Access to this full-text is provided by Wiley.
Content available from Journal of Geophysical Research: Solid Earth
This content is subject to copyright. Terms and conditions apply.
1. Introduction
Modeling the crustal architecture of a volcano is of fundamental importance; interactions between magmatic and
hydrothermal systems play a central role in volcanic unrest and eruption (e.g., Chouet & Matoza,2013; Pritchard
etal.,2019; Wilks etal.,2020). However, these systems are often poorly characterized, both due to their complex-
ity and to the difficulty of probing the crust in sufficient detail. Further motivation for investigation is that
volcanic-hosted hydrothermal systems may be a source of geothermal energy production (Reinsch etal.,2017) or
metal-rich brines (Blundy etal.,2021).
Abstract Understanding the crustal structure and the storage and movement of fluids beneath a volcano
is necessary for characterizing volcanic hazard, geothermal prospects and potential mineral resources. This
study uses local earthquake traveltime tomography to image the seismic velocity structure beneath Nabro, an
off-rift volcano located within the central part of the Danakil microplate near the Ethiopia-Eritrea border. Nabro
underwent its first historically documented eruption in June 2011, thereby providing an opportunity to analyze
its post-eruptive state by mapping subsurface fluid distributions. We use a catalog of earthquakes detected
on a temporary seismic array using machine learning methods to simultaneously relocate the seismicity and
invert for the three-dimensional P- and S-wave velocity structures (VP, VS) and the ratio between them (VP/
VS). Overall, our model shows higher than average P- and S-wave velocities, suggesting the presence of
high-strength, solidified intrusive magmatic rocks in the crust. We identify an aseismic region of low VP, low
VS, and high VP/VS ratio at depths of 6–10kmb.s.l., interpreted as the primary melt storage region that fed the
2011 eruption. Above this is a zone of high VS, low VP, and low VP/VS ratio, representing an intrusive complex
of fractured rocks partially saturated with over-pressurized gases. Our observations identify the persistence of
magma in the subsurface following the eruption, and track the degassing of this melt through the crust to the
surface. The presence of volatiles and high temperatures within the shallow crust indicate that Nabro is a viable
candidate for geothermal exploration.
Plain Language Summary Understanding the structure of the crust and the distribution and
movement of fluids beneath a volcano allows for the assessment of volcanic hazard, geothermal potential
and possible mineral extraction. To identify different regions of the crust and differentiate between fluids, we
use the fact that the speed of seismic waves depends on the material they are traveling through. For example,
seismic waves will travel through magma (molten, or liquid, rock) at lower speeds than in the surrounding rock.
The focus of this study is Nabro volcano in Eritrea, which erupted in 2011. We use earthquakes that have been
automatically detected following the eruption to image the structure of the crust in the form of 3D variations
in seismic wave speeds. This identifies a volume of magma stored at depths of 6–10km below sea level, which
fed the eruption. Above this, we observe a region of rocks that are likely remnants of earlier eruptions at Nabro,
with fractures containing gases at high pressure. The source of this high pressure is the release of gas from the
magma storage zone. The presence of hot fluids means Nabro could be used as a source of geothermal power in
the future.
GAUNTLETT ETAL.
© 2023. The Authors.
This is an open access article under
the terms of the Creative Commons
Attribution License, which permits use,
distribution and reproduction in any
medium, provided the original work is
properly cited.
Seismic Tomography of Nabro Caldera, Eritrea: Insights
Into the Magmatic and Hydrothermal Systems of a Recently
Erupted Volcano
M. Gauntlett1 , T. Hudson1 , J.-M. Kendall1 , N. Rawlinson2 , J. Blundy1 , S. Lapins3 ,
B. Goitom3, J. Hammond4 , C. Oppenheimer5 , and G. Ogubazghi6
1Department of Earth Sciences, University of Oxford, Oxford, UK, 2Bullard Laboratories, Department of Earth Sciences,
University of Cambridge, Cambridge, UK, 3School of Earth Sciences, University of Bristol, Bristol, UK, 4Department of
Earth and Planetary Sciences, University of London, Birkbeck, UK, 5Department of Geography, University of Cambridge,
Cambridge, UK, 6Eritrea Institute of Technology, May-Nefhi, Eritrea
Key Points:
• 3D seismic modeling reveals the
structure of the magmatic and
hydrothermal systems beneath Nabro
volcano in Eritrea
• The primary melt storage region
feeding the 2011 eruption is located at
depths of 6–10km below sea level
• Degassing from the magma storage
zone causes overpressure in partially
saturated, fractured intrusive complex
above
Supporting Information:
Supporting Information may be found in
the online version of this article.
Correspondence to:
M. Gauntlett,
miriam.gauntlett@linacre.ox.ac.uk
Citation:
Gauntlett, M., Hudson, T., Kendall,
J.-M., Rawlinson, N., Blundy, J., Lapins,
S., etal. (2023). Seismic tomography
of Nabro caldera, Eritrea: Insights
into the magmatic and hydrothermal
systems of a recently erupted volcano.
Journal of Geophysical Research: Solid
Earth, 128, e2022JB025742. https://doi.
org/10.1029/2022JB025742
Received 5 OCT 2022
Accepted 1 MAY 2023
Author Contributions:
Conceptualization: M. Gauntlett, J.-M.
Kendall
Data curation: S. Lapins, B. Goitom, G.
Ogubazghi
Formal analysis: M. Gauntlett, T.
Hudson, N. Rawlinson, J. Blundy, J.
Hammond
Funding acquisition: J.-M. Kendall,
J. Hammond, C. Oppenheimer, G.
Ogubazghi
Investigation: M. Gauntlett, T. Hudson,
J.-M. Kendall, J. Blundy, J. Hammond
Methodology: N. Rawlinson, S. Lapins
10.1029/2022JB025742
RESEARCH ARTICLE
1 of 21
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
2 of 21
Seismic body wave tomography is a powerful geophysical tool used to image the Earth's interior on various scales.
It has been applied locally to deforming volcanoes in order to understand their subsurface active magmatic and
hydrothermal processes (e.g., Chiarabba & Moretti,2006; Greenfield etal.,2016; Korger & Schlindwein,2014;
Koulakov etal.,2021; Patane et al., 2002; Wilks etal.,2020). Seismic velocities of rocks are influenced by a
multitude of factors, including lithology, fractures, temperature, the presence of fluids and gases, fluid saturation
and porosity. Knowledge of the seismic velocity structure can therefore help to identify melt-bearing regions,
hydrothermal fluids and over-pressurized gases beneath a volcano (e.g., Lin,2013; Londoño & Sudo, 2003;
Vanorio etal.,2005; Wilks etal.,2020). These observations are crucial to understanding volcanic unrest, charac-
terizing future volcanic hazard and assessing geothermal potential.
Using seismic data collected following the 2011 eruption of Nabro, we present high-resolution tomographic images
of the crust with the aim of improving understanding of the post-eruptive state and dynamics at depth beneath the
volcano. Nabro volcano is located on the central part of the Danakil microplate near the Ethiopia-Eritrea border in
the Afar region. The Afar depression is part of the East African Rift System, an active continental rift system, and
contains the triple junction between Arabian, Nubian and Somalian plates (Hammond etal.,2011). Most of the
active volcanism associated with continental rifting is found along the central rift axis. However, Nabro is offset
from the axis of spreading and is thus known as an “off-rift” or “off-axis” volcano (Barberi etal.,1974; Wiart
& Oppenheimer,2005). Together with the neighboring caldera of Mallahle, Nabro makes up the Bidu Volcanic
Complex. Nabro and Mallahle are characterized by large calderas, thought to have been formed circa 130 and 295
ka ago, respectively (Oppenheimer etal.,2019). The alignment of the volcanic centers bears NE-SW, striking
obliquely to the NW-SE trend of the Red Sea (Goitom etal.,2015; Wiart & Oppenheimer,2005). Nabro is the
largest volcano in the Nabro Volcanic Range (NVR), which runs in a NNE-SSW direction from Bara’Ale volcano
in Ethiopia to the Kod Ali formation in the Red Sea (Wiart & Oppenheimer,2005). Nabro's summit is 2248m
above sea level, and its caldera reaches a diameter of 8km (Wiart & Oppenheimer,2005). It remains unclear
how magma is supplied to off-rift volcanoes such as Nabro; further, their role in accommodating extension is
not well understood (Maccaferri etal.,2014). Indeed, the propagation of strain transfer from the Aden and Red
Sea plate boundaries into the Afar region south of Nabro is complex and transient, with active faults distributed
over hundreds of thousands of square kilometres (Manighetti etal.,2001). Possible explanations for the NVR's
extensive off-rift magmatism include reactivation of an older, pre-rift structure (Barberi etal.,1974) or localized
diapiric upwellings from depth (Hammond etal.,2013).
On 12 June 2011, Nabro volcano underwent its first eruption on historical record—the last dated activity occurred
within the caldera circa 23 ka ago (Oppenheimer etal.,2019). The volcano was unmonitored at the time of the
eruption, with no geophysical surveillance networks operating in Eritrea. The eruption resulted in seven fatalities
and displaced some 12,000 people (Goitom etal., 2015). The explosive activity generated significant tephra
clouds and released 4.5Tg of SO2 into the atmosphere within the first 15days, producing the largest stratospheric
aerosol perturbation since the 1991 Pinatubo eruption (Fromm etal.,2014; Theys etal.,2013). Since the erup-
tion, Nabro has been identified as one of the main geothermal prospects in Eritrea due to increased fumarolic
activity at the surface (Yohannes,2012).
Geodetic modeling suggests that a shallow, NW–SE-trending dyke fed the eruption, which triggered slip on
parallel normal faults, consistent with the orientation of vents within the crater (Goitom etal.,2015). Petrological
analysis by Donovan etal.(2018) identifies two distinct batches of magma, one more primitive and the other high
in sulfur and water content. The authors propose that the latter batch underwent isobaric crystallisation in a stor-
age region at ∼5–7km depth below sea level (b.s.l.), while the more primitive batch rose rapidly to the shallow
crust from depth. In the months following the eruption, Nabro experienced subsidence at a slowly decaying rate
(J. Hamlyn etal.,2018; J. E. Hamlyn etal.,2014). By inverting the deformation field, J. Hamlyn etal.(2018)
propose a best-fitting deflating Mogi source at 6.4±0.3km depth b.s.l..
A temporary seismic network was established around Nabro in the aftermath of the eruption, operational from
31 August 2011 until October 2012. Lapins etal. (2021) trained and validated a novel deep learning model
on this data in order to automatically detect phase arrivals. When deployed, the deep learning model signifi-
cantly augmented the seismic catalog analyzed in previous studies. From this catalog, Lapins(2021a) calculates
hypocenter locations, local and moment magnitudes, path/site attenuation effects and b-values. A key result
is that seismicity beneath Nabro lies above and below an inferred, aseismic magma storage zone at depths of
6–9kmb.s.l., consistent with the modeled Mogi source from previous studies (Goitom etal.,2015; J. E. Hamlyn
etal.,2014) and petrological inferences about magma storage depths (Donovan etal.,2018). Events below the
Project Administration: J.-M. Kendall,
J. Hammond, C. Oppenheimer
Resources: B. Goitom
Software: N. Rawlinson
Supervision: T. Hudson, J.-M. Kendall
Writing – original draft: M. Gauntlett,
T. Hudson, N. Rawlinson, J. Blundy, J.
Hammond, C. Oppenheimer
Writing – review & editing: M.
Gauntlett, T. Hudson, J.-M. Kendall, N.
Rawlinson, J. Blundy, J. Hammond, C.
