Towards The Geological Parametrization of Seismic

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

Geophys. J. Int. (2023) 234, 1447–1462 https://doi.org/10.

1093/gji/ggad140
Advance Access publication 2023 March 29
GJI Seismology

Towards the geological parametrization of seismic tomography

Victor C. Tsai , Christian Huber and Colleen A. Dalton


Department of Earth, Environmental and Planetary Sciences, Brown University, Providence, RI 02912, USA. E-mail: victor tsai@brown.edu

Accepted 2023 March 27. Received 2022 August 31; in original form 2023 March 14

SUMMARY
Seismic tomography is a cornerstone of geophysics and has led to a number of important

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


discoveries about the interior of the Earth. However, seismic tomography remains plagued by
the large number of unknown parameters in most tomographic applications. This leads to the
inverse problem being underdetermined and requiring significant non-geologically motivated
smoothing in order to achieve unique answers. Although this solution is acceptable when using
tomography as an explorative tool in discovery mode, it presents a significant problem to use of
tomography in distinguishing between acceptable geological models or in estimating geolog-
ically relevant parameters since typically none of the geological models considered are fit by
the tomographic results, even when uncertainties are accounted for. To address this challenge,
when seismic tomography is to be used for geological model selection or parameter estimation
purposes, we advocate that the tomography can be explicitly parametrized in terms of the
geological models being tested instead of using more mathematically convenient formulations
like voxels, splines or spherical harmonics. Our proposition has a number of technical diffi-
culties associated with it, with some of the most important ones being the move from a linear
to a non-linear inverse problem, the need to choose a geological parametrization that fits each
specific problem and is commensurate with the expected data quality and structure, and the
need to use a supporting framework to identify which model is preferred by the tomographic
data. In this contribution, we introduce geological parametrization of tomography with a few
simple synthetic examples applied to imaging sedimentary basins and subduction zones, and
one real-world example of inferring basin and crustal properties across the continental United
States. We explain the challenges in moving towards more realistic examples, and discuss the
main technical difficulties and how they may be overcome. Although it may take a number
of years for the scientific program suggested here to reach maturity, it is necessary to take
steps in this direction if seismic tomography is to develop from a tool for discovering plausible
structures to one in which distinct scientific inferences can be made regarding the presence or
absence of structures and their physical characteristics.
Key words: Inverse theory; Tomography; Seismic tomography; Theoretical seismology.

waves travelling through the inner core (Cao et al. 2005; Shearer
1 I N T RO D U C T I O N
et al. 2011).
Seismic tomography, the imaging of 3-D Earth structure with seis- A primary reason that there remain controversies about the inter-
mic waves, has expanded enormously since the early studies of pretation of seismic imaging results is that the images are inevitably
Aki et al. (1977) and Dziewonski et al. (1977). It has resulted in smoothed, typically as a consequence of inadequate data coverage,
countless discoveries including deep subducting slabs (van der Hilst and thus require interpretation as to whether the images approxi-
et al. 1997), lower mantle large low-shear-wave velocity provinces mate the anticipated geological structures or not. Given this issue,
(Su et al. 1994) and numerous oil-producing reservoirs (Karcher a number of studies have attempted to address the interpretation
1987). Seismic tomography has also resulted in more controver- problem in a variety of ways. A number of purely mathematical ap-
sial findings, such as the debated seismic manifestation of mantle proaches to improving interpretability despite the need to regularize
plumes (Montelli et al. 2004, 2006; van der Hilst & de Hoop 2005, inversions have been proposed, including techniques based on sin-
2006; Hwang et al. 2011), the debated interpretation of near-trench gular value decomposition (Jackson 1972), well-posed stochastic
seismic anisotropy as trench parallel mantle flow (Russo & Silver extensions (Jordan & Franklin 1971), incorporating a priori infor-
1994; Song & Kawakatsu 2012) and the debated discovery of shear mation (Jackson 1979; Jordan & Sverdrup 1981), total variation


C The Author(s) 2023. Published by Oxford University Press on behalf of The Royal Astronomical Society. This is an Open Access

article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which
permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.
1447
1448 V.C. Tsai et al.

regularization (Rudin et al. 1992), subspace inversion (Bodin & philosophy implicitly favoured in this decision is one of maximiz-
Sambridge 2009; Fang et al. 2019), sparsity regularization (Candes ing the opportunity for discovery. Specifically, this parametrization
& Wakin 2008; Candes et al. 2008) and homogenization in upscaling choice maximizes the range of specific structures that could result
(Backus 1962; Capdeville et al. 2010). In contrast, other approaches from inversion and minimizes the potential bias from any a pri-
have proposed to address the interpretation problem by posing the ori expectations for what structures may exist. When knowledge is
problem with different constraints. For example, Nataf & Ricard poor or potentially unreliable, as it is when one is attempting to im-
(1996) performed geodynamic simulations to predict seismic data age a structure for the first time, this type of general mathematical
and compare with tomographic results; Zelt & Smith (1992) and parametrization is an excellent choice and has a solid epistemolog-
Zelt (1999) posed minimum parameter inversions with as simple of ical justification (Curd & Cover 1998).
a geometry as possible; Magistrale et al. (2000) created a rule-based However, this maximal degree of freedom philosophy is not with-
reference model for basins in southern California based on fitting out its issues and limitations. Perhaps the two most significant issues
tomographic and geotechnical data; Sun & Li (2015) used clustering are that (1) a unique solution often does not exist for the imaging
to improve the geological interpretability of gravity and crosswell problem with many degrees of freedom and (2) the solutions can
seismic data; Khan et al. (2008), Giraud et al. (2017), Koelemeijer be unphysical or unable to distinguish between physical alterna-
et al. (2018) and Astic & Oldenburg (2019) used a Bayesian frame- tives that are of interest, thereby making it very difficult to infer
work to estimate geodynamic and petrophysical parameters from relevant geological or geodynamic parameters. The first problem,

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


seismic, gravity, magnetic and electromagnetic data, respectively; which is also known as the underdetermined problem, is due to the
many studies in the exploration seismology context like Dadash- limitations of the imaging data (e.g. finite number and uneven distri-
pour et al. (2008), Dessa et al. (2009) and Hubbard et al. (2014) bution of events and stations in the seismological imaging context)
used a combination of migrated seismic reflection and refraction and is often dealt with using regularization or smoothing techniques
data and well logs to subsequently infer structural geology models, (Aster et al. 2013), which reduces the effective number of model
including inferring reservoir parameters; Arnold & Curtis (2018) parameters to be solved for by imposing smoothness or continuity
used interrogation theory to pose specific geological questions us- constraints on the model. This approach is successful in reducing
ing seismic data; a number of groundwater studies such as Linde the underdeterminism that underlies the imaging problem, but often
et al. (2015) and Laloy et al. (2018) imposed statistical constraints causes the solution to be unphysical or ambiguous with regards to
to improve the realism of inverted hydrological networks; and Muir plausible alternatives (second problem). Many physical structures
& Tsai (2020a) introduced a geometric level-set parametrization for are not expected to be smooth or otherwise satisfy the continuity
tomography that more naturally allows for sharp geological bound- properties that are imposed in the regularization process. More-
aries to be imaged. Although all of these studies make progress over, even when the number of grid cells is large and the amount
towards better interpretation, none of them solve the underlying is- of smoothing modest, there are often artefacts of the mathematical
sue that the information provided by seismic tomographic imaging parametrization in the final image that do not satisfy certain physi-
is not the information that is usually of most interest from the point cal principles that may be expected (e.g. Lin et al. 2014; Hosseini
of view of the geological processes involved. et al. 2020).
To more directly take steps in the direction of better interpretabil- To be concrete about these two limitations to standard tomo-
ity of seismic tomography, here we introduce the concept of geo- graphic imaging, we provide a simple geologically motivated ex-
logical parametrization of tomography, a natural extension of the ample in Fig. 1. The example is specific to imaging a sedimentary
geometric parametrization of Muir & Tsai (2020a) and to some basin with traveltime tomography but is archetypical of a broader
extent a generalization of the Zelt & Smith (1992) philosophy. In class of problems. In this example, a sedimentary basin is assumed
Section 2, we describe some of the main issues with standard to- to form due to a series of horst and graben faulting events and de-
mography parametrizations. In Section 3, we introduce the central position events. The velocities within each layer are assumed to be
idea of geological parametrization of tomography. We then provide constant, and only first-arrival P-wave traveltimes are assumed to be
2 simple synthetic examples of the approach applied to sedimen- measured. Fig. 1(a) depicts the series of four geological events that
tary basins and subduction zones and one real-world example of the occurred to form the eventual structure to be imaged using con-
approach applied to estimating basin and crustal properties across ventional transmission straight-ray traveltime tomography. To be
the continental United States. In Section 4, we delve into some specific about the imaging capabilities, in this example we assume
of the main differences between geological tomography and more there are 50 stations placed uniformly along the surface and as-
standard approaches and explain some of the main challenges of sume there are 39 sources evenly spaced along the other boundaries
geological tomography that must be dealt with that are unique from (Fig. 1b). Although this setup is idealized, it has the features needed
the challenges of other approaches. In Section 5, we summarize the to demonstrate some of the limitations of standard parametriza-
findings. tions. In this example, there are 1950 (39 × 50) ray paths and also
1950 pieces of traveltime data, and the ray path density is shown in
Fig. 1(b) for a model that is pixelated with a grid that is 200 by 100
pixels. Although the ray path density is relatively high in the upper
portion of the model, with an average of more than 20 crossing
2 T H E I S S U E W I T H S TA N D A R D
rays crossing each pixel, performing an inversion with the 20000
T O M O G R A P H Y PA R A M E T R I Z AT I O N S
pixels is clearly underdetermined since 20 000 is far in excess of the
The majority of tomographic inversions parametrize the region of 1950 pieces of data, a result that is well understood from an inverse
interest in terms of mathematically convenient structures such as theory standpoint (e.g. Aster et al. 2013). Even in the upper portion
pixels, voxels, grid cells, Voronoi cells, splines or spherical har- of the model where the ray path density is high, there is significant
monics (Lay & Wallace 1995; Liu & Gu 2012; Fang et al. 2019). trade-off between model parameters due to equal sensitivity along
Typically, a very large number of cells or degrees of freedom are each ray path. As a result, the unsmoothed tomographic problem is a
allowed in order to capture the most general possible structure. The linear inverse problem but with a nearly singular matrix inverse that
Geological tomography 1449

