The September 19, 2017 (MW 7.

1),
intermediate-depth Mexican earthquake: a slow and
energetically inefficient deadly shock
1 1 1
Aron Mirwald , Vı́ctor M. Cruz-Atienza , John Dı́az-Mojica , Arturo
1 1 1 2
Iglesias , Shri K. Singh , Carlos Villafuerte and Josué Tago

cruz@geofisica.unam.mx

Vı́ctor M. Cruz-Atienza

1
Instituto de Geofı́sica, Universidad

Nacional Autónoma de México

2
Facultad de Ingenierı́a, Universidad

Nacional Autónoma de México

This article has been accepted for publication and undergone full peer review but has not been through
the copyediting, typesetting, pagination and proofreading process, which may lead to differences be-
tween this version and the Version of Record. Please cite this article as doi: 10.1029/2018GL080904

2019
c American Geophysical Union. All Rights Reserved.
Abstract. We investigate dynamic source parameters of the MW 7.1 Puebla-

Morelos intermediate depth earthquake (IDE) (h = 57 km) of September 19,

2017, that devastated Mexico City. Our simple, elliptical source model, cou-

pled with a new Particle Swarm Optimization algorithm, revealed rupture

propagation within the subducted Cocos plate, featuring a high stress drop

(∆τ = 14.2 ± 5.8 MPa) and a remarkably low radiation efficiency (ηr =

0.16 ± 0.09). Fracture energy was large (G = (1.0 ± 0.3) × 1016 J), produc-

ing a slow dissipative rupture (Vr /Vs = 0.34±0.04) with scaling-consistent

radiated energy (Er = (1.8 ± 0.9) · 1015 J) and energy-moment ratio

(Er /M0 = 3.2×10− 5). About 84% of the available potential energy for the

dynamic rupture was dissipated in the focal region, likely producing friction-

induced melts in the fault-core of 0.2-1.2 cm width due to heat production

(700 − 1200◦ C temperature rise). Such source features seem to be a uni-

versal signature of IDEs.

Keypoints:

• Dynamic source inversion of the September 19, (Mw7.1) 2017, intermediate-

depth Mexican earthquake with a Particle Swarm Optimization algorithm

• Highly dissipative event with a slow rupture propagation and scaling-

consistent radiated energy.

• The temperature rise associated with the specific fracture energy likely

produced friction-induced melts in the fault-core

2019
c American Geophysical Union. All Rights Reserved.
1. Introduction

On September 19th 2017, an MW 7.1 intermediate-depth normal-faulting earthquake

(IDE) occurred in the subducting Cocos plate at 57 km depth, with epicentral distance of

114 km from Mexico City [Cruz-Atienza et al., 2017; Singh et al., 2018] (Figure 1). The

peak ground acceleration (PGA) at the rock reference site of Ciudad Universitaria (CU)

was 57 cm/s2 , twice the value recorded at that site during the 1985 devastating earth-

quake (29 cm/s2 ). The earthquake resulted in 369 deaths, of which more than two third

perished in the capital, where 57 buildings suffered total or partial collapse. Devastation

produced by an IDE in the country is not unusual. Several past intraslab earthquakes, as

the great MW 8.2 rupture offshore the states of Oaxaca and Chiapas on September 8, 2017

[Okuwaki and Yagi , 2017; Suárez et al., 2019], have also caused large damage in other

regions of the country (see Singh et al. [2018] for a historical overview of IDEs in Mexico).

Despite the hazard that IDEs represent in Mexico and other regions of the globe, the

physics of this kind of events has challenged the seismological community for decades

[Frohlich, 2006]. At focal depths below 30 kilometers, brittle failure should be inhibited

due to the high normal stresses and frictional stability [Scholz , 1998; Green and Hous-

ton, 1995]. Yet most IDEs are characterized by unstable slip radiating high frequencies

[Frohlich, 2006]. Dehydration reactions at those depths have often been invoked to explain

brittle fracture [Hacker et al., 2003]. Water release from mineralogical phase transitions

could counteract the confining pressure directly (dehydration embrittlement) [Jung et al.,

2019
c American Geophysical Union. All Rights Reserved.
2004] or indirectly (dehydration driven stress-transfer) [Ferrand et al., 2017], enabling a

brittle fracture.

However, it is also possible that brittle fracture is not responsible for IDEs. Thermal

shear runaway in the focal region can generate highly localized ductile deformation and

slip instabilities [Kanamori et al., 1998; Prieto et al., 2012]. This mechanism was shown

to be plausible by numerical simulations [Braeck and Podladchikov , 2007; John et al.,

2009] and geological evidence [Di Toro et al., 2005; John et al., 2009; Andersen et al.,

2008]. Recently, local and global seismological studies have shown that most of IDEs have

a low radiation efficiency, suggesting that thermal shear runaway is the leading source

mechanism of IDEs [Prieto et al., 2013; Dı́az-Mojica et al., 2014; Nishitsuji and Mori ,

2013; Prieto et al., 2017].

Advances in source imaging currently permit the retrieval of some dynamic rupture

parameters directly from ground motion records. After the pioneering work of Peyrat and

Olsen [2004], a few other efforts have led to improved methodologies for inverting the fric-

tion and stress changes on seismogenic faults [Di Carli et al., 2010; Ruiz and Madariaga,

2013; Dı́az-Mojica et al., 2014; Twardzik et al., 2014; Herrera et al., 2017]. Since these

methods model the spontaneous rupture process to explain the observed seismograms (by

solving the elastodynamic equations coupled with a fault-friction constitutive law), they

allow assessment of fundamental properties of the source, such as the energy partition

and its relation with the rupture kinematics.