Oppenheimer
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
3 of 21
reservoir are thought to result from small pulses of magma or volatile migration, while events above the aseismic
zone could reflect outgassing processes, migrations of fluid or melt into the reservoir or intense fracturing as a
result of the observed subsidence (Lapins,2021a). The patterns of seismicity to the northeast of Nabro indicate
that deeper fluid or magmatic processes have triggered movement on a shallower fault, suggesting that fluids may
play an important role in regional extensional processes (Lapins,2021a).
Lapins(2021a) notes that one of the major limitations of their study is the lack of a well-constrained velocity model.
Here, we apply seismic tomography methods to the data set from Lapins etal.(2021) to derive a more accurate veloc-
ity model. We then use this new model to jointly carry out 3D P-wave (VP), S-wave (VS) and VP/VS tomography and
earthquake hypocenter relocation, in order to yield further insight into the subsurface processes responsible for the
seismicity and surface deformation at Nabro. By interpreting these tomographic images, we aim to characterize the
migration and distribution of volcanic fluids in Nabro's active magmatic system. These results have particular rele-
vance for the assessment of Nabro's geothermal energy potential, as well as its future seismic and volcanic hazard.
2. Data
2.1. Network and Data Collection
The seismic data used in the tomographic inversion were collected by a temporary local seismic network deployed
in the aftermath of Nabro's 2011 eruption. Eight three-component broadband 30 s Güralp seismometers (five
CMG-6TD and three CMG-40TD) were provided by SEIS-UK to monitor Nabro's post-eruptive state (J. E. Hamlyn
etal.,2014). The network was fully operational from 31 August 2011 until October 2012. However, one of the 40TD
stations, NAB6, was damaged due to flooding and thus was inoperable, producing no usable data (Lapins,2021a).
NAB7 also had frequent data gaps but was still used for phase arrival picking and event location, and therefore we
also use it in our tomographic inversions, along with the other six stations. The data were all initially recorded at
100Hz sample frequency and then switched to 50Hz sample frequency early in October 2011 (Lapins,2021a).
Manually picking seismic phase arrivals is time-consuming, and can be especially difficult in volcanic settings
due to the fact that volcano-tectonic earthquakes tend to be fairly low magnitude (<4) events. A previous manual
analysis of the Nabro seismic data only covered the time period 31 August–31 December 2011 (Goitom,2017),
which left eight months of data unpicked. Therefore, a new deep learning model for automated phase arrival
detection based on a convolutional neural network, known as U-GPD, was applied to the seismic data from the
temporary network around Nabro (Lapins etal.,2021). The U-GPD model is trained and validated using 35days
of manually picked data from Goitom(2017). To overcome issues surrounding the use of a small training set,
they use transfer learning on an existing deep learning model for phase arrival detection (Ross etal.,2018) trained
using millions of phase arrivals from earthquakes in Southern California. The resulting U-GPD transfer learning
model was shown to outperform two existing, comprehensively trained models, PhaseNet (Zhu & Beroza,2019)
and GPD (Ross etal.,2018), and the existing manual catalog in terms of pick error and number of phase arriv-
als detected (Lapins etal., 2021). When the automated phase arrival picks and original manual pick times are
compared, the root mean square deviation for P-wave picks is 0.038s and for S-wave picks it is 0.053s (Figure
6 in Lapins(2021a)). The model is far more efficient than manual phase picking, processing the 14months of
seismic data in less than 4hr. The three output channels of the U-GPD model give the probability of a P-wave
arrival, S-wave arrival or neither (noise), respectively. A P or S prediction probability must exceed a threshold
value of 0.4 to be identified as a true phase arrival detection (Lapins etal.,2021).
Events are then located using NonLinLoc, a probabilistic, nonlinear hypocenter location package (Lomax
etal.,2000). The 1D starting velocity model used in the NonLinLoc inversion is based on the crustal structure of
the Afar region deduced from wide-angle controlled-source seismology and assuming a VP/VS ratio of 1.76, the
approximate average for continental crust (Ginzburg etal.,1981; Goitom,2017). This produces an initial catalog
of 31,387 events with at least four P-wave arrivals and one S-wave arrival that are available for subsequent use in
seismic tomography inversions (Lapins etal.,2021).
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
4 of 21
2.2. Data Selection
Seismic tomography is highly reliant on accurate traveltime picks, and therefore we restrict the catalog produced
by U-GPD and located in NonLinLoc based on the inversion statistics and event properties. We select earthquakes
with at least four P and four S phases, azimuthal gaps that are less than 180° and location errors of less than 2km,
reducing the catalog to 11,319 earthquakes (Figure1).
The U-GPD deep learning model does not include explicit pick uncertainties, but following Lapins(2021a), we
associate errors to the picks based on the probability of being a true arrival. If this probability exceeds 0.85, a
pick error of 0.05s is assigned to it. Pick arrivals with probabilities 0.7–0.85, 0.55–0.7, and 0.4–0.55 are assigned
pick errors of 0.1, 0.2, and 0.3s, respectively. The tomography results are independent of the absolute values of
Figure 1. The final event catalog after relocation in a joint inversion for velocity structure and earthquake location. Thin
black lines represent the caldera rims of Nabro and Mallahle calderas. Seismic stations from the temporary seismic network
are plotted as inverted yellow triangles, excluding the inoperational station (NAB6). The dark orange line shows the extent
of the 2011 eruption lava flow, and the orange star shows the location of the vent region (J. E. Hamlyn etal.,2014). The
cross-sections show the catalog projected into the longitude-depth and latitude-depth planes, with the Mogi source from J. E.
Hamlyn etal.(2014) represented by a purple star. The seismicity is colored by depth below sea level (b.s.l.) and the histogram
uses bins of 1km depth b.s.l. The inset shows a regional map of the Afar Triple Junction. Red triangles represent Holocene
volcanoes recorded in the Smithsonian catalog “Volcanoes of the World” database. The white star indicates the location of
Nabro. Dashed lines are political borders. DM: Danakil Microplate, RS: Red Sea Rift, GOA: Gulf of Aden Rift, MER: Main
Ethiopian Rift, ARZ: Afar Rift Zone, NP: Nubian Plate.
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
5 of 21
these errors due to the application of regularization (see Section3.4); rather, it is the relative difference in the
pick errors that matters, giving less weight to the picks we are less confident in during the tomographic inversion.
3. Methodology
3.1. Tomographic Method
To investigate the subsurface velocity structure at Nabro, we make use of an iterative nonlinear tomo-
graphic inversion package, FMTOMO (Fast-Marching TOMOgraphy) (de Kool et al., 2006; Rawlinson &
Sambridge,2004a,2004b). FMTOMO inverts seismic traveltime data to constrain 3D VP and VS structure. The
package has been adapted by Pilia etal.(2013) to permit the fully nonlinear relocation of hypocenters and to
solve directly for VP/VS structure. It has been applied in a variety of tectonic settings, using either passive or active
source datasets, or a combination of the two (e.g., Brikke,2010; Korger & Schlindwein,2014; Pilia etal.,2013;
Rawlinson etal.,2006; Rawlinson & Kennett,2008; Wilks etal.,2020; Zenonos etal.,2019).
The key innovation of FMTOMO is the use of an efficient, consistent and robust grid-based eikonal solver known
as the fast marching method (FMM) (Sethian,1996; Sethian & Popovici,1999) to solve the forward problem of
predicting traveltimes in a 2D or 3D heterogeneous, layered medium. Since the subsurface structure beneath volca-
noes is often highly heterogeneous, this makes FMTOMO an appropriate choice for our study of Nabro. Further-
more, FMTOMO can invoke traveltime reciprocity when solving the forward problem. The FMM source points
are interchanged with the receivers, and the eikonal solver computes traveltimes from each receiver location to all
the other grid points, so that the complete traveltime field for each receiver is available rather than for each source.
Typically, most of the computing time of FMTOMO is dedicated to calculating these traveltime fields, and a large
ratio between the number of sources and receivers means that invoking the reciprocity principle can lead to a
significant increase in efficiency. We refer the reader to de Kool etal.(2006) for a more detailed overview of FMM.
FMTOMO defines the seismic velocity field with a regular 3D grid of nodes, which are used as the control verti-
ces of a mosaic of cubic B-spline volume elements. Cubic B-spline functions preserve continuity of the second
derivative whilst also being defined in terms of local basis functions, meaning that changing the velocity value of
one node will only affect the velocities at nodes in the immediate vicinity. This creates a smoothly varying, locally
controlled velocity continuum. Cubic B-spline functions can also be rapidly evaluated, which is useful, since the
multi-stage FMM requires several evaluations of the spline function.
The next step of the algorithm solves the linearized problem of matching observed and predicted traveltimes, that
is, finding model parameters that best satisfy the data. In this case, the data are the arrival time residuals, and
the unknowns are the grid of vertices which control the pattern of the cubic B-spline velocity field. FMTOMO
implements the gradient-based subspace inversion scheme of Kennett etal.(1988), which minimizes the objec-
tive function:
()=
1
2
[(()−)−1
(()−)+(−0)−1
(−0)+]
,
(1)
where the vector m represents the model vector of unknown velocity parameters that are adjusted during the
inversion process, g(m) are the predicted traveltime residuals associated with the model defined by m, dobs are the
observed residuals, m0 is the reference model, Cd is the data covariance matrix and Cm is the a priori model covar-
iance matrix. S(m) is minimized when the model traveltimes most closely resemble the observed traveltimes. The
subspace method locally minimizes the objective function by projecting the quadratic approximation of S(m) onto
an n-dimensional subspace of the full model space. In this case we choose a maximum of n=20 orthogonal search
directions, with singular value decomposition used to reduce the size of the subspace based on the magnitude of
the singular values. Regularization constraints are applied to address solution non-uniqueness: the damping term
encourages the search for models to remain within the vicinity of the reference model, whilst the smoothing term
minimizes the amount of structural variation required to satisfy the observational constraint. Further information
on the inversion scheme and how it is implemented can be found in Rawlinson etal.(2006).
Many local earthquake tomography algorithms rely on a linearized approach to the tomographic inversion (e.g.,
Evans et al., 1994). However, the hypocenter location problem is more strongly nonlinear than the velocity
recovery problem, meaning that the linearized approximation leads to poor results in regions where there is
significant velocity heterogeneity and/or where source locations are not well constrained (Pilia et al., 2013).
Since the computational cost of having a fully nonlinear inversion scheme for both velocity structure and source
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
6 of 21
location would be huge, we opt for a compromise approach using a fully nonlinear source relocation algorithm,
which exploits the grid-based nature of FMM. The availability of the complete traveltime field for each receiver
means that a fully nonlinear grid search for the best source location can be done efficiently, regardless of how
complex the velocity model is (Pilia etal.,2013). The objective function minimized in the grid search is given in
Supporting InformationS1.
Although the source relocation algorithm is fully nonlinear, the use of a linearized velocity inversion scheme
means that an iterative approach is needed to account for the trade-off between velocity variations and hypocenter
locations. The source and velocity inversions are done sequentially. Sources are first relocated using P- and
S-arrival times via the nonlinear grid search method. Next, VP and VS are updated using the new locations, which
involves two steps: (a) the solution of the forward problem using FMM; (b) an inversion for velocity parameters
using the subspace inversion scheme. This is undertaken separately for P-wave and S-wave velocity structure,
but both VP and VS models must be updated between each relocation as both P- and S-arrival times are used to
constrain hypocenter location. We then repeat the entire process of source relocation and velocity inversion. In
this case, an acceptable level of convergence is attained after six iterations.
