a)
noise. (b)

(a) (b)

Figure 3. Panels (a) and (b) show the SM3 model from Cao et al., 2018 at x = 0 km and z = 1 km,
respectively. The black dots represent the MT receivers.

The whole simulation domain for the forward modeling and the inversion are the
same, and their size are 40 km × 40 km × 40 km in the X, Y, and Z directions. The thickness
of the air layer is 20 km for this model. The forward modeling domain is discretized into
743,603 elements. The inversion domain extends from –4 km to 4 km in the horizontal
direction and from 0 km to 4 km in the vertical direction. The mesh of the inversion domain
is discretized into 171,370 elements, and the total number of elements in the whole mesh
is 480,795. Both the initial model and the reference model are set to be a homogeneous
half-space with a resistivity of 50 Ωm. The standard error is set to 0.02·sqrt  Zxy · Zyx  . The


roughness matrix is calculated using the 20 surrounding elements of each inversion cell.
 Figure 3. Panels (a) and (b) show the SM3 model from Cao et al., 2018 at x = 0 km and z = 1 km,
respectively. The black dots represent the MT receivers.

The whole simulation domain for the forward modeling and the inversion are the
same, and their size are 40 km × 40 km × 40 km in the X, Y, and Z directions. The thickness
Minerals 2022, 12, 734 of the air layer is 20 km for this model. The forward modeling domain is discretized into 7 of 24
743,603 elements. The inversion domain extends from –4 km to 4 km in the horizontal
direction and from 0 km to 4 km in the vertical direction. The mesh of the inversion do-
main is discretized into 171,370 elements, and the total number of elements in the whole
We
meshsetisthe maximum
480,795. number
Both the of iterations
initial model and thetoreference
be 10, and the are
model threshold
set to beRMS for the inversion
a homogene-
is
ous half-space with a resistivity of 50 Ωm. The standard error is set to 0.02 ∙
1.05.
𝑠𝑞𝑟𝑡(|𝑍
We ∙ 𝑍𝑦𝑥 |). Thethe
𝑥𝑦 terminate roughness
inversionmatrix is five
after calculated usingwhen
iterations the 20the
surrounding
convergenceelements
stallsofas shown
each
in inversion
Figure 4. Thecell. We set
RMS the maximum
rapidly decreasenumber
from theof initial
iterations to be
value of 10, andtothe
16.83 threshold
1.06. The runtime is
RMS
9.2 h for
andthe theinversion
memory is consumption
1.05. is 137.5 GB.

Figure 4.4.Convergence
Figure Convergenceplot forfor
plot the the
SM3SM3
model. The black
model. dasheddashed
The black line indicates the threshold
line indicates RMS.
the threshold RMS.

We
Theterminate
anisotropicthe inversion
inversionafter fiveare
results iterations
shownwhen the convergence
in Figures stalls as
5–8. The shape shown
and conductivity
in Figure 4. The RMS rapidly decrease from the initial value of 16.83 to 1.06.
values in the anisotropic X and Y directions can be recovered well, and the anisotropic The runtime
is 9.2 h and the memory consumption is 137.5 GB.

coefficient between the resistivity in the X and Y directions (defined by log10 ρ x /ρy ) is
The anisotropic inversion results are shown in Figures 5–8. The shape and conduc-
(b) source of the magnetotelluric field is the plane wave
also reasonably recovered. The main
tivity values in the anisotropic X and Y directions can be recovered well, and the aniso-
which propagates vertically. Therefore, the resistivity in the Z direction only has limited
tropic coefficient between the resistivity in the X and Y directions (defined by
influence
log10 (𝜌𝑥 ⁄𝜌on the MT response. As a result, the resolution of the recovered vertical resistivity
Minerals 2022, 12, x FOR PEER REVIEW 𝑦 )) is also reasonably recovered. The main source of the magnetotelluric 8 of field
26
is
is the plane wave which propagates vertically. Therefore, the resistivity in the Z direction results
relatively lower compared to that of the horizontal resistivities. [37]. Our inversion
are
onlyinhas
a good
limited agreement with
influence on the Cao et al., 2018
MT response. As [15].
a result, the resolution of the recovered
vertical resistivity is relatively lower compared to that of the horizontal resistivities. [37].
Our inversion results are in a good agreement with Cao et al., 2018 [15].
(a) fitting is shown in Figure
The data (b) 9, we choose the Zxy and Zyx components at the
frequency of 5 Hz as an example. We can see that the observed data compare well to the
predicted data.

(c) (d)

(e) (f)

Figure 5. The inverted principal conductivity along the X-axis. Panels (a) and (b) show the true
Figure 5. The inverted principal conductivity along the X-axis. Panels (a) and (b) show the true
model and the recovered model at x = 0 km. Panels (c) and (d) show the true model and recovered
model
model at y =the
and recovered
0 km. Panels (e)model
and (f) at x =the
show 0 km.
true Panels (c) the
model and andrecovered
(d) showmodel
the true
at z =model
1 km. and recovered
model at y = 0 km. Panels (e) and (f) show the true model and the recovered model at z = 1 km.
(a) (b)
Minerals 2022, 12, 734 Figure 5. The inverted principal conductivity along the X-axis. Panels (a) and (b) show the true
8 of 24
model and the recovered model at x = 0 km. Panels (c) and (d) show the true model and recovered
model at y = 0 km. Panels (e) and (f) show the true model and the recovered model at z = 1 km.

(a) (b)

(c) (d)

(e) (f)

Minerals 2022, 12, x FOR PEER REVIEW 9 of 26

Figure 6.6. The

Figure The inverted principalconductivity
inverted principal conductivityalong
along the
the Y-axis.
Y-axis. Panels
Panels (a) (a)
andand
(b) (b)
showshowthe the
truetrue
model and recovered model at x = 0 km. Panels (c) and (d) show the true model and
model and recovered model at x = 0 km. Panels (c) and (d) show the true model and recovered recovered model
model at y = 0 km. Panels (e) and (f) show the true model and the recovered model
at y = 0 km. Panels (e) and (f) show the true model and the recovered model at z = 1 km. at z = 1 km.

(a) (b)

(c) (d)

(e) (f)

Figure7.7. The
Figure The inverted
inverted principal
principalconductivity
conductivityalong
alongZ-axis.
Z-axis.Panels (a)(a)
Panels andand
(b) (b)
show the the
show truetrue
model
model
and the recovered model at x = 0 km. Panels (c) and (d) show the true model and the recovered
and the recovered model at x = 0 km. Panels (c) and (d) show the true model and the recovered model
model at y = 0 km. Panels (e) and (f) show the true model and the recovered model at z = 1 km.
at y = 0 km. Panels (e) and (f) show the true model and the recovered model at z = 1 km.
3.2. Two-blocks Model
To test the effectiveness of our developed inversion algorithm and study the influ-
ence of electrical anisotropy, we design a two-blocks model as shown in Figure 10a,b.
These two anisotropic anomalies with the resistivities of 𝜌𝑥 /𝜌𝑦 /𝜌𝑧 = 100 Ω𝑚/10 Ω𝑚/
Minerals 2022,
Minerals 12,12,
2022, 734x FOR PEER REVIEW 9 of 24
10 of 26

(a) (b)
Minerals 2022, 12, x FOR PEER REVIEW 10 of 26

(a) (b)

(c) (d)

(c) (d)

(e) (f)

(e) (f)

Figure8.
Figure 8. The
Theinverted
invertedanisotropic
anisotropiccoefficient. Panels
coefficient. (a) and
Panels (a)(b)and
show
(b)the true the
show model and
true the recov-
model and the
ered model at x = 0 km. Panels (c) and (d) show the true model and the recovered model at y = 0 km.
recovered model at x = 0 km. Panels (c) and (d) show the true model and the recovered model at
Panels (e) and (f) show the true model and the recovered model at z = 1 km.
y = 0 km. Panels (e) and (f) show the true model and the recovered model at z = 1 km.