2019
c American Geophysical Union. All Rights Reserved.
 In this work, we present the dynamic source inversion of the 2017 Puebla-Morelos earth-

quake to investigate the physics of the source process that originated the devastating

ground motion in central Mexico. To this purpose, we improved the methodology intro-

duced by Dı́az-Mojica et al. [2014], by implementing a new Particle Swarm Optimization

algorithm [Eberhart and Kennedy, 1995] that takes advantage of parallel computing, high

convergence rates and a thorough exploration of the model space around the best solutions.

2. Dynamic Source Inversion

2.1. Source Model

We model the earthquake rupture as a frictional process on a fault plane embedded in

a 3-D elastic full space. Sliding begins when the shear traction in the nucleation zone

(NZ) exceeds the static fault strength. In nature, the traction builds up gradually due to

tectonic loading. Here, the shear traction in the NZ, which we assume as a 3 km-radius

circular patch, is prescribed slightly higher than the static strength of the fault so that

the spontaneous rupture initiates with the simulation. To stop rupture propagation, we

assume a barrier model that makes the strength of the rock infinite outside the source.

The rupture area has an elliptical shape, which is a reasonable constraint for moderate

sized IDEs that generally show simpler and localized slip distributions [Di Carli et al.,

2010; Ruiz and Madariaga, 2013; Dı́az-Mojica et al., 2014].

We adopt the linear slip-weakening friction law [Ida, 1972] that depends on three con-

stitutive parameters: the static (µs ) and dynamic (µd ) friction coefficients, and the slip

weakening distance (Dc ). We assume a constant fault normal traction equal to the litho-

static pressure at 60 km depth (1,564 MPa), although this is irrelevant for the rupture

2019
c American Geophysical Union. All Rights Reserved.
process because seismic radiation only depends on the relative stress changes. We fix µd

equal to 0.5 and invert for the change of friction coefficient from the static to the dynamic

levels, ∆µ. We also invert for Dc , the last constitutive parameter defining our friction

model.

The source model is determined by nine parameters: the initial shear tractions in the NZ

and on the fault outside the NZ, the change of the friction coefficient, the slip weakening

distance, and the five geometrical parameters of the ellipse (lengths of its two semiaxes,

two coordinates of its center with respect to the NZ, and its rotation angle). From models

parameterized in this way, we can derive physical quantities involved in the rupture process

[Dı́az-Mojica et al., 2014] such as the fracture energy, G, the radiated energy, Er , and,

hence, the radiation efficiency, ηr = Er /(Er +G), which tells how dissipative is the rupture

process.

2.2. Inversion Method

The dynamic source inversion problem has been solved using heuristic strategies such

as the neighborhood (e.g. Di Carli et al. [2010]; Ruiz and Madariaga [2013]) and genetic

[Dı́az-Mojica et al., 2014] algorithms. In this work we introduce a new Particle Swarm

Optimization (PSO) algorithm to invert the rupture dynamics, which has the remarkable

capacity to thoroughly sample the neighborhood of optimal solutions. Problem solving

may be seen as a population-wide phenomenon, emerging from the individual behaviors

of the individuals through their interactions. In any case, populations are organized ac-

cording to some sort of communication structure, often thought of as a social network.

In PSO a number of simple entities −the particles− are placed in the model space, and

2019
c American Geophysical Union. All Rights Reserved.
each evaluates the objective (or misfit) function at its current location. Each particle

then determines its movement through the space by combining some aspect of the his-

tory of its own current and best locations with those of one or more members of the

swarm, with some random perturbations. Eventually the swarm as a whole, like a flock of

birds collectively foraging for food, is likely to move close to an optimum of the objective

function [Poli et al., 2007]. The PSO has proven to be very efficient for a large class

of multiparametric nonlinear problems [Blum and Li, 2008]. Our PSO algorithm follows

Suganthan [1999] with and increasing size of the models’ vicinity to prevent premature

convergence. Its parameters, listed in the Supplementary Table 1, were chosen based on

a parametric study we performed for the 9-dimensional Langerman function [Jamil and

Yang, 2013]. The size of the population we used was 420 particles (i.e. models), and

we ran the inversions for 50 generations, resulting in 420 × 50 = 21,000 dynamic-source

tested models per inversion. Since we performed 12 inversions per auxiliary fault plane as

explained later in section Results, we explored a total of 504,000 dynamic source models

to generate our preferred solution. In Table 1 we provide the searching ranges of the

model space. To avoid wasting computational time when solving unfeasible models, we

constrained the parameters in such way that the initial traction was always larger than

the static strength in the NZ, and lower than that in the rest of the fault.

To quantify the quality of the source models we use a misfit function that involves the

cross-correlation between observed and synthetic seismograms [Dı́az-Mojica et al., 2014]:

X  τmax + τc cross(xobs − xsyn )


f= − , (1)
2τc auto(xobs ) + auto(xsyn )

2019
c American Geophysical Union. All Rights Reserved.
where the sum is over the three components for all seismic stations. The first term on

the right-hand side accounts for the time shift between the signals that depends on the

characteristic time, τc , which is approximately equal to the source-duration halftime (in

this case τc = 7s) [Dı́az-Mojica et al., 2014], and τmax is the absolute value of the time shift

for the maximum cross-correlation coefficient. The second term involves the maximum

of the cross-correlation and auto-correlation functions, and accounts for the similarity of

waveform and amplitude.