Following this, we use the final hypocenter locations (as determined by the “joint” inversion for VP, VS and
earthquake location) in the modified FMTOMO algorithm developed by Pilia etal.(2013) in order to calculate
VP/VS. This procedure inverts S-P differential traveltimes for VP/VS structure along the ray paths from the S-wave
model. We assume that (a) each S-wave path between two points has a corresponding P-wave path; (b) the P-
and S-wave paths taken between two different points are identical and have similar Fresnel zones. Under these
assumptions, the inverse problem is linear, as any lateral heterogeneity will cause a divergence of the P- and S-ray
paths (Eberhart-Phillips & Reyners,2012; Thurber,1993; Walck,1988). The method requires common P- and
S-arrival times, so rays with only S-phases or P-phases will be removed. In our case, this does not result in signif-
icant data loss between the calculation of the VP and VS models as compared to the VP/VS model, as the U-GPD
model phase association method has already discarded rays which only have S-arrivals (Lapins etal.,2021). See
Supporting InformationS1 for more detail on how the problem is formulated within a linear framework, and Pilia
etal.(2013) for a full description of the direct inversion of S-P differential traveltimes.
We choose to directly invert S-P differential traveltimes rather than dividing the P-wave model by the S-wave model to
obtain VP/VS. S-wave data coverage tends to be poorer than P-wave data coverage, and is usually noisier, due to S-wave
arrivals being more difficult to pick. The imposition of relatively arbitrary regularization constraints on the amplitude
of anomalies means that the resulting S-wave solution models are comparatively smoother than P-wave models. For
interpretation of individual P- and S-wave models, the absolute amplitude being correct is less relevant than the over-
all pattern of anomalies. However, when dividing the models to obtain VP/VS, the amplitude of P- and S-wave velocity
anomalies directly influences the VP/VS model. If the S-wave model is smoother, the final VP/VS model obtained from
direct division can inherit smaller wavelength features from the P-wave model, as shown by Pilia etal.(2013).
One potential drawback of our approach is that the VP/VS model cannot be derived explicitly from the VP and VS
models but due to solution non-uniqueness, we argue that our inversion produces the optimum model of each
type. Any inconsistencies must be viewed in the context of model uncertainty, which is unavoidable when under-
taking an inversion with noisy data that is unevenly distributed. Indeed, synthetic tests show that the assumptions
inherent to this technique have less effect on the results than ad hoc regularization choices (Pilia etal.,2013).
A flow chart detailing the full tomographic workflow can be found in Figure S10 in Supporting InformationS1.
3.2. 1D Model Selection
Under the assumption of weak nonlinearity, velocity perturbations cannot move too far from the unperturbed
model. Hence, the starting reference model should ideally be as close to the true solution as possible to avoid
the inversion becoming trapped in a local minimum that is far from the correct model. Initial test 3D inversions
using a basic three-layer velocity model from Ginzburg etal.(1981) as a starting reference model do not resolve
much detail, further demonstrating the need for a reference model that more closely reflects the local velocity
structure at Nabro. Therefore we use FMTOMO in a ‘quasi-1D’ inversion to develop and refine a suitable 1D
velocity model to use as a reference model for subsequent 3D inversions, following the method of Wilks(2016).
Because FMTOMO requires there to be at least two velocity grid nodes in any particular direction when the grid
is defined, we cannot explicitly invert for 1D velocity structure. Instead, we perform “quasi-1D” inversions using
a simplified velocity grid, with two nodes defined in both latitude and longitude, spanning 13.15°–13.55°N,
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
7 of 21
41.5°–41.9°E. 23 nodes are defined in the depth direction, and the grid spans
−3–20kmb.s.l., resulting in velocities defined at ∼1km depth increments.
We then invert for VP and VS and relocate hypocenters, and calculate the aver-
age velocity at each nodal depth across the four nodes. FMTOMO automati-
cally generates a boundary layer of two additional nodes at the grid limits, so
it is important to exclude the nodes that make up this padding when averaging
over the nodes.
When a velocity model with a number of discrete, homogeneous layers is
used to locate earthquakes, they tend to cluster at the velocity discontinuities
(J. E. Hamlyn etal.,2014). Thus we smooth out the sharp discontinuities in
the three-layer regional model with a Gaussian filter and use the result as a
starting model. The resulting P- and S-wave “quasi-1D” velocity models are
plotted in Figure2.
For the P-wave velocity inversion, the data variance is reduced from
0.0997 to 0.0273 s
2. For the S-wave velocity inversion, the variance is
reduced from 0.264 to 0.0469s
2. This suggests that this new 1D model
is “closer” to the solution model and represents the seismic structure at
Nabro more accurately than the three-layer regional model from Ginzburg
etal.(1981).
To investigate how sensitive the 3D solution model is to the initial model,
we perturb each value of the new 1D model randomly by up to 10% prior
to inversion and carry out 3D tomographic inversions. We find that the
solution models show broadly similar structure, with differences only
occurring outside of the data resolution limits determined by synthetic
resolution tests described in Section4 (i.e., the models differ in small-scale
structure or in regions of poor data coverage). Results of these perturbation
tests are plotted in Figure S1 in Supporting InformationS1. Analysis of
the inversion statistics shows the same or higher traveltime residuals and
variances for the solution models using the perturbed 1D model compared
to their starting model, as expected, with percentage deviations that are
less than 16% as shown in Table S1 in Supporting InformationS1. There-
fore we conclude that our choice of 1D starting models is robust, and use
these models as the starting model in all of the subsequent VP and VS 3D
inversions.
3.3. Defining a Grid for 3D Inversions
Before undertaking 3D inversions, we first define an inversion grid, which
describes the velocity model in terms of cubic B-spline functions. We also
define a propagation grid, which represents a discrete sampling of the
velocity field for use in the grid-based eikonal solver employed during the
forward step and during the nonlinear relocation of events. Both grids are
comprized of a 3D set of nodes that span 13.15°N–13.55°N, 41.5°E–41.9°E,
and −3.0–20.0km in depth below sea level. The node spacing for the propa-
gation grid is chosen to be ∼0.5km and for the inversion grid it is ∼1km. The
relocation code also carries out a sub-cell search, dicing the initial cells by a
factor of 10, and therefore 50m is the smallest separation distance between
the relocated earthquakes. The inversion grid spacing is sufficiently small to capture features that are constrained
by the data, noting that smoothing is applied to control the wavelength of recovered features. At half the spacing
of the inversion grid, the propagation grid is sufficiently fine to render errors in the forward prediction of travel-
times sufficiently small that they will not influence the inversion results. The 1D model determined in Section3.2
is used as a starting model.
Figure 2. The VP and VS velocity model output from a quasi-1D inversion (red
circles and navy circles respectively), compared with the smoothed three-layer
starting models (red and navy lines, respectively).
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
8 of 21
3.4. Optimizing Data-Model Fit
The aim of the tomographic inversion procedure is to find a model that is similarly smooth and as close to the
initial model as possible (to satisfy local linearity), whilst still satisfying the data. Therefore, three parameters
need to be minimized: data fit, model variance and model roughness. The solution model variance is a measure
of the difference between the starting model and the final model, and model roughness is a measure of how much
complexity exists in the final model itself (based on the second spatial derivative). We optimize these using the
smoothing and damping parameters, η and ϵ. These regularization parameters also affect the data fit, that is,
the difference between the observed data and the final solution model predictions, quantified by the variance of
the traveltime residuals.
Through numerous inversions for 3D velocity structure using different values of the regularization parameters, we
plot trade-off curves to find the damping and smoothing parameters that give the best compromise between data
fit, model variance and model roughness. This process is done separately for the VP, VS and VP/VS models; see the
Supplementary Information for further detail. For VP the parameters are ϵ=3 and η=10, for VS they are ϵ=10
and η=50 and for VP/VS, ϵ=100 and η=20 (see Figure S2 in Supporting InformationS1).
3.5. Earthquake Relocations
After an initial inversion using the optimal damping and smoothing parameters, we examine the output relocated
seismicity, plotted in Figure S3 in Supporting InformationS1. We find that many events have been relocated
substantial distances. A considerable number of earthquakes migrate to the inversion grid boundaries and above
the topography line, which indicates that the initial locations of these earthquakes are poorly constrained. The
mean relocation offset is 1.61km. Therefore, we identify events that are relocated by distances greater than 3km,
and remove them from the catalog. We then repeat the inversion procedure with this reduced catalog of 8,893
events in order to improve the inversion stability.
The inversion results using the new subset of events show improved data fits over the full catalog. For VP the
data variance of the final solution model is reduced from 0.0218s
2 using the full catalog to 0.0101s
2 using the
subset of events. The data variance of the final VS solution model decreases from 0.0298 to 0.0170s
2. VP/VS
also experiences a reduction in the variance from 0.0386 to 0.0283s
2. Furthermore, events are now relocated by
reduced offsets: the mean relocation offset for the reduced catalog is 0.91km. We note that some earthquakes
remain located above the topography line after this process. We attribute this to shallow events being more poorly
constrained and thus erroneously relocated in the air. The average depth uncertainty across all events located
above sea level is ±4.99km, calculated following Wilks etal. (2020)and described in Text S1 in Supporting
InformationS1. The shallowest depth an earthquake is relocated to is 2.49km above sea level. Therefore, within
error, the location of these earthquakes is consistent with their true locations being in the very shallow subsurface.
We therefore use this reduced catalog in the following section to carry out resolution tests and produce the final
solution model.
4. Results
4.1. Checkerboard Resolution Tests
The tomographic inversion problem is non-unique, with many different models able to satisfy the data. Different
factors that help constrain the solution of this inverse problem include the path coverage of the data, data noise
and the choice of implicit and explicit regularization. Thus, assessing the solution robustness is challenging, yet
crucial for comprehensive evaluation of the spatial resolution of the estimated model.
Synthetic reconstruction tests are typically used to investigate the robustness of tomographic models. During these
tests, a heterogeneous synthetic model is formulated and the forward problem is solved using this model, with
the identical source-receiver configurations to the observational data set. This produces a synthetic traveltime
data set. The same inversion method used for the experimental data is applied to the synthetic data set in order to
reconstruct the synthetic model. The most common input structure is a “checkerboard” structure overlain on the
starting model, with alternating positive and negative velocity anomalies making up the “checkers” (e.g., Glahn
etal.,1993; Hearn & Clayton,1986; Rawlinson & Sambridge,2003). Regions where the checkerboard pattern is
recovered are considered to be well resolved.
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
9 of 21
The parameters we choose for our checkerboard tests are: VP perturbations set to ±0.5km/s from the initial model
(±7.35−11.7%) and VS perturbations set to±0.2km/s from the initial model (±5.18−8.69%). It is instructive to
generate checkerboards with different scale lengths of perturbation—this amounts to altering the number of grid
nodes, N, that are perturbed simultaneously. For example, a checkerboard of size N=2 will have nodes perturbed
in pairs. Increasing the value of N will increase the size of each checkerboard element. In Figure3, we present
checkerboard tests with size N=2, 4, 8, corresponding to scale lengths of ∼2, ∼4, and ∼8 km respectively, to
assess the model resolution at different scales.
If no noise is added to the synthetic data set, the result will give an indication of the optimal spatial resolution.
However, since data noise is present in all seismic datasets, noise with a Gaussian distribution is often added, with
a standard deviation equal to that of noise estimates obtained from the data. However, it is important to note that
estimating data uncertainty is often subjective, and it is not clear that the actual noise distribution takes a Gauss-
ian form (Rawlinson & Spakman,2016). Bearing this in mind, we add Gaussian noise with a standard deviation
of 0.05s, representing half a cycle of the dominant frequency of the microseisms (∼10Hz).
For our synthetic checkerboard test, we carry out six iterations for VP/VS and source location, with results plotted
in Figure4. The output checkerboards of scale length ∼4 and ∼8km show that VP/VS anomalies on these scales
are well-resolved within the seismic network in map view (Figures4a and4b). Both checkerboards are resolvable
down to 10km depth b.s.l.. The finest scale checkerboard (∼2km, Figure4c) is much less well-defined.
These tests indicate the limits of resolution of our data set—they demonstrate that VP/VS anomalies can be
robustly detected above depths of 10 km within the seismic array and on scale lengths greater than ∼2 km.
