Full text

Turn on search term navigation

Solid Earth, 3, 339354, 2012 www.solid-earth.net/3/339/2012/ doi:10.5194/se-3-339-2012 Author(s) 2012. CC Attribution 3.0 License.

S. C. Stahler1, K. Sigloch1, and T. Nissen-Meyer2

1Dept. of Earth and Environmental Sciences, Ludwig-Maximilians-Universitat (LMU), Mnchen, Germany

2Institute of Geophysics, Swiss Federal Institute of Technology (ETH), Zrich, Switzerland

Correspondence to: S. C. Stahler ([email protected])

Received: 18 May 2012 Published in Solid Earth Discuss.: 2 July 2012Revised: 15 September 2012 Accepted: 2 October 2012 Published: 7 November 2012

Abstract. Triplicated body waves sample the mantle transition zone more extensively than any other wave type, and interact strongly with the discontinuities at 410 km and 660 km. Since the seismograms bear a strong imprint of these geo-dynamically interesting features, it is highly desirable to invert them for structure of the transition zone. This has rarely been attempted, due to a mismatch between the complex and band-limited data and the (ray-theoretical) modelling methods. Here we present a data processing and modelling strategy to harness such broadband seismograms for nite-frequency tomography. We include triplicated P-waves (epi-central distance range between 14 and 30 ) across their entire broadband frequency range, for both deep and shallow sources. We show that is it possible to predict the complex sequence of arrivals in these seismograms, but only after a careful effort to estimate source time functions and other source parameters from data, variables that strongly inuence the waveforms. Modelled and observed waveforms then yield decent cross-correlation ts, from which we measure nite-frequency traveltime anomalies. We discuss two such data sets, for North America and Europe, and conclude that their signal quality and azimuthal coverage should be adequate for tomographic inversion. In order to compute sensitivity kernels at the pertinent high body wave frequencies, we use fully numerical forward modelling of the seismic wave-eld through a spherically symmetric Earth.

Solid Earth

Triplicated P-wave measurements for waveform tomography of the mantle transition zone

1 Introduction

The mantle transition zone (MTZ) is of great interest geodynamically, since its properties determine the extent to which material and heat gets exchanged between the upper and lower mantle. In the seismological view, the MTZ extends from the discontinuity at 410 km depth to one at 660 km both discontinuities are characterized by marked jumps in seismic velocity. The sharpness and topographic undulations of these discontinuities can be linked to mineral physics laboratory experiments in order to infer material properties and mantle rheology.

The seismic waves that sample the MTZ most extensively are regional body waves, i.e., refracted waves that turn within the MTZ. Travelling only moderate distances of [Delta1] 1500

3100 km (i.e. 1429 ), these waves are recorded strong and clear on seismic stations, delivering by far the highest signal to noise ratio of any wave type that could be used to study the transition zone. Yet these phases have rarely been used in general, and for seismic tomography in particular, in sharp contrast to teleseismic phases ([Delta1] > 30 ). The reason is that these regional waves generate more complex signals than teleseismic ones as they have interacted more extensively with the MTZ discontinuities. Such observations do not lend themselves to abstraction into isolated pulses and the associated, idealized modelling by ray theory. Conversely we may suspect that if we succeed at modelling and inverting these waveforms, we will be able to learn a great deal about the mantle structures that are leaving such a strong imprint on them. Here we demonstrate that this should be possible.

Published by Copernicus Publications on behalf of the European Geosciences Union.

340 S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms

Regional P and S waves are commonly termed triplicated waves, since every interaction with a discontinuity spawns three distinguishable phases. We investigate triplicated P-waves, which occur at epicentral distances of 14 to 29 , and sample the MTZ in some interval halfway between source and receiver. Our aim is to use them in nite-frequency waveform inversion for transition zone structure.