Following Di Carli et al. [2010], the forward problem is solved in two steps. First,

the dynamic rupture is simulated using a highly accurate 3D traction-at-slip-node finite-

difference method [Dalguer and Day, 2007] as compared with most of the well-established

numerical methods [Harris et al., 2018], in a cubic domain with 50 km length per side and

a grid size of 400 m. We verified that such grid size virtually produces the same solutions

as for 300 m in a wide range of constitutive parameters including those explored in our

inversions (e.g. rupture time errors smaller than 10% in all cases). The time step was

adjusted according to the stability criterion. The cube contains the fault-plane where the

friction law is imposed. Around the cube, the outgoing seismic energy is absorbed with

a Perfectly Matched Layer [Marcinkovich and Olsen, 2003] to simulate an infinite space.

Second, the output source kinematics is used to propagate the wave-field up to the stations

by convolving the slip-rate functions with the double-couple Green’s functions computed

using a discrete wave-number method [Bouchon, 1981] in a layered medium appropriate

for the region [Campillo et al., 1996]. The dynamic source inversion methods are very

2019
c American Geophysical Union. All Rights Reserved.
time consuming. In our case, each forward problem took about 15 minutes in a single 2.4

GHz Intel Xeon processor. Each inversion was run in parallel with the PSO algorithm

over 210 processors. Considering a population of 420 models and 50 generations, each

inversion lasted about 25 hours.

2.3. Data and Focal Mechanism

We used accelerograms of 6 strong motion stations belonging to the National Univer-

sity of Mexico (UNAM) with an epicentral distance smaller than 110 km (Figure 1a).

We selected the stations avoiding those with large site-effects, while maintaining a good

azimuthal coverage. The seismograms were aligned with the theoretical P-wave arrival

times predicted by the regional velocity model [Campillo et al., 1996]. Then, the accelero-

grams were bandpass filtered between 0.05 and 0.15 Hz and integrated twice to obtain

the displacements that were inverted. The inverted frequency band contains the corner

frequency of the event, fc ∼0.08 Hz (Supplementary Figure 1) (consistent with the source

duration of ∼13 s reported by the USGS), and thus essential information of the source

finiteness. We choose the upper cutoff frequency to avoid unmodeled effects due to the

simplicity of both the source model and the 1D velocity structure, that prevent us to solve

for source details that are not critical for capturing the overall physics of the rupture pro-

cess.

For an ω 2 Brune source model, the radiated energy contained in frequencies f < 2 ∗ fc

(i.e. below 0.16 Hz in this case) is less than 42% of the total [Singh and Ordaz , 1994].

However, since low frequencies are intrinsically linked to high frequencies through the

downscale causal relationship given by the elastodynamic and fault constitutive equations

2019
c American Geophysical Union. All Rights Reserved.
governing the propagating crack, the source energy partition is reasonably well solved

provided that the fault mechanical model we assumed is a good proxy. This has been

proven to be so by Dı́az-Mojica et al. [2014] and can be seen in Supplementary Figure

S1 for this event, where the spectrum of our preferred source model above 0.16 Hz is

consistent with two other functions determined independently and slightly above the ω 2

Brune’s prediction.

Since it was not possible to determine unambiguously the fault plane using the small

aftershocks sequence, we inverted the source process for both auxiliary planes (Figure 1).

The inversions yielded similar model parameters for both planes, so we decided to discuss

only the results for the fault dipping to the south (φ = 108◦ , δ = 47◦ , λ = −98◦ ), which

yielded 20% better fits with the observed waveforms.

3. Results

Because of the stochastic nature of the optimization method, we decided to run multiple

inversions in the same model space. Although the best-misfit values from all inversions

were of the same order, we noticed that the associated parameters were different to some

extent and that they were interdependent. As expected, (a) larger stress-drops were gen-

erally associated with smaller rupture areas and larger Dc , and (b) lower rupture velocities

were often associated with larger Dc . Ruiz and Madariaga [2013] noticed similar trade-

offs between these parameters in their dynamic source inversions. However, the variation

range of the preferred models from our inversions is relatively small. More importantly,

there are quasi-invariant meta-parameters across the solutions such as the radiated en-

ergy, fracture energy, radiation efficiency and rupture velocity. This means that all models

2019
c American Geophysical Union. All Rights Reserved.
essentially share the same energy partition and rupture kinematics.

Our solution models correspond to the best-fit individuals from 12 independent inver-

sions, whose misfit values are smaller than 0.15. We took as the final, preferred solution

(Table 1) the average of these models along with their standard deviation per parameter.

The seismograms fits are shown in Figure 2, where we compare the observed waveforms

(black curves) with the average waveforms from the selected models (red curves) along

with the associated standard deviations (pink band). Despite the simplicity of the source

model and the velocity structure, most of the waveforms and amplitudes are well repro-

duced in the three components, which gives us confidence to our solution models.

In Figure 3a we show the average final slip distributions from the 12 selected models

and the geometry of the best-fit solution whose misfit value is 0.11 (white dashed ellipse).

Rupture predominantly propagates to the northwest and updip with an average final slip

of 2.1 ± 0.3 m, which is a reasonable value for earthquakes of this size. If we project the

rupture path from the NZ to the left extremity of the rupture into the Earth’s surface,

rupture directivity points roughly towards Mexico City (Figure 1a), suggesting that the

anomalously high ground accelerations and velocities observed in that direction could be

due to source directivity [Singh et al., 2018]. It is worth mentioning that a consistent

direction of rupture directivity was found for all solutions obtained from the inversions

done for the other auxiliary fault plane, dipping to the north (not shown).

