RESEARCH ARTICLE Three-Dimensional Inversion of Magnetotelluric Data for

10.1029/2020JB020562
a Resistivity Model With Arbitrary Anisotropy
Key Points:
Wenxin Kong1,2, Handong Tan1 , Changhong Lin1 , Martyn Unsworth1,2,
• A three-dimensional non-linear
Benjamin Lee2 , Miao Peng1 , Mao Wang1, and Tuo Tong1
conjugate gradient inversion
algorithm is implemented for 1
School of Geophysics and Information Technology, China University of Geosciences, Beijing, China, 2Department of
the inversion of magnetotelluric
(MT) data to produce an electrical Physics, University of Alberta, Edmonton, AB, Canada
resistivity model with arbitrary
anisotropy
• Inversion examples of synthetic and Abstract Electrical anisotropy is increasingly recognized as an important aspect of the resistivity
real MT data validate the usefulness
models required to explain magnetotelluric (MT) observations. However, a limited number of practical
of the algorithm
• Real MT data inversion obtains an MT inversion algorithms that can consider anisotropy have been published to date. To address this
anisotropic model which is more problem, we have developed a three-dimensional (3-D) MT inversion algorithm that recovers a 3-D
consistent with the mapped geology
resistivity model that considers arbitrary electrical anisotropy. The inversion uses the same inversion
than the corresponding isotropic
model algorithm as the widely used ModEM inversion algorithm, and a novel forward modeling algorithm to
consider the anisotropic Earth. The algorithm was tested on both synthetic and field MT data. Inversions
considered both a completely general anisotropy tensor with six components and approximations with
Supporting Information:
Supporting Information may be found
less parameters. Synthetic inversions show that the two horizontal components of resistivity and the
in the online version of this article. anisotropy strike can be well recovered, while the vertical component of resistivity is poorly resolved,
primarily because current flow in MTs is dominantly horizontal. The synthetic examples confirm the
Correspondence to: limitation of the axial anisotropic inversion technique when applied to MT data produced by a resistivity
H. Tan and C. Lin, model with arbitrary anisotropy. The synthetic inversions also showed that inversion of data from an
[email protected]; isotropic model will not result in an artificially anisotropic model. Compared to the isotropic inversion
[email protected]
model of the real MT data, the anisotropic model clearly shows some features that are consistent with the
mapped geology. As expected, the results showed that a given data set can be fit by a range of models, with
Citation:
an inherent trade-off from 3-D heterogeneity to 3-D anisotropy. This uncertainty can be reduced with the
Kong, W., Tan, H., Lin, C., Unsworth,
M., Lee, B., Peng, M., et al. (2021).
use of prior information in the inversion.
Three-dimensional inversion of
magnetotelluric data for a resistivity Plain Language Summary The resistivity model of real Earth generally contains anisotropic
model with arbitrary anisotropy. features. Field magnetotelluric (MT) data may be significantly influenced by these features. To interpret
Journal of Geophysical Research: Solid
Earth, 126, e2020JB020562. https://doi. such data, inversion algorithms which can consider electrical anisotropy are highly demanded. However, a
org/10.1029/2020JB020562 limited number of practical MT inversion algorithms that can consider anisotropy have been published to
date. To address this problem, we have developed a three-dimensional (3-D) MT inversion algorithm that
Received 10 JUL 2020 recovers a 3-D resistivity model that considers arbitrary electrical anisotropy. The algorithm was tested to
Accepted 4 AUG 2021
be effective on both synthetic and field MT data. Compared to the isotropic inversion models of the tested
MT data, the resolved anisotropic models are more reliable.

1. Introduction
With recent advances in the power of computers, three-dimensional (3-D) inversion has become a wide-
ly used tool for magnetotelluric (MT) data interpretation (Miensopust, 2017). Some of the most widely
used 3-D inversion strategies include the model-space Occam method, the data-space Occam method, the
Gauss-Newton (GN) method, the Gauss-Newton with the conjugate gradient (GN-CG) method, the non-lin-
ear conjugate gradient (NLCG) method, the quasi-Newton (QN) method and some variants modified from
them (Siripunvaraporn, 2012 and references therein), among which the Occam and the GN-CG methods
are derived from the GN method. Each of these strategies is feasible for 3-D inversion of MT data and the
efficiency of a particular algorithm depends largely on the code implementation. Our inversion utilizes the
NLCG algorithm which has low memory requirements and a relatively fast convergence rate (Newman &
Alumbaugh, 2000).

© 2021. American Geophysical Union. With a few exceptions, previously developed algorithms for the 3-D inversion of MT data only considered
All Rights Reserved. an Earth model with an isotropic resistivity structure. While this assumption is often valid, there are some

KONG ET AL. 1 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

geological settings where strong electrical anisotropy is present. This can be caused by a range of structures
that include (a) the preferred orientations of fracture porosity, (b) fluidized/melt-bearing zones in the crust,
and (c) hydrogen diffusion in shear-aligned olivine crystals in the mantle (Wannamaker, 2005). This study
addresses the need to consider electrical anisotropy in 3-D MT inversions.

Bostick and Smith (1962) may be the first paper raising the question of possible anisotropy effects in MT
data. In the past 30 years, research into anisotropy effects in MT has been mainly focused on one-dimen-
sional (1-D) and two-dimensional (2-D) anisotropic data analysis (Martí, 2014). For the case of 1-D ani-
sotropy, a number of researchers have made significant contributions in this area, for example, Pek and
Santos (2002, 2006). For forward modeling, Pek and Verner (1997) developed a 2-D MT finite difference
approximation for an anisotropic structure, while Li (2002) used a finite element approach. MT inversion
algorithms were first developed for 2-D anisotropic models using the NLCG method (Pek et al., 2011) and
the GN method (Li et al., 2003). Research in this area also included the 2-D GN inversion of Chen and Weck-
mann (2012), the artificial neural network inversion of Montahaei and Oskooi (2014) and the model-space
Occam inversion of Huang et al. (2016). The most widely used 2-D anisotropic inversion algorithms are the
finite difference code of Baba et al. (2006), which was modified from the 2-D isotropic inversion of Rodi
and Mackie (2001), and the finite element approach of Key (2016) with applications published by Feucht
et al. (2017), Ye et al. (2019), and Yin et al. (2014). Both of these 2-D inversion methods are based on an axial
anisotropic resistivity model, which solves for the three axial resistivities ρx, ρy, and ρz.

For the 3-D MT anisotropic problem, the first attempts at forward modeling were published in the 1990s.
Martinelli and Osella (1997) applied the Rayleigh-Fourier method to the MT forward problem with a 3-D
anisotropic resistivity structure. Weidelt (1999) used the finite difference method to address a similar prob-
lem. After that, there was a two-decade gap without any significant progress on 3-D anisotropic forward
modeling and inversion of MT data for a 3-D anisotropic resistivity model. In recent years this issue has
been revisited with a number of publications focused on the solution of 3-D MT forward problem for an
anisotropic Earth model (Han et al., 2018; Kong et al., 2018; Liu et al., 2018; Xiao et al., 2018, 2019). A
commercial software package called Comsol Multiphysics has been used to simulate the electromagnetic
fields for an arbitrary anisotropic model (Häuserer & Junge, 2011; Löwer & Junge, 2017). In addition, Cao
et al. (2018) implemented the limited-memory Broyden–Fletcher–Goldfarb–Shanno (LBFGS) inversion.
This is a limited-memory version of the QN method for inverting MT data and assumes a 3-D axial aniso-
tropic model.

In order to better approximate the real Earth, we have implemented a 3-D MT inversion algorithm that can
solve for an arbitrary anisotropic model with six parameters. For computational efficiency and to lower
memory requirements, the NLCG strategy is applied to iteratively solve the 3-D MT anisotropic inverse
problem, which avoids the explicit or direct computation and storage of a key component of the inversion:
the Jacobian matrix. This anisotropic inversion algorithm was developed from the widely used 3-D isotropic
inversion software package ModEM (Egbert & Kelbert, 2012; Kelbert et al., 2014). The inversion used the
ModEM algorithm, coupled with a novel algorithm to compute the forward MT response of a 3-D aniso-
tropic Earth. To validate the algorithm, we compared our results with those of Cao et al. (2018). Then, both
synthetic and real MT data are used to further test the inversion algorithm. As with other inversion studies,
this approach cannot overcome the fact that many MT data sets can be fit with either an anisotropic model
or a 3-D isotropic model (Martí, 2014; Weidelt, 1999). Distinguishing between these two options requires
careful analysis and the use of external constraints, such as the mapped geology as described in this study.