The
Figure data
8. The fittinganisotropic
(a) inverted is shown in FigurePanels
(b)coefficient. 9, we(a)
choose
and (b)the
(c) showZxy
theand
true Zmodel
(d) yx components at the
and the recov-
ered model at
frequency ofx5= Hz
0 km.
asPanels (c) and (d)
an example. Weshow
canthe true
see model
that the and the recovered
observed model at ywell
data compare = 0 km.
to the
Panels (e) and
predicted (f) show the true model and the recovered model at z = 1 km.
data.

(a) (b) (c) (d)

(e) (f) (g) (h)

(e) (f) (g) (h)

Figure 9. Data fitting of Zxy and Zyx components at the frequency of 5 Hz. Panels (a) and (b) show
the real and imaginary parts of the observed Zxy component. Panels (c) and (d) show the real and
imaginary parts of the observed Zyx component. Panels (e) and (f) show the real and imaginary parts
of the predicted Zxy component. Panels (g) and (h) show the real and imaginary parts of predicted
Zyx component.
Figure 9.
Figure 9. Data
Data fitting
fittingof
ofZZxyxyand
andZZyx components
yx componentsat the frequency
at the of 5of
frequency Hz. Panels
5 Hz. (a) and
Panels (b) show
(a) and (b) show
the real and imaginary parts of the observed Zxy component. Panels (c) and (d) show the real and
the real and imaginary parts of the observed Zxy component. Panels (c) and (d) show the real and
imaginary parts of the observed Zyx component. Panels (e) and (f) show the real and imaginary parts
imaginary parts of the observed Zyx component. Panels (e) and (f) show the real and imaginary parts
of the predicted Zxy component. Panels (g) and (h) show the real and imaginary parts of predicted
of
Zyxthe predicted Zxy component. Panels (g) and (h) show the real and imaginary parts of predicted
component.
Zyx component.
Minerals 2022, 12, 734 10 of 24

3.2. Two-Blocks Model

To test the effectiveness of our developed inversion algorithm and study the influence
of electrical anisotropy, we design a two-blocks model as shown in Figure 10a,b. These two
anisotropic anomalies with the resistivities of ρ x /ρy /ρz = 100 Ωm/10 Ωm/100 Ωm and
ρ x /ρy /ρz = 1000 Ωm/10, 000 Ωm/1000 Ωm are embedded in the homogeneous half-space
with a resistivity of 300 Ωm. Both of these anomalies are 10 km× 10 km × 5 km. The central
positions of these two bodies are (−7.5, 0, 4.5) km and (7.5, 0, 4.5) km, respectively. The
full impedance tensor data for 12 frequencies, evenly spaced in the logarithmic space from
0.001 Hz to 10 Hz, at 169 sites are used as the observed data. We contaminate the
Minerals 2022, 12, x FOR PEER REVIEW 11synthetic
of 26
data with 2% Gauss random noise. In Appendix C, we also study the influence of different
noise levels on the inversion results.

(a) (c)

(b) (d)

Figure 10. Panels

Figure Panels (a)
(a)and
and(b)
(b)show
showthethetrue
truemodel
model atat
y =y 0=km and
0 km z = z6 =km.
and Panels
6 km. (c) and
Panels (d) show
(c) and (d) show
the isotropic inversion results at y = 0 km and z = 6 km. The black dots represent MT receivers.
the isotropic inversion results at y = 0 km and z = 6 km. The black dots represent MT receivers.

The size
The sizeofofthe
thewhole
whole simulation
simulationdomain
domain for the
forforward
the forward modeling and inversion
modeling are
and inversion
bothboth
are 200 200
km km
× 200× km
200 ×km
100×km. 100The
km.domain
The domain extends from −100
extends fromkm to 100
−100 kmkm in the
to 100 km in
horizontal direction and from −20 km to 80 km in the vertical direction.
the horizontal direction and from −20 km to 80 km in the vertical direction. The mesh The mesh for for-
ward
for modeling
forward is discretized
modeling into 834,627
is discretized into cells.
834,627 Forcells.
the inversion, the mesh in
For the inversion, thethe inver-
mesh in the
inversion domain is discretized into 188,389 cells, and the total number of mesh is The
sion domain is discretized into 188,389 cells, and the total number of mesh is 417,504. 417,504.
inversion
The domain
inversion extends
domain fromfrom
extends −20 km
−20tokm 20 to
km20inkm theinhorizontal direction
the horizontal and from
direction and 0from
km to 20 km in the vertical direction. The initial and reference
0 km to 20 km in the vertical direction. The initial and reference models models are set to be 100 Ωm
are set
homogeneous half-space. The standard error is set to be 0.02 ∙ 𝑠𝑞𝑟𝑡(|𝑍𝑥𝑦 ∙ 𝑍𝑦𝑥 |) . The 
to be
100 Ωm homogeneous half-space. The standard error is set to be 0.02·sqrt Zxy · Zyx  . The
roughness matrix is calculated by the 20 surrounding elements of each cell. The maximum
roughness matrix is calculated by the 20 surrounding elements of each cell. The maximum
number of inversion iterations is set to 10, and the threshold RMS for the inversion is 1.05.
number of inversion iterations is set to 10, and the threshold RMS for the inversion is 1.05.
The model difference between the true model and the inverted model is defined as
The model difference between the true model and the inverted model is defined as
follows [38]:
follows [38]: s
N𝑚m 𝑖𝑛𝑣 inv true 2 2
∑𝑁


∑ i = (𝑚m𝑖 i −−𝑚m
1
𝑡𝑟𝑢𝑒
𝑖 i ) (25) (25)
δ𝛿== √ 𝑖=1
, ,
𝑁N𝑚m
inv true
where m𝑚i 𝑖𝑖𝑛𝑣and
where andm𝑚 i 𝑖 areare
𝑡𝑟𝑢𝑒 thethe
inverted
inverted model
modeland andthethetrue
truemodel.
model.
The isotropic inversion results
results are shown
shown in in Figure
Figure 10c,d.
10c,d. From
From this
this figure,
figure, wewe can
can see
that the isotropic inversion produces some artifacts surrounding the anisotropic
see that the isotropic inversion produces some artifacts surrounding the anisotropic con- conductive
body.
ductiveMoreover, the shape
body. Moreover, theand
shapelocation of the conductive
and location body body
of the conductive are poorly recovered.
are poorly recov- The
shape and shape
ered. The locationandof location
the anisotropic resistive body
of the anisotropic can bebody
resistive recovered
can beinrecovered
this scenario, which
in this
reveals
scenario,that the reveals
which effect ofthat
electrical anisotropic
the effect on anisotropic
of electrical the resistiveonbody is smaller
the resistive than
body is the
smaller thanbody
conductive the conductive
[36]. The body [36]. The
anisotropic anisotropic
inversion inversion
results results in
are shown areFigures
shown in 11Fig-
and 12.
ures 11
From and two
these 12. From these
figures, wetwo
canfigures,
see thatwethecan see that
shapes and the shapes and
locations locations
of these two of these
anisotropic
two anisotropic
anomalies anomalies
are well recovered areinwell
therecovered
principal in the principal
X and X and
Y directions. Y directions.
The anisotropicThe an-
coefficient
isotropic coefficient
between between
the resistivity in thetheprincipal
resistivityXinand
theYprincipal
directionsX and Y directions
is also reasonably is also rea-
recovered.
sonably recovered. Considering the resistivity in the Z direction cannot be effectively re-
covered in anisotropic inversion, we calculate the model difference 𝛿 using the resistivity
in the X and Y direction. The model difference 𝛿 is 0.94 and 0.75 for isotropic and aniso-
tropic inversion, respectively. With anisotropic inversion, we can obtain a better data fit-
ting for this model.
Minerals 2022, 12, 734 11 of 24