Finite-frequency modelling as originally conceived by Dahlen et al. (2000) is feasible across the entire relevant frequency range of body waves, but has been limited to interpreting direct and reected teleseismic phases (Montelli et al., 2004; Sigloch et al., 2008; Zhou et al., 2004, 2006; Tian et al., 2011), due to the applicability limitations of paraxial ray tracing for computing sensitivity kernels. Here we present kernels that overcome this limitation.They are obtained from fully numerical forward computations of the seismic waveeld, using the spectral element code of (Nissen-Meyer et al., 2007b). Exploitation of the near-spherical symmetry of the 3-D Earth uses the currently available computational resources very efciently, allowing access to the entire relevant broadband range of the wave-eld (0.03 to 1 Hz), like the original Dahlen method.

Triplicated body waves carry a very strong imprint of their interaction with transition zone discontinuities, which is both an advantage and a challenge. Not only do the triplicated arrivals overlap each other in time, due to the nite-frequency nature of real data, but for shallow earthquakes they additionally overlap the depth phases pP and sP, which get triplicated themselves. Hence an integral part of modelling the waveforms is the careful estimation of source parameters (since they determine the shape of the synthetic Greens function), and of the source time function.

The differential moveout of the triplicated phases has since long been used to derive one-dimensional velocity models(Grand, 1984), which required a large number of seismic stations in a narrow azimuth range. Modelling of individual triplicated waveforms has been successfully used to sample localized heterogeneities in the MTZ (Tajima and Grand, 1995; Melbourne and Helmberger, 2002; Tajima et al., 2009). A few studies that included picked arrival times into ray theoretical inversions demonstrated the potential of triplicated waves for tomography (Grand, 2002), but so far they have rarely been used. Only very recently, Zhu et al. (2012) used triplicated phases, amongst others, for a European regional tomography, albeit at lower frequencies.

Regional body waves are complementary to all phases currently used for MTZ studies. Teleseismic body waves also offer good signal-to-noise ratios, but have comparably low sensitivity in the MTZ, which they traverse at steep angles.Hence they constrain the MTZ beneath sources and/or station, whereas triplicated waves sample it extensively midway between sources and receivers. So it would be highly benecial to combine them in one inversion.

Other methods for sampling transition zone discontinuities have much lower signal-to-noise ratios (SNR). Receiver functions exploit P-to-S or S-to-P converted energy of body waves, which require stacking numerous seismograms. Their migration from the time domain to depth depends on a velocity structure model, which either must have been obtained independently or is neglected, with corresponding systematic errors in the result.

PP and SS precursors, i.e., body waves reected at the undersides of MTZ discontinuities, have also been used in structural studies, (Shearer and Masters, 1992; Thomas and Billen, 2009; Deuss et al., 2006; Deuss, 2009). Their SNR is low, so that only the rare strong earthquakes that are recorded on seismic arrays can be used. Distinguishing between the signal of an undulated discontinuity and a volumetric velocity perturbation is challenging (Chaljub and Tarantola, 1997), as for receiver functions.

Surface waves have lower image resolution than body waves. Only their higher modes have signicant sensitivity to the MTZ, but higher modes carry little energy, and are more difcult to process. Hence surface wave tomography is largely limited to depths above the MTZ.

We start with a discussion of the nature of triplicated body waves, and of their expression in actual seismic broadband data (Sect. 2.1). Section 2.2 and 2.3 demonstrate how we model these waveforms; this covers the computations of Greens functions and the inversion for source time functions and source mechanisms. We then explain the concept of waveeld kernels (Sect. 3), and show how passband-ltering to different frequency bands signicantly increases the resolution in the transition zone, by comprehensively exploiting information across the entire broadband range (Sect. 3.3).The sizeable and well-instrumented continents of Europe and North America currently offer the most favourable source-receiver combinations of criss-crossing regional body waves for tomography. Section 4 discusses the two data sets that we have assembled for these two regions, with a focus on the USArray. We conclude with a discussion of the results, and the prospects of inverting these data for MTZ structure (Sects. 5 and 6).