Figure 3. Input VP/VS checkerboard models of differing scale lengths: (a) ∼8km; (b) ∼4km; (c) ∼2km. Slices are taken through the maximum perturbations of each
checkerboard in latitude, longitude and depth, as indicated by the white dashed lines. Seismic stations are plotted as yellow triangles.
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
10 of 21
Similar resolution is observed in the results of synthetic inversions for VP and
VS structure (see Figures S5 and S7 in Supporting InformationS1).
The lack of recovery of anomalies at depth can be explained by 99% of the
seismic events in our data set occurring between the surface and ∼10km
depth b.s.l.; thus, there are very few paths available to resolve structure below
this depth. The amplitude recovery of the input perturbations varies–in places
the amplitude of anomalies is underestimated, whereas it is overestimated in
certain regions (e.g., the negative anomaly in the longitude cross-section of
Figure4a). Outside the seismic array, the amplitude recovery is particularly
weak, as expected.
4.2. Inverting for 3D Velocity Structure
4.2.1. Inversion Statistics
The inversion statistics in Table1 show that the RMS arrival time residu-
als and the data variance are reduced for the final P-wave, S-wave and VP/
VS solution models as compared to the starting models. The normalized χ
2
value is the result of a statistical test for how well a model compares to actual
observed data. In theory, it should be equal to 1 if all the data are satisfied to
Figure 4. Output VP/VS checkerboard models of differing scale lengths: (a) ∼8km; (b) ∼4km; (c) ∼2km. Slices are taken through the maximum perturbations of each
checkerboard in latitude, longitude and depth, as indicated by the white dashed lines. Seismic stations are plotted as yellow triangles.
Table 1
Inversion Statistics for the Final VP, VS and VP/VS Inversions
Parameter Starting model Solution model Reduction (%)
RMS residuals
VP0.166s 0.104s 37.3
VS0.182s 0.130s 28.6
VP/VS0.192 0.168 12.5
Variance
VP0.0276s
20.0110s
260.1
VS0.0332s
20.0170s
248.8
VP/VS0.0364 0.0283 22.3
χ
2
VP4.59 1.39 69.7
VS9.60 4.66 51.5
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
11 of 21
the level of the noise. For the VP solution model, it is reduced to χ
2≈1, but for VS it is only reduced to χ
2=4.66.
However, as detailed by Rawlinson etal.(2010), tomographic inversions usually do not have χ
2=1 due to (a)
estimation of data uncertainties being difficult to quantify; (b) the use of a regular and smooth model parameter-
ization; (c) application of ad hoc regularization to stabilize the inversion, which suppresses some structures that
are needed to satisfy the data; (d) the assumptions and approximations made when solving the forward problem.
Thus the range of models that can be retrieved is limited. Despite this, the final data fit is a significant improve-
ment on the starting 1D models for all the solution models, indicating that recovered lateral heterogeneities are
generally required by the data and hence are physically meaningful within the limits of data resolution, as esti-
mated from the synthetic tests.
4.2.2. Velocity Structure
Figure5 shows east-west and north-south cross-sections through the final VP, VS, and VP/VS solution models. The
cross-sections are taken through the center of the caldera, passing directly through the vent location of the 2011
eruption (J. E. Hamlyn etal.,2014). For clarity, we only plot the anomalies and seismicity below the topography
line. The original figures can be seen in Figure S8 in Supporting Information S1. Depth slices are plotted in
Figure6.
The dominant feature in the VS solution model (Figures5c and 5d, Figures6e–6h) is a region of high VS and
high levels of seismicity within the caldera outline, extending downwards from the surface to 6km b.s.l.. At
6kmb.s.l., an aseismic region of low VS is observed, extending downwards to ∼10kmb.s.l.. Close to the surface,
the high VS region is surrounded to the north and south by very low VS areas, and to the east and west by less
pronounced low VS areas, all of which are aseismic.
The VP model shows more heterogeneity compared to the VS model. The difference is particularly notice-
able in the perpendicular east-west and north-south cross-sections taken through the center of the caldera
(Figures5a and5b). A low VP structure extending from 41.7°–41.8°E dips from east to west from a depth of
4kmb.s.l. to 10 kmb.s.l.. Above this, there is a region of high VP. This region contains two low VP anom-
alies extending ∼2–3km in depth and ∼0.2° in longitude, which is around the limit of data resolution. In
the depth slice at 1kmb.s.l., a low VP region extends from 13.44°–13.48°N, aligned N-S (Figure6a). This
region extends in depth down to 10kmb.s.l., as observed in the longitude cross-section (Figure5a), and is
seismically active.
Following the joint inversion for VP and VS structure, we calculate the average ratio of the VP and VS models across
all velocity grid nodes, which is 1.77. We use this reference value to adjust our color scale in the VP/VS plots in
Figures5 and6; red colors correspond to ratios higher than the reference and blue colors to ratios lower than the
reference.
Figure 5. Cross-sections through the final VP (a and d), VS (b and e) and VP/VS (c and f) solution models at a longitude of 41.7°E and latitude of 13.36°N (the center of
the caldera). The VP and VS models are plotted as percentage deviations from the initial 1D model. The VP/VS ratio model is plotted as absolute values, with the center
of the color bar corresponding to the reference VP/VS value–the average ratio of the VP and VS models across all velocity grid nodes. Earthquakes within ±1km of the
displayed section are indicated by black dots. The yellow stars mark the vent location of the 2011 eruption. As discussed in the text, we only show the earthquakes and
anomalies below the topography line for the purposes of clarity, with the full solution plotted in Figure S8 in Supporting InformationS1.
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
12 of 21
Figure 6.
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
13 of 21
At depths of 6–10km b.s.l., a region of high VP/VS ratio (as high as 1.9) is observed in the longitude and latitude
cross-sections (Figures5e and 5f). This region correlates with low VS and low VP anomalies, and is aseismic.
Above this high VP/VS region, there is an area of high seismicity and very low VP/VS ratio (as low as 1.5), extend-
ing from depth 0kmb.s.l. down to 6kmb.s.l. and lying within the caldera outline (Figures6i–6l). The strongest
low VP/VS ratios are seen between 0 and 2kmb.s.l. and correspond to high VS values. Close to the surface, high
VP/VS ratios are observed again, with the ratio reaching 2.0 in places. These high VP/VS regions all exhibit low
levels of seismicity.
As described in Section3, we obtain the VP/VS solution model by directing inverting S-P differential traveltimes.
We also plot the solution model obtained from simply dividing the VP solution model by the VS solution model
(Figure S9 in Supporting InformationS1). The direct division model shows largely the same pattern of velocity
anomalies within the data resolution limits determined by our synthetic tests (Section4). Differences in amplitude
are seen, but this is expected due to solution non-uniqueness and the imposition of regularization constraints.
Thus, we are confident that our method produces a model that is consistent with direct division, and that our
subsequent interpretations of these VP/VS ratio anomalies would remain the same if we had chosen to directly
divide the VP solution model by the VS solution model instead.
5. Discussion
5.1. Defining a Local 1D Velocity Model at Nabro
Compared to the three-layer velocity model developed for the Afar region in previous studies, our refined
local 1D velocity model has faster P- and S-wave velocities in the uppermost 10km. High crustal velocities
in a volcanic setting are typically attributed to the presence of solidified, high-strength intrusive magmatic
rocks, such as the cumulates and dykes at Mount Etna (Aloisi etal., 2002), a plutonic body at Mount St.
Helens (Lees,1992), an old lateral dyke system at Tungurahua volcano (Molina etal., 2005) and the solid
andesitic cores of the volcanic complexes of Soufriére and Centre Hills, Montserrat (Paulatto etal., 2010;
Shalev etal., 2010). At Nabro, analysis of inclusions in erupted products from 2011 suggests that these are
derived from older and more primitive basalt (Donovan etal.,2018). The presence of xenocryst material in
the erupted magmas leads Donovan etal.(2018) to conclude that the subsurface crustal structure beneath the
caldera is composed of a series of sills and older eruptive products. This provides supporting evidence that the
elevated crustal velocities directly beneath Nabro reflect intrusions, potentially remnants of earlier episodes of
magmatism.
Below 10km, the refined model shows negligible variation from the regional model, as expected due to the fact
that the vast majority of seismic events originate above 10kmb.s.l.. Using this refined local 1D model as the
starting model for tomographic inversions results in solution models that better fit the traveltime data.
5.2. Earthquake Detection Using U-GPD
This study presents the first seismic tomography results from an earthquake catalog detected using machine
learning methods. The inversion statistics from this study (RMS residuals, variance and χ
2) are of the same order
of magnitude as those from previous FMTOMO studies that relied on manually picked catalogs of earthquakes
(e.g., Pilia etal.,2013; Wilks etal.,2020), demonstrating that our use of the deep learning model U-GPD to pick
the seismic arrivals has not adversely affected the stability and robustness of the inversion. The method therefore
has the ability to detect seismic events with sufficient accuracy to be used successfully in an FMTOMO tomo-
graphic inversion. This has implications for future tomographic studies at volcanoes: the efficiency of U-GPD's
phase arrival picking method means that far more events can be detected than previously possible, enabling more
timely exploitation of such data for the purposes of seismic tomography.
Figure 6. Cross sections at depths of 1, 3, 5, and 7km below sea level (b.s.l.) through the final (a–d) VP (e–h) VS, and (i–l) VP/VS ratio velocity models. The VP and VS
models are plotted as percentage deviations from the initial 1D model. The VP/VS ratio model is plotted as absolute values, with the center of the color bar corresponding
to the reference VP/VS value–the average ratio of the VP and VS models across all velocity grid nodes. Earthquakes within ±0.5km of the displayed section are plotted as
black dots and seismic stations are plotted as yellow inverted triangles. The orange stars mark the vent location of the 2011 eruption. The gray dashed lines represent the
latitude and longitude cross-sections depicted in Figure5.
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
14 of 21
5.3. Interpretation of VP/VS Variations
Seismic velocity variations reflect a variety of physical parameters, including rock characteristics (composition,
porosity, fractures, mineralogy), saturation conditions, presence of fluids (gases or liquids), temperature, and
pressure. The interplay of these diverse influences makes it difficult to interpret observed seismic anomalies.
Considering the ratio of compressional velocity to shear velocity, VP/VS, enables greater constraint to be placed
on the cause of seismic velocity variations. This is particularly helpful when attempting to constrain the loca-
tion of fluids in the subsurface of a volcano, because the VP/VS ratio is sensitive to the type of fluid present and
can distinguish between regions of partial melt or hydrothermal fluids, both of which are encountered beneath
volcanoes. In saturated or partially saturated rocks, the content and physical state of fluids has a greater effect on
P-wave velocities than S-wave velocities (Vanorio et al.,2005). Fluid phase transitions induce changes in fluid
compressibility and thus bulk modulus (Ito etal.,1979; Wang & Nur,1986). Shear moduli are little affected
by fluid phase transitions and hence S-wave velocities change insignificantly due to a density effect, meaning
that low VP/VS ratios tend to characterize gas-bearing rocks (i.e., those with high fluid compressibility) whereas
liquid-bearing rocks (with low fluid compressibility) are characterized by high VP/VS ratios (Vanorio etal.,2005).
5.3.1. High VP/VS Anomaly at 6–10km Depth
In Figure5, the high VP/VS (>1.9) region at 6–10kmb.s.l. coincides with pronounced low VS (<3.8kms
−1)and
low VP (∼ 6.6kms
−1) anomalies as compared to the starting model. This correlation between the P- and S-wave
velocity models in a region of high VP/VS ratio suggests a region of elevated temperature (Sanders etal.,1995).