Considering the resistivity in the Z direction cannot be effectively recovered in anisotropic

inversion, we calculate the model difference δ using the resistivity in the X and Y direction.
Minerals 2022, 12, x FOR PEER REVIEW 12 of 26
The model difference δ is 0.94 and 0.75 for isotropic and anisotropic inversion, respectively.
Minerals 2022, 12, x FOR PEER REVIEW 12 of 26
With anisotropic inversion, we can obtain a better data fitting for this model.

(a) (b)
(a) (b)

(c) (d)
(c) (d)

(e) (f)
(e) (f)

Figure 11. Anisotropic inversion results at y = 0 km. Panels (a), (c), and (e) show the true model for
Figure 11. Anisotropic inversion results at y = 0 km. Panels (a), (c), and (e) show the true model
Figure 11. Anisotropic
the principal inversion
X, Y directions and results at y = 0 km.
the anisotropic Panels (a),
coefficient. (c), and
Panels (b), (e)
(d),show theshow
and (f) true model for
the corre-
for
the the principal
principal
sponding X, Y directions
X, Y directions
inversion results. and the anisotropic
and the anisotropic coefficient.
coefficient. Panels Panels
(b), (d), and (b),the
(f) show (d), and (f) show the
corre-
corresponding
sponding inversioninversion
results. results.

(a) (b)
(a) (b)

(c) (d)
(c) (d)

(e) (f)
(e) (f)

Figure 12. Anisotropic inversion results at z = 6 km. Panels (a), (c), and (e) show the true model for
the principal X and Y directions, and the anisotropic coefficient. Panels (b), (d), and (f) show the
corresponding inversion results.
 Minerals 2022, 12, x FOR PEER REVIEW 13 of 26

Minerals 2022, 12, 734 12 of 24

Figure 12. Anisotropic inversion results at z = 6 km. Panels (a), (c), and (e) show the true model for
the principal X and Y directions, and the anisotropic coefficient. Panels (b), (d), and (f) show the
corresponding inversion results.

The convergence plots for the isotropic and anisotropic inversion are shown in Fig-
The convergence plots for the isotropic and anisotropic inversion are shown in Figure
ure 13. We terminate both inversions after five iterations when the convergence stalls.
13. We terminate both inversions after five iterations when the convergence stalls. For the
For
isotropic inversion,inversion,
the isotropic the RMS
the RMS decrease decrease
from from
the initial theofinitial
value value
23.57 to 1.97. of
The23.57 to 1.97. The
runtime
runtime is 5.6 h and the memory consumption is 45.6 GB. For the anisotropic
is 5.6 h and the memory consumption is 45.6 GB. For the anisotropic inversion, the RMS inversion, the
RMS decrease from the initial value of 23.57 to 1.12. The runtime is 9.7
decrease from the initial value of 23.57 to 1.12. The runtime is 9.7 h and the memory con-h and the memory
consumption
sumption is 138.7 is 138.7
GB. IfGB.
theIfmodel
the model
space space
method method
was used wasforused for the anisotropic
the anisotropic inversion,inversion,
the memory requirement would be 4759.6 GB.
the memory requirement would be 4759.6 GB.

Figure 13.
Figure 13.Convergence
Convergenceplotplot
for the
forisotropic and anisotropic
the isotropic inversion. inversion.
and anisotropic The black dashed line indi-
The black dashed line
cates the threshold RMS.
indicates the threshold RMS.
We show the data fitting for isotropic and anisotropic inversion in Figure 14. The Zxy
We show the data fitting for isotropic and anisotropic inversion in Figure 14. The Zxy
and Zyx components at the frequency of 0.46 Hz are chosen as an example. It is clear that
and Zyx components at the frequency of 0.46 Hz are chosen as an
Minerals 2022, 12, x FOR PEER REVIEW example.
14 of 26 It is clear that
the data fitting for the anisotropic inversion is better than that of the isotropic inversion.
the data fitting for the anisotropic inversion is better than that of the isotropic
We know that MT only has limited sensitivity to the vertical resistivity. As a result,
inversion.
we may(a) exclude the vertical
(b) resistivity from
(c)
the model parameters.
(d)
We examine the effect
of including/excluding the vertical resistivity in the inversion parameters, as shown in
Figures 15 and 16. We can see that the inversion results are almost the same when includ-
ing and excluding the vertical resistivity from the model parameters. When excluding the
vertical resistivity from the inversion model parameters, we assume its value is the same
as the initial horizontal resistivity. The model difference 𝛿 are 0.75 and 0.74 for the inver-
sion with(e) including/excluding
(f) the vertical
(g)resistivity from(h)the model parameters. There-
fore, we do not consider the influence of vertical resistivity in the later section.

(i) (j) (k) (l)

Figure 14. Data fitting for Zxy and Zyx components at the frequency of 0.46 Hz. Panels (a) and (b)
Figure 14. Data fitting for Z observed
show the real and imaginary parts ofxy
and ZZyxxy component.
components at the frequency of 0.46 Hz. Panels (a) and (b)
Panels (c) and (d) show the real and
show theparts
imaginary realofand imaginary
observed parts Panels
Zyx component. (e) and (f)Zshow
of observed xy component. Panels parts
the real and imaginary (c) and
of (d) show the real and
predicted Zxy component for isotropic inversion. Panels (g) and (h) show the real and imaginary
imaginary parts of observed Zyx component. Panels (e) and (f) show the real and imaginary parts of
parts of predicted Zyx component for isotropic inversion. Panels (i) and (j) show the real and imagi-
nary parts ofZ
predicted xy component
predicted for isotropic
Zxy component inversion.
for anisotropic inversion.Panels (g)and
Panels (k) and(l)(h) show
show the
the real real and imaginary parts
and
imaginary
of predictedparts of
Zyxpredicted Zyx component
component for anisotropic
for isotropic inversion.
inversion. Panels (i) and (j) show the real and imaginary
parts of(a)
predicted Zxy component for anisotropic inversion. Panels (k) and (l) show the real and
(b)
imaginary parts of predicted Zyx component for anisotropic inversion.