(a) x1 x2 (b) Raypath Density


s1 20
15
Graben 1 v0 10
5
0

h1
Deposition 1 (c) Smoothed Inversion

x3 x4

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


s2

Graben 2

(d) More Discretized Inversion


v2
v1
Deposition 2

Figure 1. Schematic for sedimentary basin process and standard inversion. (a) Schematic shows a series of 4 geological events that form a sedimentary basin.
(b) Ray path density (colours) for the 50 stations (red circles) and 39 sources (cyan squares) along the boundary. (c) Schematic result of a smoothed traditional
tomographic inversion with 200 grid cells. (d) Schematic result of a more coarsely discretized traditional tomographic inversion with 50 grid cells.

leads to a lack of a practical solution. There are two potential stan- framework for imaging based on geological parametrizations. This
dard approaches. One can either discretize the model less finely or framework builds naturally on a previous ‘geometric tomography’
apply smoothing, both of which reduce the effective number of pa- parametrization introduced by Muir & Tsai (2020a) in which the to-
rameters and thus reduce the underdeterminism of the inversion. As mographic model is parametrized in terms of geometric boundaries.
shown schematically in Figs 1(c) and (d), though, neither solution is Comparisons between all three versions for parametrizing tomog-
able to capture key features of the original geological model. Most raphy (i.e. standard, geometric and geological) are shown later in
problematic is that the sharp geological boundaries between faulted this section.
blocks are no longer apparent, either having been smoothed away or The central requirement of the geological parametrization frame-
displaying primarily the pixilation of the less discretized grid cell work is for the parameters of the tomographic problem to be directly
parametrization. Importantly from an inference and epistemolog- related to the geological processes responsible for the structure be-
ical perspective, the fault locations, fault offsets and depositional ing imaged. As an example, for the sedimentary basin example
amounts are not determined nor easily estimated with the resulting shown in Fig. 1, the most natural parameters for the geological
images. Even seismic observations with better sensitivity to strong parametrization would be the locations, orientations and offsets on
velocity gradients than the path-integrated traveltimes used in this the faults, the amount of sediment deposited, and the seismic ve-
example, such as scattered and reflected waves, suffer from finite locities of each of the geological layers. As we will discuss later
resolution, which would introduce error into estimates of geologi- (Section 4.1), such a parametrization is inevitably a simplification
cal parameters. Thus, potentially some of the information that is of of reality and also makes certain assumptions (such as constant
most interest and why the imaging might have been performed in layer velocities and how they relate to the depositional events) that
the first place are not straightforwardly estimated. may in some cases be poor. Thus, an important challenge in us-
ing such a parametrization is ensuring that the assumptions are
commensurate with the a priori knowledge available (Section 4.1).
Explicitly choosing a parametrization that is geologically informed
3 G E O L O G I C A L PA R A M E T R I Z AT I O N
distinguishes the procedure from the interrogation theory approach
OF TOMOGRAPHY
(Arnold & Curtis 2018; Zhao et al. 2022) in which tomographic
gridding is still used but specific geological questions are posed.
3.1 The central idea and requirements
This geological approach is also distinguished from purely geomet-
Given the desire to be able to more directly infer geologically and ric approaches (Zelt & Smith 1992; Muir & Tsai 2020b) by hav-
geodynamically interesting parameters and the deficiencies of stan- ing parameters that are not solely geometric, like fault slip (which
dard mathematically gridded tomography, here we propose a new can affect multiple aspects of the geometry simultaneously or other
1450 V.C. Tsai et al.

physical variables that affect the velocities in a non-geometric way). between the fully unconstrained approach of traditional tomography
We discuss the relation with other approaches further in Section and the fully geodynamic approach proposed by others (e.g. Nataf
4.6. We emphasize that in using the geological parametrization ad- & Ricard 1996; Khan et al. 2008) when the framework is to be used
vocated here, the nature of the specific geological parametrization to robustly and quantitatively estimate geological information.
(i.e. what the unknown parameters are) will be strongly dependent
on the geological setting as well as the a priori information known
about the situation (see Section 4.1 for further details).
A second requirement that we impose with this imaging frame-
3.2 Example sedimentary basin inversion with geological
work is that the geological parameters be distinguishable by the
tomography
available data and imaging method, at least in theory, as could
be explored for example using synthetic data. As a simple exam- Before discussing the various challenges inherent in proceeding
ple, if synthetic tests or physical arguments demonstrate that the with the geological tomography framework as proposed, we first
presence or absence of a fault in a particular location cannot be present a proof-of-principle demonstration of its implementation as
distinguished from the data (e.g. due to insufficient ray or station applied to the simple sedimentary basin example from Fig. 1 and
coverage), then such a fault should not be included as a possibility compare the results with those of traditional tomography and ge-
within this framework and its parameters should not be inverted for. ometric tomography. Although this example is idealized in many

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


The ability to distinguish such a fault feature may depend on other ways, including the 2-D nature of the problem, the dense spac-
choices made about the parametrization. This requirement ensures ing of stations and sources and the assumed uniform velocities of
that the tomographic inversion is well posed, without a significant the different layers, it nonetheless nicely highlights the differences
null space. This condition further limits the types of models that between the different approaches, including both the expected ad-
can be expressed within this framework, and the limitations may vantages as well as some of the challenges of geological tomography
be more or less severe depending on the availability of a priori that are distinct from using more traditional approaches. For the ex-
information. We also return to this point in Section 4.1. This second ample shown in Fig. 2, we further assume that the graben forming
requirement distinguishes the suggested approach from the more faults have a set angle from the horizontal that is known a priori (for
general related idea of using forward modelling, for example us- example, due to assumed Andersonian faulting (Anderson 1905) or
ing geodynamic numerical software, to predict seismic velocities other observational constraints independent from the seismic data
and subsequently compare them with seismic data or tomographic to be used for imaging) and that these faults continue with constant
models (e.g. Nataf & Ricard 1996; Khan et al. 2008; Koelemeijer slip to infinite depth. With these additional assumptions, there are
et al. 2018). In those works, the underlying model parameters are 10 parameters that uniquely determine the ‘geological’ model: the
not expected to be uniquely determined, with significant trade-offs four horizontal surface locations of the two pairs of faults (x1 , x2 ,
between them that may be explored in a Bayesian sense or explored x3 , x4 ), the 2 vertical offsets for the 2 graben forming events (s1 , s2 ),
only to a very limited extent. In such approaches, a main question the 1 thickness of the first deposition event (h1 ) and the 3 seismic
is in understanding how the parameters of interest may or may not velocities of the background block (v0 ), the first sediment layer (v1 )
be sensitive to the (seismic) data being analysed. In contrast, in our and the second sediment layer (v2 ) (see Fig. 1a). In this example, we
suggested methodology we advocate for the parametrization to be assume the second deposition event to fill to form a flat horizontal
chosen such that uncertainties and trade-offs are minimal and arise free surface so that the thickness of the second deposition event is
primarily from the implementation of solving the inverse problem not a free parameter to be solved for. In Fig. 2(a), we show one pos-
with noisy or imperfect data. The difference may be subtle, but sible specific realization which we then calculate synthetic P-wave
we believe that making difficult choices in the simplification of the traveltime data for (Fig. 3a).
geological model to be determined is essential to this type of ge- Inversions of this noise-free traveltime data set using the three
ological tomography being useful (and serves a different purpose different tomographic parametrizations are shown in Figs 2(b)–(d),
than the geodynamic comparisons that have been done previously). with Fig. 2(b) showing the traditional traveltime tomography result
Not satisfying this second requirement would necessitate mathe- (parameters are 200 unknown gridded velocities), Fig. 2(c) showing
matical regularization which would be counter to the central goal an inversion with the depths of the 2 geometric boundaries being
of the geological parametrization approach. the unknowns (parameters are the 18 unknown depths plus 3 un-
As a result of the second requirement, rather than solve for first- known velocities), and Fig. 2(d) showing an inversion with the same
principles parameters such as chemical composition or thermal 10 parameters that are used to generate the geological model. For
characteristics, we advocate for solving for parameters that directly the traditional tomography result, smoothing is required to obtain
influence the imaging results such as seismic velocities (rather than a unique solution, and standard Tikhonov regularization is utilized
composition) and fault offsets (rather than driving forces) except with both zeroth and first-order regularization terms added (see
when robust theory can uniquely translate more fundamental vari- Supporting Information). For the geometric tomography result, an
ables into directly observable ones (see Section 3.3 for an example). iterative linearized inversion is performed, with updates alternating
As a further example of the advocated philosophy, in the sedimen- between the boundary depths and internal velocities. The geometric
tary basin example, total fault offsets and total depositional amounts inversion is highly non-linear, and the final solution obtained is sen-
are the parameters to invert for, rather than faulting rates and dura- sitive to the initial guess, due to many local minima (see Supporting
tions and depositional rates and durations since, within the context Information).
of the provided information and assumed model, there is com- Finally, for the geological parametrization, the inversion is per-
plete trade-off between rates and durations, with only the product formed stepwise, with a slightly different set of 12 ‘stencil’ pa-
constrained. Again, there is a subtle distinction, and which param- rameters (see Fig. 3b) that is then mapped onto the original 10
eters are directly distinguishable depends on the degree of a priori parameters (see Fig. 3c). This choice is made because the 12 pa-
knowledge, the physics considered and the quality and quantity rameters with a fixed stencil are more directly relatable to the data
of the data. However, we believe this choice is a good compromise and there is a convenient one-to-one mapping between those 12
Geological tomography 1451

Slowness
(1/v)
Parameters
10 input geologic parameters:
(a)
v0=3 km/s, v1=1 km/s
'True' Model v2=0.5 km/s, h1=4 km
Realization s1=6 km, s2=1.5 km
x1=1 km, x2=17 km,
x3=4 km, x4=19 km