2 Regional P-waves

2.1 Triplications

The term triplication refers to three seismic wave phases that have similar ray parameters and arrive closely spaced in time.The concept is rooted in ray theory. While our data processing and kernel computations are targeted at interpreting the full waveforms, it is useful to consider their ray theoretical approximation rst, in order to appreciate the nature of trip-licated waves. It assumes that the wave from one source to one receiver travels along an innitesimally narrow ray path, which is described by the eikonal equation (e.g., Kennett,

Solid Earth, 3, 339354, 2012 www.solid-earth.net/3/339/2012/

S. C. Stahler et al.: Finite-frequency measurements of triplicated

waveforms

341

S. C. Stahler et al.: Finite Frequency Measurements of triplicated

waveforms

Fig. 1. Observed waveforms and nominal ray paths of regional P

-waves,

for a

magnitude

6.4

earthquake

off

the coast of Mexico

the waveform clusters

gure. Our automated clustering algorithm sorts the different kinds of triplicated phases into different groups, based on the dissimilarity the broadband waveforms. The waveform below each seismogram group is the waveform stack. The groups coincide with the distance bins predicted by theory (see text for details). This demonstrates that we are dealing with a robust signal that waveformshould be able to interpret.

02:33:05, Lat: 17.61 Lon: -105.50, Depth: 6 km). The ray coloring

left

gure

corresponds

the

coloring

Fig. 1. Observed waveforms and nominal ray paths of regional P-waves, for a magnitude 6.4 earthquake off the coast of Mexico (2008/09/24 02:33:05, Lat: 17.61 Lon: 105.50, Depth: 6 km). Ray colouring in the left gure corresponds to the colouring of the waveform clusters in the right gure. Our automated clustering algorithm sorts the different kinds of triplicated phases into different groups, based on similarity of the broadband waveforms. The waveform below each seismogram group is the waveform stack. This empirical sorting into groups coincides with the epicentral distance bins predicted by theory (see text for details). This demonstrates that we are dealing with a robust signal that waveform tomography should be able to interpret.

Fig. 1. Observed waveforms and nominal ray paths of regional P-waves, for a magnitude 6.4 earthquake off the coast of Mexico (2008/09/24 02:33:05, Lat: 17.61 Lon: -105.50, Depth: 6 km). The ray coloring in left gure corresponds to the coloring of the waveform clusters in right gure. Our automated clustering algorithm sorts the different kinds of triplicated phases into different groups, based on the dissimilarity of the broadband waveforms. The waveform below each seismogram group is the waveform stack. The groups coincide with the epicentral distance bins predicted by theory (see text for details). This demonstrates that we are dealing with a robust signal that waveform tomography should be able to interpret.

(1)

pr2/(}c(r))2 1

(}) =

[integraldisplay]

(2)

c(r)

p1 (}c(r))2/r2

Fig. 2. Triplication of P-wave traveltimes in IASP91. The triplication C to D is produced by the discontinuity at 410 km depth, the one from E to F by the one at 660 km. The AB triplication originates from the 210 km discontinuity, which is only of second order in IASP91. This triplication can hardly be observed, especially since it overlaps with the one from 410 km.

at the discontinuity and then to one traveling below the traveltime-distance relation T ( ), which isas well, a pattern called triplication emerges (Fig. 2).

175 The discontinuities in seismic velocity in the upper mantle410 and 660 km depth in the IASP91 (Kennett and 1991) reference model, henceforth termed 410 and 660 duce triplications: Neglecting crustal phases and the developed 210-km discontinuity, we receive one P

180 to a distance of 14 , which has traveled through the mantle. At 14 another one arrives 8 seconds laterin Fig. 2), being reected at the 410. This branch into one reected path (DC), and one refracted below The refraction branch continues to 28 distance (E),

185 the uppermost-mantle branch meets with the reected21.5 (C) and ends there. From 17.5 (F) on there are reection and refraction branches from the 660. The reection branch (EF) meets with the 410-refraction (DF) at 27.8 (E), where they both end. Only the 660