(c) (d)
 (i) (j) (k) (l)

Minerals 2022, 12, 734 13 of 24

We know that MT only has limited sensitivity to the vertical resistivity. As a result,
we may exclude the vertical resistivity from the model parameters. We examine the effect
of including/excluding the vertical resistivity in the inversion parameters, as shown in
Figure 14. Data fitting for Zxy and Zyx components at the frequency of 0.46 Hz. Panels (a) and (b)
Figures 15 and 16. We can see that the inversion results are almost the same when including
show the real and imaginary parts of observed Zxy component. Panels (c) and (d) show the real and
and excluding
imaginary the
parts of vertical
observed Zyxresistivity
component.from
Panelsthe model
(e) and parameters.
(f) show the real andWhen excluding
imaginary parts of the
vertical resistivity from the inversion model parameters, we assume its value
predicted Zxy component for isotropic inversion. Panels (g) and (h) show the real and imaginary is the same as
the initial
parts horizontal
of predicted resistivity.
Zyx component forThe model
isotropic difference
inversion. Panels δ are 0.75
(i) and (j)and
show0.74 for the
the real and inversion
imagi-
nary including/excluding
with parts of predicted Zxy component for anisotropic
the vertical resistivityinversion.
from thePanels
model (k)parameters.
and (l) show the real and we
Therefore,
imaginary
do parts ofthe
not consider predicted Zyx component
influence of verticalfor anisotropicininversion.
resistivity the later section.

(a) (b)

(c) (d)

(e) (f)

Minerals 2022, 12, x FOR PEER REVIEW

Figure15.
Figure 15. Anisotropic
Anisotropic inversion
inversion results
results at
at yy ==00km.
km.Panels
Panels(a)
(a)and
and(b)
(b)show
showthe
theinverted 15 of in
invertedresistivity
resistivity 26
in the
the principal
principal X direction
X direction when
when including
including andand excluding
excluding thethe vertical
vertical resistivity
resistivity in the
in the inversion.
inversion. Panels
Panels (c) and (d) show the inverted resistivity in the principal Y direction when including and
(c) and (d) show the inverted resistivity in the principal Y direction when including and excluding
excluding
the vertical the verticalinresistivity
resistivity in thePanels
the inversion. inversion. Panels
(e) and (e) and
(f) show the(f) show the
inverted inverted coefficient
anisotropic anisotropic
coefficient when including and excluding the vertical resistivity
when including and excluding the vertical resistivity in the inversion. in the inversion.

(a) (b)

(c) (d)

(e) (f)
Figure 16. Cont.
Minerals 2022, 12, 734 14 of 24

(e) (f)

Figure16.
Figure 16.Anisotropic
Anisotropicinversion
inversionresults
results
at at
z =z 6= km.
6 km. Panels
Panels (a)(a)
andand
(b)(b) show
show thethe inverted
inverted resistivity
resistivity in
in the principal X direction when including and excluding the vertical resistivity in the inversion.
the principal X direction when including and excluding the vertical resistivity in the inversion. Panels
Panels (c) and (d) show the inverted resistivity in the principal Y direction when including and
(c) and (d) show the inverted resistivity in the principal Y direction when including and excluding
excluding the vertical resistivity in the inversion. Panels (e) and (f) show the inverted anisotropic
the vertical resistivity in the inversion. Panels (e) and (f) show the inverted anisotropic coefficient
coefficient when including and excluding the vertical resistivity in the inversion.
when including and excluding the vertical resistivity in the inversion.
3.3.Isotropic
3.3. Isotropicand
andAnisotropic
AnisotropicBlocks
BlocksModel
ModelwithwithTopography
Topography
Tofurther
To furthertest
test the
the performance
performance of the
of the developed
developed algorithm,
algorithm, we design
we design an isotropic
an isotropic and
and anisotropic
anisotropic blocksblocks
model,model,
as shownas shown
in Figurein Figure
17. The17. The background
background model ismodel is a homo-
a homogeneous
geneous half-space
half-space with the of
with the resistivity resistivity
300 Ωm.of The Ωm.principal
300three The three principal of
resistivities resistivities of the
the anisotropic
block are 𝜌𝑥 /𝜌𝑦 /𝜌 𝑧 = 10 Ωm/100 Ωm/10 Ωm,
block are ρ x /ρy /ρz = 10 Ωm/100 Ωm/10 Ωm, and its central location is (0, −7.5, 5.5)iskm.
anisotropic and its central location (0,
−7.5, 5.5) km. The resistivity of the isotropic block is 1000 Ωm, and its central
The resistivity of the isotropic block is 1000 Ωm, and its central location is (0, 7.5, 5.5) km. location is
Minerals 2022, 12, x FOR PEER REVIEW 16 of 26
(0, 7.5, 5.5) km. The full impedance tensor data for 12 frequencies, evenly
The full impedance tensor data for 12 frequencies, evenly spaced in the logarithmic16scale
Minerals 2022, 12, x FOR PEER REVIEW spaced in the
of 26
logarithmic
from 0.01 Hz scale
to 100from 0.01
Hz, at 169Hz to are
sites 100used
Hz, at
for169 sites areWe
inversion. used
addfor2%inversion.
Gauss randomWe add 2%
noise
toGauss randomdata.
the synthetic noiseTheto the syntheticindata.
topography The topography
the inversion domain in the inversion
is shown in Figuredomain
18. is
shown in Figure 18. (b)
(a) (b)
(a)

Figure 17. Panels (a) and (b) show the isotropic and anisotropic blocks model at x = 0 km and z = 6
Figure
Figure
km, 17.Panels
17. Panels(a)
respectively. (a)and
The and (b)dots
(b)
black show
show theisotropic
the isotropic
represent and
and
the MT anisotropicblocks
anisotropic
receivers. blocksmodel
modelatatxx==00km
kmand
andzz==66
km,respectively.
km, respectively.The
Theblack
blackdots
dotsrepresent
representthe
theMT
MTreceivers.
receivers.

Figure 18. The

Figure18. The topography
topography in
in the
the inversion
inversion domain
domain (with
(with exaggeration
exaggerationin
inthe
the vertical
verticaldirection).
direction).The
The
Figure
black 18.
blackdots The topography
dotsindicate
indicatethe
theMT in the inversion
MTreceivers.
receivers. domain (with exaggeration in the vertical direction). The
black dots indicate the MT receivers.
The whole simulation domain extends from −100 km to 100 km in the horizontal di-
The
rection whole
and fromsimulation
−20 km to domain
80 km inextends fromdirection.
the vertical −100 km to 100
The km in the
inversion horizontal
domain di-
extends
rection and from −20 km to 80 km in the vertical direction. The inversion domain extends
from −20 km to 20 km in the horizontal direction and from −1.16 km to 20 km in the vertical
from −20 km to 20 km in the horizontal direction and from −1.16 km to 20 km in the vertical
Minerals 2022, 12, 734 15 of 24

The whole simulation domain extends from −100 km to 100 km in the horizontal
direction and from −20 km to 80 km in the vertical direction. The inversion domain extends
from −20 km to 20 km in the horizontal direction and from −1.16 km to 20 km in the vertical
direction. The mesh for forward modeling is discretized into 824,132 cells. The inversion
domain is discretized into 200,934 cells, and the total number of cells for the inversion
mesh is 644,824. The initial and reference models of the inversion are a homogeneous
half-space with the resistivity of 100 Ωm. The standard error is set to 0.02·sqrt  Zxy · Zyx  .
 

The maximum number of inversion iterations is set to 10, and the threshold RMS for the
inversion is 1.05.
Minerals 2022, 12, x FOR PEER REVIEW 17 of 26
The convergence plot is shown in Figure 19, we can see that the RMS decreases from
21.72 to 1.05 after five iterations, the runtime is 8.2 h and the memory consumption is 99.2 GB.