Furthermore, the anomalous region is approximately aseismic. In an examination of a deep cluster of seismicity
at Nabro, Lapins(2021a) finds that the highest attenuation is observed at station NAB1 where the raypaths travel
directly through this aseismic, high VP/VS anomaly, suggesting that the region attenuates S-waves strongly. Anom-
alously low QS at depth is usually attributed to the presence of partial melt (Sanders etal., 1995). The region
also coincides with the location of a Mogi source inferred by J. E. Hamlyn etal.(2014) to explain the observed
post-eruptive surface subsidence at Nabro. Petrological analysis by Donovan etal. (2018) finds that most melt
inclusions in erupted products from the 2011 eruption were entrapped at 5–10km depth b.s.l.. This represents the
storage location of an older body of melt which was remobilized and erupted when an intrusion of fresh melt rose
through the crust and mingled with the older melt (Donovan etal.,2018). The estimated depth of this melt body
is consistent with the depth of our observed high VP/VS anomaly.
Previous tomographic studies at volcanoes have interpreted similar regions with high VP/VS ratios as delineating
magmatic storage zones. For example, an anomalous body with low P-wave velocity, low S-wave velocity and
VP/VS ratio >1.84 is observed by Lin etal.(2014) at 8–11km depth beneath the Klauea volcano in Hawaii, and
interpreted as a crustal magma reservoir beneath the volcanic pile. At Nevado del Ruiz volcano in Colombia,
a region of high VP/VS ratios (>1.80) at 2–10km is inferred to be an intrusive body of magmatic origin that
included partial melt zones associated with low S-wave velocity anomalies (Londoño & Sudo,2003). Greenfield
etal.(2016) observe a pair of prominent anomalies with low P- and S-wave velocities and VP/VS ratios >1.82 at
depths of 5 and 9kmb.s.l. beneath Askja volcano, Iceland, which are interpreted as the primary magma storage
regions in the upper crust.
Thus, we interpret the high VP/VS anomaly as the storage location of melt that fed the 2011 eruption. Only a
small fraction (5%–20%) of the stored melt is typically erupted at the surface (Greenfield & White,2015; White
etal.,2019). This is seen in the results of a post-eruption seismic velocity study at Mount St. Helens, which finds
that there is a persistent high VP/VS region at 4–13km b.s.l., interpreted as the primary upper-middle crustal
magma reservoir and indicating that a significant amount of melt remains in the crust (Kiser etal.,2016). Simi-
larly, our observation of a high VP/VS region at depths 6–10kmb.s.l. suggests that melt is still stored at these
depths, which could feed future eruptions.
5.3.2. Low VP/VS Region
Above the inferred magma storage region, our results show a region of low VP/VS ratio (1.5–1.7), colocated with
high S-wave velocities (>3.6kms
−1) and low P-wave velocities compared to the starting model (region B in
Figure7). This anomaly extends from depths of ∼5kmb.s.l. to the surface. The high VS values can be explained
by lithology—high-strength, solidified intrusive magmatic rocks are expected to show high seismic velocities
(e.g., Aloisi etal., 2002; Lees, 1992; Lees, 2007; Molina etal., 2005). However, the VP/VS ratio in intrusive
igneous rocks is expected to be higher than typical continental crust (Christensen,1996), whereas the VP/VS ratio
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
15 of 21
we measure in this region is low—down to 1.5 in places—and so there must be another factor acting to reduce
the VP/VS ratio. It has been shown that the velocities of P- and S-waves in rocks are strongly affected by the satu-
ration conditions of the rock, particularly whether the rock is saturated with gas, liquid or a mixture thereof (Ito
etal.,1979; Toksöz etal.,1976). A geothermal regime near the water-steam transition has low P-wave velocities
but normal S-wave velocities: the presence of gas reduces the bulk modulus and causes a decrease in P-wave
velocity, without significantly altering the propagation of shear waves (Walck, 1988). The addition of a small
amount of gas in a water-brine mixture can lower the velocity of P-waves significantly (Toksöz etal.,1976). We
observe low P-wave velocities coincident with the lowest VP/VS ratio in our model, providing evidence for the
presence of gas in the rocks in this region. Furthermore, calculations of seismic attenuation in P- and S-waves
at Nabro show that P-wave attenuation is significantly higher than S-wave attenuation across all seismic stations
(Lapins,2021a), which is generally attributed to partial saturation of a compressible fluid in cracks, fractures or
pores (Amalokwu etal.,2014; Hauksson & Shearer,2006; Winkler & Nur,1979). Further evidence in support of
the existence of gases in the upper subsurface is the significant post-eruption fumarolic activity observed at Nabro
(Yohannes,2012). Petrological analysis of melt inclusions erupted in 2011 indicates that melt-fluid separation
occurred at depths of up to 18kmb.s.l., generating CO2 rich fluids (Donovan etal.,2018). The magma storage
zone described in Section5.3.1 coincides with a Mogi source; the deformation model invokes deflation, which is
explained by the outgassing of magma at depth (J. Hamlyn etal.,2018). The low VP/VS region could reflect the
degassing pathways between the magma and the surface fumaroles.
A velocity reversal is seen in the 1D VS model between 3 and 4kmb.s.l., where velocity values decrease with
depth (Figure2). The trend in VP does not reverse here, but the rate of change of velocity with depth decreases.
The location of these reversals is coincident with the lowest VP/VS ratio seen in the model. Density, resistivity and
sonic velocity logs that go through velocity reversals are generally interpreted as departures of the effective stress
from normal compaction trends (Bowers,2002; Hottmann & Johnson, 1965; Pikington, 1988). Overpressure
is one explanation for this. If the fluid pressure is higher than the normal hydrostatic fluid gradient for a given
depth, it prevents the effective stress from increasing with depth as it usually would (Vanorio etal.,2005). To be
over-pressurized, the gas present in the low VP/VS region would have to be trapped and experiencing expansion,
uplift, compaction, temperature increase or a combination of these factors; all of which are possible in the context
Figure 7. The VP/VS (left image), VP (top right image) and VS (bottom right image) solution models plotted along the 41.7°E cross-section beneath Nabro. To aid
visualization, contours are added to the VP and VS models for velocities greater than 4.8 and 3.0km/s, respectively. (a) The aseismic high VP/VS, low VP and low VS
magma storage region is described in Section5.3.1. (b) The very low VP/VS and high VS region is described in Section5.3.2; we interpret this as a zone of intrusive
rock, potentially in a stacked sill structure. The region hosts fractured, partially saturated rocks with low VP values. The yellow star marks the vent location of the 2011
eruption. Earthquakes located within ±1km of the cross-section are plotted as gray circles. The yellow dashed arrow indicates a potential degassing pathway from the
region of partial melt to the surface, following the densest clusters of seismicity.
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
16 of 21
of a recently active volcano. Indeed, the overpressure could be driven by the magma storage zone directly below
the low VP/VS region that extends from depths of ∼5kmb.s.l. to the surface (J. Hamlyn etal.,2018).
Nabro's caldera outline matches the region of high VS and low VP/VS well (Figure6), while the low VP region is
slightly smaller. Therefore, we propose that the crustal structure within the caldera is formed of intrusive rock,
potentially with layers of stacked sills from previous eruptions. This would elevate the S-wave velocity.
During the June 2011 eruption of Nabro, magma ascended to the surface via a NW-SE-oriented dyke (Goitom
etal.,2015). Volcanic conduits have associated damage zones, common to all shallow magmatic systems beneath
volcanoes (Afanasyev etal.,2018), and contain fragmental infills related to prior intrusions, eruptions and steam
explosions (Blundy etal.,2021). These conduits therefore have high porosities, within which fluids can be stored.
Dyke emplacement also causes fracturing and creates permeable pathways for fluid transport and accumulation
(e.g., Bakker etal.,2016; Brown etal.,2007). Our results show that this region is highly seismogenic, pointing
to the presence of fractures and cracks enabling fluid migration that drives pore pressure increases and leads
to abundant seismicity. Thus, the most likely route from the degassing magma storage region to the surface is
along the conduit that fed the 2011 eruption, as it will be formed from highly permeable and damaged rock. This
explains why the low VP region (indicating the presence of gas) is less horizontally extensive than the high VS
region.
Similar conclusions are reached at Aluto volcano, where a region of low VP/VS ratio is interpreted as the signa-
ture of an over-pressurized gas volume within a hydrothermal system (Wilks etal., 2017, 2020). A seismo-
genic zone of low VP/VS ratios coincident with low P-wave velocities is also observed at Campi Flegrei, and
explained as over-pressurized gases accumulating at the top of dyke intrusions (Chiarabba & Moretti, 2006;
Vanorio etal., 2005). At Nevado del Ruiz volcano, the upper part of a high P- and S-wave velocity anomaly
(0–2km depth) is characterized by low VP/VS ratios (<1.68) and described as a steam-dominated geothermal
system by Londoño and Sudo(2003). Multiple studies of Mammoth Mountain, California, have identified a low
VP/VS region from depths of -3–2km, attributed to the presence of CO2 distributed in oblate spheroid pores, which
supplies gas-rich thermal springs at the surface (Dawson etal., 2016; Foulger etal.,2003; Julian et al., 1998;
Lin,2013).
Our observations have shown that the degassing of partial melt influences the upper crustal substructure beneath
Nabro. The coupling of the shallow heat source to volatile transport above, as well as the presence of fumaroles at
the surface, suggests the possibility of a high-temperature geothermal system similar to those hosted by volcanic
s in the Main Ethiopian Rift (Pürschel etal.,2013).
5.3.3. Shallow, High VP/VS Anomalies
In volcanic settings, high VP/VS values at depth are typically associated with the presence of melt. However, the
high temperatures and pressures necessary for sustaining partial melt post-eruption are unlikely to prevail very
close to the surface (depths 0–5km below surface), and so other explanations have been proposed: for example,
steam condensates that manifest at shallow depths off the main volcanic edifice, where temperatures are reduced
(Aster & Meyer,1988; Chiarabba & Moretti,2006; Vanorio etal.,2005; Wilks etal.,2020). These condensates
may form brines that migrate toward the surface along fracture networks, explaining extensive fumarolic activ-
ity at the surface (Hudson etal.,2022; MacQueen etal.,2021). Alternatively, high VP/VS anomalies have been
attributed to the penetration of meteoric water into the volcanic cone through fractures (Bushenkova etal.,2019;
Koulakov etal.,2021).
The shallowest part of our model shows high VP/VS anomalies (>1.9) to the north and south of the caldera that
coincide with low VP and low VS anomalies. These anomalies are in the very shallow subsurface, meaning that
they are constrained by only one seismic station. Therefore, in the absence of further geophysical constraints,
such as magneto-telluric surveys, fluid sampling and analysis or well-log data, it is possible that these anomalies
are artifacts.
5.4. The Crustal Substructure Beneath Nabro Caldera
Nabro is an active volcano that experienced subsidence and seismicity following its 2011 eruption. Figure7
provides an overview of the results and interpretations of the seismic anomalies identified in this study, as
summarized below.
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
17 of 21
An aseismic region of low VP, low VS and high VP/VS ratio at depths of 6–10km (region A in Figure7) likely
represents a storage region of partially molten material, and is consistent with the results of previous geodetic,
petrological and seismic studies. It is likely that partial melt has remained stored here post-eruption, as only a
small fraction of the total volume of melt stored in a reservoir is generally erupted at the surface (Greenfield &
White,2015; White etal.,2019). The cause of surface subsidence is likely degassing of volatiles from this magma
storage region (Donovan etal.,2018; J. E. Hamlyn etal.,2014).
Above the zone of partial melt, we observe a region of abundant seismicity and high VS (region B in Figure7),
which we interpret as a zone of intrusive rocks from previous eruptions. These could exist in a stacked sill
structure, the fine details of which we are unable to resolve with our tomographic model. This region contains
the conduit that fed the 2011 eruption, which is formed of fractured, cracked rocks partially saturated with
over-pressurized gases (a mixture of CO2 and H2O), leading to low VP and very low VP/VS ratio. The magma
storage region below is a likely cause of the overpressure in these gases. The influence of this shallow heat source
on the outflow of gases through the subsurface means that Nabro is likely to have geothermal potential that may
be exploited as an energy resource.