190 tion branch can be measured at distances exceedingWith this conceptual knowledge, consider the real Fig. 1. It shows broadband P-waveforms in a 20 s dow, starting 5 s before the rst arrival according to

If the velocity gradient with respect to depth is smooth, and [Delta1] increase monotonously with decreasing }, i.e. stations at a larger distance from the earthquake record the P-phase at later times. If the velocity prole contains strong gradients, or even discontinuities with positive velocity jumps, d[Delta1]d}

changes sign, so that rays of smaller } arrive closer to the source1. Because d(})d} < 0 in either case, two rays arrive at the same distance at different times. If dcdr < 0, [Delta1](}) is still continuous. The turning points of [Delta1](}) mark the transition (with decreasing }) rst from a ray travelling above the discontinuity to one being critically reected at the discontinuity and then to one travelling below it. In the traveltime-distance relation ([Delta1]), which is also continuous, a pattern called trip-lication emerges (Fig. 2).

1The literature does not provide any hard constraints for a trip-lication to occur. However, it is clear that a discontinuity in c(r) or dc/dr will be sufcient.

www.solid-earth.net/3/339/2012/ Solid Earth, 3, 339354, 2012

at the discontinuity and then to one traveling below it. In the traveltime-distance relation T ( ), which is continuous as well, a pattern called triplication emerges (Fig. 2).

175 The discontinuities in seismic velocity in the upper mantle (at 410 and 660 km depth in the IASP91 (Kennett and Engdahl, 1991) reference model, henceforth termed 410 and 660), produce triplications: Neglecting crustal phases and the weakly developed 210-km discontinuity, we receive one P phase up

180 to a distance of 14 , which has traveled through the upper mantle. At 14 another one arrives 8 seconds later (point D in Fig. 2), being reected at the 410. This branch bifurcates into one reected path (DC), and one refracted below (DE). The refraction branch continues to 28 distance (E), while

185 the uppermost-mantle branch meets with the reected one at21.5 (C) and ends there. From 17.5 (F) on there are also reection and refraction branches from the 660. The 660-reection branch (EF) meets with the 410-refraction branch (DF) at 27.8 (E), where they both end. Only the 660 refrac-

190 tion branch can be measured at distances exceeding 27.8 . With this conceptual knowledge, consider the real data in Fig. 1. It shows broadband P-waveforms in a 20 s window, starting 5 s before the rst arrival according to IASP91.

Fig. 2. Triplication of P-wave traveltimes in IASP91. The triplication C to D is produced by the discontinuity at 410 km depth, the one from E to F by the discontinuity at 660 km. The AB triplication originates from the 210 km discontinuity, which is only of second order in IASP91. This triplication can hardly be observed, especially since it overlaps with the one from 410 km.

2001). Strictly speaking, the approximation only holds for the case of a non-dispersive medium. Even when the medium is not intrinsically dispersive, velocity heterogeneities may introduce a frequency-dependent group velocity, which is the motivation for nite-frequency tomography.

In a layered velocity model c(r), the traveltime and the angular distance [Delta1] of a direct phase (e.g., a P-wave) depend only on the ray parameter }.

[Delta1](}) =

[integraldisplay]

342 S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms

4 S. C. Stahler et al.: Finite Frequency Measurements of triplicated waveforms

A B C D

10.

Fig. 3. Schematic composition of a triplicated broadband P-wave signal: (A) Arrival times (top row) and nominal ray paths (bottom row) of the direct P-waves. Three pulses arrive, two of which are refracted above and below the discontinuity (here 410), and the third reected by it.(B) If the earthquake is shallow, additional depth phases arrive (pP and sP), which get triplicated themselves. (C) Real-world seismograms do not resemble sequences of dirac pulses. Rather, the pulses of the Greens functions are dispersed and the source time function (bottom row) convolves into all of them. (D) The sum of all contributions is the predicted seismogram (top), which may be compared to the observed seismogram (bottom row), in order to extract measurements for waveform tomography. Even though the waveform is complex, it can be modelled sufciently well using a layered background model.