(b)
Traditional 200 gridded velocities,
1 for each 1x1 km pixel
Gridded
Inversion

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


18 depths + 3 velocities:
(c)
9 depths for each of 2 boundaries
Geometric v0=3.28±0.05 km/s
Inversion v1=0.86±0.01 km/s
(e.g. Muir & Tsai 2020)
v2=0.50±0.01 km/s

10 recovered geologic parameters:


(d)
v0=3.000±0.000 km/s, v1=1.01±0.02 km/s
Geologic v2=0.498±0.003 km/s, h1=4.00±0.01 km
Inversion s1=6.00±0.01 km, s2=1.50±0.01 km
x1=1.1±0.2 km, x2=17.2±0.6 km,
x3=3.8±0.6 km, x4=18.9±0.2 km

Figure 2. Comparison between three different inversion parametrizations for a sedimentary basin example. (a) Example of a specific realization of geological
events to form a sedimentary basin to be imaged (input model). (b) Inversion with traditional gridded parametrization. (c) Inversion with geometric parametriza-
tion (e.g. Muir & Tsai 2020a). (d) Inversion with geological parametrization. Estimated parameters are listed to the right, with (a) listing the parameters of the
input model. Uncertainties are estimated as described in the Supporting Information.

parameters and the 10 original parameters (note that 4 of the pa- uniquely determined (see Fig. 3c). This method of estimating pa-
rameters are identical, 2 new binary parameters are introduced to rameters also allows for estimation of each parameter’s errors (see
map the position of the faults onto the real axis and the other pa- Section 4.3 for more details) and it reveals, in this case, that there
rameters have clear relationships). There are multiple ways in which is a very different magnitude of error on the different parameters
inversion for the geological model could have been performed. One (see uncertainties listed in Fig. 2(d)). For example, with the method
could have performed a full Bayesian MCMC inversion or a gra- described above, the fault horizontal location errors for x2 and x3 are
dient based iterative solver (see Section 4 for further discussion), close to 1 km, related to the spacing of the sources, whereas the fault
but for this example we choose to show a step-wise best-fitting horizontal location errors for x1 and x4 are an order of magnitude
inversion in which parameters are solved for in a specific order to smaller, and related more to the spacing of the stations. The fact
minimize the amount of trade-off between other parameters. For ex- that some geological parameters (e.g. locations x1 and x4 ) are very
ample, given the source–receiver array geometry, the background accurately determined, even though other parts of the inversion are
velocity v0 is easily determined with low error and essentially no less effective, is a general characteristic of this type of inversion
trade-off with other parameters and is therefore the first parameter that may be desirable. We note that the inversion routine could be
solved for (see Fig. 3d). Two of the horizontal locations of the faults added to and the error on some of the parameters further reduced,
(x1 ’ and x4 ’) and the shallowest layer velocity v2 are the next most for example by adding a perturbative routine after the routine as
easily determined parameters, after which x2 ’ and x3 ’ are deter- described. However, given our focus on the new parametrization
mined. The inversion routine becomes more complex at this stage rather than the specific inversion procedure, we do not perform
but the thicknesses h1 ’, h2 ’, h3 ’ and stencil flags b1 and b2 are deter- this extra step in the example. Regardless of the specific imple-
minable along with v1 (see Supporting Information), finishing the mentation, we expect the recovery to typically be imperfect and
inversion for the 12 stencil parameters. Once these 12 stencil pa- the data not perfectly fit primarily due to limitations in the model
rameters are determined, the 10 original geological parameters are discretization and array and source geometry, even when the data
1452 V.C. Tsai et al.

(a) Observed Travel Time Record Section is noiseless. Since the same physics is assumed that generates
the structure, in some sense the geological inversion performance
35 shown here is a best-case scenario; nonetheless, the idealized ex-
30 ample is useful to compare the strengths and weaknesses of the
Travel time (s) 25 proposed approach. If data uncertainties were too large, the geo-
logical inversion would not be guaranteed to produce reasonable
20
results, but with small enough data uncertainties, the results would
15
be similar. A full analysis of how the results depend on data uncer-
10 tainties would be appropriate for future work.
5 There are a few key differences between the 3 inversion results
0 that we would like to highlight. First and most importantly, by con-
10 20 30 40 50
Station # struction, the geological inversion (Fig. 2d) has exactly the same
parametrization as the geological events assumed to produce the
structure and hence captures exactly the elements of the geological
processes that were assumed. Furthermore, parameters of geological
(b) Stencil for Geologic Inversion
interest, such as fault slip and sediment depositional amount, have

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


x1' x2' x3' x4'
h'3 v2 v0 or v2
been directly estimated. Notably, the geological inversion is able to
v0 or v2
h2' accurately determine all 10 geological parameters, up to a certain
v0 or v1 v0 or v1 error. In contrast, it is unclear where the geological unit bound-
h'1 v1
aries should be placed in the standard inversion result (Fig. 2b),
v0
and although the geometric inversion (Fig. 2c) captures the distinct
geological boundaries, the parametrization is not able to capture
the precise shapes or provide useful constraints on fault slip mag-
b1 binary variable: b2 binary variable:
nitudes. Secondly, the data misfits achieved are similar for the ge-
True if v1 and v2 fill left section True if v1 and v2 fill right section
ological inversion [root mean square error (RMSE) of 0.24 s] and
False if v0 fills left section False if v0 fills right section
the traditional gridded inversion (RMSE of 0.26 s), with the geo-
12 stencil variables: v0, v1, v2, x1' , x'2, x3' , x4' , h1' , h2' , h'3, b1, b2 metric inversion misfit being somewhat higher but still relatively
low (RMSE of 0.58 s). These misfit comparisons are not entirely
meaningful since all three inversions could achieve lower misfits
(c) Mapping from Stencil to Geologic Variables in a variety of ways. For example, lower misfit could be achieved
in the gridded inversion by reducing the strength of regularization,
v0, v1, v2 v0, v1, v2
in the geometric inversion by increasing the number of parameters,
h'
1
h1 and in the geological inversion by using a finer grid search on cer-
h1' +h3' s1 tain parameters. In all cases, there are significant trade-offs in both
computational time and robustness of the subsequent result, and
h2' s2
it is unclear that solutions with lower misfits would be preferable.
If b1=0, x2'
x1 Thirdly, although the parametrizations chosen are independent of
If b1=1, x1' the data type used, we observe that the three models obtained by
If b2=0, x3' fitting traveltime data as in this example will have varying successes
x2
If b2=1, x4' in predicting other data like full-waveform seismic data. Although
If b1=0, x1'
we do not perform full-waveform synthetics here, it is clear that the
x3 geologically inverted model would fit these alternative data better
If b1=1, x'2
since it is a better representation of the assumed reality. In contrast,
If b2=0, x4'
x4 the gridded model’s large-scale smoothness blurs certain sharp ar-
If b2=1, x3' rivals and the small-scale blockiness adds sharp arrivals where they
should not exist. The geometric model is similar to the geological
(d) Inversion Order model, but is not nearly as accurate since the allowed geometry
4
Left block: 6
was not flexible enough to capture the assumed reality. This may
1 2 3
h'2 or h1' , h3' , v1 v be the most important conclusion since it suggests that the more
⇒ 5 ⇒ 1
v0 ⇒ x1' , x4' , v2 ⇒ x2' , x3' ⇒⇒ b1, b2 ⇒ h2' , v1 physically realistic model parametrization is more robust to mak-
⇒ ⇒
Right block: h1' , h3' , v1 ing predictions of data types that were not observed. This trait is a
hallmark of successful scientific models and to some extent is the
h2' or h1' , h'3, v1
entire goal of the scientific method (Curd & Cover 1998).
Figure 3. Steps for geological inversion of sedimentary basin example.
We close this subsection by commenting on how the geological
(a) Traveltime record section. Each line represents one of the 39 sources, model could be made more complex. Since the example shown was
with the observed traveltimes at the 50 stations tabulated across the x-axis. an idealization from even a simple real sedimentary basin, one may
Synthetics are using a straight-ray assumption. (b) Stencil for the geological wonder whether this limits the geological tomography approach to
inversion with 12 variables. ‘Left’ section refers to the block between x1 ’ only the simplest of structures. Although adding model complex-
and x2 ’, and ‘right’ section refers to the block between x3 ’ and x4 ’. (c) ity creates certain challenges in how to perform the inversion (see
Mapping from stencil to geological variables. (d) Schematic of the order in Section 4.1), there is no inherent issue in adding justified model
which inversion occurs. Numbers are referred to in the text and details are complexity. For example, one could straightforwardly add vertical
provided in the Supporting Information. gradients in velocity to all of the layers so that there were three
Geological tomography 1453

additional parameters to the example of this section. If the assumed complicated than the example shown. In principle, all of the
gradients are sufficiently weak, these three additional parameters additional complexity could be incorporated into a geological
can simply be included and determined at the end of the inversion parametrization and we suggest that future work should focus on
(solved for last), with no further changes to the geological inver- how to best do this. The purpose of the present example is to
sion routine. When gradients are allowed to be large and non-linear, demonstrate proof of principle and highlight some of the key dif-
the inversion routine would need to be changed, but there is no ferences in a scenario where a more complex physical model (i.e.
inherent problem in adding such gradients. On the other extreme a scenario where smooth seismic velocity gradients arise from the
end of the complexity spectrum, where geological information is thermal boundary layer adjacent to the subducted plate) is used
abundant, one might be interested in allowing for other geolog- within the geological parametrization (compared to the relatively
ical processes including 3-D ductile deformation, listric faulting, simple physical model assumed in the sedimentary basin example),
rollover anticlines, fault-bend folds, fault-propagation folds or half and for which mathematical approaches like sparsity regularization
grabens (e.g. Waldron & Snyder 2020). Again, there is no inherent (Candes et al. 2008) would not be effective. As with the previous
limitation in including such processes when warranted. In princi- example, the physics assumed in the geological inversion is ex-
ple, all of the processes included in structural geology modelling actly the physics used to create the example, making the results a
software such as GemPy (de la Varga et al. 2019) could be in- best-case scenario for the performance of the geological inversion.
corporated. However, only if the data is sufficiently good or the a Inversions of the traveltime data set from Fig. 4(b) using the