2. Methodology
2.1. Mathematical Representation of Electrical Anisotropy

MT data are generally acquired at the Earth’s surface, or seafloor, and then used to infer the subsurface
electrical structure, which is primarily defined by the distribution of the electrical resistivity (or the recipro-
cal quantity conductivity). If the Earth has an isotropic structure, the resistivity is a scalar quantity, which
means it is represented by a single number. If the Earth has an anisotropic resistivity, it is represented by a
tensor, which can be written as a symmetric and positive definite matrix of order three.

KONG ET AL. 2 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Two different definitions of the resistivity (or conductivity) tensor have been proposed (Pek & Santos, 2002;
Yin, 2000). In this study, the definition of Pek and Santos (2002) is applied, which is shown as follows
  xx  xy  xz   x 0 0 
   
  yx  yy  yz   Rz   S  Rx   D  Rz   L   0  y 0  Rz  L  Rx  D  Rz  S  (1)
  zy  zz  0 0  z 
 zx 
where ρx, ρy and ρz are three axial resistivities; αS, αD and αL are three rotation angles that describe the spatial
relationship between the coordinate system and axial directions; Rx   and Rz   are rotation matrices.
Readers are referred to Pek and Santos (2002) for details. Explicit expressions for the entries of the symmet-
ric and positive definite resistivity tensor are given in Appendix A.

For the simplest case of anisotropy, the principal axes are aligned with the coordinate system and the resis-
tivity tensor has just three non-zero entries on the diagonal. In more complicated cases, there are situations
when the off-diagonal elements can be non-zero. The distribution of the non-zero elements depends on the
anisotropic model being considered. For example, the resistivity tensor has only three diagonal non-zero
entries in the vertical anisotropic model of Li (2002), of which the two horizontal axial resistivities are the
same but differ from the vertical one.

2.2. Magnetotelluric Forward Problem

When the electrical resistivity structure of the Earth is defined, the spatial distribution of electric and mag-
netic fields can be computed by solving Maxwell’s equations. MT data are generally measured at the Earth’s
surface, so the electric and magnetic fields at the MT stations should be computed and used to calculate
the predicted MT data, for example, the impedance and tipper data. The forward problem is to calculate the
predicted MT data at these MT stations from a defined electrical conductivity model. Thus, the calculated,
or predicted, MT data are often called model responses.

In MT, the displacement current can be ignored since the frequencies used are very low. Then the frequency
i t
domain Maxwell’s equations can be simplified as follows, assuming a harmonic time dependence of e
  E  i0 H (2)

 H  E (3)
where E and H are the electric and magnetic fields, respectively;  is the angular frequency (equal to 2 f ,
where f is the frequency); 0 is the magnetic permeability of free space;  is the conductivity, which is a
scalar for the isotropic case but a tensor for the anisotropic case. Note that the conductivity of the air layer
is close to zero, but to achieve numerical stability, it is generally assumed to be a very small value (e.g.,
1e−10 S/m) in both the isotropic and anisotropic cases.

For the 3-D MT forward problem, these equations cannot be solved analytically. Only a numerical solu-
tion can be found, which is an approximation to the exact solution. To obtain this numerical solution, the
modeling domain is discretized. Commonly used discretization methods include integral equations, finite
elements and finite differences (see the reviews by Avdeev, 2005 and Börner, 2010). In this study, the stag-
gered-grid finite difference scheme was used, which allows Equations 2 and 3 to be discretized and rear-
ranged into a large linear algebra system of the form
Sm h  b (4)
where Sm is the so-called coefficient matrix, which is a square matrix of order np. Here np equals the sum
of the number of Hx, Hy and Hz components sampled in the modeling domain. Sm is sparse (i.e., most entries
are zero) and symmetric which means that only half of the non-zero entries need to be stored. Sm is equiva-
lent to the expression VSm in Egbert and Kelbert (2012). h is an unknown vector with length np composed
of the magnetic field components at the staggered sampling locations; b is a known vector related to the
boundary magnetic fields, which are computed by the 2-D anisotropic code of Pek and Toh (2000).

Note that the inversion presented in this study uses the previously published ModEM inversion algo-
rithm, coupled with a new 3-D MT forward modeling algorithm. For details of the 3-D anisotropic forward

KONG ET AL. 3 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

modeling algorithm and its validation, readers are referred to our previous forward modeling study (Kong
et al., 2018), which was modified from the isotropic code of Tan et al. (2003) and further modified to be
compatible with the modular implementation of ModEM (Kelbert et al., 2014). In addition, the forward
modeling responses computed with our 3-D code are compared with the 1-D analytical solution of Pek and
Santos (2002) and the 2-D quasi-analytical solution of Qin et al. (2013) in Figures S1–S4. These comparisons
give additional validation of the accuracy of the anisotropic forward modeling algorithm. The accuracy of
the forward code was found to be within 5% for typical parameterizations. Additional details of the forward
modeling are included in Section S3.

2.3. Magnetotelluric Inverse Problem

The inverse problem determines a subsurface electrical resistivity model that fits the acquired MT data.
In this paper, a 3-D anisotropic MT inversion algorithm has been implemented following the computa-
tional recipes of Egbert and Kelbert (2012) and the modular programming framework of ModEM (Kelbert
et al., 2014). Note that we are using the existing ModEM inversion algorithm combined with a new aniso-
tropic forward modeling code.

2.3.1. Objective Function

Mathematically, the solution of the inverse problem is sought by minimizing the following objective function

 
T
  m   d  f  m  Cd1 conjg d  f  m      m  m0  Cm1  m  m0 
T
(5)

where m is the model parameter vector; m0 represents the prior model, which allows the user to include
prior information such as measurements collected in wells or obtained from geological or other geophysical
data; d is the observed data vector; f  m  is the model response vector; Cd is the data covariance matrix, a
diagonal matrix with its diagonal entries being the square of data errors (off-diagonal entries are set to zero
because data errors at different stations are assumed to be uncorrelated); Cm is the model covariance matrix
where a specific smoothing scheme can be applied to the subsurface structure; λ is the regularization factor;
T represents the transpose operation; conjg represents the conjugate operation, with which Equation 5 is
applicable to both real (e.g., apparent resistivity and phase) and complex (e.g., impedance) data given that
the value of   m  must be a real number.

The first-order partial derivative of Equation 5 with respect to the model m gives an expression for the gra-
dient vector

 
g  m   2real J T Cd1 conjg  d  f  m     2 C 1
m  m  m0  (6)

where J is the Jacobian (or sensitivity) matrix, which is derived from the first-order partial derivative of
model response f  m  with respect to the model m; real represents taking the real part operation (applicable
to both real and complex MT data).

Following Egbert and Kelbert (2012), Equation 5 can be reduced to

 
T
  m   d  f  m   C 1 conjg d  f  m     m T m
 (7)
  d  
by the affine linear transformation of the model parameter m  Cm1/2  m  m0 . Then the gradient vector of
Equation 7 can be derived as follows


g  m   2real C1/2 T 1

m J Cd conjg d  f  m
   
   2m (8)

With this transformation, the inversion of Cm is not required at every iteration, which introduces a lot of
freedom to define a model covariance that can include prior information (Siripunvaraporn, 2000). In addi-
tion, the initial model is set equal to the prior model, resulting in m  0. Also, Kelbert et al. (2008) suggested
that this transformation can achieve preconditioning comparable to that used in the previous NLCG inver-
sion algorithms (Haber et al., 2000; Lin et al., 2008; Newman & Alumbaugh, 2000; Rodi & Mackie, 2001)
without additional computational cost.