Minerals 2022, 12, x FOR PEER REVIEW 17 of 26

Figure 19.
Figure 19. Convergence
Convergence plot for the
the isotropic
isotropic and
and anisotropic
anisotropic blocks
blocks model
model with
with topography.
topography. The
The
black dashed line indicates the threshold RMS.
black dashed line indicates the threshold RMS.

The(a)inversion results are shown

(b) in Figures 20 and 21, we can see that the shape and
Figure 19. Convergence plot for the isotropic
position of the isotropic and anisotropic blocks and anisotropic blocks model
are well recovered. Thewith topography.
inversion The
does not
black dashed line indicates the threshold RMS.
produce obvious artifacts for the isotropic anomalous body.

(a) (b)
(c) (d)

(c) (d)

(e) (f)

(e) (f)

Figure 20. Anisotropic inversion results at x = 0 km. Panels (a), (c), and (e) show the true model for
principal X and Y directions, and the anisotropic coefficient. Panels (b), (d), and (f) show the aniso-
tropic inversion results for the principal X and Y directions, and the anisotropic coefficient.

Figure20.
Figure 20. Anisotropic inversion
inversion results
results at
at xx==00km.
km.Panels
Panels(a), (c),
(a), and
(c), and(e)(e)
show
show thethe
true model
true modelfor
principal
for X and
principal Y directions,
X and andand
Y directions, the anisotropic coefficient.
the anisotropic Panels
coefficient. (b), (d),
Panels (b),and
(d),(f) show
and the aniso-
(f) show the
tropic inversion
anisotropic results
inversion for the
results forprincipal X and
the principal Y directions,
X and and and
Y directions, the anisotropic coefficient.
the anisotropic coefficient.
Minerals 2022, 12, 734 16 of 24
Minerals 2022, 12, x FOR PEER REVIEW 18 of 26

(a) (b)

(c) (d)

(e) (f)

Figure 21.
Figure 21.Anisotropic
Anisotropicinversion results
inversion at z = 5.5
results at zkm. Panels
= 5.5 km.(a), (c), and
Panels (e)(c),
(a), show the(e)
and trueshow
modelthe
fortrue model
the principal X and Y directions, and the anisotropic coefficient. Panels (b), (d), and (f) show the
for the principal
anisotropic X and
inversion Y directions,
results and the
for the principal X andanisotropic coefficient.
Y directions, Panels (b),
and the anisotropic (d), and (f) show the
coefficient.
anisotropic inversion results for the principal X and Y directions, and the anisotropic coefficient.
The data fitting is shown in Figure 22, we choose the Zxy and Zyx components at the
The data
frequency fitting
of 0.285 Hz is
as shown in Figure
an example. We can 22,see
wethat
choose the Zxy and
the observed data Z yx components
compare well at the
frequency
Minerals 2022, 12, x FOR PEER REVIEW of 0.285 Hz
with the predicted data. as an example. We can see that the observed 19 data
of 26 compare well with
the predicted data.
(a) (b) (c) (d)

(e) (f) (g) (h)

Figure 22. Data fitting for Zxy and Zyx components at the frequency of 0.285 Hz. Panels (a) and (b)
Figure 22. Data fitting for Z observed
show the real and imaginary parts ofxy
and ZZyxxy component.
components at the frequency of 0.285 Hz. Panels (a) and
Panels (c) and (d) show the real and
(b) show the real and imaginary parts of observed Z xy component.
imaginary parts of the observed Zyx component. Panels (e) and (f) show Panelsparts
the real and imaginary (c) and (d) show the real
of the predicted Zxy component. Panels (g) and (h) show the real and imaginary parts of the pre-
and imaginary parts of the observed Zyx component. Panels (e) and (f) show the real and imaginary
dicted Zyx component.
parts of the predicted Zxy component. Panels (g) and (h) show the real and imaginary parts of the
4. Conclusions
predicted Zyx component.
In this paper, we develop a triaxial anisotropic 3-D MT inversion algorithm with un-
structured tetrahedral mesh. To reduce the memory requirement, we transform the algo-
rithm from the model space to the data space and derived the inversion workflow using
the Gauss–Newton approach. In the data-space inversion, the model update direction can
be effectively solved by the parallel direct solver.
We compare the efficiency of the model-space and data-space methods by an exam-
Minerals 2022, 12, 734 17 of 24

4. Conclusions
In this paper, we develop a triaxial anisotropic 3-D MT inversion algorithm with
unstructured tetrahedral mesh. To reduce the memory requirement, we transform the
algorithm from the model space to the data space and derived the inversion workflow
using the Gauss–Newton approach. In the data-space inversion, the model update direction
can be effectively solved by the parallel direct solver.
We compare the efficiency of the model-space and data-space methods by an example.
We found that the data-space algorithm requires much less memory and computation
time than the model space method when the number of model parameters is large. We
demonstrate the effectiveness and applicability of the anisotropic inversion algorithm by
several synthetic models. The electrical structures in the principal X and Y directions can
be well recovered, but the resistivity of the principal Z direction cannot be reasonably
recovered by the current inversion algorithm. We will apply the developed anisotropic
inversion algorithm in data space to the field data in the future research.

Author Contributions: Conceptualization, J.X., H.C. and X.H.; methodology, J.X. and H.C.; software,
H.C. and J.X.; validation, S.H. and M.L.; formal analysis, H.C. and J.X.; resources, X.H. and H.C.;
writing—original draft preparation, J.X.; writing—review and editing, H.C.; supervision, X.H. and
H.C. All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by National Natural Science Foundation of China, grant number
41974089.
Data Availability Statement: Not applicable.
Acknowledgments: We acknowledge the support of this work from the computation center at China
University of Geosciences (Wuhan).
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A
The data weighted matrix Wd usually is used with nonuniform and uniform data
variance [39], and other forms. We tested the effect of Wd with nonuniform and uniform
data variance using the synthetic data (with 2% Gauss random noise) using the two-blocks
model in Section 3.2.
The data weighted matrix Wd can be written as follows:
 
1/e1
 1/e2 
Wd =  , (A1)
 
..
 . 
1/e Nd

where ei is the data variance of the ith observed data.

The nonuniform data variance can be written as follows [18,20,24]:

q 
e Zij = ε·  Zxy · Zyx , i, j = x, y. (A2)

where ε is error floor, we choose ε = 0.02 in this case.