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


priori information sufficiently strong would including this range traditional gridded inversion is shown in Fig. 4(c) and using the
of processes lead to a unique solution (see requirement 2 from geological parametrization is shown in Fig. 4(d). The geological
Section 3.1), so we would argue that such complex possibilities inversion is also performed stepwise, beginning with estimating
should be limited only to those situations, and not used when the x1 (step 1), then solving for vL and vC (step 2), re-estimating x1
information is more limited. For example, well studied sedimentary (step 3), solving for vM and h1 (step 4), solving for h2 (step 5),
basins like the Los Angeles Basin (Wright 1991) or Seattle Basin estimating θ and iteratively solving for x1 and θ (step 6) and then
(Troost & Booth 2008) may benefit from such detailed parametriza- solving for R (step 7) (see Supporting Information for details). As
tion, whereas simpler geological parametrizations may be more ap- can be observed in Fig. 4, despite having an order of magnitude
propriate for other basins like the Jakarta Basin (Cipta et al. 2018) more parameters (200 instead of 8), the traditional gridded inver-
that have less information available. sion approach cannot resolve the subtle velocity gradient caused
by the mantle wedge cooling whereas the geological parametriza-
tion can resolve it and use it to estimate the subduction rate R (see
Fig. 4d). In this example, there is a 26 per cent error on the estimate
3.3. Example subduction zone inversion with geological
of this rate, due to the subtlety of the velocity signal, but the physics
tomography
can be determined (one can infer the shape of the isotherms and
The use of geological parametrizations is not limited to sedimentary constrain the velocity of the downgoing plate) and the error can
basins or constant velocity regions, so here we also demonstrate an be estimated. We reiterate that there are multiple ways to imple-
application to subduction zone imaging in which part of the phys- ment the geological type of inversion, and the method described
ical model is more complex. In this example (Fig. 4a), a simple uses a practical compromise between computational cost and accu-
subduction zone geometry is assumed with trench location (x1 ), a racy. We note that the error could be lessened if further iteration
constant thickness oceanic lithosphere (h1 ) of velocity vL , a con- on other parameters were performed and if the data quality were
stant thickness continental crust (h2 ) of velocity vC , and a constant high enough. On the other hand, if the data quality were worse or
dip angle (θ ). For example, this could approximate an old and cold if the parametrization were too crude a representation of the reality
oceanic plate (hence approximately constant thickness) subducting of the subduction zone, then the estimated error would increase and
at near constant rate under a uniform continental crust. The only added iteration may not be warranted. Finally, we comment that
additional complexity we allow in this example is that the mantle again the example provided is a simplistic one that could be made
wedge is cooled by contact with the cold slab. We assume a diffu- more realistic in a number of ways. For example, one could add a
sive cooling model of the form T = TM + (T0 − TM ) · erfc( 2√yκt ) half-space cooling model for both the lithosphere and crustal parts
where TM is the mantle temperature (assumed 1300 ◦ C), T0 is the of the model (rather than being constant velocity), and one could
surface temperature (assumed 0 ◦ C), y is the distance perpendicular have ages of these blocks be the parameters of the model rather
to the subducting plate surface, t is the time since it subducted, and κ than the constant thicknesses assumed here. One could also im-
= 10−6 m2 s–1 is a known value for the thermal diffusivity. In order prove the physical realism of the subduction model by allowing for
to translate the model into a seismic velocity prediction, we assume warming of the subducted plate and related velocity changes, or by
a simple linear scaling of −0.75 per cent per 100 ◦ C (Cammarano using a more accurate temperature-to-seismic velocity relationship
et al. 2003) and that t = z/R, where z is the subducted distance or accounting for expected phase changes. We do not implement
and R is the subduction rate. The 8 model parameters are therefore such added complexities partly because there would be additional
h1 , h2 , x1 , θ , R, vL , vC and vM . A schematic is shown in Fig. 4(a), and choices needed, but there are no fundamental limitations that would
a specific realization is shown in Fig. 4(b). As with the sedimentary prevent such modifications of the geological parametrization once
basin example, this subduction zone example is very idealized from the physics to be considered is agreed upon.
the expected reality. For example, we assume that the plate velocity
remains constant over time, we ignore the effects of partial melting,
dehydration reactions, and other phase transitions on seismic veloc-
3.4. Continental United States basin and crustal inversion
ity in the crust and mantle, and finally we do not consider the effect
example with geological tomography
of temperature on seismic velocity outside the thermal boundary
layer between the subducted plate and the mantle wedge. Includ- We conclude this section with one non-synthetic example of de-
ing these effects will make the true situation significantly more termining basin and crustal structure across the continental United
1454 V.C. Tsai et al.

(a)

(b)

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


(c)

(d)

Figure 4. Comparison of inversions for a subduction zone example. (a) Schematic of a subduction zone with the geological processes and their associated
parameters described. (b) Example of a specific realization of geological events to form a subduction zone to be imaged (input model). (c) Inversion with
traditional gridded parametrization. (d) Inversion with geological parametrization. Estimated parameters are listed to the right, with (a) listing the parameters
of the input model. Uncertainties are estimated as described in the Supporting Information.

vS(z) tary basins are known to have seismic shear velocities that increase
with depth approximately as a power-law function of depth, vs (z)
vb(z)=v0(z/z0)a = v0 (z/z0 )a (e.g. Boore & Joyner 1997; Boore et al. 2011; Tsai
hb & Atiganyanun 2014), with the reference velocity v0 (at depth z0 )
Depth (z)

and power-law exponent a being related to the geological history


vc of the basin including the deposition rate (e.g. Audet & Fowler
hc 1992; Schneider et al. 1996). We therefore choose the three basin
parameters to be v0 , a and basin depth, hb . The crustal parameters
vm are average crustal shear-wave velocity, vc , crustal thickness, hc and
average upper mantle shear wave velocity, vm , for a total of 6 param-
eters at every location for which the 22 (11 periods for Love wave
Figure 5. Schematic 1-D velocity structure for a simple geological model and 11 periods for Rayleigh waves) phase velocity measurements
describing a basin with a power-law velocity structure, a constant velocity are available. Although the 6-parameter basin and crust package is
crust and a constant velocity mantle. yet again a significant idealization of the real crust, we find we are
able to determine key geological information such as basin depth,
States that demonstrates the feasibility of applying geological to- basin power-law exponent and crustal thickness that are useful phys-
mography to real data. In this example, we utilize surface-wave ical constraints on the geological evolution of the North American
dispersion maps across the continental United States from Ekstrom continent (see below). We note that surface-wave dispersion data are
(2017) to invert for three sedimentary basin parameters including rarely used alone to determine layer boundaries such as basin depth
basin depth and three crustal-scale parameters including crustal and crustal thickness due to their known smearing of sensitivity at a
thickness as shown schematically in Fig. 5. Specifically, sedimen- wide range of depths, but the geological parametrization allows for
Geological tomography 1455

these features to be successfully determined (see below). We also upper mantle velocities (Fig. 6e) in the tectonically active and highly
note that our parametrization shares some similar features to previ- deformed western United States and faster average crustal veloci-
ous basin studies like Magistrale et al. (2000) who define the basin ties, thicker crust and slower upper mantle velocities in the stable
depth and velocity increase as parameters; these previous studies eastern United States (e.g. Shen & Ritzwoller 2016). The Yellow-
do not perform a tomographic inversion as done here and instead stone hotspot track is clearly identified as the slowest upper mantle
use tomographic results as inputs to their model construction. velocities in Fig. 6(e), consistent with other studies (e.g. Schmandt
Results from the inversion of the dispersion data are shown in et al. 2012). The Rio Grande Rift is also evident in the slow upper
Fig. 6. This geological inversion was also performed stepwise, first mantle velocities (e.g. Lin et al. 2014). Our crustal thickness results
determining an average crustal model (vc , hc and vm ) using a linear depart from those of other studies in a few interesting ways. We
Dix inversion (Haney & Tsai 2015; Haney & Tsai 2017; Haney observe thickened crust underlying the Cascades Range, unlike in
et al. 2020) with the Rayleigh-wave data from 10 to 40 s period, and most studies (e.g. Shen & Ritzwoller 2016), but consistent with the
separately estimating a by fitting the Love-wave data from 5 to 20 s results of Haney et al. (2020), likely because our method images the
period to a power-law (Tsai & Atiganyanun 2014) and estimating slab interface as the crust–mantle boundary and treats the subducted
hb through inversion of the same Love-wave data with the linear Juan de Fuca plate and weak continental Moho as part of the crust
Dix inversion (Haney & Tsai 2020). The best-fitting v0 and updated (Haney et al. 2020). We also image a thickened Sierra Nevada that
crustal velocity vc were then determined using all of the Rayleigh- is not apparent in Shen & Ritzwoller (2016) but which would be