2019
c American Geophysical Union. All Rights Reserved.
 Although rupture velocity is remarkably low across large parts of the fault (average

speed of 0.34 ± 0.04 of the shear wave speed (Table 1)), it is also highly variable (Figure

3b). Just after nucleation, rupture propagates with speed around 0.5vs along ∼10 km

updip from the NZ and then slows down rapidly, specially along-strike, where it almost

stops a few kilometers to the northwest (i.e. with speeds of 0.1 − 0.2vs ). Rupture finally

accelerates around 20 km from the NZ to reach again values around 0.5vs close to the

northwest fault extremity. We notice that the relatevely-high front velocity just after

nuclation may be an artifact of the highly stressed NZ required to sustain the rupture

process in the whole fault.

Dynamic crack theory predicts that the faster the rupture front, the larger is the radi-

ation efficiency. This is shown in Figure 4a for the three different fracture modes (gray

lines). Since fracture energy for a given model is almost constant across the fault because

of the slip-weakening friction law (i.e. it is constant except in a narrow belt next to the

fault edge where the final slip is smaller than Dc ), rupture speed variations should map

bulk regions where radiation of seismic energy (or the seismic-moment rate) is enhanced.

Figure 3c confirms such expectation by revealing two maxima of the peak slip-rate that

spatially correlate with fault regions where Vr is maximum (compare with panel b). Ac-

cordingly, the average moment rate function (black line in Figure 3d) shows the associated

bumps centered at 1 and 9 s, as well as a total source duration close to 16 s. Although the

stress drop in the nuclation zone has been inverted, we notice that the timing and width

of the first peak could be biased to some extend by the initial kick imposed to initiate

rupture. However, as seen in Figures 3b and 3d, such effect should not be preponderant in

2019
c American Geophysical Union. All Rights Reserved.
the relatively long (∼13 s long) rupture history after the first 2-3 s of rupture propagation.

Our preferred model exhibits a stress drop of 14.9 ± 5.6 MPa, which is high compared

with typical values for shallow earthquakes, but consistent with IDEs in Mexico [Garcı́a

et al., 2004] and at global scale, which have a mean value around 15 MPa [Prieto et al.,

2012; Poli and Prieto, 2016]. From the model parameters, we derived the radiated energy

Er = (1.8 ± 0.9) · 1015 J, the fracture energy G = (1.04 ± 0.3) · 1016 J and a remarkably

low radiation efficiency ηr = 0.16 ± 0.09. All the inverted and derived parameters from

our preferred model are reported in Table 1.

4. Discussion

Similar to the analysis of the 2011 Zumpango IDE (Mw 6.5) in Guerrero, Mexico [Dı́az-

Mojica et al., 2014], the inversion of the Puebla-Morelos event revealed that the average

rupture speed (vr /vs ∼ 0.34) and the radiation efficiency (ηr ∼0.16) were also remarkably

low (Figure 4a). As expected for intraslab Mexican earthquakes, the stress drop (∆τ ∼14

MPa) was relatively high. Similar results for IDEs have been reported in the Japan sub-

duction zone [Nishitsuji and Mori , 2013], and more recently at a global scale [Poli and

Prieto, 2016] and below the Wyoming Craton, USA, in a completely different tectonic

setting [Prieto et al., 2017], suggesting that slow, inefficient processes characterize the

source of IDEs.

Although such rupture properties are typical of tsunami earthquakes (see Figure 4a),

the 2017 Mw 7.1 shock produced a Fourier acceleration spectrum at CU between 1 and

2019
c American Geophysical Union. All Rights Reserved.
2 s two times larger than observed for previous Mexican IDEs with similar magnitude

reduced to the same hypocentral distance [Singh et al., 2018]. This shows that despite the

highly dissipative source process and the slow rupture propagation, the event generated

high radiation of relatively short period waves. This is opposite to tsunami earthquakes

which are characterized by large MS − MW disparities and anomalously low values of the

ratio Er /M0 [Newman and Okal , 1998; Ammon et al., 2006]. In the case of the Puebla-

Morelos IDE, our inversion yielded Er /M0 = 3.2 × 10− 5, which is consistent with the

value found from the dynamic source inversion of the Mw 6.5 Zumpango IDE in the region

[Dı́az-Mojica et al., 2014], and with those of many earthquakes at a global scale (Figure

4b).

Careful examination of the waveforms does not reveal a preponderant source directivity

as clearly seen for a similar IDE in the region [Singh et al., 2014]. As a matter of fact, in the

absence of strong directivity such as in our preferred source model, strong motions below

1 Hz can be satisfactorily explained with the dynamic rupture properties described in last

section. The peak focal-particle acceleration in the two patches with maximum slip rates

(Figure 3c) are significantly higher than 0.5g (i.e., between 415 and 585 cm/s2 ). These

values in the source region are consistent with most of the Peak Ground Accelerations

(PGA) observed at the stations including RABO, the closest site (hypocentral distance

of 65 km), for periods longer than 1 s as shown in Figure 4c. Model underestimations

in the strike-parallel direction (i.e. at stations YAIG and FTIG) are certainly due to the

source model simplicity, specially at FTIG towards the southeast, where even kinematic

source inversions poorly fit the waveforms [Melgar et al., 2018]. Two main features of our

2019
c American Geophysical Union. All Rights Reserved.
moment rate function are consistent with previous estimates (Figure 3d): the duration of

about 16 s and two peaks in the moment release (between 1 and 4 s, and between 5 and 12

s). Although the source time functions are remarkably different among each other in the

timing of the main peaks, they share similar amplitudes except for the solution reported by

Melgar et al. [2018], which significantly underestimates the scalar moment. Integration of

the moment rate functions gives scalar moments of 1.0520 Nm (Mw = 7.28), 3.519 Nm (Mw

= 6.96), 7.6319 Nm (Mw = 7.18) and 5.819 Nm (Mw = 7.11), for our preferred model, and

the solutions from Melgar et al. [2018], Ye [2018] and Vallée and Duet [2016], respectively.