The uniform data variance can be written as [24]:

  

e Zij = max 3.5%·  Zxy − Zyx /2 , i, j = x, y. (A3)

These inversion results are shown in Figures A1 and A2, and we reset the color bar for
a higher contrast. The RMS for Wd with nonuniform and uniform data variance are 1.12
and 1.17, and the model difference δ are 0.75 and 0.76, respectively.
We can see that these two anisotropic anomalies are both well recovered for Wd
with nonuniform and uniform data variance. However, the conductive artifact blows the
 The uniform data variance can be written as [24]:
𝑒(𝑍𝑖𝑗 ) = 𝑚𝑎𝑥{3.5% ∙ (|𝑍𝑥𝑦 − 𝑍𝑦𝑥 |⁄2)}, 𝑖, 𝑗 = 𝑥, 𝑦. (A3)

These inversion results are shown in Figures A1 and A2, and we reset the color bar
for a higher contrast. The RMS for 𝐖𝑑 with nonuniform and uniform data variance are
Minerals 2022, 12, 734 18 of 24
1.12 and 1.17, and the model difference 𝛿 are 0.75 and 0.76, respectively.
We can see that these two anisotropic anomalies are both well recovered for 𝐖𝑑
with nonuniform and uniform data variance. However, the conductive artifact blows the
resistive body,forfor𝐖W
resistive body, 𝑑 with uniform data variance, more seriously than that for 𝐖 with
d with uniform data variance, more seriously than 𝑑that for Wd with
nonuniform data variance. Moreover,
nonuniform data variance. Moreover, the the
anisotropic coefficient
anisotropic for 𝐖𝑑for
coefficient with
Wd nonuni-
with nonuniform
form data variance is better recovered.
data variance is better recovered.
(a) (b)

(c) (d)

(e) (f)
(e) (f)

Figure A1. Anisotropic inversion results at y = 0 km. Panels (a) and (b) show the inverted resistivity
Figure A1. Anisotropic inversion results at y = 0 km. Panels (a) and (b) show the inverted resistivity
in the principal X direction for 𝐖𝑑 with nonuniform and uniform data variance. Panels (c) and (d)
in the principal X direction
show the inverted resistivity forprincipal
in the Wd with nonuniform
Y direction for 𝐖𝑑and
withuniform data
nonuniform variance.
and Panels (c) and
uniform data
(d) show Panels
variance. the inverted resistivity
(e) and (f) show the in the principal
inverted Y direction
anisotropic for 𝐖
coefficient for withnonuniform
W𝑑d with nonuniform andand uniform
Minerals 2022, 12, x FOR PEER REVIEW
uniform data variance. 21 of 26
data variance. Panels (e) and (f) show the inverted anisotropic coefficient for Wd with nonuniform
and uniform data variance.

(a) (b)

(c) (d)

(e) (f)

Figure A2. Anisotropic inversion results at z = 6 km. Panels (a) and (b) show the inverted resistivity
Figure A2. Anisotropic inversion results at z = 6 km. Panels (a) and (b) show the inverted resistivity
in the principal X direction for 𝐖𝑑 with nonuniform and uniform data variance. Panels (c) and (d)
show the inverted resistivity in the principal Y direction for 𝐖𝑑 with nonuniform and uniform data
variance. Panels (e) and (f) show the inverted anisotropic coefficient for 𝐖𝑑 with nonuniform and
uniform data variance.

Appendix B
Minerals 2022, 12, 734 19 of 24

in the principal X direction for Wd with nonuniform and uniform data variance. Panels (c) and
(d) show the inverted resistivity in the principal Y direction for Wd with nonuniform and uniform
data variance. Panels (e) and (f) show the inverted anisotropic coefficient for Wd with nonuniform
and uniform data variance.

Appendix B
We also carried out some numerical tests on the choices of qi and c using the synthetic
data (with 2% Gauss random  noise)

of the two-blocks model in Section 3.2, and the standard
error is set to 0.02·sqrt  Zxy · Zyx  . The final RMS for the case with qi = 0.05, 0.1, 0.4, 0.8
are 1.05, 1.06, 1.09, 1.15, and the model difference δ are 0.78, 0.75, 0.75, 0.76. The final RMS
for c = 1, 2, 3 are 1.12, 1.06, 1.04, and the model difference δ are 0.75, 0.75, 0.79.
We can see from Figures A3 and A4 that the recovered background model is inaccurate
when qi = 0.05, but these two anisotropic bodies can be reasonably recovered. Although
these inversion results are similar, the background model can be better recovered when
qi increases from 0.1 to 0.8. Taking the above factors into consideration, we believe that
qi = 0.8 is appropriate.
Minerals 2022, 12, x FOR PEER REVIEW We can see from Figures A5 and A6 that these inversion results are similar, but the
22 of 26
larger value of c, the worse the inverted background model. As a result, we choose c = 1 in
this case.
(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

FigureA3.
Figure A3.Anisotropic
Anisotropicinversion
inversion results
results at
at yy =
= 00 km
km for different q𝑞i𝑖,,and
for different and c 𝑐==1.1.Panels
Panels(a)–(d)
(a)–(d)show
show
the inverted resistivity in the principal X direction for 𝑞 𝑖 = 0.05, 0.1, 0.4, 0.8. Panels (e)–(h) show
the inverted resistivity in the principal X direction for qi = 0.05, 0.1, 0.4, 0.8. Panels (e)–(h) show the the
inverted resistivity in principal Y direction for 𝑞𝑖 = 0.05, 0.1, 0.4, 0.8. Panels (i)–(l) show the inverted
inverted resistivity in principal Y direction for qi = 0.05, 0.1, 0.4, 0.8. Panels (i)–(l) show the inverted
anisotropic coefficient for 𝑞𝑖 = 0.05, 0.1, 0.4, 0.8.
anisotropic coefficient for qi = 0.05, 0.1, 0.4, 0.8.

(a) (b) (c) (d)

(e) (f) (g) (h)

 Figure A3. Anisotropic inversion results at y = 0 km for different 𝑞𝑖 , and 𝑐 = 1. Panels (a)–(d) show
Minerals 2022, 12, 734 the inverted resistivity in the principal X direction for 𝑞𝑖 = 0.05, 0.1, 0.4, 0.8. Panels (e)–(h) show the 20 of 24
inverted resistivity in principal Y direction for 𝑞𝑖 = 0.05, 0.1, 0.4, 0.8. Panels (i)–(l) show the inverted
anisotropic coefficient for 𝑞𝑖 = 0.05, 0.1, 0.4, 0.8.

(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

Minerals 2022, 12, x FOR PEER REVIEW 23 of 26

Figure A4. Anisotropic inversion results at z = 6 km for different 𝑞 , and 𝑐 = 1. Panels (a)–(d) show
𝑖
Figure A4. Anisotropic inversion results at z = 6 km for different qi , and c = 1. Panels (a)–(d) show
the inverted resistivity in the principal X direction for 𝑞𝑖 = 0.05, 0.1, 0.4, 0.8. Panels (e)–(h) show the
the inverted resistivity in the principal X direction for qi = 0.05, 0.1, 0.4, 0.8. Panels (e)–(h) show the
invertedresistivity
inverted resistivity inin principal
principal Y direction
Y direction qi =𝑞0.05,
for for 𝑖 = 0.05,
0.1, 0.1,
0.4, 0.4,
0.8. 0.8. Panels
Panels (i)–(l)
(i)–(l) showshow the inverted
the inverted
anisotropic coefficient for 𝑞 = 0.05, 0.1,
anisotropic coefficient for qi =𝑖 0.05, 0.1, 0.4, 0.8.0.4, 0.8.

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

FigureA5.
Figure A5.Anisotropic
Anisotropic inversion
inversion results
results at y =at0 ykm= for
0 km c, and qi 𝑐,= and
for different
different 𝑞𝑖 = 0.6.
0.6. Panels Panels
(a)–(c) show(a)–(c)
show the inverted resistivity in the principal X direction for 𝑐 = 1,2,3. Panels (d)–(f)
the inverted resistivity in the principal X direction for c = 1, 2, 3. Panels (d)–(f) show the inverted show the in-
verted resistivity
resistivity in the principal
in the principal Y direction
Y direction for c = 1,for 2, 3.𝑐Panels
= 1,2,3. Panels
(g)–(i) (g)–(i)
show the show theanisotropic
inverted inverted aniso-
tropic coefficient
coefficient 2, 3.𝑐 = 1,2,3.
for c = 1,for

(a) (b) (c)

 Figure A5. Anisotropic inversion results at y = 0 km for different 𝑐, and 𝑞𝑖 = 0.6. Panels (a)–(c
show the inverted resistivity in the principal X direction for 𝑐 = 1,2,3. Panels (d)–(f) show the in
Minerals 2022, 12, 734 21 of 24
verted resistivity in the principal Y direction for 𝑐 = 1,2,3. Panels (g)–(i) show the inverted aniso
tropic coefficient for 𝑐 = 1,2,3.