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


wave data. See Supporting Information for details. consistent with Airy isostasy. Perhaps most interestingly, we image
We highlight a few interesting basin features of Figs 6(a)–(c). a shallower crustal thickness across eastern Montana and western
First, the location and existence of all major, deep sedimentary North Dakota (∼40 km compared with ∼50 km from Shen & Ritz-
basins are accurately determined (Frezon et al. 1983; Mooney & woller (2016)) which we suggest may be due to the fact that receiver
Kaban 2010), including the very thick Mississippi Embayment and functions from this region observe two strong discontinuities (e.g.
Gulf of Mexico sedimentary sequence (Stetson 1953; Davis 2017), Song et al. 2022) and may be caused by a thick package of rocks
the Williston Basin (Carlson & Anderson 1965), the Denver Basin that have properties intermediate to purely crustal or purely mantle
(Nelson & Santus 2011), the Central Valley of California (Faunt rock types.
et al. 2010) and even the relatively shallow Chesapeake Bay sedi- The features imaged here should be treated as preliminary find-
ments (Newell et al. 2004). Shallower basins are not well imaged ings to be verified with other data sets with better sensitivities to the
due to limitations of the data (with 5 s being the shortest periods various features described. However, we believe this preliminary ap-
available, and thus being insensitive to the shallowest ∼1–2 km), plication of geological tomography to the continental United States
though some of these like the Michigan Basin and the Rome Trough helps demonstrate the feasibility and usefulness of the framework,
(West Virginia to Pennsylvania) are mapped into the average crustal despite the numerous simplifications of reality that must be assumed
velocity and are clearly observed in Fig. 6(d). (Higher frequency to apply the framework. Unlike in the previous 2 synthetic exam-
data would be required to properly invert for these basins.) The ples, we reiterate that this example was not based on synthetic data
Williston Basin further stands out as a relatively shallow (2–4 km) produced with the same parametrization used for the inversion so
basin with the strongest velocity gradient (a = 0.4–0.6) (Fig. 6b). that the data is not perfectly compatible with the assumed model;
Although the Mississippi Embayment and Gulf of Mexico has likely nonetheless, the geological inversion provides a sensible and useful
unrealistically higher average basin velocity than most other basins result. The geological model assumed was also very simplified, but
in our results, the staggeringly large depth extent (>10 km) in the future work could attempt a similar but more realistic and complex
deepest part of the on-land portion is dramatically highlighted; the geological parametrization.
shallow-most low velocity Gulf sediments are not well represented
in this basin average and would require assuming a multi-part basin
(e.g. as in Section 3.2) to resolve (which is not done here). In some 4 CHALLENGES TO ADDRESS IN
cases, there are trade-offs between the apparent depth and veloc- GEOLOGICAL TOMOGRAPHY
ity contrast within the basin, with surface waves best constraining FRAMEWORK
only the product for shallow basins. One potentially new finding
From the three examples (Section 3.2–3.4), many of the differ-
which has not been commented on previously is the relatively deep
ences in the design and performance of the inversions have been
(>4 km) basin in Central Oregon and part of Northern California
demonstrated. Notably, geological tomography allows parameters
that trends north–south directly to the east of the Cascades Range.
of geological or geodynamic interest to be directly estimated from
A similar feature was observed in the somewhat shallow (1–3 km)
the inversion, such as fault slip in the sedimentary basin example,
density maps of Lin et al. (2014) and we suggest this robustly im-
subduction rate the subduction zone example, and basin depth in the
aged joint basin feature to be due to low density, low velocity ash
continental United States example. However, if geological tomog-
deposits from prior Cascades eruptions, potentially mixed with sed-
raphy is to become commonly used, it is important to also consider
imentary deposits. These basin deposits are not observed in surface
some of the main challenges with this new approach, focusing on
geology maps (e.g. Frezon et al. 1983; Mooney & Kaban 2010)
ones that differ from those of traditional tomography. We summarize
due to more recent volcanic flow deposits completely covering and
the main benefits and challenges of the three different approaches
burying them. Most boreholes in this region have focused on the
to parametrizing tomography in Table 1.
more economically relevant near-shore basins west of the Cascade
Range (Wermiel 1987) and have not previously discussed the find-
ings described here as a major tectonic feature.
4.1. Every problem is different
We next describe a few interesting features of the crustal-scale
features in Figs 6(d)–(f). Generally, our maps are consistent with One of the benefits of standard gridded tomography is that it can
previous studies of the continental United States with slower aver- be applied easily to new data sets and new regions. In contrast, in
age crustal velocities (Fig. 6d), thinner crust (Fig. 6f) and slower the geological tomography approach, every problem may need to
1456 V.C. Tsai et al.

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


Figure 6. Continental United States inversion results for basin and crustal-scale properties from surface-wave dispersion measurements. (a) Basin Vs at z0 =
1 km depth, v0 (see Fig. 5). (b) Basin power-law exponent, a (see Fig. 5). (c) Basin depth. (d) Crustal Vs , vc . (e) Mantle Vs , vm . (f) Crustal thickness.

be evaluated independently, with a particular eye to the physics, When a priori information is abundant, there is a question of
the data, the a priori information and the degree to which various how to best incorporate this a priori information. For example, in
potential geological parametrizations are commensurate with the a sedimentary basin setting, what are the 5 most important pro-
data and a priori information. Different physics may be more im- cesses to include in a parametrization or, if the data warrants it,
portant in different situations, for example, with the sedimentary what are the 10 most important processes to include? Answering
basin processes (e.g. graben faulting and deposition) and subduc- such questions takes significant expert knowledge of both the geo-
tion zone processes (e.g. subduction and mantle wedge cooling) logical processes as well as the imaging technique (e.g. seismic ray
being quite different. More subtly, every application has different theory, full-waveform inversion, microtremor methods, etc.) and is
a priori information available and different data quality and array perhaps the most difficult challenge for implementing geological
geometry. Due to our criterion that the parameters of the geolog- tomography. Different choices can be tested with synthetic exam-
ical model be uniquely determinable with the data (and a priori ples, but the difficulty of including all aspects of reality in such
information) available, this implies that new synthetic tests must examples precludes this being a perfect solution. Thus, geological
be performed for every new data set to determine which geological tomography may require significantly better education and deeper
parametrization is appropriate. The parametrization should ideally geological insight on the part of the user than has traditionally been
be only as complex as warranted, no more and no less. In cases needed for tomographers. Although it is difficult to generalize, one
where data noise is significant, the synthetic tests must account for possibility for approaching complex problems (e.g. where many of
this or else there is a danger that the geological inversion will yield the geological processes are known and should be included) is to
useless or misleading results. perform the inversion in steps that go from simpler to more complex.
Geological tomography 1457

Table 1. Comparison of the three different tomographic parametrizations.


Tomographic Geometric
parametrization Standard gridded (e.g. Muir & Tsai 2020a) Geological (this work)
Description • Pixel, voxel, spline, spherical • Geometric boundaries explicitly • Geologic/physical parameters
harmonic or cell based estimated estimated
Benefits • Flexible and general • Model geometries estimated • Physical parameters of interest
directly estimated directly, including errors
• Maximizes discovery opportunities • Includes appropriate a priori
information
• Unbiased • Fewer parameters
• Linear or weakly non-linear • Overdetermined
inversion
• Straightforward implementation • Model selection is straightforward
Challenges • Does not respect geology/physics • Computationally inefficient, • Physics must be understood or
requires care in numerical tested

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


implementation
• Physical parameters difficult to • Highly non-linear • Non-linear
estimate
• Many parameters • Local minima common • Every problem is different
• Underdetermined • Expert knowledge required
• No physical priors
• Needs regularization

For example, one might always want to begin the imaging process (e.g. Rawlinson & Sambridge 2004) or finite-frequency effects (e.g.
with standard gridded tomography to get a better sense for the limi- Zhou et al. 2004). In contrast, one of the challenges with geological
tations of the data set, and then proceed next to the simplest geolog- inversion is that it is generally more strongly non-linear and there-
ical parametrization that is consistent with what one expects based fore must rely on a range of different types of algorithms. As an
on the gridded approach. Not only would this simplest geological example, in the sedimentary basin example shown in Section 3.2
parametrization utilize as few parameters as possible, but it should (see Fig. 3d), step 1 was linear (with a subset of the data), step 2
also utilize the ones that likely have the most explanatory power and used a (non-linear) Newton method, step 3 used a data visualization
least trade-off, similar to in the examples shown in Figs 3, 4 and 6. technique to infer parameters, step 4 used a linear inversion within
Once these parameters are estimated, additional complexity can be a (non-linear) grid search, step 5 used a decision tree and step 6
added. This approach can also help reduce the computational needs had three different cases, each with a linear inversion within a grid
since it uses fewer resources to proceed stepwise than as a joint search. There are many other possible approaches to performing the
inversion. Nonetheless, one may continue to refine estimates of the same inversion. For example, steps 4, 5 and 6 could have been com-
first parameters as warranted. In most cases, a priori information bined into a 4-parameter grid search, and a Newton method could
can also be used to provide initial priors on the likely possible range have been used in step 6. In one extreme, one could have performed
for the various parameters (e.g. Jackson 1972). In some cases, the a full grid search for all 10 geological parameters or for the 12 sten-
a priori information is strong and can help make binary decisions cil parameters we initially inverted for. All of these choices have
about whether certain processes or parameters should be included implications for computational time, and this last approach would
in the inversion whereas in other cases weaker a priori information not have been computationally feasible on the computer architecture
should only loosely guide choices, leaving the decision to the inver- used, although it may have been feasible if codes were parallelized
sion based on how well the data are fit. Although there are Bayesian and high-performance computing resources were utilized (e.g. Leng
frameworks that could help weigh the data versus a priori informa- et al. 2016). As is the norm for non-linear problems, it is difficult
tion (e.g. Tarantola 2005) and incorporate conditional priors (e.g. to generalize to assess whether certain inversion techniques will
Denison et al. 2002), these frameworks may be difficult to evalu- be accurate and robust in all scenarios, and such choices are best
ate and it may be appropriate to use the more intuitive approach decided by an expert with the data at hand (e.g. Zelt 1999).
described above. We also emphasize that, however detailed, the ge- Although the specific steps of the stepwise inversions shown in
ological parametrization chosen will inevitably be a simplification Fig. 7 again depend on the specific problem, it is a general feature
of reality and therefore will always be imperfect in ways that the that the parameters are not equivalent (unlike in gridded inversion
synthetic examples shown in Section 3 are not. Again, the degree where every parameter is a gridded velocity), and some parame-
to which the parametrizations are successful can lead to criteria for ters should be more readily and robustly determined with less trade
whether they should be preferred or not (see Section 4.4). off with other parameters. In the proposed approach, one should
solve for these parameters first, similar to the discussion in Sec-
tion 4.1 except within the inversion step itself. Information about
4.2. Non-linear instead of linear inversion the geological structure thus builds up step by step throughout the
Another benefit of standard gridded inversion is that the straight- inversion as shown in Fig. 7(a) for the sedimentary basin exam-
ray inversion is strictly linear in the gridded (velocity) parameters, ple and in Fig. 7(b) for the subduction zone example. Within a
and can be solved straightforwardly with linear matrix algorithms Bayesian framework, one can think of the results of the previous
(e.g. Aster et al. 2013); furthermore, gridded inversion remains only step (posterior) to act as a prior for the next step of the inversion
weakly non-linear with added complexity such as non-straight rays (e.g. Denison et al. 2002; Tarantola 2005). Unfortunately, the most
1458 V.C. Tsai et al.