Most interplate earthquakes have radiation efficiencies larger than 0.5 [Venkataraman

and Kanamori, 2004] (Figure 4a), which implies G < Er . Since ηr ∼ 0.16 for the Puebla-

Morelos event, then G = 5.8Er in this case. This means that ∼84% of the available

potential energy for the dynamic process of faulting was dissipated in the focal region.

Energy dissipation during an earthquake may involve different processes such as off-fault

fracturing, plastic deformation or heat production. Among them, heat production is

certainly a preponderant dissipative mechanisms that may translate into pseudotachylytes

generation (friction-induced melts) in the fault-core [Kanamori et al., 1998; Di Toro et al.,

2005]. Since most of the mechanical energy during rupture is converted into work against

frictional stresses, it has been shown that about 90% of the energy partition goes into

heat production for most large earthquakes [Cocco et al., 2006]. Could fault-core melting

have happened during the Puebla-Morelos event? At the tip of the rupture front (i.e. in

the stress breakdown zone), the temperature rise, ∆T , is related to the specific fracture

energy, Gc (or breakdown work density, namely Wb = G/A), as [Prieto et al., 2013]

2019
c American Geophysical Union. All Rights Reserved.
 Gc
∆T = , (2)
Cρw

where C ∼ 1 J/g ◦ C is the heat capacity, ρ ∼3230 kg/m3 and w is the fault-core width.

This equation assumes that the whole fracture energy is dissipated into heat. In our

model, Gc is equal to (1.7 ± 0.9) · 107 J/m2 in average. Figure 4d shows the temperature

rise for different widths predicted by Equation 2. According to thermal models of the

subducted Cocos plate, the temperature of the slab where the earthquake took place is

between 600 − 700◦ C [Manea and Manea, 2011; Perry et al., 2016]. The melting tem-

perature of peridotite and gabbroic rocks at those depths (i.e. at 1.5 - 2 GPa) range

between 1400◦ and 1800◦ C [Nielsen et al., 2010; Philpotts and Ague, 2009; Takahashi and

Scarfe, 1985; Katz et al., 2003]. Therefore, we expect melting to occur at fault core widths

between 0.2-1.2 cm (see red line in Figure 4d). Field observations show that exhumed

fault-core widths are variable, but most fault veins and pseudotachylites related to sub-

duction earthquakes are between 0.1-3 cm [Andersen and Austrheim, 2006; Obata and

Karato, 1995]. Therefore, the temperature rise at the rupture front (estimated between

700 − 1200◦ C) may indeed have resulted in rock melting as suggested for other IDEs

[Prieto et al., 2013]. Note that we did not consider the heat generated by the total slip

involving the dynamic friction level in large parts of the fault, as done by Kanamori et al.

[1998], that would yield to a higher temperature rise.

Our heat production estimates suggest that melting may have happened at the rupture

tip of the Puebla-Morelos earthquake (absolute temperatures at the tip of 1300−1900◦ C),

supporting the idea that a thermal runaway process could be a preponderant mechanism

2019
c American Geophysical Union. All Rights Reserved.
enhancing unstable slip through lubrication effects despite the high confining pressures of

the focal region. Dehydration from water-rich minerals could also have played a role in

the source region, which doesn’t preclude the existence of the thermal runaway process

[John et al., 2009].

5. Conclusions

We performed the dynamic source inversion of the 2017 intermediate-depth Puebla-

Morelos earthquake with a new Particle Swarm Optimization algorithm. Our results

indicate that the rupture was highly dissipative and remarkably slow. More than 84%

of the available potential energy for the dynamic rupture propagation dissipated in the

focal region. Considering the preexistent slab temperature and heat production during

the rupture process, creation of pseudotachylytes (friction-induced melts) in the fault core

is a plausible process that could enhance the slip instability by lubricating the fault-core.

Despite the large energy dissipation and the slow rupture propagation, our dynamic model

explains most of the observed PGAs below 1 Hz (i.e. the excitation of relatively short

period waves), which were responsible of the extreme ground motion observed in Mexico

City. Independent estimates of IDEs energy balances in Guerrero and different tectonic

settings across the globe suggest that slow, inefficient ruptures with scaling-consistent

energy radiation may by a universal feature of most earthquakes at intermediate depths.

Acknowledgments. This work was supported by the CONACyT grant 2017-01-6471,

the UNAM-PAPIIT grant IG100617, the Japan-Mexico SATREPS project funded by

JST/JICA/UNAM/CONACyT, and the graduate-school scholarships program of CONA-

2019
c American Geophysical Union. All Rights Reserved.
CyT. We thank the ”Servicio Sismológico Nacional” and the ”Instituto de Ingenirı́a”,

both belonging to the ”Universidad Nacional Autónoma de México” (UNAM), for the

acceleration records used in this work. Seismic records can be obtained upon request

to these institutions through the following websites: http://www.ssn.unam.mx/contacto/

and http://aplicaciones.iingen.unam.mx/acelerogramasrsm/redacelerografica.aspx.