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

Figure A6. Anisotropic inversion results at z = 6 km for different c, and qi = 0.6. Panels (a)–(c) show
the inverted resistivity in the principal X direction for c = 1, 2, 3. Panels (d)–(f) show the inverted
resistivity in the principal Y direction for c = 1, 2, 3. Panels (g)–(i) show the inverted anisotropic
coefficient for c = 1, 2, 3.

Appendix C
We test the effects of noise level on the inversion results in this section. In this study, 2%,
5%, and 8% Gauss random noise are added to synthetic observed data for the two-blocks
model in Section 3.2.
The inversion results are shown in Figures A7 and A8. We can see that although
the shape and location of these two anisotropic anomalies are well recovered for the
resistivity in the principal X and Y directions, and the anisotropic coefficient, the recovered
background model is distorted when the noise increases. The convergence stalls after five
iterations for all cases. The RMS are 1.12, 2.65, and 4.24, and the model difference δ are 0.75,
0.76, and 0.77 when 2%, 5%, and 8% random noise are added to observed data.
 blocks model in 3.2.
The inversion results are shown in Figures A7 and A8. We can see that although the
shape and location of these two anisotropic anomalies are well recovered for the resistivity
in the principal X and Y directions, and the anisotropic coefficient, the recovered back-
ground model is distorted when the noise increases. The convergence stalls after five iter-
Minerals 2022, 12, 734
ations for all cases. The RMS are 1.12, 2.65, and 4.24, and the model difference 22
𝛿 ofare
24
0.75,
0.76, and 0.77 when 2%, 5%, and 8% random noise are added to observed data.

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

(g) (h) (i)

FigureA7.
Figure A7.Anisotropic
Anisotropic inversion
inversion results
results at yat=y0 =km.
0 km. Panels
Panels (a)–(c)
(a)–(c) showshow the inverted
the inverted resistivity
resistivity in in
theprincipal
the principalXXdirection
directionforfor
2%,2%,
5%,5%,8% 8% noise
noise level.
level. Panels
Panels (d)–(f)
(d)–(f) showshow
the the inverted
inverted resistivity
resistivity in in
Minerals 2022, 12, x FOR PEER REVIEW 25 of 26
theprincipal
the principalYYdirection
direction
forfor 2%,
2%, 5%,5%,8%8% noise
noise level.
level. Panels
Panels (g)–(i)
(g)–(i) showshow the inverted
the inverted anisotropic
anisotropic
coefficientfor
coefficient for2%,
2%, 5%,
5%, 8%8% noise
noise level.
level.

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

FigureA8.
Figure A8.Anisotropic
Anisotropic inversion
inversion results
results at zat= z6 =km.
6 km. Panels
Panels (a)–(c)
(a)–(c) showshow the inverted
the inverted resistivity
resistivity in in
the principal X direction for 2%, 5%, 8% noise level. Panels (d)–(f) show the inverted
the principal X direction for 2%, 5%, 8% noise level. Panels (d)–(f) show the inverted resistivity in resistivity in
theprincipal
the principalYYdirection
directionforfor 2%,
2%, 5%,5%,
8%8% noisenoise level.
level. Panels
Panels (g)–(i)
(g)–(i) showshow the inverted
the inverted anisotropic
anisotropic
coefficient for 2%, 5%, 8% noise level.
coefficient for 2%, 5%, 8% noise level.

References
1. Hübert, J.; Lee, B.M.; Liu, L.; Unsworth, M.J.; Richards, J.P.; Abbassi, B.; Cheng, L.; Oldenburg, D.W.; Legault, J.M.; Rebagliati,
M. Three-dimensional imaging of a Ag-Au-rich epithermal system in British Columbia, Canada, using airborne z-axis tipper
Minerals 2022, 12, 734 23 of 24