The velocity discontinuities in seismic velocity in the upper mantle (at 410 and 660 km depth in the IASP91 reference model (Kennett and Engdahl, 1991), henceforth termed 410 and 660, produce triplications: Neglecting crustal phases and the weakly developed 210-km discontinuity, we receive one P phase up to a distance of 14 , which has travelled through the upper mantle. At 14 another P-phase arrives 8 seconds later (point D in Fig. 2), being reected at the 410. This branch bifurcates into one reected path (DC), and one refracted below (DE). The refraction branch continues to 28 distance (E), while the uppermost-mantle branch meets with the reected one at 21.5 (C) and ends there. From17.5 (F) on there are also reection and refraction branches from the 660. The 660-reection branch (EF) meets with the 410-refraction branch (DF) at 27.8 (E), where they both end.

Only the 660 refraction branch can be measured at distances exceeding 27.8 .

With this conceptual knowledge, consider the real data in

Fig. 1. It shows broadband P-waveforms in a 20 s window, starting 5 s before the rst arrival predicted by IASP91. The 692 waveforms from stations in the western United States are aligned (VanDecar and Crosson, 1990) and clustered by similarity as described by Sigloch and Nolet (2006). This clustering results purely from the waveform shapes and does not use meta-information such as station distance or azimuth from the source. The clusters show a clear zonation by distance, caused by the triplications from the transition zone discontinuities. The gure can best be understood by rst considering the most distant group of P-waveforms (pink). The stations are mainly located beyond 28 epicentral distance, so that the signal contains no triplications. The secondary pulse, arriv-

10.

410

20.

Fig. 3. Schematic composition of a triplicated broadband P-wave signal: A) Arrival times (top row) and nominal ray paths (bottom row) of the direct P-waves. Three pulses arrive, two of which are refracted above and below the discontinuity (here 410), and the third reected by it. B) If the earthquake is shallow, additional depth phases arrive (pP and sP), which get triplicated themselves. C) Real-world seismograms do not resemble sequences of dirac pulses. Rather, the pulses of the Greens functions are dispersed and the source time function (bottom row) convolves into all of them. D) The sum of all contributions is the predicted seismogram (top), which may be compared to the observed seismogram (bottom row), in order to extract measurements for waveform tomography. Even though the waveform is complex, it can be modelled sufciently well using a layered background model.

The 692 waveforms from stations in the western USA are

195 aligned (VanDecar and Crosson, 1990) and clustered by similarity as described by Sigloch and Nolet (2006). This clustering results purely from the waveform shapes and does not use meta-information such as the station distance or azimuth from the source. The clusters show a clear zonation by dis-

200 tance, caused by the triplications from the transition zone discontinuities.

The gure can best be understood by rst considering the most distant group of P-waveforms (pink). The stations are mainly located beyond 28 epicentral distance, so that the

205 signal contains little energy attributed to triplications. The secondary pulse, arriving 7 s after the rst, can therefore be recognized as a surface reection (pP or sP).

In the distance bin closest to the source, from 15 to

18 (red), the triplicated phases overlap so that just one broad-

210 ened pulse arrives. In the second group (yellow), the arrivals are further spread so that a rst arriving pulse, a surface reected phase (at 5 s) and multiple triplicated phases can be distinguished. Additionally, we see an arrival 10 s after the

rst one, which is a reection from the 660 (FE branch in

215 g. 2). In the third group (green), this reection is already closer to the rst arrival, and the rst-arriving pulse is a superposition of the DE branch from the transition zone and the BC phase from the upper mantle and depth phases. The next group (dark blue) is quite heterogeneous and marks the

220 transition from 410 triplications to 660 triplications.

225 hence correspond to P (two overlapping refractions) and pP

(another two refractions). In the next two groups (grass green and purple), the refracted phase from above and below the 660 begin to separate, so that they spill into pP once more. The most distant group (pink) shows clearly separated P and

230 pP pulses no longer affected by triplications. The shape of this teleseismic waveform does not change much out to

85 distance.