References

Ammon, C. J., H. Kanamori, T. Lay, and A. A. Velasco, The 17 july 2006 java tsunami

earthquake, Geophysical Research Letters, 33 (24), 2006.

Andersen, T. B., and H. Austrheim, Fossil earthquakes recorded by pseudotachylytes in

mantle peridotite from the alpine subduction complex of corsica, Earth and Planetary

Science Letters, 242, 5872, 2006.

Andersen, T. B., K. Mair, H. Austrheim, Y. Y. Podladchikov, and J. C. Vrijmoed, Stress

release in exhumed intermediate and deep earthquakes determined from ultramafic pseu-

dotachylyte, Geology, 36 (12), 995–998, 2008.

Blum, C., and X. Li, Swarm intelligence in optimization, in Swarm Intelligence, pp. 43–85,

Springer, 2008.

Bouchon, M., A simple method to calculate green’s functions for elastic layered media,

Bulletin of the Seismological Society of America, 71 (4), 959–971, 1981.

Braeck, S., and Y. Podladchikov, Spontaneous thermal runaway as an ultimate failure

mechanism of materials, Physical Review Letters, 98 (9), 095,504, 2007.

Campillo, M., S. Singh, N. Shapiro, J. Pacheco, and R. Hermann, Crustal structure south

of the mexican volcanic belt, base don group velocity dispersion, Geofı́sica Internacional,

2019
c American Geophysical Union. All Rights Reserved.
 35 (4), 1996.

Cocco, M., P. Spudich, E. Tinti, et al., On the mechanical work absorbed on faults dur-

ing earthquake ruptures, in Radiated Energy and the Physics of Earthquakes Faulting,,

AGU, 2006.

Cruz-Atienza, V. M., S. K. Singh, and M. O. Schroeder, ¿qué ocurrió el 19 de septiembre

de 2017 en méxico?, Revista del Colegio de Ingenieros Civiles de México, 580, 20–24,

2017.

Dalguer, L. A., and S. M. Day, Staggered-grid split-node method for spontaneous rupture

simulation, Journal of Geophysical Research: Solid Earth, 112 (B2), 2007.

Di Carli, S., C. François-Holden, S. Peyrat, and R. Madariaga, Dynamic inversion of

the 2000 tottori earthquake based on elliptical subfault approximations, Journal of

Geophysical Research: Solid Earth, 115 (B12), 2010.

Di Toro, G., S. Nielsen, and G. Pennacchioni, Earthquake rupture dynamics frozen in

exhumed ancient faults, Nature, 436 (7053), 1009, 2005.

Dı́az-Mojica, J., V. M. Cruz-Atienza, R. Madariaga, S. K. Singh, J. Tago, and A. Iglesias,

Dynamic source inversion of the m6. 5 intermediate-depth zumpango earthquake in

central mexico: A parallel genetic algorithm, Journal of Geophysical Research: Solid

Earth, 119 (10), 7768–7785, 2014.

Eberhart, R., and J. Kennedy, A new optimizer using particle swarm theory, in Micro

Machine and Human Science, 1995. MHS’95., Proceedings of the Sixth International

Symposium on, pp. 39–43, IEEE, 1995.

Ferrand, T. P., et al., Dehydration-driven stress transfer triggers intermediate-depth earth-

quakes, Nature communications, 8, 15,247, 2017.

2019
c American Geophysical Union. All Rights Reserved.
Frohlich, C., Deep earthquakes, Cambridge university press, 2006.

Garcı́a, D., S. K. Singh, M. Herráiz, J. F. Pacheco, and M. Ordaz, Inslab earthquakes of

central mexico: Q, source spectra, and stress drop, Bulletin of the Seismological Society

of America, 94 (3), 789–802, 2004.

Green, H. W., and H. Houston, The mechanics of deep earthquakes, Annual Review of

Earth and Planetary Sciences, 23 (1), 169–213, 1995.

Hacker, B. R., S. M. Peacock, G. A. Abers, and S. D. Holloway, Subduction factory 2. are

intermediate-depth earthquakes in subducting slabs linked to metamorphic dehydration

reactions?, Journal of Geophysical Research: Solid Earth, 108 (B1), 2003.

Harris, R., et al., A suite of exercises for verifying dynamic earthquake rupture codes, Seis-

mological Research Letters, 89, 1146–1162, doi:https://doi.org/10.1029/2017JB015388,

2018.

Herrera, C., S. Ruiz, R. Madariaga, and P. Poli, Dynamic inversion of the 2015 jujuy

earthquake and similarity with other intraslab events, Geophysical Journal Interna-

tional, 209 (2), 866–875, 2017.

Ida, Y., Cohesive force across the tip of a longitudinal-shear crack and griffith’s specific

surface energy, Journal of Geophysical Research, 77 (20), 3796–3805, 1972.

Jamil, M., and X.-S. Yang, A literature survey of benchmark functions for global optimiza-

tion problems, Int. Journal of Mathematical Modelling and Numerical Optimisation, 4,

150194, doi:10.1504/IJMMNO.2013.055204, 2013.