KONG ET AL. 4 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

We applied the same objective function as Equation 7 in the 3-D NLCG inversion of anisotropic MT data.
The same mathematical symbols can still be used for the anisotropic inversion as for the isotropic one. How-
ever, to mitigate confusion, component-wise symbols corresponding to each anisotropic model parameter
are considered. Because there are six different model parameters in the arbitrary anisotropic case (Pek &
Santos, 2002; Yin, 2000), the model parameter vector m is defined as


m T  m Tx m Ty m Tz mST m TD m TL  (9)

where the column vector m x is composed of the axial resistivity (or conductivity) along the X axis of each
block in the subsurface part of the modeling domain; m y and m z represent the axial resistivities along the
Y and Z axes, respectively; mS is the strike angle; m D is the dip angle; m L is the slant angle. It is obvious
that the expansion of the model parameter vector and other vectors of the same dimension, for example,
the gradient and the search direction vectors in the NLCG scheme, will make the anisotropic inversion
require significantly more computation time and memory than the isotropic inversion. This represents a
significant challenge for the 3-D anisotropic inversion of MT data. However, this challenge can be overcome
with recent development of computer software and hardware. A similar MPI framework to ModEM has
been implemented for the 3-D anisotropic forward and pseudo-forward computations of different periods
and polarizations. The code structure is illustrated in Figure 1. Typical run times of days or weeks can be
achieved for the anisotropic inversion of typical MT data sets. Details are listed in Table 1 in Section 3.2 and
in Figure S5

With the expansion of the model parameter vector, the affine linear transformation is then rewritten as
 Cm1/2 
  mx  mx0 
 Cm1/2  m  m 
  y y0 
 Cm1/ 2   mz  mz0 
m
  m  m  (10)
 Cm1/2  S S0 
 Cm1/2   mD  m D0 
  
  m  mL0 
Cm1/2   L 
 
1/2
where we use the same model covariance (i.e., Cm used in the isotropic version of ModEM) for each kind
of model parameter. For the implementation details of the model covariance, readers are referred to Kelbert
et al. (2008).

For the model parametrization in the inversion, we use the natural logarithms of the three axial resistivities
and the three rotation angles measured in radians. The largest permitted numerical ranges are −16.12–
16.12 for the logarithm axial resistivities, that is, ln (10−7)–ln (107), considering the electrical resistivity
range 10−7–107 Ωm of rocks and other common Earth materials (Simpson & Bahr, 2005), and −6.28–6.28
(i.e., −2π–2π) for angles in radians. If the magnitude of one model parameter is much larger than the oth-
ers, its model norm will dominate the others in the objective function. Thus, the smoothing of the other pa-
rameters can become unstable, which results in an NLCG search direction that will not uniformly smooth
all parameters. Because of the small magnitude difference between the logarithm resistivity and anisotropy
angle in radians, it is acceptable to use the same regularization parameter % for both kinds of parameters as
was done in the 2-D arbitrary anisotropic inversion of Pek et al. (2011).

2.3.2. Calculation of the Jacobian

For almost all MT inversion algorithm strategies, the calculation of the Jacobian matrix J is an essential
component. Egbert and Kelbert (2012) used a new forward mapping scheme ψ h  m  , m for the model  
response function f  m . From the perspective of our experience, this method with the chain rule makes
the derivation of J much easier than those described in the previous publications (e.g., Lin et al., 2008;
Newman & Alumbaugh, 2000) and it was used in this study. Details of the Jacobian calculation are included
in Section S4.

KONG ET AL. 5 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 1. Schematic overview of ModEM (modified from Kelbert et al., 2014). The shading areas show where the
modifications to ModEM were made in the anisotropic inversion. (a) Model space module; (b) Model Mapping module;
(c) Forward solver module; (d) Electromagnetic field interpolation module; (e) Data functionals module; (f) Solver
sensitivity module.

Table 1
Run Time for Each P (x,y,z,S,0,0) Inversion of Two Synthetic Data Sets With Different λani
Synthetic data set 1 λani 0 0.3 1 3 10 30 100 300
Run time 4.16 (h) 3.38 3.76 3.78 4.00 4.01 4.83 5.39
λani 1E3 3E3 1E4 3E4 1E5 1E6 1E7 1E8
Run time 6.03 6.51 6.92 6.58 6.96 6.77 7.19 7.13
Synthetic data set 2 λani 0 0.3 1 3 10 30 100 300
Run time 5.75 (h) 7.94 8.31 8.03 8.09 8.82 8.77 14.86
λani 1E3 3E3 1E4 3E4 1E5 1E6 1E7 1E8
Run time 12.52 18.50 18.43 19.44 19.02 16.53 15.88 16.17

KONG ET AL. 6 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

2.3.3. Inversion Strategy

The NLCG inversion strategy used here is the same as that described by Kelbert et al. (2014) in a pseu-
do-code form and the line search technique mentioned in that pseudo-code can be found in Nocedal and
Wright (2006). Therefore, we will not repeat the details here and provide just a brief introduction. The goal
of the NLCG inversion is to find a search direction and a step length in each iteration of the inversion. The
product of these two components is exactly the update for the model m in one inversion iteration. The
search direction indicates the way to the minimum of the objective function (i.e., where to go). The initial
search direction is defined by the negative gradient, which is also called the steepest descent direction. Then
it is updated herein by the scheme of Polak and Ribière (1969). One of the key factors that make each NLCG
algorithm different may be the technique used to search for the best step length. This step length means how
far the minimization should go along the search direction in one inversion iteration. To find the best step
length, several trial values may be tested. For each of them, one forward computation is needed, which is a
highly time-consuming process. In this study the step length was calculated by an enhanced backtracking
line search procedure based on interpolation of the known objective function values and gradients, which
was detailed described in Nocedal and Wright (2006).

Another important concept to consider in the inversion is the anisotropy penalty applied in the 2-D aniso-
tropic research of Pek et al. (2011), who used the following formulas

       
 k  k  
2
 k  k   
2
  ln  x  ln  y    ln  y  ln  z  
  
 ani  m    kM1    

   
2
 

ln   k   ln   k   
  z x  

 2 1 1  

 
ln  x 
k



     
k k k 



  kM1  ln  x  ln  y  ln  z    1 2 1   ln  y 
1 1 2  

 
k (11)


  
 ln   k 
z

 Wani 0 
 mT  m
0 0
 
where Wani is a square matrix of order 3M (M is the number of subsurface grid cells, i.e., the number of
model parameters in isotropic case).

Taking the affine linear transformation (10) into consideration, Equation 11 can be modified as
 Wani 0  1/2
   
T
  m   C1/2 m  m
   C m  m0 (12)
ani 0 0
m
0 m
 
The gradient of the anisotropy penalty (12) can be derived as follows
  Wani 0  1/2

m
ani 1/2
 2Cm 
0


 C m  m0
0 m

 (13)

Note that Wani is a 3 & 3M matrix.

We investigated three different formulas to include the anisotropy penalty in the objective function (7).
These were

 
T
  m   d  f  m   C 1 conjg d  f  m     m T m   
ani ani  m 
   (14)
  d  

 
T
  m   d  f  m   C 1 conjg d  f  m      m T m   
ani ani  m  
    (15)
  d   

 
  T
T
  m   d  f  m   C 1 conjg d  f  m  
   m  
m m  ani  (16)
  d   1  ani 
ani 

KONG ET AL. 7 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

where λani is a regularization factor similar to λ, which can be automat-

ically reduced in ModEM when the difference between the root-mean-
square (r.m.s.) values of two adjacent inversion iterations is less than a
specified tolerance.

For the objective function (14), if the damping factor λani is large enough,
ρx, ρy and ρz should be approximately equal to each other in order to min-
imize the objective function, which is the isotropic case with a tensor
representation. Therefore, the aim of introducing the anisotropy penal-
ty term is to suppress anisotropy in the resistivity structure that is, not
required by the data (Pek et al., 2011). However, if λani is too large, the
inversion may not be able to converge. Therefore, a similar decrease for
λ would be preferred as for λani. Then, the objective functions (15) and
(16) are better than (14). Basically, these two formulas (15) and (16) can
achieve the same effect by adjusting λani or λ according to our numerical
experiment results, while the former is more intuitive. Therefore, only
the objective function (15) will be considered in the following inversions.

3. Synthetic Inversion Tests

Figure 2. The SM3 model used in Cao et al. (2018). (a) Plan view; (b) To validate our implementation of this algorithm, we compared our 3-D
Section view at the location shown by the dashed line in panel (a). The inversion result with that of Cao et al. (2018) based on their axial aniso-
black stars denote the magnetotelluric stations between −4 and 4 km in tropic synthetic data and carried out some inversions for several synthetic
both horizontal directions (X and Y), 441 in total.
MT data sets that were computed using non-axial anisotropic models.