Previous studies are in agreement with our interpretations of the velocity structure at Nabro. By inverting the
deformation field from satellite InSAR images, J. Hamlyn etal.(2018) explain the subsidence as the deflation of
a Mogi source located at 6.4±0.3km depth b.s.l.. This coincides with Donovan etal.(2018)'s findings that most
melt inclusions in erupted lava were trapped at 5–10km depth b.s.l.. Their petrological study of erupted products
from the 2011 eruption concludes that distinct batches of magma were stored in sills and mixed together prior
to eruption (Donovan etal., 2018). A study of the post-eruption seismicity also identifies an aseismic magma
storage zone at depths of 6–9kmb.s.l (Lapins,2021a). Numerous fumaroles have been observed at Nabro after
the eruption, which has led to its identification as a region of geothermal interest (Yohannes,2012).
5.5. Comparisons With Other Volcanoes
As an off-rift caldera in the under-studied East African Rift System that erupted months prior to the seismic
deployment, there are no previous seismic tomography studies that allow for direct comparison to Nabro.
However, it is still instructive to examine a few examples of volcanoes that share certain similar features with it.
Koryaksky volcano in Kamchatka erupted months before the seismic events used in the tomographic study of
Bushenkova etal.(2019) were recorded. Despite the different tectonic setting, the tomographic images show a
similar structure to Nabro. At depth, a high VP/VS anomaly represents a magma storage region. Above this, there
is a low VP/VS anomaly associated with a vertical seismicity cluster, marking the pathway of fluid ascent. Another
actively erupting volcano, Klauea in Hawaii, shows elevated VP and VS at depth, interpreted as representing the
high-velocity cumulates of the volcanic core (Lin etal.,2014). An anomalous body of low VP, low VS and high
VP/VS at 8–11km depth is explained as a crustal magma reservoir. Both of these features are also observed at
Nabro. The Klauea images also show a region of low VP/VS above the magma reservoir, but this is not interpreted.
Aluto volcano is also located in the East African Rift System, situated in the Main Ethiopian Rift. The seis-
mic tomography study of Wilks etal.(2020) finds a large low velocity, high VP/VS zone at depths of 4–9km,
interpreted as a more ductile and melt-bearing region. Away from the volcano, there are shallow, localized high
VP/VS regions, representing steam condensates which may form brines that migrate to the surface. A hydrother-
mal system with very low VP/VS is observed at shallow depths, hosting gases exsolved from the deeper melt body.
These features are broadly similar to what is seen beneath Nabro. The main difference is that Nabro's low VP/
VS region extends to greater depths. Despite a recent increase in surface deformation, Aluto has been quiescent
for thousands of years. Therefore, the crustal substructure of Aluto could represent a ‘steady-state’ situation
for volcanoes in the region, from which Nabro has been disturbed due to its recent eruption and the ascent of
magmatic fluid from depth.
5.6. Limitations and Future Work
A fundamental limitation on our results is the resolution of the tomographic inversion, likely caused by the small
size of the seismic network deployed at Nabro. Jointly inverting for velocity structure and event relocation with
such a small seismic array is challenging. This is reflected in the large relocation offsets observed in particular for
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
18 of 21
events outside of the aperture of the seismic network, which subsequently are removed from the catalog. Indeed,
checkerboard sensitivity tests demonstrate that outside of the seismic array and deeper than ∼10kmb.s.l., we
cannot recover synthetic velocity perturbations. Recovery of velocity anomalies on scales of <2 km is also
limited. Therefore, we restrict our interpretations of the tomographic images to velocity anomalies occurring on
scales >4km, within the seismic network and in the uppermost 10km of the crust.
We are also limited in our interpretations by the lack of other observations at Nabro. Studies such as magneto-telluric
surveys, geochemical analyses of volcanic fluids and well-log data would all help to provide further constraints
on the interpretations presented here. The resolution of seismic velocity structure is poor at depths greater than
10kmb.s.l., due to the distribution of seismicity being mostly located at shallower depths. Thus it is difficult to
form broader conclusions about magmatic processes in the mid-lower crust.
Future work could involve attempts to probe the lower crust at Nabro, for example, through receiver function
analysis (e.g., Hammond,2014; Janiszewski etal.,2020) or an investigation of seismic anisotropy using shear-
wave splitting (e.g., Nowacki etal.,2018), in order to understand how off-rift magmatism at Nabro is sustained
and supplied. The application of U-GPD to seismic datasets from other volcanoes, particularly those in a
post-eruptive state, would also provide useful points of comparison to this study.
6. Conclusion
We use a seismic catalog created by a deep learning model for automating phase arrival detection to invert for the
earthquake locations and the 3D velocity structure beneath Nabro caldera, an off-rift volcano in the Afar region.
This has produced the first tomographic images of the volcano, which was unmonitored before its explosive
eruption in June 2011.
The main findings of the tomographic study are: (a) an aseismic region of low VP, low VS and high VP/VS at
depths of 6–10kmb.s.l., interpreted as the primary melt storage region that fed the 2011 eruption; (b) a region
of high seismicity, very low VP/VS ratio and low VP, representing a zone of partially saturated rocks containing
gases that are over-pressurized due to degassing from the magma storage zone directly below; (c) general high
VP and VS beneath the volcanic edifice, pointing to the existence of high-strength, solidified intrusive magmatic
rocks.
Our results have demonstrated that deep learning models are an efficient way to obtain earthquake catalogs for
the purposes of seismic tomography at volcanoes. Although our model cannot elucidate the origins of magma
supply to Nabro at depths exceeding 10kmb.s.l., it does illustrate that this off-rift volcano has a similar shallow
magmatic plumbing system to other hydrothermally active, restless volcanoes. The observations are consistent
with the existence of a melt storage region at 6–10kmb.s.l. beneath Nabro. We have also uncovered a region
that is high in volatile content, coupled to the degassing magmatic system, indicating that Nabro should be
considered a region of geothermal interest. Our study highlights the need for further geophysical studies at
Nabro.
Data Availability Statement
The raw seismic data used in this study are from the Nabro Urgency Array (Hammond etal., 2012), publicly
available through IRIS Data Services (http://service.iris.edu/fdsnws/dataselect/1/). Full code to reproduce the
U-GPD transfer learning model, perform model training, run the U-GPD model over continuous sections of
data and use model picks to locate events in NonLinLoc (Lomax etal., 2000) are available at https://github.
com/sachalapins/U-GPD, with the release (v1.0.0) associated with this study archived and available through
Zenodo (Lapins,2021b). The arrival time picks for the initial event catalog produced by the U-GPD model, as
well as the station metadata, are also archived in a Zenodo repository (Lapins,2022). The FMTOMO package
is freely available to download at http://rses.anu.edu.au/∼nick/fmtomo.html (Rawlinson & Sambridge,2004a).
Files containing the final VP, VS, and VP/VS models, and the relocated event catalog are available through Zenodo
(Gauntlett etal.,2023).
Figures and maps were plotted using Generic Mapping Tools (GMT) version 6 (Wessel et al., 2019) licensed
under LGPL version 3 or later, available at https://www.genericmapping-tools.org.
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
19 of 21
References
Afanasyev, A., Blundy, J., Melnik, O., & Sparks, S. (2018). Formation of magmatic brine lenses via focused fluid-flow beneath volcanoes. Earth
and Planetary Science Letters, 486, 119–128. https://doi.org/10.1016/j.epsl.2018.01.013
Aloisi, M., Cocina, O., Neri, G., Orecchio, B., & Privitera, E. (2002). Seismic tomography of the crust underneath the Etna volcano, Sicily. Phys-
ics of the Earth and Planetary Interiors, 134(3–4), 139–155. https://doi.org/10.1016/s0031-9201(02)00153-x
Amalokwu, K., Best, A. I., Sothcott, J., Chapman, M., Minshull, T., & Li, X.-Y. (2014). Water saturation effects on elastic wave attenuation in
porous rocks with aligned fractures. Geophysical Journal International, 197(2), 943–947. https://doi.org/10.1093/gji/ggu076
Aster, R., & Meyer, R. (1988). Three-dimensional velocity structure and hypocenter distribution in the Campi Flegrei caldera, Italy. Tectonophys-
ics, 149(3–4), 195–218. https://doi.org/10.1016/0040-1951(88)90173-4
Bakker, R. R., Fazio, M., Benson, P.M., Hess, K.-U., & Dingwell, D. B. (2016). The propagation and seismicity of dyke injection, new experi-
mental evidence. Geophysical Research Letters, 43(5), 1876–1883. https://doi.org/10.1002/2015gl066852
Barberi, F., Bonatti, E., Marinelli, G., & Varet, J. (1974). Transverse tectonics during the split of a continent: Data from the Afar rift. Tectono-
physics, 23(1–2), 17–29. https://doi.org/10.1016/0040-1951(74)90108-5
Blundy, J., Afanasyev, A., Tattitch, B., Sparks, S., Melnik, O., Utkin, I., & Rust, A. (2021). The economic potential of metalliferous sub-volcanic
brines. Royal Society Open Science, 8(6), 202192. https://doi.org/10.1098/rsos.202192
Bowers, G. L. (2002). Detecting high overpressure. The Leading Edge, 21(2), 174–177. https://doi.org/10.1190/1.1452608
Brikke, N. E. A. (2010). 3D seismic traveltime tomography of the central South Island.
Brown, R., Kavanagh, J., Sparks, R., Tait, M., & Field, M. (2007). Mechanically disrupted and chemically weakened zones in segmented dike
systems cause vent localization: Evidence from kimberlite volcanic systems. Geology, 35(9), 815–818. https://doi.org/10.1130/g23670a.1
Bushenkova, N., Koulakov, I., Senyukov, S., Gordeev, E. I., Huang, H.-H., El Khrepy, S., & Al Arifi, N. (2019). Tomographic images of magma
chambers beneath the Avacha and Koryaksky volcanoes in Kamchatka. Journal of Geophysical Research: Solid Earth, 124(9), 9694–9713.
https://doi.org/10.1029/2019jb017952
Chiarabba, C., & Moretti, M. (2006). An insight into the unrest phenomena at the Campi Flegrei caldera from Vp and Vp/Vs tomography. Terra
Nova, 18(6), 373–379. https://doi.org/10.1111/j.1365-3121.2006.00701.x
Chouet, B. A., & Matoza, R. S. (2013). A multi-decadal view of seismic methods for detecting precursors of magma movement and eruption.
Journal of Volcanology and Geothermal Research, 252, 108–175. https://doi.org/10.1016/j.jvolgeores.2012.11.013
Christensen, N. I. (1996). Poisson’s ratio and crustal seismology. Journal of Geophysical Research, 101(B2), 3139–3156. https://doi.
org/10.1029/95jb03446
Dawson, P., Chouet, B., & Pitt, A. (2016). Tomographic image of a seismically active volcano: Mammoth Mountain, California. Journal of
Geophysical Research: Solid Earth, 121(1), 114–133. https://doi.org/10.1002/2015jb012537
de Kool, M., Rawlinson, N., & Sambridge, M. (2006). A practical grid-based method for tracking multiple refraction and reflection phases in
three-dimensional heterogeneous media. Geophysical Journal International, 167(1), 253–270. https://doi.org/10.1111/j.1365-246x.2006.03078.x
Donovan, A., Blundy, J., Oppenheimer, C., & Buisman, I. (2018). The 2011 eruption of Nabro volcano, Eritrea: Perspectives on magmatic
processes from melt inclusions. Contributions to Mineralogy and Petrology, 173(1), 1–23. https://doi.org/10.1007/s00410-017-1425-2
Eberhart-Phillips, D., & Reyners, M. (2012). Imaging the Hikurangi Plate interface region, with improved local-earthquake tomography.