John, T., S. Medvedev, L. H. Rüpke, T. B. Andersen, Y. Y. Podladchikov, and H. Aus-

trheim, Generation of intermediate-depth earthquakes by self-localizing thermal run-

away, Nature Geoscience, 2 (2), 137, 2009.

2019
c American Geophysical Union. All Rights Reserved.
Jung, H., H. W. Green Ii, and L. F. Dobrzhinetskaya, Intermediate-depth earthquake

faulting by dehydration embrittlement with negative volume change, Nature, 428 (6982),

545, 2004.

Kanamori, H., and E. E. Brodsky, The physics of earthquakes, Reports on Progress in

Physics, 67 (8), 1429, 2004.

Kanamori, H., D. L. Anderson, and T. H. Heaton, Frictional melting during the rupture

of the 1994 bolivian earthquake, Science, 279 (5352), 839–842, 1998.

Katz, R. F., M. Spiegelman, and C. H. Langmuir, A new parameterization of hydrous

mantle melting, Geochem. Geophys. Geosyst., 4 (9), 2003.

Manea, V. C., and M. Manea, Flat-slab thermal structure and evolution beneath central

mexico, Pure and Applied Geophysics, 168 (8-9), 1475–1487, 2011.

Marcinkovich, C., and K. Olsen, On the implementation of perfectly matched layers in a

three-dimensional fourth-order velocity-stress finite difference scheme, Journal of Geo-

physical Research: Solid Earth, 108 (B5), 2003.

Melgar, D., X. Pérez-Campos, L. Ramirez-Guzman, Z. Spica, V. H. Espı́ndola, W. C.

Hammond, and E. Cabral-Cano, Bend faulting at the edge of a flat slab: The 2017

mw7. 1 puebla-morelos, mexico earthquake, Geophysical Research Letters, 45 (6), 2633–

2641, 2018.

Newman, A. V., and E. A. Okal, Teleseismic estimates of radiated seismic energy: The

e/m 0 discriminant for tsunami earthquakes, Journal of Geophysical Research: Solid

Earth, 103 (B11), 26,885–26,898, 1998.

Nielsen, S., P. Mosca, G. Giberti, G. Di Toro, T. Hirose, and T. Shimamoto, On the

transient behavior of frictional melt during seismic slip, Journal of Geophysical Research:

2019
c American Geophysical Union. All Rights Reserved.
 Solid Earth, 115 (B10301), 2010.

Nishitsuji, Y., and J. Mori, Source parameters and radiation efficiency for intermediate-

depth earthquakes in northeast japan, Geophysical Journal International, 196 (2), 1247–

1259, 2013.

Obata, M., and S.-i. Karato, Ultramafic pseudotachylite from the balmuccia peridotite,

ivrea-verbano zone, northern italy, Tectonophysics, 242, 313–328, 1995.

Okuwaki, R., and Y. Yagi, Rupture process during the mw 8.1 2017 chiapas mexico

earthquake: Shallow intraplate normal faulting by slab bending, Geophysical Research

Letters, 44 (23), 2017.

Perry, M., G. A. Spinelli, I. Wada, and J. He, Modeled temperatures and fluid source

distributions for the mexican subduction zone: Effects of hydrothermal circulation and

implications for plate boundary seismic processes, Geochemistry, Geophysics, Geosys-

tems, 17 (2), 550–570, 2016.

Peyrat, S., and K. Olsen, Nonlinear dynamic rupture inversion of the 2000 western tottori,

japan, earthquake, Geophysical research letters, 31 (5), 2004.

Philpotts, A., and J. Ague, Principles of igneous and metamorphic petrology, Cambridge

University Press, Second Edition(ISBN 978-0-521-88006-0), 2009.

Poli, P., and G. A. Prieto, Global rupture parameters for deep and intermediate-depth

earthquakes, Journal of Geophysical Research: Solid Earth, 121 (12), 8871–8887, 2016.

Poli, R., J. Kennedy, and T. Blackwell, Particle swarm optimization. an overview, Swarm

Intell, doi:10.1007/s11721-007-0002-0, 2007.

Prieto, G. A., G. C. Beroza, S. A. Barrett, G. A. López, and M. Florez, Earthquake

nests as natural laboratories for the study of intermediate-depth earthquake mechanics,

2019
c American Geophysical Union. All Rights Reserved.
 Tectonophysics, 570, 42–56, 2012.

Prieto, G. A., M. Florez, S. A. Barrett, G. C. Beroza, P. Pedraza, J. F. Blanco, and

E. Poveda, Seismic evidence for thermal runaway during intermediate-depth earthquake

rupture, Geophysical Research Letters, 40 (23), 6064–6068, 2013.

Prieto, G. A., B. Froment, C. Yu, P. Poli, and R. Abercrombie, Earthquake rupture below

the brittle-ductile transition in continental lithospheric mantle, Science Advances, 3 (3),

e1602,642, 2017.

Ruiz, S., and R. Madariaga, Kinematic and dynamic inversion of the 2008 northern iwate

earthquake, Bulletin of the Seismological Society of America, 103 (2A), 694–708, 2013.

Scholz, C. H., Earthquakes and friction laws, Nature, 391 (6662), 37, 1998.

Singh, S. K., and M. Ordaz, Seismic energy release in mexican subduction zone earth-

quakes, Bull. Seismol. Soc. Am., 84, 15331550, 1994.

Singh, S. K., X. Pérez-Campos, V. Espı́ndola, V. Cruz-Atienza, and A. Iglesias, Intraslab

earthquake of 16 june 2013 (mw5.9), one of the closest such events to mexico city,

Seismological Research Letters, 85 (2), 268–277, 2014.