3.1. Validation of the Code

The model SM3 used in Cao et al. (2018) is shown in Figure 2 and consists of an anisotropic block
(3.6 & 3.6 & 1.0 km) with three axial resistivity values of 10/30/50 Ωm. The top of the block is at a depth of
0.5 km in a 100 Ωm isotropic half space. The full impedance tensor data were calculated for this resistivity
model at five frequencies (50, 10, 5, 1, and 0.1 Hz). The impedance data were then contaminated with 2%
Gaussian noise. The data standard deviation was assumed to be 2% of Z xy  Z yx . The model grid used for
both our NLCG inversion and the LBFGS inversion of Cao et al. (2018) was 43 & 43 & 43 (including 10 air
layers), which was the same discretization as for the forward computation of the synthetic data. Specific
grid parameters are defined in Section S5. The NLCG inversion was started from an initial model of 50 Ωm
isotropic half space (equal to the prior model herein). The regularization parameters λ and λani were set to
1 and 0, respectively.
After 13 inversion iterations, the r.m.s. misfit value had been reduced from 14.7 to 0.95 (shown in Fig-
ure S7). This value is expected to be around unity for a reasonable statistical fit between the observed and
predicted MT data (Figure S8 shows the data fit for off-diagonal apparent resistivities). Figure 3 shows the
comparison between our inversion result and that of Cao et al. (2018). It can be observed that our inversion
model is consistent with the axial anisotropic inversion result of Cao et al. (2018). In particular, the position
and two horizontal axial resistivity values of the target body are recovered satisfactorily (notable artifacts
are observed just below the conductor and not reduced after applying less vertical smoothing). However,
the vertical resistivity is poorly resolved, that is, just small deviations from the starting model, which is as
expected for MT where the electric current flow is primarily horizontal. The weak sensitivity of the MT
data to the vertical resistivity was previously noted in the 1-D and 2-D cases of anisotropic inversion (Pek &
Santos, 2006; Pek et al., 2011). MT data are generally not sensitive to the vertical conductivity, except where
there are strong, vertical conduction pathways to the surface (Evans et al., 2005).

This comparison validates our algorithm for the axial anisotropic case. To further test its utility, we have
used our code to invert two synthetic MT data sets generated from non-axial anisotropic models. The au-
thors recognize that tests on many resistivity models are needed to fully validate a 3-D MT inversion code.
In this paper we focus on (a) a 3-D model containing an anisotropic block and (b) a 1-D model containing

KONG ET AL. 8 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 3. Comparison of the inversion model SM3 of Cao et al. (2018) and the result of the inversion algorithm
developed in this paper. (a) True model; (b, d, and f) Inversion model of Cao et al.; (c, e, and g) Inversion model of
this study. Panels (a–e) show vertical slices for ρx (the first column), ρy (the second column) and ρz (the third column).
Panels (f–g) show horizontal slices for these three axial resistivities. The black dashed box indicates the location of the
target.

an anisotropic layer. These two models illustrate important classes of resistivity model, and were used to
validate the inversion code. An important check was to verify that anisotropy was not introduced by the in-
version when it was not present in the original model. More complex anisotropic models will be considered
in future studies.

KONG ET AL. 9 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 4. Three-dimensional (3-D) block model. (a) Plan view; (b) Section view; (c) Horizontal mesh. The magnetotelluric stations are denoted by black dots
(225 in total). The black dashed lines indicate the location of the 3-D block.

3.2. Synthetic MT Data Set 1

The model for this test is shown in Figure 4. It contains a 3-D anisotropic block (10 & 10 & 4 km) that is,
larger and deeper than that found in the SM3 model. The block is embedded in a 300 Ωm isotropic half
space, with its top at a depth of 2 km. The three axial resistivities for the block are 10/100/10 Ωm and the
strike angle was set to αS = 30°. The angles and resistivities are based on the model of Pek et al. (2011). The
mesh size is 52 & 52 & 37 (including 10 air layers).

The full impedance (denoted as Z) and tipper data (denoted as T) were calculated for this model at 12
logarithmically spaced frequencies in the range 0.003–100 Hz. 2% Gaussian noise was then added to these
data, and the data standard deviation was set to 2% of Z xy  Z yx for the impedance and 0.02 for the tipper.
The starting model was chosen to be a 300 Ωm isotropic half space. The model covariance values were set
to 0.5 for the X, Y and Z directions. This value is higher than the standard value of 0.3 in ModEM, but was
found to give a better recovered model. The regularization parameter λ was set to a starting value 10 and a
range of values of λani from 0 to 108 were investigated. For each inversion, 25 tasks (equal to the sum of the
product of the numbers of polarizations and periods plus 1 master node) were run on 1 cluster node with
20 logical CPUs (equal to the product of 2 physical CPUs and 10 cores per physical CPU) of model Intel(R)
Xeon(R) CPU E5-2680 v2 @ 2.80 GHz. Run times for each inversion are listed in Table 1 and in Figure S5 .
The table and figures show that the anisotropic inversion can be implemented in an MPI environment with
realistic run times.

The isotropic inversion model of the anisotropic synthetic data set 1 is shown in Figure S9, which shows a
complex subsurface structure with a few inhomogeneous features resolved in both the 3-D block and the
host rock.

In the most general case, the inversion solves for six model parameters in each cell (three axial resistivity
values and three angles that define the orientation of the principal components). However, this full inver-
sion for six parameters is computationally demanding, especially when topography is included. Therefore
inversions solving for just three and four parameters were investigated to see if they could fit the data
with reduced computational effort. To make it clear what is being inverted, we introduce the notation P
(x,y,z,S,D,L) where P is short for parameter and x, y, z, S, D, and L indicate each anisotropic parameter. The
following inversions used the Z data for the case when no restriction is applied to the anisotropy (i.e., λani
is set to 0).

KONG ET AL. 10 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 5. Synthetic data set 1: Comparison of the horizontal slices at z = 3.0 km of different inversion models. (a) True model; (b) P (x,y,z,0,0,0) inversion
model, (c) P (x,y,z,S,0,0) inversion model, and (d) P (x,y,z,S,D,L) inversion model with λani = 0; (e) P (x,y,z,S,0,0) inversion model with λani = 1,000. The black
dashed box indicates the location of the target. The objective function is defined in Equation 15.

1. P (x,y,z,0,0,0): The first inversion solved for just the three axial resistivities ρx, ρy, ρz with the three angles
forced to zero and is denoted by P (x,y,z,0,0,0), where the “0” indicates the parameter at that index lo-
cation does not vary during the inversion and is equal to its initial value. This inversion converged to a
misfit of 1.23 and the model contains some artifacts as the inversion can only fit the MT data by placing
inhomogeneities in both the 3-D block and surrounding half space (Figures 5b and 6b). ρx is generally
well resolved while ρy and ρz are poorly resolved. These limitations arise because the inversion does not
have the flexibility to correctly reproduce the true resistivity structure. Data fit is illustrated in Figure S2
(1b–16b), and a wide range λani values are considered.
2. P (x,y,z,S,0,0): The second inversion solved for ρx, ρy, ρz and αS. In this inversion, the angles αD and αL
were forced to be zero implying a horizontal orientation for the anisotropic structure. The inversion
achieved a misfit of 1.04 and recovered the original model more reliably than the previous inversion, in
terms of both the resistivity parameters ρx, ρy and the angle parameter αS (Figures 5c and 6c). However,
ρz is still poorly resolved due to the fact that MT induces currents that flow horizontally and are relative-
ly insensitive to this parameter. This example reveals the limitations of the axial anisotropic inversion
when applied to data from a more complicated anisotropic resistivity structure.

KONG ET AL. 11 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 6. Synthetic data set 1: Comparison of the vertical slices at x = 0.0 km and y = 0.0 km of different inversion models. (a) True model; (b) P (x,y,z,0,0,0)
inversion model, (c) P (x,y,z,S,0,0) inversion model, and (d) P (x,y,z,S,D,L) inversion model with λani = 0; (e) P (x,y,z,S,0,0) inversion model with λani = 1,000. The
black dashed box indicates the location of the target.

3. P (x,y,z,S,D,L): The inversion model with general anisotropy considered in the inversion, that is, P
(x,y,z,S,D,L), is shown in Figures 5d and 6d. ρx and ρy are resolved quite well. Note that αS and αL share
half of the theoretical value for αS, which is defined as a horizontally deflected angle when αD and αL are
zero. This is understandable because αS and αL are both horizontally deflected angles in this case where
αD is almost 0. This is caused by the inherent non-uniqueness of geophysical inversion, which is more
severe in the anisotropic case.

It is well known that there is an inherent trade-off between fitting data with model heterogeneity and aniso-
tropy (Wannamaker, 2005). This effect is examined by varying the degree of anisotropy allowed in the model
by increasing λani. As shown in Figures 5c and 6c, the P (x,y,z,S,0,0) inversion resolves the true model quite
well, although some inhomogeneous features were still observed in the region surrounding the 3-D block.
As λani increases to a relatively large value, this phenomenon is reduced as the model is forced to be smooth-
er. Figures 5e and 6e shows the P (x,y,z,S,0,0) inversion model for the Z data with λani = 1,000. Compared
to the inversion model with λani = 0 (Figures 5c and 6c), ρy is relatively homogeneous in the region outside
the block and the recovered resistivity is less than the true value. It can also be seen that ρz is better resolved
than in the inversion with λani = 0 due to applying the extra model constraint of anisotropy. Geological in-
formation can provide external information about the degree of anisotropy expected, and could influence
the choice of λani for the inversion of field MT data.