Geophysical Journal International, 190(2), 1221–1242. https://doi.org/10.1111/j.1365-246x.2012.05553.x
Evans, J. R., Eberhart-Phillips, D., & Thurber, C. (1994). User’s manual for SIMULPS12 for imaging Vp and Vp/Vs; a derivative of the” thurber”
tomographic inversion SIMUL3 for local earthquakes and explosions (Tech. Rep.). US Geological Survey.
Foulger, G. R., Julian, B., Pitt, A., Hill, D., Malin, P., & Shalev, E. (2003). Three-dimensional crustal structure of Long Valley caldera,
California, and evidence for the migration of CO2 under Mammoth Mountain. Journal of Geophysical Research, 108(B3). https://doi.
org/10.1029/2000jb000041
Fromm, M., Kablick, G., III, Nedoluha, G., Carboni, E., Grainger, R., Campbell, J., & Lewis, J. (2014). Correcting the record of volcanic
stratospheric aerosol impact: Nabro and Sarychev Peak. Journal of Geophysical Research: Atmospheres, 119(17), 10–343. https://doi.
org/10.1002/2014jd021507
Gauntlett, M., Hudson, T., Kendall, J.-M., Rawlinson, N., Blundy, J., Lapins, S., etal. (2023). Nabro 3D velocity model produced by the FMTOMO
code. Zenodo. https://doi.org/10.5281/zenodo.7701958
Ginzburg, A., Makris, J., Fuchs, K., & Prodehl, C. (1981). The structure of the crust and upper mantle in the Dead Sea rift. Tectonophysics,
80(1–4), 109–119. https://doi.org/10.1016/0040-1951(81)90144-x
Glahn, A., Granet, M., & Group, R. G. T. (1993). Southern Rhine Graben: Small-wavelength tomographic study and implications for the dynamic
evolution of the Graben. Geophysical Journal International, 113(2), 399–418. https://doi.org/10.1111/j.1365-246x.1993.tb00896.x
Goitom, B. (2017). The Nabro volcano, tectonic framework and seismic hazard assessment of Eritrea (Unpublished doctoral dissertation).
University of Bristol.
Goitom, B., Oppenheimer, C., Hammond, J. O., Grandin, R., Barnie, T., Donovan, A., etal. (2015). First recorded eruption of Nabro volcano,
Eritrea, 2011. Bulletin of Volcanology, 77(10), 1–21. https://doi.org/10.1007/s00445-015-0966-3
Greenfield, T., & White, R. S. (2015). Building Icelandic igneous crust by repeated melt injections. Journal of Geophysical Research: Solid
Earth, 120(11), 7771–7788. https://doi.org/10.1002/2015jb012009
Greenfield, T., White, R. S., & Roecker, S. (2016). The magmatic plumbing system of the Askja central volcano, Iceland, as imaged by seismic
tomography. Journal of Geophysical Research: Solid Earth, 121(10), 7211–7229. https://doi.org/10.1002/2016jb013163
Hamlyn, J., Wright, T., Walters, R., Pagli, C., Sansosti, E., Casu, F., etal. (2018). What causes subsidence following the 2011 eruption at Nabro
(Eritrea)? Progress in Earth and Planetary Science, 5(1), 1–15. https://doi.org/10.1186/s40645-018-0186-5
Hamlyn, J. E., Keir, D., Wright, T. J., Neuberg, J. W., Goitom, B., Hammond, J. O., etal. (2014). Seismicity and subsidence following the 2011
Nabro eruption, Eritrea: Insights into the plumbing system of an off-rift volcano. Journal of Geophysical Research: Solid Earth, 119(11),
8267–8282. https://doi.org/10.1002/2014jb011395
Hammond, J. (2014). Constraining melt geometries beneath the Afar Depression, Ethiopia from teleseismic receiver functions: The anisotropic
H-κ stacking technique. Geochemistry, Geophysics, Geosystems, 15(4), 1316–1332. https://doi.org/10.1002/2013gc005186
Hammond, J., Goitom, B., Kendall, J. M., & Ogubazghi, G. (2012). Nabro urgency array. International Federation of Digital Seismograph
Networks. https://doi.org/10.7914/SN/4H_2011
Hammond, J., Kendall, J.-M., Stuart, G., Keir, D., Ebinger, C., Ayele, A., & Belachew, M. (2011). The nature of the crust beneath the Afar triple
junction: Evidence from receiver functions. Geochemistry, Geophysics, Geosystems, 12(12). https://doi.org/10.1029/2011gc003738
Hammond, J., Kendall, M., Stuart, G., Ebinger, C., Bastow, I. D., Keir, D., etal. (2013). Mantle upwelling and initiation of rift segmentation
beneath the Afar Depression. Geology, 41(6), 635–638. https://doi.org/10.1130/g33925.1
Acknowledgments
The authors thank Michael Bostock and
Greg Waite for their helpful editorial
comments. We are also grateful for the
comments from Ivan Koulakov and
an anonymous reviewer, which helped
improve this paper. The seismic data were
collected with funding from the Natural
Environment Research Council (NERC)
project NE/J012297/1 (“Mechanisms
and implications of the 2011 eruption of
Nabro volcano, Eritrea”). The UK seismic
instruments and data management facil-
ities were provided under loan number
976 by SEIS-UK at the University of
Leicester. The facilities of SEIS-UK are
supported by NERC under Agreement
R8/H10/64. Author MG was supported
by a Doctoral Training Partnership
studentship from NERC [NE/S007474/1].
Author SL was supported by a GW4+
Doctoral Training Partnership studentship
from the Natural Environment Research
Council (NERC) [NE/L002434/1].
Author BG was funded by the Engi-
neering and Physical Sciences Research
Council (EPSRC) and the School of Earth
Sciences at the University of Bristol. We
gratefully acknowledge the cooperation
we received from the Eritrea Institute
of Technology, Eritrean government,
Southern and Northern Red Sea Admin-
istrations, local sub-zones and village
administrations. We thank the Department
of Mines, Ministry of Energy and Mines
for their continued support throughout the
project. Special thanks go to Zerai Berhe,
Mebrahtu Fisseha, Michael Eyob, Ahmed
Mohammed, Kibrom Nerayo, Asresehey
Ogbatsien, Andemichael Solomon, and
Isaac Tuum. We thank Alem Kibreab for
vital help in facilitating the fieldwork. We
would also like to acknowledge helpful
comments from Tarje Nissen-Meyer and
Lara Wagner on the initial results of this
study. The tomographic inversions were
carried out using the University of Oxford
Advanced Research Computing service;
we thank Andrew Walker for his assis-
tance with this. IRIS Data Services are
funded through the Seismological Facil-
ities for the Advancement of Geoscience
(SAGE) Award of the National Science
Foundation under Cooperative Support
Agreement EAR-1851048.
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
20 of 21
Hauksson, E., & Shearer, P.M. (2006). Attenuation models (qp and qs) in three dimensions of the southern California crust: Inferred fluid satura-
tion at seismogenic depths. Journal of Geophysical Research, 111(B5). https://doi.org/10.1029/2005jb003947
Hearn, T. M., & Clayton, R. W. (1986). Lateral velocity variations in southern California. I. Results for the upper crust from Pg waves. Bulletin
of the Seismological Society of America, 76(2), 495–509.
Hottmann, C., & Johnson, R. (1965). Estimation of formation pressures from log-derived shale properties. Journal of Petroleum Technology,
17(06), 717–722. https://doi.org/10.2118/1110-pa
Hudson, T. S., Kendall, J.-M., Pritchard, M. E., Blundy, J. D., & Gottsmann, J. H. (2022). From slab to surface: Earthquake evidence for fluid
migration at Uturuncu volcano, Bolivia. Earth and Planetary Science Letters, 577, 117268. https://doi.org/10.1016/j.epsl.2021.117268
Ito, H., DeVilbiss, J., & Nur, A. (1979). Compressional and shear waves in saturated rock during water-steam transition. Journal of Geophysical
Research, 84(B9), 4731–4735. https://doi.org/10.1029/jb084ib09p04731
Janiszewski, H. A., Wagner, L. S., & Roman, D. C. (2020). Aseismic mid-crustal magma reservoir at cleveland volcano imaged through novel
receiver function analyses. Scientific Reports, 10(1), 1–9. https://doi.org/10.1038/s41598-020-58589-0
Julian, B. R., Pitt, A., & Foulger, G. (1998). Seismic image of a CO2 reservoir beneath a seismically active volcano. Geophysical Journal Inter-
national, 133(1), F7–F10. https://doi.org/10.1046/j.1365-246x.1998.1331540.x
Kennett, B., Sambridge, M., & Williamson, P. (1988). Subspace methods for large inverse problems with multiple parameter classes. Geophysical
Journal International, 94(2), 237–247. https://doi.org/10.1111/j.1365-246x.1988.tb05898.x
Kiser, E., Palomeras, I., Levander, A., Zelt, C., Harder, S., Schmandt, B., etal. (2016). Magma reservoirs from the upper crust to the Moho
inferred from high-resolution Vp and Vs models beneath Mount St. Helens, Washington State, USA. Geology, 44(6), 411–414. https://doi.
org/10.1130/g37591.1
Korger, E., & Schlindwein, V. (2014). Seismicity and structure of the 85 E volcanic complex at the ultraslow spreading Gakkel Ridge from local
earthquake tomography. Geophysical Journal International, 196(1), 539–551. https://doi.org/10.1093/gji/ggt390
Koulakov, I., Komzeleva, V., Smirnov, S. Z., & Bortnikova, S. B. (2021). Magma-fluid interactions beneath Akutan Volcano in the Aleutian
arc based on the results of local earthquake tomography. Journal of Geophysical Research: Solid Earth, 126(3), e2020JB021192. https://doi.
org/10.1029/2020jb021192
Lapins, S. (2021a). Detecting and characterising seismicity associated with volcanic and magmatic processes through deep learning and
the continuous wavelet transform (Unpublished doctoral dissertation). University of Bristol. Retrieved from http://hdl.handle.net/1983/
ea90148c-a1b2-47ae-afad-5dd0a8b5ebbd
Lapins, S. (2021b). Python notebooks to accompany paper ’A little data goes a long way: Automating seismic phase arrival picking at Nabro
volcano with transfer learning’. Zenodo. https://doi.org/10.5281/zenodo.4558154
Lapins, S. (2022). Nabro volcano event catalogue from Lapins etal., 2021, JGR Solid Earth. Zenodo. https://doi.org/10.5281/zenodo.7398824
Lapins, S., Goitom, B., Kendall, J.-M., Werner, M. J., Cashman, K. V., & Hammond, J. O. (2021). A little data goes a long way: Automating
seismic phase arrival picking at Nabro volcano with transfer learning. Journal of Geophysical Research: Solid Earth, 126(7), e2021JB021910.
https://doi.org/10.1029/2021jb021910
Lees, J. M. (1992). The magma system of Mount St. Helens: Non-linear high-resolution P-wave tomography. Journal of Volcanology and
Geothermal Research, 53(1–4), 103–116. https://doi.org/10.1016/0377-0273(92)90077-q
Lees, J. M. (2007). Seismic tomography of magmatic systems. Journal of Volcanology and Geothermal Research, 167(1–4), 37–56. https://doi.
org/10.1016/j.jvolgeores.2007.06.008
Lin, G. (2013). Seismic investigation of magmatic unrest beneath Mammoth Mountain, California, USA. Geology, 41(8), 847–850. https://doi.
org/10.1130/g34062.1
Lin, G., Amelung, F., Lavallée, Y., & Okubo, P.G. (2014). Seismic evidence for a crustal magma reservoir beneath the upper east rift zone of
Kilauea volcano, Hawaii. Geology, 42(3), 187–190. https://doi.org/10.1130/g35001.1
Lomax, A., Virieux, J., Volant, P., & Berge-Thierry, C. (2000). Probabilistic earthquake location in 3D and layered models. In Advances in seismic
event location (pp.101–134). Springer.