(a) (b) 4.3. Error analysis


The examples show that the errors on the different parameters can be
1 vastly different. A natural outcome of the geological inversion is that
different parameters have different uncertainties. If the inversion is
performed stepwise, as done in the examples, these uncertainties
can be straightforwardly estimated since there are clear relation-
ships between the parameters estimated and the steps themselves.
1 2
For example, on each Newton step, an uncertainty on the specific
parameter being solved for can be estimated with the inverse Hes-
sian on that variable (e.g. Aster et al. 2013); on a grid search step,
uncertainties are again easily associated with the numerically ap-
2 3 & 4 proximated Hessian; in other steps, the grid spacing or step size can
be associated with the uncertainty in a parameter (e.g. in step 3 of
the sedimentary basin inversion). Although these stepwise uncer-
tainties are conditional uncertainties rather than full uncertainties

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


(e.g. Aster et al. 2013), the full uncertainties are dominated by the
3 5 conditional uncertainties due to the order of inversion; moreover,
the full uncertainties could be calculated at the end of the inversion
by rerunning the specific inversion step for a range of values for
other parameters. In some cases, the ‘true’ parameter uncertainty
4 & 5 is smaller than the step size used in the inversion, in which case
6
an upper bound on the uncertainty of that parameter is determined.
Generally, this ability of geological inversion to directly estimate
the uncertainties on geological parameters of interest is one of the
main expected benefits of the approach. As mentioned in Section
6 7 2, standard gridded tomography will always be useful for cases
where the geological processes are unknown (for example, when
one encounters a new region without any geological or other imag-
ing data). However, quantitative inferences about the geological
Figure 7. Schematic showing how information is obtained step by step
throughout the geological inversion. (a) Sedimentary basin example. (b)
processes involved are best estimated with the geological inversion
Subduction zone example. Numbers correspond to numbers described in approach.
the text (Sections 3.2 and 3.3, respectively).

4.4. Model selection for different models


straightforward way to know which parameters are better deter-
mined is to make use of uncertainty quantification from a full grid One of the key ways in which the geological inversion described here
search (e.g. Tarantola 2005; Sullivan 2015). Since such a full grid can be useful is in determining which model framework amongst
search is computationally inefficient and often practically impos- two or more frameworks is better or more parsimoniously fits
sible, one must inevitably use less direct expert knowledge of the the data. To accomplish this comparison between different model
situation to make educated guesses as to what some likely possi- frameworks, including those with potentially quite different num-
bilities might be, and that smaller set of design possibilities may bers of parameters, a model selection framework is useful and can
each be tested to determine which is better. Even this smaller set of be applied. Model selection has been used in many other con-
architectures may be challenging to test since the implementation texts, including to choose between different seismic hazards models
of each one can take significant effort on the part of the researcher. (Scherbaum et al. 2009), to decide whether data is sufficient to con-
Finally, we note that since inversion is non-linear, a Markov- clude that the Earth has a core (Muir & Tsai 2020b), to determine
Chain Monte Carlo (MCMC) approach may be particularly effective the best structural model in engineering (Beck & Yuen 2004) and
(e.g. Bodin & Sambridge 2009). Since MCMC inversion approaches to decide the appropriate basis for fluid imaging (Grauer et al.
rely only on many forward solves, without requiring inverses to be 2017). Similar to how model selection is argued to help in the ge-
taken, and is more efficient than a full grid search, the computational ometric parametrization framework (Muir & Tsai 2020a), model
time required is similar to that needed when applying MCMC to selection using various information criterion (e.g. Claeskens 2016),
linear inverse problems while simultaneously potentially being more using posterior predictive performance checks like leave-one-out
robust than determining a single best-fitting model. Since generic cross validation (LOO-CV; Vehtari et al. 2017), using Bayes fac-
MCMC or even sophisticated samplers like Hamiltonian Monte tors (Weinberg 2012) or using Bayesian evidence (Sambridge et al.
Carlo (Betancourt 2018) or affine invariant samplers (Goodman & 2006) can determine which geological model framework is prefer-
Weare 2010) may still be prohibitively expensive, when applied to able.
geological inversion, we would still advocate that the sampling be Using a model selection framework is especially important for
done stepwise, in the same manner as described above, to reduce geological tomography when it is unclear which of a variety of dif-
the computational burden of exploring the entire parameter space. ferent simplifications of the complex reality is most parsimonious
For example, for the sedimentary basin example, this would imply in terms of explaining the data. Unlike in traditional tomography,
sampling and performing MCMC on the background velocity first where one can choose to ignore the physics in favour of having
with a subset of the data, after which the distribution on this first a maximally generic and flexible set of basis functions, in using
parameter can be fixed for sampling of the other parameters. a geological parametrization, the user must grapple with the fact
Geological tomography 1459

that likely no reasonable geological parametrization can express all discussed in Section 4.2, one can certainly use a Bayesian frame-
possible images of the Earth. Although some may see this as a draw- work to implement a geological inversion, and the two approaches
back of the approach, this has advantages from an epistemological may converge in the distant future, when significantly more infor-
standpoint, since a large subset of possibilities can be ruled out mation is known about the Earth’s interior. However, until that time,
by the nature of the inversion, and the inversion and user can thus the approaches should be thought of as being distinctly different.
focus only on the smaller subset of physically plausible structures On the other hand, the philosophical approach taken in inter-
to distinguish between (e.g. Curd & Cover 1998). Moreover, having rogation theory is to pose a specific question that one is inter-
to choose simplifications that do not capture all of reality force the ested in having answered by an imaging data set. Again, there is
user to focus their attention to those parameters and processes which some similarity to the geological tomography approach where the
are possible to distinguish from each other and away from those for value of each of the geological parameters of interest is estimated.
which the data have less explanatory power over (and for which the The main difference is that geological tomography estimates all
null space is large). of the geological parameters of interest simultaneously by forcing
the parametrization to be geological; in contrast, within the inter-
rogation framework, one would need to ask separate questions for
each parameter and the tomography itself still relies on traditional
4.5. Extension to other data types
parametrizations. Within interrogation theory, the estimates of cer-

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


Although the examples shown are only for transmission straight-ray tain parameters are not used in the estimates of the other parameters,
traveltime tomography and for one surface-wave dispersion data whereas it is an integral part of geological tomography for estimates
set, the geological parametrization framework is one that applies to of certain parameters (with more trade-off) to depend on other pa-
arbitrary data types and arbitrarily complex forward solvers. Com- rameters (that have less trade-off), with the order of determination
putational issues aside, it would be straightforward to include other being important.
seismic phases including both P and S reflected, refracted or scat- In some sense, geological tomography can be thought of as a
tered arrivals using the geological parametrizations discussed here, middle ground between a fully Bayesian approach and the interro-
and one can also straightforwardly perform full-waveform anelas- gation approach. In geological tomography, an intermediate num-
tic inversion within the same framework. One can also apply the ber of geological parameters to estimate is chosen that carefully
framework to non-seismic data, including gravity data, electromag- weighs a trade-off between complexity and simplicity. Furthermore,
netic data and neutrino data. However, due to the vastly different it achieves an intermediate result of quantitatively and jointly esti-
sensitivities of those data types to the structures of interest, the spe- mating as many parameters as are well determined.
cific implementation of an inversion routine would likely change
and even the sets of geological parameters that are best to solve for
may be different. As an example, while traveltime data alone cannot
5 C O N C LU S I O N S
distinguish between density anomalies independently from elastic
moduli, if gravity data were included jointly in an inversion then it In this contribution, we have introduced the concept of the geo-
is likely that density should be added as another parameter (or as logical parametrization of tomography, a natural extension of the
multiple parameters) to express the fact that the data are robustly geometric parametrization approach of Muir & Tsai (2020a). The
sensitive to it. key requirement is for the parameters of the inversion to be directly
related to the geological processes being imaged while still being
uniquely determinable within the context of the available data and a
priori information. We found that performing inversions for geolog-
4.6. Relationship with other alternative approaches
ical parameters is feasible in synthetic tests for both a sedimentary
Given some of the similarities between geological tomography and basin example and a subduction zone example and the geologi-
alternative approaches like Bayesian tomography (Khan et al. 2008; cal inversion addresses some of the main limitations of standard
Koelmeijer et al. 2018) and tomography with interrogation theory gridded tomography. Preliminary application to a real dispersion
(Arnold & Curtis 2018), we provide some further explanation about data set at the continental United States scale demonstrates the
how those different approaches may relate to geological tomogra- usefulness of the approach for making inferences about geological
phy. features including basin history and crustal thickness. Although we
The philosophical approach taken in Bayesian tomography is to believe there to be great promise in geological tomography allow-
perform full geodynamic simulations to the best accuracy possible ing for more useful scientific inferences regarding the geological
and to compare predictions of seismic observables for an ensem- processes involved, there are numerous challenges that also must
ble of geodynamic simulations with the observed seismic data sets. be addressed for geological tomography to become widespread.
Through a Bayesian (sometimes MCMC) approach, one can then Amongst the most important challenges are (1) the need for expert
make inferences about the likelihood of whether geodynamic pa- knowledge of both the geology and (seismic) data set to pose a
rameters take on certain values or not. At face value, then, the parsimonious geological inversion, (2) the move from a linear or
geological tomography approach shares some of the philosophy weakly non-linear inverse problem to a strongly non-linear one and
of the Bayesian approach. The key difference is that there are of- (3) the setup of an inversion to uniquely solve for complex sets of
ten significant trade-offs between geodynamic parameters that are geological parameters, particularly when abundant information is
not allowed within the stricter requirements of geological inversion, available. We have provided guidance on each of these challenges
and moreover exploring the full range of geodynamic possibilities is but fully expect there to be an extensive development time period
typically not computationally feasible. In contrast, in the geological over which more detailed answers to these challenges are elaborated
tomography approach, one must make difficult choices in order to upon and specialized for different geological contexts. Once geo-
prevent the underdetermined and trade-off problems and expresses logical tomography is mature, tomography and Earth imaging will
a less ambitious philosophy than that of Bayesian tomography. As finally be able to move beyond presenting qualitatively plausible
1460 V.C. Tsai et al.