Space does not permit all the inversion models and fits to data to be displayed. These are included in the
Data Set S1–S8 where the inversion is shown for the Z data and Z + T data for all values of λani. Inversions
included are: P (x,y,z,0,0,0), P (x,y,z,S,0,0), P (x,y,z,S,D,0) and P (x,y,z,S,D,L). Some conclusions of these
inversions are as follows:

1. Adding tipper was not found to significantly change the recovered model, that is, the Z and Z + T models
were quite similar.
2. For most inversions, ρx, ρy, and αS are quite well resolved, while ρz is poorly resolved as expected unless
for large values of λani.

KONG ET AL. 12 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 7. Synthetic data set 2: Comparison of the horizontal slices at z = 3.0 km of different inversion models. (a) True model; (b) P (x,y,z,0,0,0) inversion
model with λani = 0; (c) P (x,y,z,0,0,0) inversion model with λani = 30; (d) P (x,y,z,S,0,0) inversion model with λani = 0; (e) P (x,y,z,S,0,0) inversion model with
λani = 30.

3.3. Synthetic MT Data Set 2

The model for this test is a 1-D layered resistivity model rather than a 3-D model. This model was chosen
to investigate the question of how MT data can distinguish heterogeneity from anisotropy. In this model an
anisotropic layer replaced the 3-D block at the same depth and with the same anisotropic model parameters.
All the other settings for the computation of the synthetic MT data and inversion were kept the same as
for the previous model. Run times for each inversion of Synthetic data set 2 can be found in Table 1 and in
Figure S5

1. The P (x,y,z,0,0,0) inversion is shown in Figures 7b and 8b with λani = 0 and the final misfit was 1.29. The
depth of the layer is resolved quite well and the true value of ρy is better recovered than that of ρx. However,
ρy shows more spatial variations (i.e., model is heterogeneous) than ρx. Again, ρz is poorly resolved, as a
result of the horizontal current flow making the MT data relatively insensitive to this parameter.

This inversion was run with λani = 30 and a misfit of 1.61 was obtained after 160 iterations. The final model
is shown in Figures 7c and 8c, and it contains more heterogeneity than that obtained with λani = 0.

2. The P (x,y,z,S,0,0) inversion gives a slightly better model than P (x,y,z,0,0,0) as expected and both ρx and
ρy are relatively well recovered (Figures 7d and 8d). The angle αS has a value in the range 15–20° which
is lower than the true value of 30°. It appears that the anisotropy in the measured MT data is partly
reproduced by spatial variations in ρx and ρy (i.e., heterogeneity) rather than αS (anisotropy). Again, ρz
is poorly resolved, as a result of the horizontal current flow in MT making the data insensitive to this
parameter.

For this example the 3-D MT inversions show that the true model resistivity is relatively well recovered in
the center of the grid of MT stations. Model resolution is lowest at the edge of the grid, as routinely observed

KONG ET AL. 13 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 8. Synthetic data set 2: Comparison of the vertical slices at x = 0.0 km and y = 0.0 km of different inversion models. (a) True model; (b) P (x,y,z,0,0,0)
inversion model with λani = 0; (c) P (x,y,z,0,0,0) inversion model with λani = 30; (d) P (x,y,z,S,0,0) inversion model with λani = 0; (e) P (x,y,z,S,0,0) inversion model
with λani = 30. The black dashed lines indicate the location of the target.

in MT inversions of field data. An additional test with a limited number of additional MT stations distribut-
ed outside the dense grid improved the model recovery, as illustrated in Figures S10 and S11.

This inversion was run with λani = 30 and a misfit of 1.32 was obtained after 130 iterations. The model is
shown in Figures 7e and 8e. The model still shows some heterogeneity, but the anisotropy is recovered more
reliably than the inversion with λani = 0.

As with data set 1, many inversions were carried out, including P (x,y,z,0,0,0), P (x,y,z,S,0,0), P (x,y,z,S,D,0)
and P (x,y,z,S,D,L) with both the Z data and Z + T data. For a 1-D model the vertical magnetic field is zero
and therefore T is zero. It is still worth inverting T as this still provides a constraint on the inversion model.
These inversions with a range of values for λani are presented in the Data Set S9–S16. It should be noted that
model resolution at the edge of the grid of MT stations is reduced, compared to the center. This results in
artifacts, such as the dip of the low resistivity layer at the edge of the array of MT stations. This phenomenon
is a consequence of the physics of MT. This is common in isotropic MT inversion and is a greater problem
in anisotropic MT inversion because of the greater number of parameters. Interpretation of models should
consider this effect and it reflects the inherent trade-off between model heterogeneity and model anisotropy
well known to MT practitioners (Wannamaker, 2005).

3.4. Anisotropic Inversion of Isotropic Data Sets

An important test of our inversion algorithm is to investigate if it can correctly invert MT data from a 3-D
isotropic resistivity model. The anisotropic 3-D block for Synthetic data set 1 was modified to be isotropic
with the resistivity of 10 Ωm and labeled Synthetic data set 3. Similarly, the anisotropic layer model was
made isotropic with the resistivity of 10 Ωm and labeled Synthetic data set 4.

The inversion control settings were kept the same as in the above anisotropic inversions and λani = 0. Fig-
ures 9 and 10 show the inversion models for both data sets. It can be observed that two horizontal resis-
tivities 'x and 'y are well resolved, while vertical resistivity ρz is still poorly resolved. Though poor, ρz has a
relatively better resolution in the 3-D case than the 1-D one. No features indicating anisotropy, that is, the
significant difference between horizontal resistivities and the large values of anisotropy angles, are discov-
ered in the anisotropic inversion models of the isotropic data sets. It indicates that the anisotropic inversion
algorithm can be used to invert isotropic data without introducing significant anisotropic features.

Only inversion models of the Z data are shown here. Inversion models of Z + T data and more inversion
details like the data fits can be found in the Data Set S18 and S20.

KONG ET AL. 14 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 9. Synthetic data set 3 and 4: Anisotropic inversion models of isotropic data. Horizontal slices at z = 3.0 km. (a) True model, and (b) P (x,y,z,S,D,L)
inversion model of Synthetic data set 3 for the isotropic 3-D block model; (c) True model, and (d) P (x,y,z,S,D,L) inversion model of Synthetic data set 4 for the
isotropic layer model. The black dashed lines indicate the location of the target.

4. Inversion of Field MT Data

4.1. Field MT Data and Inversion Setup

The field MT data set consists of 57 stations acquired as part of an exploration program for geothermal
resources in the Rocky Mountain Trench in British Columbia. The survey was performed in a region with
highly deformed metamorphic rocks displaying anisotropic textures and containing sulfides and graphite
detected by optical and scanning electron microscopy (Finley, 2020). The combination of (a) anisotropic

KONG ET AL. 15 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 10. Synthetic data set 3 and 4: Anisotropic inversion models of isotropic data. Vertical slices at x = 0.0 km and y = 0.0 km. (a) True model, and (b) P
(x,y,z,S,D,L) inversion model of Synthetic data set 3 for the isotropic 3-D block model; (c) True model, and (d) P (x,y,z,S,D,L) inversion model of Synthetic data
set 4 for the isotropic layer model. The black dashed lines indicate the location of the target.

grain textures and (b) possibly interconnected sulfides and graphite along rock foliation motivated separate
isotropic and anisotropic inversions of the MT data. The MT data were inverted to obtain 3-D isotropic and
anisotropic resistivity models (Lee, 2020). The spatial distribution of the MT stations is shown in Figures 12
and 13.
Full impedance tensor data were measured at all 57 stations and tipper data were available at 44 stations.
Data at 23 frequencies in the range 320–0.0125 Hz were considered. These data were inverted with both the
isotropic ModEM software package of Kelbert et al. (2014) and the anisotropic code developed here. Imped-
ance data were assigned an error floor of 5% of Z xy  Z yx while tipper data were assigned a constant error
of 0.02. The corresponding model grid consists of 74 & 134 & 83 cells in the X, Y, Z directions, respectively.
The core cells have a size of 50 & 50 m and it increases with a factor of 1.5 for the padding area. The model
covariance was set to 0.5. The regularization parameter λ was set to 10. All inversions start from an isotropic
half space of 50 Ωm.