Hand-picked arrival times of the different triplication branches have been used in a few tomographic studies

235 (Grand, 1994, 2002), although this approach fails to take into account the broadband character of seismic waves. As Fig. 1 demonstrates, the triplications are often not clearly separated, due to several factors. The nite duration of the source rupture has a lowpassing effect. Earths intrinsic attenuation dis-

240 perses the P-wave pulse, which does not contain much energy above 1 Hz. Additionally, waves of a nite wavelength are inuenced by scatterers off of the direct path if the path difference is less than /2.

2.2 Waveform modelling

245 We download broadband seismograms from the IRIS and

ORFEUS data management centers. The waveforms are corrected are bandpassed between 0.01 Hz and 3.5 Hz, detrended, and transfered to ground displacement by deconvolving the instrument responses. A time window is cropped

250 from 10 s before to 25 s after the theoretical P arrival. Noisy seismograms are singled out by a clustering algorithm and removed manually.

In order to do cross-correlation measurements for waveform tomography, it is necessary to compare a synthetic waveform

ing 7 s after the rst one, can therefore be recognized as a reection from the ocean surface near the source (pwP).

In the distance bin closest to the source, from 15 to

18 (red), the triplicated phases overlap so that just one broadened pulse arrives. In the second group (yellow), the arrivals are further spread so that a rst arriving pulse, a surface reected phase (at 5 s) and multiple triplicated phases can be distinguished. Additionally, we observe an arrival 10 s after

the rst one, which is a reection from the 660 (FE branch in Fig. 2). In the third group (green), this reection is already closer to the rst arrival, and the rst-arriving pulse is a superposition of the DE branch from the transition zone and the BC phase from the upper mantle and depth phases. The next group (dark blue) is quite heterogeneous and marks the transition from 410 triplications to 660 triplications.

In the cyan group, beginning at 22 , the refractions above and below the 660 arrive almost at the same time and collapse into a single pulse. The reected phase is usually quite weak and not visible at all here. The two clearly separated arrivals hence correspond to P (two overlapping refractions) and pwP (another two refractions). In the next two groups (grass green and purple), the refracted phase from above and below the 660 begin to separate, so that they spill into pwP once more. The most distant group (pink) shows clearly separated P and pwP pulses no longer affected by triplications. The shape of this teleseismic waveform does not change much out to

85 distance.

Hand-picked arrival times of the different triplication branches have been used in a few tomographic studies (Grand, 1994, 2002), although this approach fails to take into account the nite frequencies of seismic waves. As Fig. 1 demonstrates, the triplications are often not clearly separated,

Solid Earth, 3, 339354, 2012 www.solid-earth.net/3/339/2012/

S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms 343

due to several factors. The nite duration of the source rupture has a low-passing effect. Earths intrinsic attenuation disperses the P-wave pulse, which does not contain much energy above 1 Hz. Additionally, waves of a nite wavelength are inuenced by scatterers off of the direct path, if the detour to reach the scatterer is less than /2, which is the denition of the Fresnel zone. This argument is continued in Sect. 3.

2.2 Waveform modeling

We download broadband seismograms from the IRIS and ORFEUS data management centres. The waveforms are corrected are band-passed between 0.01 Hz and 3.5 Hz, detrended, and transferred to ground displacement by deconvolving the instrument responses. A time window is cropped from 10 s before to 25 s after the theoretical P arrival. Noisy seismograms are singled out by a clustering algorithm and removed manually.

In order to do cross-correlation measurements for waveform tomography, it is necessary to compare a synthetic waveform to the observed seismogram. Synthetic seismo-grams are calculated using the reectivity method of Fuchs and Mller (1971). As a reference model we use IASP91 (Kennett and Engdahl, 1991), together with the density and intrinsic attenuation of PREM (Dziewoski, 1981).