Londoño, J. M., & Sudo, Y. (2003). Velocity structure and a seismic model for Nevado del Ruiz Volcano (Colombia). Journal of Volcanology and
Geothermal Research, 119(1–4), 61–87. https://doi.org/10.1016/s0377-0273(02)00306-2
Maccaferri, F., Rivalta, E., Keir, D., & Acocella, V. (2014). Off-rift volcanism in rift zones determined by crustal unloading. Nature Geoscience,
7(4), 297–300. https://doi.org/10.1038/ngeo2110
MacQueen, P., Gottsmann, J., Pritchard, M. E., Young, N., Ticona J, F., Ticona, E. T., & Tintaya, R. (2021). Dissecting a Zombie: Joint analysis
of density and resistivity models reveals shallow structure and possible sulfide deposition at Uturuncu Volcano, Bolivia. Frontiers of Earth
Science, 755. https://doi.org/10.3389/feart.2021.725917
Manighetti, I., Tapponnier, P., Courtillot, V., Gallet, Y., Jacques, E., & Gillot, P.-Y. (2001). Strain transfer between disconnected, propagating rifts
in Afar. Journal of Geophysical Research, 106(B7), 13613–13665. https://doi.org/10.1029/2000jb900454
Molina, I., Kumagai, H., Le Pennec, J.-L., & Hall, M. (2005). Three-dimensional P-wave velocity structure of Tungurahua Volcano, Ecuador.
Journal of Volcanology and Geothermal Research, 147(1–2), 144–156. https://doi.org/10.1016/j.jvolgeores.2005.03.011
Nowacki, A., Wilks, M., Kendall, J.-M., Biggs, J., & Ayele, A. (2018). Characterising hydrothermal fluid pathways beneath Aluto volcano,
Main Ethiopian Rift, using shear wave splitting. Journal of Volcanology and Geothermal Research, 356, 331–341. https://doi.org/10.1016/j.
jvolgeores.2018.03.023
Oppenheimer, C., Khalidi, L., Gratuze, B., Iverson, N., Lane, C., Vidal, C., etal. (2019). Risk and reward: Explosive eruptions and obsidian lithic
resource at Nabro volcano (Eritrea). Quaternary Science Reviews, 226, 105995. https://doi.org/10.1016/j.quascirev.2019.105995
Patane, D., Chiarabba, C., Cocina, O., De Gori, P., Moretti, M., & Boschi, E. (2002). Tomographic images and 3D earthquake locations of the
seismic swarm preceding the 2001 Mt. Etna eruption: Evidence for a dyke intrusion. Geophysical Research Letters, 29(10), 135–141. https://
doi.org/10.1029/2001gl014391
Paulatto, M., Minshull, T., Baptie, B., Dean, S., Hammond, J. O., Henstock, T., etal. (2010). Upper crustal structure of an active volcano
from refraction/reflection tomography, Montserrat, Lesser Antilles. Geophysical Journal International, 180(2), 685–696. https://doi.
org/10.1111/j.1365-246x.2009.04445.x
Pikington, P. (1988). Uses of pressure and temperature data in exploration and now developments in overpressure detection. Journal of Petroleum
Technology, 40(05), 543–549. https://doi.org/10.2118/17101-pa
Pilia, S., Rawlinson, N., Direen, N., Cummins, P.R., & Balfour, N. (2013). Structural controls on localized intraplate deformation and seismicity
in Southern Australia: Insights from local earthquake tomography of the Flinders Ranges. Journal of Geophysical Research: Solid Earth,
118(5), 2176–2190. https://doi.org/10.1002/jgrb.50168
Pritchard, M., Mather, T., McNutt, S. R., Delgado, F., & Reath, K. (2019). Thoughts on the criteria to determine the origin of volcanic unrest as
magmatic or non-magmatic. Philosophical Transactions of the Royal Society A, 377(2139), 20180008. https://doi.org/10.1098/rsta.2018.0008
Journal of Geophysical Research: Solid Earth
GAUNTLETT ETAL.
10.1029/2022JB025742
21 of 21
Pürschel, M., Gloaguen, R., & Stadler, S. (2013). Geothermal activities in the main Ethiopian rift: Hydrogeochemical characterization of
geothermal waters and geothermometry applications (Dofan-Fantale, Gergede-Sodere, Aluto-Langano). Geothermics, 47, 1–12. https://doi.
org/10.1016/j.geothermics.2013.01.001
Rawlinson, N., & Kennett, B. (2008). Teleseismic tomography of the upper mantle beneath the southern Lachlan Orogen, Australia. Physics of
the Earth and Planetary Interiors, 167(1–2), 84–97. https://doi.org/10.1016/j.pepi.2008.02.007
Rawlinson, N., Reading, A. M., & Kennett, B. L. (2006). Lithospheric structure of Tasmania from a novel form of teleseismic tomography. Jour-
nal of Geophysical Research, 111(B2). https://doi.org/10.1029/2005jb003803
Rawlinson, N., & Sambridge, M. (2004a). Multiple reflection and transmission phases in complex layered media using a multistage fast marching
method. Geophysics, 69(5), 1338–1350. https://doi.org/10.1190/1.1801950
Rawlinson, N., & Sambridge, M. (2004b). Wave front evolution in strongly heterogeneous layered media using the fast marching method.
Geophysical Journal International, 156(3), 631–647. https://doi.org/10.1111/j.1365-246x.2004.02153.x
Rawlinson, N., Sambridge, M. (2003). Seismic traveltime tomography of the crust and lithosphere. Advances in Geophysics, 46, 81–199.
Rawlinson, N., & Spakman, W. (2016). On the use of sensitivity tests in seismic tomography. Geophysical Journal International, 205(2), 1221–
1243. https://doi.org/10.1093/gji/ggw084
Rawlinson, N., Tkalčić, H., & Reading, A. M. (2010). Structure of the Tasmanian lithosphere from 3D seismic tomography. Australian Journal
of Earth Sciences, 57(4), 381–394. https://doi.org/10.1080/08120099.2010.481325
Reinsch, T., Dobson, P., Asanuma, H., Huenges, E., Poletto, F., & Sanjuan, B. (2017). Utilizing supercritical geothermal systems: A review of past
ventures and ongoing research activities. Geothermal Energy, 5(1), 1–25. https://doi.org/10.1186/s40517-017-0075-y
Ross, Z. E., Meier, M.-A., Hauksson, E., & Heaton, T. H. (2018). Generalized seismic phase detection with deep learningshort note. Bulletin of
the Seismological Society of America, 108(5A), 2894–2901. https://doi.org/10.1785/0120180080
Sanders, C., Ponko, S., Nixon, L., & Schwartz, E. (1995). Seismological evidence for magmatic and hydrothermal structure in Long Valley
caldera from local earthquake attenuation and velocity tomography. Journal of Geophysical Research, 100(B5), 8311–8326. https://doi.
org/10.1029/95jb00152
Sethian, J. A. (1996). A fast marching level set method for monotonically advancing fronts. Proceedings of the National Academy of Sciences,
93(4), 1591–1595. https://doi.org/10.1073/pnas.93.4.1591
Sethian, J. A., & Popovici, A. M. (1999). 3-d traveltime computation using the fast marching method. Geophysics, 64(2), 516–523. https://doi.
org/10.1190/1.1444558
Shalev, E., Kenedi, C., Malin, P., Voight, V., Miller, V., Hidayat, D., etal. (2010). Three-dimensional seismic velocity tomography of Montserrat
from the SEA-CALIPSO offshore/onshore experiment. Geophysical Research Letters, 37(19). https://doi.org/10.1029/2010gl042498
Theys, N., Campion, R., Clarisse, L., Brenot, H., Gent, J. v., Dils, B., etal. (2013). Volcanic SO2 fluxes derived from satellite data: A survey
using OMI, GOME-2, IASI and MODIS. Atmospheric Chemistry and Physics, 13(12), 5945–5968. https://doi.org/10.5194/acp-13-5945-2013
Thurber, C. (1993). Local earthquake tomography: Velocities and V_p/V_s-Theory. In Seismic tomography: Theory and practice (pp.563–583).
Toksöz, M. N., Cheng, C., & Timur, A. (1976). Velocities of seismic waves in porous rocks. Geophysics, 41(4), 621–645. https://doi.
org/10.1190/1.1440639
Vanorio, T., Virieux, J., Capuano, P., & Russo, G. (2005). Three-dimensional seismic tomography from P wave and S wave microearthquake
travel times and rock physics characterization of the Campi Flegrei Caldera. Journal of Geophysical Research, 110(B3), B03201. https://doi.
org/10.1029/2004jb003102
Walck, M. C. (1988). Three-dimensional VP/VS variations for the Coso region, California. Journal of Geophysical Research, 93(B3), 2047–2052.
https://doi.org/10.1029/jb093ib03p02047
Wang, Z., & Nur, A. (1986). Effect of temperature on wave velocities in sands and sandstones with heavy hydrocarbons. 56th ann. Internat. Mtg.,
Soc. Expl. Geophys.
Wessel, P., Luis, J., Uieda, L., Scharroo, R., Wobbe, F., Smith, W. H., & Tian, D. (2019). The generic mapping tools version 6. Geochemistry,
Geophysics, Geosystems, 20(11), 5556–5564. https://doi.org/10.1029/2019gc008515
White, R. S., Edmonds, M., Maclennan, J., Greenfield, T., & Agustsdottir, T. (2019). Melt movement through the Icelandic crust. Philosophical
Transactions of the Royal Society A, 377(2139), 20180010. https://doi.org/10.1098/rsta.2018.0010
Wiart, P., & Oppenheimer, C. (2005). Large magnitude silicic volcanism in north Afar: The Nabro Volcanic range and Ma’alalta volcano. Bulletin
of Volcanology, 67(2), 99–115. https://doi.org/10.1007/s00445-004-0362-x
Wilks, M. (2016). A seismological investigation into tectonic, magmatic and hydrothermal processes at Aluto and Corbetti, two restless volcanoes
in the Main Ethiopian Rift (Unpublished doctoral dissertation). University of Bristol.
Wilks, M., Kendall, J.-M., Nowacki, A., Biggs, J., Wookey, J., Birhanu, Y., etal. (2017). Seismicity associated with magmatism, faulting and
hydrothermal circulation at Aluto Volcano, Main Ethiopian Rift. Journal of Volcanology and Geothermal Research, 340, 52–67. https://doi.
org/10.1016/j.jvolgeores.2017.04.003
Wilks, M., Rawlinson, N., Kendall, J.-M., Nowacki, A., Biggs, J., Ayele, A., & Wookey, J. (2020). The coupled magmatic and hydrothermal
systems of the restless Aluto caldera, Ethiopia.
Winkler, K., & Nur, A. (1979). Pore fluids and seismic attenuation in rocks. Geophysical Research Letters, 6(1), 1–4. https://doi.org/10.1029/
gl006i001p00001
Yohannes, E. (2012). Geothermal development in Eritrea: A country update. In Proceedings 4th African Rift Geothermal Conference.
Zenonos, A., De Siena, L., Widiyantoro, S., & Rawlinson, N. (2019). P and S wave travel time tomography of the SE Asia-Australia collision
zone. Physics of the Earth and Planetary Interiors, 293, 106267. https://doi.org/10.1016/j.pepi.2019.05.010
Zhu, W., & Beroza, G. C. (2019). PhaseNet: A deep-neural-network-based seismic arrival-time picking method. Geophysical Journal Interna-
tional, 216(1), 261–273.