4.2. Inversion Results

Several anisotropic inversions were run for the real MT data example, including P (x,y,z,0,0,0), P (x,y,z,S,0,0),
and P (x,y,z,S,D,0) inversions for λani = 0 as well as P (x,y,z,S,0,0) inversion for λani = 10. Figure 11 shows
some vertical slices of the inversion models along the profile P1 shown in panel (e) in Figure 13. As can
be seen from the figure, all the resistivity models are relatively similar for inversions with 3, 4, or 5 param-
eters defining the anisotropic resistivity. Increasing the anisotropy penalty to a value of λani = 10 did not
significantly change the model, especially in terms of the anisotropic features AF2 and AF3 located beside

KONG ET AL. 16 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 11. Vertical slices through the anisotropic 3-D resistivity and rotation angle models along the profile P1 shown in panel (e) in Figure 13. (a) P
(x,y,z,0,0,0) inversion model, (b) P (x,y,z,S,0,0) inversion model and (c) P (x,y,z,S,D,0) inversion model for for λani = 0; (d) P (x,y,z,S,0,0) inversion model for
λani = 10. Black inverted triangles show locations of MT stations. AF2 and AF3 are features described in the main text. F2 and F3 are faults described in the
main text.

the fault F2, which will be discussed below in detail. Thus, we will focus on the results of the P (x,y,z,0,0,0)
inversion in the later analysis because it achieves the lowest r.m.s. misfit (see Figure 11) among all.

Figure 12 shows examples of the data misfit from the isotropic and the P (x,y,z,0,0,0) anisotropic inversions.
Panel (a) shows that the overall r.m.s. misfit of the anisotropic inversion (1.59) was lower than the r.m.s.
misfit of the isotropic inversion (2.06). Panels (b) and (c) are maps of the r.m.s. misfit at each MT station
from the isotropic and anisotropic inversions, respectively. The anisotropic inversion also obtained a lower
r.m.s. misfit at each period, and also on a station-by-station basis. Examples of apparent resistivity and
phase data from stations labeled 1 to 4 are shown in panels (d) to (g). While the data fit is slightly improved
at stations 2 and 4, the anisotropic inversion clearly improved the data fit at stations 1 and 3. For example,
as seen in panel (d), the YY apparent resistivity and phase at station 1 are much more closely fit by the an-
isotropic inversion. At station 3 (panel f), the anisotropic inversion clearly improved the fit of the XX and
YX component data. Overall, it is clear that the anisotropic inversion obtained a better fit to the observed
MT data than the isotropic inversion. Although the lower r.m.s. misfit from the anisotropic inversion is
preferred, the comparison of resistivity models in the following section is needed to conclusively choose the
preferred inversion result.

4.3. Interpretation and Discussion

The resistivity models obtained from the isotropic and anisotropic inversions are shown in Figure 13. Note
that this inversion solved for three parameters, as the inclusion of topography made the six parameter

KONG ET AL. 17 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 12. Examples of data misfit from the isotropic and anisotropic inversions of the magnetotelluric (MT) data set from the Rocky Mountain Trench. (a)
root-mean-square (r.m.s.) misfit as a function of period from the isotropic (iso) and anisotropic (aniso) inversions. r.m.s. misfit at each MT station is shown as
a colored circle for the (b) isotropic and (c) anisotropic inversions. Apparent resistivity (ρa) and phase (φ) as a function of period are shown in (d, e, f, g) for
stations 1 to 4 (labeled in b and c). See legend for symbol definitions. Tipper data fit is shown in Figure S12.

inversion computationally challenging. Horizontal slices through the resistivity models are shown in panels
(a) to (d) and vertical slices through the model along two profiles are shown in (f) to (i). The locations of
the profiles P1 and P2 are shown in panel (e). In order to focus on comparing the isotropic and anisotropic
models, not all features in the resistivity models are discussed in this study (for a full interpretation of the
resistivity models see Lee, 2020). The three highly conductive features IC3, IC4, and IC5 are conspicuous
features in the isotropic model (see panels a and f). These conductors are distinct from each other and are
embedded in the highly resistive feature IR1. In contrast, the anisotropic model contains simpler structure
in the same area. The horizontal slices in panels (b) to (d) show that instead of the isotropic features IR1,
IC3, IC4, and IC5, the anisotropic model simply contains one anisotropic feature (AF3) that is, highly con-
ductive in ρy, moderately conductive in ρx, and resistive in ρz. Features in the anisotropic resistivity model
are also in better agreement with the mapped geology (Finley, 2020; Lee, 2020). F2 is a regional-scale,
steeply dipping normal fault with an unknown amount of strike and dip separation. F3 is a regional thrust
fault predating F2 with an estimated shortening of at least 20 km. The anisotropic model contains a sharp
boundary separating the moderately anisotropic feature AF2 from the highly anisotropic AF3 which co-
incides with the mapped fault trace of F2. This resistivity boundary is clearly seen in the horizontal slices
as well as the vertical slices (panels f to i). Furthermore, the isotropic feature IC4, seen in panel (f), has a

KONG ET AL. 18 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Figure 13. Slices through the isotropic and anisotropic 3-D resistivity models obtained from inversion of the Rocky Mountain Trench magnetotelluric (MT)
data set. (a) Horizontal slice at 0 km depth through the isotropic model and anisotropic model (b, c, d). ρx, ρy, and ρz are the principal resistivities in the north,
east, and vertical directions, respectively. (e) surface traces of profiles P1 and P2. Vertical slices through the isotropic and anisotropic models along profiles P1
and P2 are shown in (f, g, h, i). Black circles and inverted triangles show locations of MT stations. IR1, IC3, IC4, IC5, AF2, and AF3 are features described in the
main text. F2 and F3 are faults described in the main text.

dip angle that is, nearly perpendicular to the inferred planes of F2 and F3. In comparison, the anisotropic
feature AF3 is bounded by F2 and F3 and may represent a distinct lithology from the hanging wall of F2
and the footwall of F3.

The large differences between the anisotropic and isotropic resistivity models demonstrate non-unique-
ness in modeling MT data, and are related to the insensitivity of MT data to the spatial scale of anisotropy
(Wannamaker, 2005; Weidelt, 1999). The concept of non-uniqueness has been demonstrated by comparing
isotropic models with repeating macroscopic features (i.e., structural anisotropy) to anisotropic models with
microscopic anisotropy (e.g., Eisel & Haak, 1999; Kellett et al., 1992; Patro & Egbert, 2011). In such cases
where the MT data fit is comparable, additional geological information is needed to choose the more real-
istic model.

For our example in the Rocky Mountain Trench, we believe the problem of non-uniqueness is can be ad-
dressed by comparing the isotropic and anisotropic models to the known geology. Even though the isotropic
inversion obtained a satisfactory r.m.s. misfit of 2.06, the model contains macroscopic features in disagree-
ment with geological mapping. We prefer the simpler geometry of features in the anisotropic model because
the anisotropic features are in better agreement with the inferred locations of faults F2 and F3 (Figures 11
and 13). This example demonstrates that in some complicated geological settings, a 3-D anisotropic inver-
sion may be necessary to obtain a more realistic model than a 3-D isotropic inversion.

5. Conclusions
We have implemented a 3-D MT nonlinear conjugate gradient inversion algorithm that determines a resis-
tivity model with arbitrary anisotropy. The algorithm is validated through the inversion model comparison
with the axial anisotropy result of Cao et al. (2018). Then it is further tested with both synthetic and real
MT data.

KONG ET AL. 19 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

The synthetic examples demonstrate that: (a) the two horizontal resistivities and the anisotropy strike can
be well resolved, while the vertical resistivity is poorly resolved; (b) the introduction of the anisotropy pen-
alty can reduce some model inhomogeneity (data set 1) and improve the inversion model that allowed the
anisotropy strike to vary (data set 2); (c) the P (x,y,z,0,0,0) inversion, that is, the axial anisotropic inversion,
has some limitations when inverting for a model with arbitrary anisotropic MT data even if the data fit is
basically as good; (d) isotropic data can be inverted with the anisotropic algorithm to obtain an isotropic
model rather than an anisotropic model with anisotropy appearing everywhere.

For the Rocky Mountain Trench MT data, the isotropic inversion obtained a model containing macroscopic
conductors and resistors, which appear to be artifacts and difficult to reconcile with the known geology.
In contrast, the anisotropic model clearly shows a simpler structure that is, consistent with the mapped
geology. The anisotropic model gives a better fit to the data than the isotropic model, although it should be
noted that this may be partly due to the additional number of free parameters.