Fig. 3 demonstrates the challenge of modelling regional P-waveforms. Within a short time window, three to ve trip-licated phases arrive (Fig. 3a). If the earthquake is shallow, as most earthquakes are, the surface-reected phases pP, sP or pwP arrive within a few seconds of P, and are themselves triplicated (Fig. 3b). For an earthquake at 20 epicentral distance, ten phases arrive within less than 9 s, triplicated by both the 410 and the 660 discontinuities. The polarity of the reected phases is negative compared to the refracted phases, and the depth phases may have reversed polarity depending on the source plane orientation. Hence the overall waveform is highly sensitive to the exact depth, mechanism, and distance of the earthquake.

Additionally, the nite duration of the source process is imprinted on the seismogram, which is the convolution of the moment rate function m(t), termed source time function

(STF), with the Greens function G (rs,rr,). Since duration of the STF is usually of several seconds, it can change the waveform completely.

If the maximum frequency in the seismograms is f

1/T , where T is the duration of earthquake rupture, then the source time function is sometimes approximated by a Dirac delta, a triangular function or a Gaussian. This is usually done in surface wave tomography and long-period waveform inversion. We use earthquakes of moment magnitudes between 5.5 and 7.0, where T is between one and several tens of seconds, and we want to invert up to dominant periods of up to 1 s. Hence this approximation is too rough for our purposes, and we need explicit estimates of each STF in order to construct the matched lter.

Fig. 4. Flowchart of the source inversion procedure as described in (Sigloch and Nolet, 2006).

2.3 Source inversion

As seen in the previous section, waveform tomography requires an inversion for the temporal and spatial parameters of the earthquake sources prior to the actual tomographic inversion. For the deconvolution of the STF, we use teleseismic P-waves rather than the complex, triplicated regional waves.We briey describe the procedure here and in Fig. 4, for details see Sigloch and Nolet (2006): We remove obviously problematic stations and align all waveforms to the arrival of the P-phase (VanDecar and Crosson, 1990). We then choose a reasonable candidate depth range to survey, 150 km for shallow events and NEIC depth [notdef]30 km for deep events.

Then we execute the following scheme for each candidate depth: First a joint deconvolution of the synthetic seismo-grams, calculated with the NEIC moment tensor M0 from the measured seismograms is done, resulting in an STF estimate m(). Source orientation is assumed to be constant

during the rupture, so that m() is identical for all compo

nents of the moment tensor. Second, with this STF, an update for the moment tensor M is calculated and the amplitudes of all stations are corrected individually. The new moment tensor M0 + M and the amplitude corrections are used to

derive a new STF estimation and this is repeated, until the

www.solid-earth.net/3/339/2012/ Solid Earth, 3, 339354, 2012

344 S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms

RMS mist between synthetics and broadband seismograms has converged.

After all depths have been treated, we manually choose the depth, at which the RMS mist is minimal and the STF does not contain any signicant negative parts, which would be unphysical. The STF and moment tensor results for this most likely depth are retained for kernel calculation and tomography.

The teleseismic Greens functions are calculated by the WKBJ code of Chapman (1978), using IASP91 as reference model. For source inversion we use only waveforms from globally distributed, reliable broadband stations (GSN and Geoscope networks). For typical earthquakes, around 40 of these stations are located within teleseismic range. Since STF deconvolution is a numerically sensitive operation, we deliberately use only this small, high-quality ensemble, rather than all available broadband stations. The spatial distribution of these permanent, international network stations is relatively even, whereas a deconvolution from all IRIS stations would always be dominated by the 1000+ receivers located in North America. Their waveforms and mists are highly correlated, so that the additional information content with regard to the source is low. However, all available stations will later be used for tomography.

We always attempt to invert for a single STF that ts all global teleseismic data, but in about one out of three earthquakes, we need to allow for two or more regional STF solutions, e.g., when the European station cluster cannot be t to the same STF as the North American cluster. This may be due to structure close to the source, like a subducting slab or be an effect of source directivity. Our view of a source time function is pragmatic: we want it to absorb all signal that is common to all seismograms, even when that signal does not derive from the source rupture sensu stricto. The remaining signal can then be interpreted as an imprint of the structure along the wave-path.