structures to presenting quantitative constraints on the underlying Cipta, A., Cummins, P., Dettmer, J., Saygin, E., Irsyam, M., Rudyanto, A.
geological processes. & Murjaya, J., 2018. Seismic velocity structure of the Jakarta Basin,
Indonesia, using trans-dimensional bayesian inversion of horizontal-to-
vertical spectral ratios, Geophys. J. Int., 215, 431–449.
AC K N OW L E D G M E N T S Claeskens, G., 2016. Statistical model choice, Annu. Rev. Stat. Appl., 3,
233–256.
This study was supported by National Science Foundation grant Curd, M. & Cover, J.A., 1998. Philosophy of Science: The Central Issues,
EAR-2011079. We thank Malcolm Sambridge, editor Andrea W.W. Norton and Company.
Morelli and two anonymous reviewers for constructive criticism. Dadashpour, M., Landro, M. & Kleppe, J., 2008. Nonlinear inversion for
VCT designed the study, developed the codes, drafted the figures estimating reservoir parameters from time-lapse seismic data, J. geophys.
Eng., 5, 54–66.
and led the writing. VCT and CH designed the subduction zone
Davis, R.A., 2017. Sediments of the Gulf of Mexico, in Habitats and Biota of
test. VCT and CAD designed the continental United States inver-
the Gulf of Mexico: Before the Deepwater Horizon Oil Spill, pp. 165–215,
sion methodology. All authors contributed to the analysis and to ed. Ward, C.H., Springer.
editing of the paper. De la Varga, M., Schaaf, A. & Wellmann, F., 2019. GemPy 1.0: open-source
stochastic geological modeling and inversion, Geosci. Model Dev., 12,
1–32.

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


D ATA AVA I L A B I L I T Y Denison, D.G.T., Holmes, C.C., Mallick, B.K. & Smith, A.F.M., 2002.
Bayesian Methods for Nonlinear Classification and Regression, John Wi-
No new data were generated or analysed in support of this research. ley and Sons.
Dispersion data are available from Ekstrom (2017). Code to imple- Dessa, J.-X., et al. 2009. Megathrust earthquakes can nucleate in the forearc
ment the various inversions are available at https://doi.org/10.5281/ mantle: evidence from the 2004 Sumatra event, Geology, 37, 659–662.
zenodo.7730731. Dziewonski, A.M., Hager, B.H. & O’Connell, R.J., 1977. Large-scale het-
erogeneities in the lower mantle, J. geophys. Res., 82, 239–255.
Ekstrom, G., 2017. Short-period surface-wave phase velocities across the
REFERENCES conterminous United States, Phys. Earth planet. Inter., 270, 168–175.
Aki, K., Christoffersson, A. & Husebye, E.S., 1977. Determination of the Fang, H., van der Hilst, R.D., de Hoop, M.V., Kothari, K., Gupta, S. & Dok-
three-dimensional seismic structure of the lithosphere, J. geophys. Res., manic, I., 2019. Parsimonious seismic tomography with Poisson Voronoi
82, 277–296. projections: methodology and validation, Seismol. Res. Lett., 91, 343–
Anderson, E.M., 1905. The dynamics of faulting, Trans. Edinburgh Geol. 355.
Soc., 8, 387–402. Faunt, C.C., Belitz, K. & Hanson, R.T., 2010. Development of a three-
Arnold, R. & Curtis, A., 2018. Interrogation theory, Geophys. J. Int., 214, dimensional model of sedimentary texture in valley-fill deposits of Central
1830–1846. Valley, California, USA, Hydrogeol. J., 18, 625–649.
Aster, R.C., Borchers, B. & Thurber, C.H., 2013. Parameter Estimation and Frezon, S.E., Finn, T.M. & Lister, J.M., 1983. Total thickness of sedimentary
Inverse Problems, 2nd edn, Elsevier. rocks in the conterminous United States, United States Geological Survey
Astic, T. & Oldenburg, D.W., 2019. A framework for petrophysically and Open-File Report 83-920, doi:10.3133/ofr83920.
geologically guided geophysical inversion using a dynamic gaussian mix- Giraud, J., Pakyuz-Charrier, E., Jessell, M., Lindsay, M., Martin, R. & Og-
ture model prior, Geophys. J. Int., 219, 1989–2012. arko, V., 2017. Uncertainty reduction through geologically conditioned
Audet, D.M. & Fowler, A.C., 1992. A mathematical model for compaction petrophysical constraints in joint inversion, Geophys., 82, ID19–ID34.
in sedimentary basins, Geophys. J. Int., 110, 577–590. Goodman, J. & Weare, J., 2010. Ensemble samplers with affine invariance,
Backus, G.E., 1962. Long-wave elastic anisotropy produced by horizontal Comm. App. Math. and Comp. Sci., 5, 65–80.
layering, J. Geophys. Res., 67, 4427–4440. Grauer, S.J., Hadwin, P.J., Sipken, T.A. & Duan, K.J., 2017. Measurement-
Beck, J.L. & Yuen, K.-V., 2004. Model selection using response mea- based meshing, basis selection, and prior assignment in chemical species
surements: bayesian probabilistic approach, J. Engineer. Mech., 130, tomography, Opt. Express, 25, 25135–25148.
192–203. Haney, M.M. & Tsai, V.C., 2015. Nonperturbational surface-wave inversion:
Betancourt, M., 2018. A conceptual introduction to Hamiltonian Monte a Dix-type relation for surface waves, Geophys., 80, EN167–EN177.
Carlo, arXiv:1701.02434v2 [stat.ME], doi:10.48550/arXiv.1701.02434. Haney, M.M. & Tsai, V.C., 2017. Perturbational and nonperturbational in-
Bodin, T. & Sambridge, M., 2009. Seismic tomography with the reversible version of Rayleigh-wave velocities, Geophys., 82, F15–F28.
jump algorithm, Geophys. J. Int., 178, 1411–1436. Haney, M.M. & Tsai, V.C., 2020. Perturbational and nonperturbational in-
Boore, D.M. & Joyner, W.B., 1997. Site amplifications for generic rock sites, version of Love-wave velocities. Geophys., 85, F19–F26.
Bull. seism. Soc. Am., 87, 327–341. Haney, M.M., Ward, K.M., Tsai, V.C. & Schmandt, B., 2020. Bulk struc-
Boore, D.M., Thompson, E.M. & Cadet, H., 2011. Regional correlations of ture of the crust and upper mantle beneath Alaska from an approximate
VS30 and velocities averaged over depths less than and greater than 30 Rayleigh-wave dispersion formula, Seismol. Res. Lett., 91, 3064–3075.
meters, Bull. seism. Soc. Am., 101, 3046–3059. Hosseini, K., Sigloch, K., Tsekhmistrenko, M., Zaheri, A., Nissen-Meyer,
Cammarano, F., Goes, S., Vacher, P. & Giardini, D., 2003. Inferring upper- T. & Igel, H., 2020. Global mantle structure from multifrequency tomog-
mantle temperatures from seismic velocities, Phys. Earth Planet. Interior, raphy using P, PP and P-diffracted waves, Geophys. J. Int., 220, 96–141.
138, 197–222. Hubbard, J., Shaw, J.H., Dolan, J., Pratt, T.L., McAuliffe, L. & Rockwell,
Candes, E.J. & Wakin, M.B., 2008. An introduction to compressive sam- T.K., 2014. Structure and seismic hazard of the Ventura Avenue anticline
pling, IEEE Signal Process. Mag., 25, 21–30. and Ventura Fault, California: prospect for large, multisegment ruptures
Candes, E.J., Wakin, M.B. & Boyd, S.P., 2008. Enhancing sparsity by in the western Transverse Ranges, Bull. seism. Soc. Am., 104, 1070–1087.
reweighted L1 minimization, J. Fourier Anal. Appl., 14, 877–905. Hwang, Y.K., Ritsema, J., van Keken, P.E., Goes, S. & Styles, E., 2011.
Cao, A., Romanowicz, B. & Takeuchi, N., 2005. An observation of PKJKP: Wavefront healing renders deep plumes seismically invisible, Geophys. J.
inferences on inner core shear properties, Science, 308, 1453–1455. Int., 187, 273–277, doi: .
Capdeville, Y., Guillot, L. & Marigo, J.-J., 2010. 2-D non-periodic homog- Jackson, D.D., 1972. Interpretation of inaccurate, insufficient and inconsis-
enization to upscale elastic media for P-SV waves, Geophys. J. Int., 182, tent data, Geophys. J. R. astr. Soc., 28, 97–109.
903–922. Jackson, D.D., 1979. The use of a priori data to resolve non-uniqueness in
Carlson, C.G. & Anderson, S.B., 1965. Sedimentary and tectonic history of linear inversion, Geophys. J. R. astr. Soc., 57, 137–157.
North Dakota part of Williston Basin, Bull. Am. Assoc. Petrol. Geol., 49, Jordan, T.H. & Franklin, J.N., 1971. Optimal solutions to a linear inverse
1833–1846. problem in geophysics, Proc. Nat. Acad. Sci., 68, 291–293.
Geological tomography 1461