In conclusion, these examples show the usefulness of the anisotropic inversion algorithm. As expected,
the inversion algorithm cannot overcome the fact that a given anisotropic data set can be fit by models that
contain a mixture of anisotropy and heterogeneity. Additional research should focus on how external con-
straints can be used to address this issue.

Appendix A
The explicit expressions of the entries of the symmetric resistivity tensor for an arbitrary anisotropic Earth
are given as follows:

  cos S cos L  sin  S cos D sin  L   x

2
 xx
(A1)
  cos S sin  L  sin  S cos D cos L   y  sin 2  S sin 2  D  z
2

  sin  S cos L  cos S cos D sin  L   x

2
 yy
(A2)
  cos S cos D cos L  sin  S sin  L   y  cos2  S sin 2  D  z
2

 zz  sin 2  D sin 2  L  x  sin 2  D cos2  L  y  cos2  D  z (A3)

 xy   cos S cos L  sin  S cos D sin  L   sin  S cos  L  cos S cos D sin  L   x
  cos S sin  L  sin  S cos D cos L   sin  S sin  L  cos S cos D cos L   y
(A4)
 sin  S cos S sin  D  z
2

 xz   cos S cos L  sin  S cos D sin  L  sin  D sin  L  x

  cos S sin  L  sin
n  S cos D cos L  sin  D cos L  y (A5)
 sin  S sin  D cos D  z

 yz   sin  S cos L  cos S cos D sin  L  sin  D sin  L  x

  sin  S sin  L  coss S cos D cos L  sin  D cos L  y (A6)
 cos S sin  D cos D  z

KONG ET AL. 20 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Data Availability Statement

The developed anisotropic inversion code is freely available to the academic community and the data sets
used in this study are openly available at https://osf.io/uwhxv/files/.

Acknowledgments References
The authors thank Gary Egbert, Anna
Kelbert, and Naser Meqbel for making Avdeev, D. B. (2005). Three-dimensional electromagnetic modelling and inversion from theory to application. Surveys in Geophysics, 26(6),
their ModEM software package availa- 767–799. https://doi.org/10.1007/s10712-005-1836-x
ble. The authors also thank Josek Pek Baba, K., Chave, A. D., Evans, R. L., Hirth, G., & Mackie, R. L. (2006). Mantle dynamics beneath the East Pacific Rise at 17°S: Insights
for use of his 2-D anisotropic forward from the Mantle Electromagnetic and Tomography (MELT) experiment. Journal of Geophysical Research, 111(B2), B02101. https://doi.
modeling code. Anna Kelbert and two org/10.1029/2004jb003598
anonymous reviewers as well as the Börner, R. (2010). Numerical modelling in geo-electromagnetics: Advances and challenges. Surveys in Geophysics, 31(2), 225–245. https://
Associate Editor Max Moorkamp are doi.org/10.1007/s10712-009-9087-x
thanked for their constructive reviews. Bostick, F., & Smith, H. (1962). Investigation of large-scale inhomogeneities in the Earth by the magnetotelluric method. Proceedings of the
This study was supported by research IRE, 50(11), 2339–2346. https://doi.org/10.1109/JRPROC.1962.287961
grants from the State Key Program of Cao, H., Wang, K., Wang, T., & Hua, B. (2018). Three-dimensional magnetotelluric axial anisotropic forward modeling and inversion.
National Natural Science Foundation of Journal of Applied Geophysics, 153, 75–89. https://doi.org/10.1016/j.jappgeo.2018.04.015
China (No. 41830429) and the National Chen, X., & Weckmann, U. (2012). Inversion of 2D magnetotelluric data with anisotropic conductivities. Paper presented at the Extended
Natural Science Foundation of China Abstract 21st Workshop.
(Nos. 41674134, 41874159). The support Egbert, G. D., & Kelbert, A. (2012). Computational recipes for electromagnetic inverse problems. Geophysical Journal International, 189(1),
provided by the China Scholarship 251–267. https://doi.org/10.1111/j.1365-246X.2011.05347.x
Council (CSC) during a visit of Wenxin Eisel, M., & Haak, V. (1999). Macro-anisotropy of the electrical conductivity of the crust: A magnetotelluric study of the German Conti-
Kong to the University of Alberta is nental Deep Drilling site (KTB). Geophysical Journal International, 136(1), 109–122. https://doi.org/10.1046/j.1365-246X.1999.00707.x
acknowledged. Many of the inversions Evans, R. L., Hirth, G., Baba, K., Forsyth, D., Chave, A., & Mackie, R. (2005). Geophysical evidence from the MELT area for compositional
were made possible by access to the controls on oceanic plates. Nature, 437, 249. https://doi.org/10.1038/nature04014
computer clusters operated by Westgrid Feucht, D. W., Sheehan, A. F., & Bedrosian, P. A. (2017). Magnetotelluric imaging of lower crustal melt and lithospheric hydration in the
and Compute Canada. rocky mountain front transition zone, Colorado, USA. Journal of Geophysical Research: Solid Earth, 122(12), 9489–9510. https://doi.
org/10.1002/2017JB014474
Finley, T. (2020). Master's thesisFault-hosted geothermal systems in southeastern British Columbia. Canada: University of Alberta.
Haber, E., Ascher, U. M., & Oldenburg, D. (2000). On optimization techniques for solving nonlinear inverse problems. Inverse Problems,
16(5), 1263–1280. https://doi.org/10.1088/0266-5611/16/5/309
Han, B., Li, Y., & Li, G. (2018). 3D forward modeling of magnetotelluric fields in general anisotropic media and its numerical implementa-
tion in Julia. Geophysics, 83(4), F29–F40. https://doi.org/10.1190/geo2017-0515.1
Häuserer, M., & Junge, A. (2011). Electrical mantle anisotropy and crustal conductor: A 3-D conductivity model of the Rwenzori Region in
western Uganda. Geophysical Journal International, 185(3), 1235–1242. https://doi.org/10.1111/j.1365-246X.2011.05006.x
Huang, Y., Hu, X., & Han, B. (2016). Forward and inverse modeling of the magnetotelluric field in 2D anisotropic media with an adaptive
finite element method. Oil Geophysical Prospecting, 51(4), 809–820. https://doi.org/10.13810/j.cnki.issn.1000-7210.2016.04.024-en
Kelbert, A., Egbert, G. D., & Schultz, A. (2008). Non-linear conjugate gradient inversion for global EM induction: Resolution studies. Geo-
physical Journal International, 173(2), 365–381. https://doi.org/10.1111/j.1365-246X.2008.03717.x
Kelbert, A., Meqbel, N., Egbert, G. D., & Tandon, K. (2014). ModEM: A modular system for inversion of electromagnetic geophysical data.
Computers and Geosciences, 66, 40–53. https://doi.org/10.1016/j.cageo.2014.01.010
Kellett, R. L., Mareschal, M., & Kurtz, R. D. (1992). A model of lower crustal electrical anisotropy for the Pontiac subprovince of the Cana-
dian shield. Geophysical Journal International, 111(1), 141–150. https://doi.org/10.1111/j.1365-246X.1992.tb00560.x
Key, K. (2016). MARE2DEM: A 2-D inversion code for controlled-source electromagnetic and magnetotelluric data. Geophysical Journal
International, 207(1), 571–588. https://doi.org/10.1093/gji/ggw290
Kong, W., Lin, C., Tan, H., Peng, M., Tong, T., & Wang, M. (2018). The effects of 3d electrical anisotropy on magnetotelluric responses:
Synthetic case studies. Journal of Environmental and Engineering Geophysics, 23(1), 61–75. https://doi.org/10.2113/jeeg23.1.61
Lee, B. (2020). Doctoral dissertationImproving Exploration for geothermal Resources with the magnetotelluric method. Canada: University
of Alberta.
Li, Y. (2002). A finite-element algorithm for electromagnetic induction in two-dimensional anisotropic conductivity structures. Geophysi-
cal Journal International, 148(3), 389–401. https://doi.org/10.1046/j.1365-246x.2002.01570.x
Li, Y., Pek, J., & Brasse, H. (2003). Magnetotelluric inversion for 2D anisotropic conductivity structures. Paper presented at the 20th Colloq.
“Electromagnetic Depth Investigations”.
Lin, C., Tan, H., & Tong, T. (2008). Three-dimensional conjugate gradient inversion of magnetotelluric sounding data. Applied Geophysics,
5(4), 314–321. https://doi.org/10.1007/s11770-008-0043-1
Liu, Y., Xu, Z., & Li, Y. (2018). Adaptive finite element modelling of three-dimensional magnetotelluric fields in general anisotropic media.
Journal of Applied Geophysics, 151, 113–124. https://doi.org/10.1016/j.jappgeo.2018.01.012
Löwer, A., & Junge, A. (2017). Magnetotelluric transfer functions: Phase tensor and tipper vector above a simple anisotropic three-dimen-
sional conductivity anomaly and implications for 3D isotropic inversion. Pure and Applied Geophysics, 174(5), 2089–2101. https://doi.
org/10.1007/s00024-016-1444-3
Martí, A. (2014). The role of electrical anisotropy in magnetotelluric responses: From modelling and dimensionality analysis to inversion
and interpretation. Surveys in Geophysics, 35(1), 179–218. https://doi.org/10.1007/s10712-013-9233-3
Martinelli, P., & Osella, A. (1997). MT forward modeling of 3-D anisotropic electrical conductivity structures using the Rayleigh-Fourier
method. Journal of Geomagnetism and Geoelectricity, 49(11–12), 1499–1518. https://doi.org/10.5636/jgg.49.1499
Miensopust, M. P. (2017). Application of 3-D electromagnetic inversion in practice: Challenges, pitfalls and solution approaches. Surveys
in Geophysics, 38(5), 869–933. https://doi.org/10.1007/s10712-017-9435-1
Montahaei, M., & Oskooi, B. (2014). Magnetotelluric inversion for azimuthally anisotropic resistivities employing artificial neural net-
works. Acta Geophysica, 62(1), 12–43. https://doi.org/10.2478/s11600-013-0164-7