Singh, S. K., E. Reinoso, D. Arroyo, M. Ordaz, V. M. Cruz-Atienza, X. Pérez-Campos,

A. Iglesias, and V. Hjorleifsdottir, Deadly intraslab mexico earthquake of 19 september

2017 (mw7.1): Ground motions and damage pattern in mexico city, Accepted in the

Seismological Research Letters, 2018.

Suárez, G., M. A. Santoyo, V. Hjorleifsdottir, A. Iglesias, C. Villafuerte, and V. M.

Cruz-Atienza, Large scale lithospheric detachment of the downgoing cocos plate: The 8

september 2017 earthquake (mw8.2)., Earth and Planetary Science Letters, 509, 9–14,

doi:https://doi.org/10.1016/j.epsl.2018.12.018, 2019.

2019
c American Geophysical Union. All Rights Reserved.
Suganthan, P. N., Particle swarm optimiser with neighbourhood operator, in Proceed-

ings of the 1999 Congress on Evolutionary Computation-CEC99 (Cat. No. 99TH8406),

vol. 3, pp. 1958–1962, IEEE, doi:10.1109/CEC.1999.785514, 1999.

Takahashi, E., and C. M. Scarfe, Melting of peridotite to 14 gpa and the genesis of

komatiite, Nature, 315, 1985.

Twardzik, C., S. Das, and R. Madariaga, Inversion for the physical parameters that control

the source dynamics of the 2004 parkfield earthquake, Journal of Geophysical Research:

Solid Earth, 119 (9), 7010–7027, 2014.

Vallée, M., and V. Duet, A new database of source time functions (stfs) extracted from

the scardec method, Phys. Earth Planet. Int., 257, 149–157, 2016.

Venkataraman, A., and H. Kanamori, Observational constraints on the fracture energy of

subduction zone earthquakes, Journal of Geophysical Research: Solid Earth, 109 (B5),

2004.

Ye, L., Personal communication, California Institute of Technology, USA, 2018.

2019
c American Geophysical Union. All Rights Reserved.
Table 1. Inverted and derived parameters of our preferred dynamic source model. Search

range for the inverted parameters, and mean values with their standard deviations for both the

inverted and derived parameters are provided.

Symbol Parameter Range Value
Inverted Dc Slip-weakening distance 0.2 - 1 m 0.6 ± 0.3 m
∆µ Increment of friction coefficient 0.015 - 0.5 0.035 ± 0.007
∆τ Stress-drop in fault 0 - 48 MPa 14.9 ± 5.6 MPa
∆τ n Stress-drop in NZ 18 - 78 MPa 55 ± 11 MPa
A Area 80 - 1250 km2 713 ± 250 km2
Derived G Fracture energy - (1.04 ± 0.3) · 1016 J
Er Radiated energy - (1.8 ± 0.9) · 1015 J
ηr Radiation efficiency - 0.16 ± 0.09
κ ∼ ratio of strain energy to Gc - 1.5 ± 0.2
Vr Rupture velocity - (0.34 ± 0.04)Vs

2019
c American Geophysical Union. All Rights Reserved.
Figure 1. Map showing significant IDEs in central Mexico. (a) Location of strong motion

stations (green squares) used for the inversion of the Puebla-Morelos event (modified from Cruz-

Atienza et al. [2017]). The straight dashed line depicts the cross section shown below. (b) Cross

section showing the tectonic setting of the region and the projection of some IDEs.

2019
c American Geophysical Union. All Rights Reserved.
Figure 2. Observed (black) and synthetic (red) seismograms in the frequency band 0.05-0.15

Hz. The black lines are the average of the 12 solution models from independent inversions and

the shaded pink areas depicts two times the associated standard deviation.

2019
c American Geophysical Union. All Rights Reserved.
Figure 3. Averages of the 12 solution models where the dashed ellipse corresponds to the

best-fit model geometry. (a) Final slip, (b) rupture velocity normalized by a shear wave speed of

4.7 km/s along with the rupture time contours (in seconds) of the best-fit model, (c) peak slip

rate and (d) moment rate function along with the standard deviation in gray. All models share

the same nucleation zone geometry (black dashed circle).

2019
c American Geophysical Union. All Rights Reserved.
Figure 4. (a) Radiation efficiency (ηr ) as a function of rupture velocity (Vr) normalized by the shear
wave speed (Vs) for different types of earthquakes including the Puebla-Morelos and Zumpango events
in central Mexico. Reported IDEs (red dots) share the same properties as tsunami earthquakes, with
very low rupture velocity and radiation efficiency. Gray lines depict theoretical predictions for the three
different fracture modes (modified from Kanamori and Brodsky [2004]). (b) Energy-moment ratios for
global earthquakes as a function of magnitude Mw taken from Kanamori and Brodsky [2004] along with
values found for the Puebla-Morelos and Zumpango IDEs. The straight lines depict the best linear fit
(solid) and the 60th percentile (dashed). (c) Horizontal-components geometric mean of the observed
and synthetic (best-fit model) PGAs below 1 Hz. (d) Temperature rise as a function of the fault-core
width according to Equation 2. The green dashed lines depict the temperature range associated to our
fracture energy uncertainty, and the red dashed lines the range where peridotites and gabbroic rocks
melt at the earthquake depth.

2019
c American Geophysical Union. All Rights Reserved.