Jordan, T.H. & Sverdrup, K.A., 1981. Teleseismic location techniques and Scherbaum, F., Delavaud, E. & Riggelsen, C., 2009. Model selection in
their application to earthquake clusters in the south-central Pacific, Bull. seismic hazard analysis: an information-theoretic perspective, Bull. seism.
seism. Soc. Am., 71, 1105–1130. Soc. Am., 99, 3234–3247.
Karcher, J.C., 1987. The reflection seismograph: its invention and use in the Schmandt, B., Dueker, K., Humphreys, E. & Hansen, S., 2012. Hot mantle
discovery of oil and gas fields, Leading Edge, 6, 10–19. upwelling across the 660 beneath Yellowstone, Earth planet. Sci. Lett.,
Khan, A., Connolly, J.A.D. & Taylor, S.R., 2008. Inversion of seismic and 331-332, 224–236.
geodetic data for the major element chemistry and temperature of the Schneider, F., Potdevin, J.L., Wolf, S. & Faille, I., 1996. Mechanical
Earth’s mantle, J. geophys. Res., 113(B9), doi:10.1029/2007JB005239. and chemical compaction model for sedimentary basin simulators,
Koelemeijer, P., Schuberth, B.S.A., Davies, D.R., Deuss, A. & Ritsema, J., Tectonophysics, 263, 307–317.
2018. Constraints on the presence of post-perovskite in Earth’s lowermost Shearer, P.M., Rychert, C.A. & Liu, Q., 2011. On the visibility of the inner-
mantle from tomographic-geodynamic model comparisons, Earth planet. core shear wave phase PKJKP at long periods, Geophys. J. Int., 185,
Sci. Lett., 494, 226–238. 1379–1383.
Laloy, E., Herault, R., Jacques, D. & Linde, N., 2018. Training-image based Shen, W. & Ritzwoller, M.H., 2016. Crustal and uppermost mantle structure
geostatistical inversion using a spatial generative adversarial neural net- beneath the United States, J. geophys. Res., 121, 4306–4342.
work, Water Resour. Res., 54, 381–406. Song, J., Gao, S.S., Liu, K.H., Sun, M., Yu, Y., Kong, F. & Mickus, K., 2022.
Lay, T. & Wallace, T.C., 1995. Modern Global Seismology, Academic Press. Crustal structure and subsidence mechanisms of the Williston Basin: new
Leng, K., Nissen-Meyer, T. & van Driel, M., 2016. Efficient constraints from receiver function imaging, Earth planet. Sci. Lett., 593,

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023


global wave propagation adapted to 3-D structural complexity: a doi:10.1016/j.epsl.2022.117686.
pseudospectral/spectral-element approach, Geophys. J. Int., 207, 1700– Song, T.-R.A. & Kawakatsu, H., 2012. Subduction of oceanic astheno-
1721. sphere: evidence from sub-slab seismic anisotropy, Geophys. Res. Lett.,
Lin, F.-C., Tsai, V.C. & Schmandt, B., 2014. 3-D crustal structure 39, doi:10.1029/2012GL052639.
of the western United States: application of Rayleigh-wave ellip- Stetson, H.C., 1953. The sediments of the western Gulf of Mexico, in The
ticity extracted from noise cross-correlations, Geophys. J. Int., 198, Sediments of the Western Gulf of Mexico. I. The continental terrace of the
656–670. western Gulf of Mexico: its surface sediments, origin and development.
Linde, N., Renard, P., Mukerji, T. & Caers, J., 2015. Geological realism in II. Chemical studies of the western Gulf of Mexico”, Papers in Physical
hydrogeological and geophysical inverse modeling: a review, Adv. Water Oceanography and Meteorology, Stetson, H.C. & Trask, P.D., Vol. 12, pp.
Resour., 86, 86–101. 1953–1905, doi:10.1575/1912/1063.
Liu, Q. & Gu, Y.J., 2012. Seismic imaging: from classical to adjoint tomog- Su, W., Woodward, R.L. & Dziewonski, A.M., 1994. Degree 12 model of
raphy, Tectonophysics, 566-567, 31–66. shear velocity heterogeneity in the mantle, J. geophys. Res., 99, 6945–
Magistrale, H., Day, S., Clayton, R.W. & Graves, R., 2000. The SCEC 6980.
southern California reference three-dimensional seismic velocity model Sullivan, T.J., 2015. Introduction to Uncertainty Quantification. Springer
version 2, Bull. seism. Soc. Am., 90, S65–S76. International Publishing.
Montelli, R., Nolet, G. & Dahlen, F.A., 2006. Comment on ‘banana- Sun, J. & Li, Y., 2015. Multidomain petrophysically constrained inver-
doughnut kernels and mantle tomography’ by van der Hilst and de Hoop, sion and geology differentiation using guided fuzzy c-means clustering,
Geophys. J. Int., 167, 1204–1210. Geophys., 80, ID1–ID18.
Montelli, R., Nolet, G., Dahlen, F.A., Masters, G., Engdahl, E.R. & Hung, Tarantola, A., 2005. Inverse Problem Theory and Methods for Model Pa-
S.-H., 2004. Finite-frequency tomography reveals a variety of plumes in rameter Estimation, SIAM.
the mantle, Science, 303, 338–343. Troost, K.G. & Booth, D.B., 2008. Geology of Seattle and the Seattle area,
Mooney, W.D. & Kaban, M.K., 2010. The North American upper man- Washington, in Landslides and Engineering Geology of the Seattle, Wash-
tle: density, composition and evolution, J. geophys. Res., 115(B12), ington, Area: Geological Society of America Reviews in Engineering Ge-
doi:10.1029/2010JB000866 ology, pp. 1–35, doi:10.1130/2008.4020(01).
Muir, J.B. & Tsai, V.C., 2020a. Geometric and level set tomography Tsai, V.C. & Atiganyanun, S., 2014. Green’s functions for surface waves in
using ensemble Kalman inversion, Geophys. J. Int., 220, 967–980, a generic velocity structure. Bull. seism. Soc. Am., 104, 2573–2578.
doi:10.1093/gji/ggz472. van der Hilst, R.D. & de Hoop, M.V., 2005. Banana-doughnut kernels and
Muir, J.B. & Tsai, V.C., 2020b. Did Oldham discover the core after all? mantle tomography, Geophys. J. Int., 163, 956–961.
Handling imprecise historical data with hierarchical Bayesian model se- van der Hilst, R.D. & de Hoop, M.V., 2006. Reply to comment by R. Mon-
lection, Seismol. Res. Lett., 91, 1377–1383. telli, G. Nolet and F.A. Dahlen on ‘banana-doughnut kernels and mantle
Nataf, H.-C. & Ricard, Y., 1996. 3SMAC: an a priori tomographic model tomography’, Geophys. J. Int., 167, 1211–1214.
of the upper mantle based on geophysical modeling, Phys. Earth planet. van der Hilst, R.D., Widiyantoro, S. & Engdahl, E.R., 1997. Evidence
Inter., 95, 101–122. for deep mantle circulation from global tomography, Nature, 386,
Nelson, P.H. & Santus, S.L., 2011. Gas, oil and water production from 578–584.
Wattenberg Field in the Denver Basin, Colorado, United States Geological Vehtari, A., Gelman, A. & Gabry, J., 2017. Practical Bayesian model evalu-
Survey Open-File Report 2011-1175. ation using leave-one-out cross validation and WAIC, Stat. Comput., 27,
Newell, W.L., Clark, I. & Bricker, O., 2004. Distribution of Holocene sedi- 1413–1432.
ments in Chesapeake Bay as interpreted from submarine geomorphology Waldron, J. & Snyder, M., 2020. Geological Structures: A Practical Intro-
of the submerged landforms, selected core holes, bridge borings and seis- duction, University of Alberta. https://openeducationalberta.ca/introduc
mic profiles, United States Geological Survey Open-File Report 2004– torystructuralgeology/
1235. Weinberg, M.D., 2012. Computing the Bayes factor from a Markov chain
Rawlinson, N. & Sambridge, M., 2004. Wave front evolution in strongly Monte Carlo simulation of the posterior distribution, Bayesian Anal., 7,
heterogeneous layered media using the fast marching method, Geophys. 737–770.
J. Int., 156, 631–647. Wermiel, D.E., 1987. Available well records and samples of onshore and off-
Rudin, L.I., Osher, S. & Fatemi, E., 1992. Nonlinear total variation based shore oil and gas exploration wells in Oregon, State of Oregon Department
noise removal algorithms, Physica D, 60, 259–268. of Geology and Mineral Industries Oil and Gas Investigations OGI–16.
Russo, R.M. & Silver, P.G., 1994. Trench-parallel flow beneath the Nazca Wright, T.L., 1991. Structural geology and tectonic evolution of the Los An-
Plate from seismic anisotropy, Science, 263, 1105–1111. geles Basin, California, in Active Margin Basins, pp. 35–134, American
Sambridge, M., Gallagher, K., Jackson, A. & Rickwood, P., 2006. Trans- Association of Petroleum Geologists: AAPG Special Volumes.
dimensional inverse problems, model comparison and the evidence, Zelt, C.A., 1999. Modelling strategies and model assessment for wide-angle
Geophys. J. Int., 167, 528–542. seismic traveltime data, Geophys. J. Int., 139, 183–204.
1462 V.C. Tsai et al.

Zelt, C.A. & Smith, R.B., 1992. Seismic traveltime inversion for 2-D crustal Text S2. Text describing implementation of geometric parametriza-
velocity structure, Geophys. J. Int., 108, 16–34. tion of tomography.
Zhao, X., Curtis, A. & Zhang, X., 2022. Interrogating subsurface structures Text S3. Text describing implementation of geological parametriza-
using probabilistic tomography: an example assessing the volume of the tion of tomography.
Irish Sea basins, J. geophys. Res., 127, e2022JB024098.
Zhou, Y., Dahlen, F.A. & Nolet, G., 2004. Three-dimensional sensitivity Please note: Oxford University Press is not responsible for the con-
kernels for surface wave observables, Geophys. J. Int., 158, 142–168. tent or functionality of any supporting materials supplied by the
authors. Any queries (other than missing material) should be di-
rected to the corresponding author for the paper.
S U P P O RT I N G I N F O R M AT I O N
Supplementary data are available at GJ I online.
Text S1. Text describing implementation of standard gridded
parametrization of tomography.

Downloaded from https://academic.oup.com/gji/article/234/2/1447/7093399 by guest on 22 April 2023

You might also like