References
1. Hübert, J.; Lee, B.M.; Liu, L.; Unsworth, M.J.; Richards, J.P.; Abbassi, B.; Cheng, L.; Oldenburg, D.W.; Legault, J.M.; Rebagliati, M.
Three-dimensional imaging of a Ag-Au-rich epithermal system in British Columbia, Canada, using airborne z-axis tipper
electromagnetic and ground-based magnetotelluric data. Geophysics 2016, 81, B1–B12. [CrossRef]
2. Wu, Y.; Han, J.; Liu, Y.; Liu, L.; Guo, L.; Guang, Y.; Zhang, Y.; Li, S. Three-dimensional electrical structures and mineralization
significance in the Shuangjianzishan ore-concentrated area, Inner Mongolia. Chin. J. Geophys. 2021, 64, 1291–1304. (In Chinese)
3. Bai, D.; Unsworth, M.J.; Meju, M.A.; Ma, X.; Teng, J.; Kong, X.; Sun, Y.; Sun, J.; Wang, L.; Jiang, C.; et al. Crustal deformation of the
eastern Tibetan plateau revealed by magnetotelluric imaging. Nat. Geosci. 2010, 3, 358–362. [CrossRef]
4. Ogawa, Y.; Ichiki, M.; Kanda, W.; Mishina, M.; Asamori, K. Three-dimensional magnetotelluric imaging of crustal fluids and
seismicity around Naruko volcano, NE Japan. Earth Planets Space 2014, 66, 158. [CrossRef]
5. Yang, W.; Jin, S.; Zhang, L.; Qu, C.; Hu, X.; Wei, W.; Yu, C.; Yu, P. The three-dimensional resistivity structures of the lithosphere
beneath the Qinghai-Tibet Plateau. Chin. J. Geophys. 2020, 63, 817–827. (In Chinese)
6. Piña-Varas, P.; Ledo, J.; Queralt, P.; Marcuello, A.; Bellmunt, F.; Hidalgo, R.; Messeiller, M. 3-D magnetotelluric exploration of
Tenerife geothermal system (Canary Islands, Spain). Surv. Geophys. 2014, 35, 1045–1064. [CrossRef]
7. Zhang, C.; Hu, S.; Song, R.; Zuo, Y.; Jiang, G.; Lei, Y.; Zhang, S.; Wang, Z. Genesis of the hot dry rock geothermal resources in the
Gonghe basin: Constraints from the radiogenic heat production rate of rocks. Chin. J. Geophys. 2020, 63, 2697–2709. (In Chinese)
8. Christensen, N.B. Difficulties in determining electrical anisotropy in subsurface investigations. Geophys. Prospect. 2000, 48, 1–19.
[CrossRef]
9. Wannamaker, P.E. Anisotropy versus heterogeneity in continental solid Earth electromagnetic studies: Fundamental response
characteristics and implications for physicochemical state. Surv. Geophys. 2005, 26, 733–765. [CrossRef]
10. Newman, G.A.; Commer, M.; Carazzone, J.J. Imaging CSEM data in the presence of electrical anisotropy. Geophysics 2010, 75,
F51–F61. [CrossRef]
11. Guo, Z.; Wei, W.; Ye, G.; Jin, S.; Jing, J. Canonical decomposition of magnetotelluric responses: Experiment on 1D anisotropic
structures. J. Appl. Geophys. 2015, 119, 79–88. [CrossRef]
12. Liu, Y.; Xu, Z.; Li, Y. Adaptive finite element modelling of three-dimensional magnetotelluric fields in general anisotropic media.
J. Appl. Geophys. 2018, 151, 113–124. [CrossRef]
13. Weidelt, P. 3-D conductivity models: Implications of electrical anisotropy. In Three-Dimensional Electromagnetics; Society of
Exploration Geophysicists: Houston, TX, USA, 1999.
14. Miensopust, M.P.; Jones, A.G. Artefacts of isotropic inversion applied to magnetotelluric data from an anisotropic Earth. Geophys.
J. Int. 2011, 187, 677–689. [CrossRef]
15. Cao, H.; Wang, K.; Wang, T.; Hua, B. Three-dimensional magnetotelluric axial anisotropic forward modeling and inversion.
J. Appl. Geophys. 2018, 153, 75–89. [CrossRef]
16. Li, Y.; Berlin, F.U.; Prag, J.P.; Brasse, H. Magnetotelluric inversion for 2D anisotropic conductivity structures Inversion methodology.
In Kolloquium Elektromagnetische Tiefenforschung; Burg Ludwigstein; Deutsche Geophysikalische Gesellschaft e.V.: Königstein,
Germany, 2003.
17. Pek, J.; Santos, F.; Li, Y. Non-linear conjugate gradient magnetotelluric inversion for 2-D anisotropic conductivities. In Proceedings
of the 24 SchmuckerWeidelt-Colloquium, Neustadt an der Weinstraße, Germany, 19–23 September 2011.
18. Yu, G.; Farquharson, C.G.; Xiao, Q.; Li, M. Two-dimensional anisotropic magnetotelluric inversion using a limited-memory
quasi-Newton method. Geophysics 2021, 87, E13–E34. [CrossRef]
19. Wang, K.; Wang, X.; Cao, H.; Lan, X.; Duan, C.; Luo, W.; Yuan, J. Magnetotelluric axial anisotropic parallelized 3D inversion based
on cross gradient structural constraint. Chin. J. Geophys. 2021, 64, 1305–1319. (In Chinese)
20. Kong, W.; Tan, H.; Lin, C.; Unsworth, M.; Lee, B.; Peng, M.; Wang, M.; Tong, T. Three-dimensional inversion of magnetotelluric
data for a resistivity model with arbitrary anisotropy. J. Geophys. Res. Solid Earth 2021, 126, e2020JB020562. [CrossRef]
21. Hauserer, M.; Junge, A. Electrical mantle anisotropy and crustal conductor: A 3-D conductivity model of the Rwenzori Region in
western Uganda. Geophys. J. Int. 2011, 185, 1235–1242. [CrossRef]
22. Liu, Y.; Junge, A.; Yang, B.; Lower, A.; Cembrowski, M.; Xu, Y. Electrically anisotropic crust from three-dimensional magnetotel-
luric modeling in the western Junggar, NW China. J. Geophys. Res. Solid Earth 2019, 124, 9474–9494. [CrossRef]
23. Siripunvaraporn, W.; Egbert, G.; Lenbury, Y.; Uyeshima, M. Three-dimensional magnetotelluric inversion: Data-space method.
Phys. Earth Planet. Inter. 2005, 150, 3–14. [CrossRef]
24. Hu, X.; Li, Y.; Yang, W.; Wei, W.; Guo, R.; Han, B.; Peng, R. Three-dimensional magnetotelluric parallel inversion algorithm using
the data-space method. Chin. J. Geophys. 2013, 56, 484–493. (In Chinese)
25. Usui, Y.; Ogawa, Y.; Aizawa, K.; Kanda, W.; Hashimoto, T.; Koyama, T.; Yamaya, Y.; Kagiyama, T. Three-dimensional resistivity
structure of Asama Volcano revealed by data-space magnetotelluric inversion using unstructured tetrahedral elements. Geophys.
J. Int. 2016, 208, 1359–1372. [CrossRef]
26. Kordy, M.; Wannamaker, P.; Maris, V.; Cherkaev, E.; Hill, G. 3-dimensional magnetotelluric inversion including topography using
deformed hexahedral edge finite elements and direct solvers parallelized on symmetric multiprocessor computers—Part II: Direct
data-space inverse solution. Geophys. J. Int. 2016, 204, 94–110. [CrossRef]
27. Pek, J.; Santos, F.A.M. Magnetotelluric inversion for anisotropic conductivities in layered media. Phys. Earth Planet. Inter. 2006,
158, 139–158. [CrossRef]
Minerals 2022, 12, 734 24 of 24

28. Schenk, O.; Gärtner, K. Solving unsymmetric sparse systems of linear equations with PARDISO. Future Gener. Comput. Syst. 2004,
20, 475–487. [CrossRef]
29. Key, K. MARE2DEM: A 2-D inversion code for controlled-source electromagnetic and magnetotelluric data. Geophys. J. Int. 2016,
207, 571–588. [CrossRef]
30. Cai, H.; Long, Z.; Lin, W.; Li, J.; Lin, P.; Hu, X. 3D multinary inversion of controlled-source electromagnetic data based on the
finite-element method with unstructured mesh. Geophysics 2021, 86, E77–E92. [CrossRef]
31. Wang, F.; Morten, J.P.; Spitzer, K. Anisotropic three-dimensional inversion of CSEM data using finite-element techniques on
unstructured grids. Geophys. J. Int. 2018, 213, 1056–1072. [CrossRef]
32. Golob, G.H.; Loan, C.F.V. Matrix Computations, 4th ed.; Johns Hopkins University: Baltimore, MD, USA, 2013.
33. Pacheco, P.S. An Introduction to Parallel Programming; Morgan Kaufmann: Burlington, VT, USA, 2013.
34. Grayver, A.V.; Streich, R.; Ritter, O. Three-dimensional parallel distributed inversion of CSEM data using a direct forward solver.
Geophys. J. Int. 2013, 193, 1432–1446. [CrossRef]
35. Usui, Y. 3-D inversion of magnetotelluric data using unstructured tetrahedral elements: Applicability to data affected by
topography. Geophys. J. Int. 2015, 202, 828–849. [CrossRef]
36. Xiang, Y.; Yu, P.; Zhang, L.; Feng, S.; Utada, H. Regularized magnetotelluric inversion based on a minimum support gradient
stabilizing functional. Earth Planets Space 2017, 69, 158. [CrossRef]
37. Cao, X.; Huang, X.; Yin, C.; Yan, L.; Zhang, B. 3D MT anisotropic inversion based on unstructured finite-element method.
J. Environ. Eng. Geoph. 2021, 26, 49–60. [CrossRef]
38. Zhang, L.; Koyama, T.; Utada, H.; Yu, P.; Wang, J. A regularized three-dimensional magnetotelluric inversion with a minimum
gradient support constraint. Geophys. J. Int. 2012, 189, 296–316. [CrossRef]
39. Menke, W.; Creel, R. Gaussian process regression reviewed in the context of inverse theory. Surv. Geophys. 2021, 42, 473–503.
[CrossRef]