KONG ET AL. 21 of 22
 21699356, 2021, 8, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2020JB020562 by Capes, Wiley Online Library on [07/06/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Solid Earth 10.1029/2020JB020562

Newman, G. A., & Alumbaugh, D. L. (2000). Three-dimensional magnetotelluric inversion using non-linear conjugate gradients. Geophys-
ical Journal International, 140(2), 410–424. https://doi.org/10.1046/j.1365-246x.2000.00007.x
Nocedal, J., & Wright, S. J. (2006). Numerical optimization (2nd ed.,). Springer.
Patro, P. K., & Egbert, G. D. (2011). Application of 3D inversion to magnetotelluric profile data from the Deccan volcanic province of west-
ern India. Physics of the Earth and Planetary Interiors, 187(1), 33–46. https://doi.org/10.1016/j.pepi.2011.04.005
Pek, J., & Santos, F. A. M. (2002). Magnetotelluric impedances and parametric sensitivities for 1-D anisotropic layered media. Computers
& Geosciences, 28(8), 939–950. https://doi.org/10.1016/S0098-3004(02)00014-6
Pek, J., & Santos, F. A. M. (2006). Magnetotelluric inversion for anisotropic conductivities in layered media. Physics of the Earth and Plan-
etary Interiors, 158(2), 139–158. https://doi.org/10.1016/j.pepi.2006.03.023
Pek, J., Santos, F. A. M., & Li, Y. (2011). Non-linear conjugate gradient magnetotelluric inversion for 2-D anisotropic conductivities. Paper
presented at the Schmucker-Weidelt-Kolloquium, Neustadt an der Weinstrasse.
Pek, J., & Toh, H. (2000). Numerical modelling of MT fields in 2-D anisotropic structures with topography and bathymetry considered. Paper
presented at the 18th Colloq “Electromagnetic depth investigations”.
Pek, J., & Verner, T. (1997). Finite-difference modelling of magnetotelluric fields in two-dimensional anisotropic media. Geophysical Jour-
nal International, 128(3), 505–521. https://doi.org/10.1111/j.1365-246X.1997.tb05314.x
Polak, E., & Ribiere, G. (1969). Note sur la convergence de méthodes de directions conjuguées. ESAIM: Mathematical Modelling and Nu-
merical Analysis, 3(16), 35–43. Retrieved from http://www.numdam.org/item/M2AN_1969__3_1_35_0/
Qin, L.-j, Yang, C.-f., & Chen, K. (2013). Quasi-analytic solution of 2-D magnetotelluric fields on an axially anisotropic infinite fault. Geo-
physical Journal International, 192(1), 67–74. https://doi.org/10.1093/gji/ggs018
Rodi, W., & Mackie, R. L. (2001). Nonlinear conjugate gradients algorithm for 2-D magnetotelluric inversion. Geophysics, 66(1), 174–187.
https://doi.org/10.1190/1.1444893
Simpson, F., & Bahr, K. (2005). Practical magnetotellurics. Cambridge University Press.
Siripunvaraporn, W. (2000). Doctoral dissertationAn efficient data-subspace two-dimensional magnetotelluric inversion and its application
to high-resolution profile across the San Andreas Faults at Parkfield, California. Oregon State University.
Siripunvaraporn, W. (2012). Three-dimensional magnetotelluric inversion: An introductory guide for developers and users. Surveys in
Geophysics, 33(1), 5–27. https://doi.org/10.1007/s10712-011-9122-6
Tan, H. D., Yu, Q. F., Booker, J., & Wei, W. B. (2003). Magnetotelluric three-dimensional modeling using the staggered-grid finite difference
method. Chinese Journal of Geophysics, 46, 705–711. https://doi.org/10.1002/cjg2.420
Wannamaker, P. E. (2005). Anisotropy versus heterogeneity in continental solid earth electromagnetic studies: Fundamental response char-
acteristics and implications for physicochemical state. Surveys in Geophysics, 26(6), 733–765. https://doi.org/10.1007/s10712-005-1832-1
Weidelt, P. (1999). 8. 3-D conductivity models: Implications of electrical anisotropy. In Three-dimensional electromagnetics (pp. 119–137).
Society of Exploration Geophysicists. https://doi.org/10.1190/1.9781560802154.ch8
Xiao, T., Huang, X., & Wang, Y. (2019). 3D MT modeling using the T–( method in general anisotropic media. Journal of Applied Geophys-
ics, 160, 171–182. https://doi.org/10.1016/j.jappgeo.2018.11.012
Xiao, T., Liu, Y., Wang, Y., & Fu, L. Y. (2018). Three-dimensional magnetotelluric modeling in anisotropic media using edge-based finite
element method. Journal of Applied Geophysics, 149, 1–9. https://doi.org/10.1016/j.jappgeo.2017.12.009
Ye, G., Unsworth, M., Wei, W., Jin, S., & Liu, Z. (2019). The lithospheric structure of the Solonker suture zone and adjacent areas: Crustal
anisotropy revealed by a high-resolution magnetotelluric study. Journal of Geophysical Research: Solid Earth, 124(2), 1142–1163. https://
doi.org/10.1029/2018JB015719
Yin, C. (2000). Geoelectrical inversion for a one-dimensional anisotropic model and inherent non-uniqueness. Geophysical Journal Inter-
national, 140(1), 11–23. https://doi.org/10.1046/j.1365-246x.2000.00974.x
Yin, Y., Unsworth, M., Liddell, M., Pana, D., & Craven, J. A. (2014). Electrical resistivity structure of the Great Slave Lake shear zone, north-
west Canada: Implications for tectonic history. Geophysical Journal International, 199(1), 178–199. https://doi.org/10.1093/gji/ggu251

References From the Supporting Information

Avdeev, D., & Avdeeva, A. (2009). 3D magnetotelluric inversion using a limited-memory quasi-Newton optimization. Geophysics, 74(3),
F45–F57. https://doi.org/10.1190/1.3114023
Barrett, R., Berry, M., Chan, T. F., Demmel, J., Donato, J., Dongarra, J., et al. (1994). Templates for the solution of linear systems: Building
blocks for iterative methods. Society for Industrial and Applied Mathematics.
Saad, Y. (2003). Iterative methods for sparse linear systems. Society for Industrial and Applied Mathematics.
Smith, J. T. (1996). Conservative modeling of 3-D electromagnetic fields, Part II: Biconjugate gradient solution and an accelerator. Geophys-
ics, 61(5), 1319–1324. https://doi.org/10.1190/1.1444055
van der Vorst, H. A. (1992). Bi-CGSTAB: A fast and smoothly converging variant of Bi-CG for the solution of nonsymmetric linear systems.
SIAM Journal on Scientific and Statistical Computing, 13(2), 631–644. https://doi.org/10.1137/0913035

KONG ET AL. 22 of 22