3 Sensitivities

3.1 Waveform tomography direct phases

The concept of arrival time of a phase, which may be picked manually by an analyst, assumes a broadband minimum-phased signal. In a heterogeneous Earth, scattering off of the direct path adds a small, frequency-dependent component to the direct waveform. Hence scattering introduces a nonlinear dispersion even if the medium is perfectly viscoelastic (Dahlen et al., 2000; Nolet, 2008). This dispersion reects the scale-dependent interaction of different nite wavelengths with the mantle heterogeneities that we want to image. Hence it embodies the information that waveform inversion captures above and beyond the ray theoretical approximation. We measure this dispersion by the method of matched ltering (Sigloch and Nolet, 2006).

A predicted waveform u(t) is synthesized from six partial Greens functions Gj (rs,rr,t), weighted by the six independent components of the moment tensor Mj , and convolved by the source time function m(t). The broadband u(t)

is immediately bandpass ltered to uk(t), through convolution with a lter response fk(t) (k is the frequency band index). Hence the nite-frequency synthetic is

uk(t) =

Xj=1Gj (rs,rr,t) [notdef] m(t) [notdef] Mj [notdef] fk(t), (3)

which must be compared to an accordingly ltered observed waveform uok(t). For this we parametrize uk(t) by the two observables (mist measures) that we want to estimate: the traveltime anomaly Tk and the amplitude anomaly Ak,

uk(t) = Akuo(t Tk). (4)

uk(t) is the matched lter, and the optimal Tk, Ak are obtained by minimizing the RMS mist betweenk(t) and

uok(t). This is equivalent to nding the time shift T that maximizes the cross correlation betweenk(t) and uok(t).

With this method, we obtain up to 2 [notdef] k frequency-

dependent mist observables Tk, Ak from one broadband P-wave seismogram. The derivative of these mists with respect to the Earth model can be expressed in so-called sensitivity kernels Ki(rx). These represent the sensitivity of the traveltime mist or the amplitude mist with respect to changes in P-wave velocity [Delta1]VP/VP at a given point rx. The traveltime anomaly Tk is then modelled as

Tk =

[integraldisplay]

[circleplustext]

Kk(rx)[Delta1]VP

VP (rx) d3rx. (5)

Obviously, this kernel has to be calculated using a reference velocity model. Equation 5 and the construction of the kernel Kk contain the assumption that the traveltime anomaly Tk originates exclusively from single scattering off of anomalies [Delta1]VP/VP, which quantify the difference between the reference model and true Earth structure. This so-called Born approximation is justied by the observation that the magnitude of lateral mantle heterogeneities [Delta1]VP/VP is small, typically on the order of a few percent, so that multiple scattering can be neglected. Kernel Kk is the rst Frchet derivative of Tk towards [Delta1]VP(rx) (Marquering et al., 1999).

Dahlen et al. (2000) proposed a fast algorithm to derive kernels for teleseismic body waves by paraxial ray tracing.This effectively uses ray theory to go beyond the limits of ray theory, since it synthesizes sensitivity kernels of nite volume from the interaction of an innity of rays. The method has been applied to teleseismic body waves with great success (Montelli et al., 2004; Sigloch et al., 2008) and was extended to surface waves (Zhou et al., 2004, 2006; Tian et al., 2011) but its scope is limited to phases without caustics, diffractions or other wave effects (Dahlen et al., 2000;Nissen-Meyer et al., 2007a).

Solid Earth, 3, 339354, 2012 www.solid-earth.net/3/339/2012/

S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms 345

3.2 Waveform tomography arbitrary phases

Triplicated waveforms, with their interactions of refracted and reected phases around the discontinuity, are not adequately modelled by the ray theoretical formalism. The caustics at turning points AF (Fig. 2) would lead to innite amplitudes at the corresponding distances. Moreover, the ray-centred approach by Dahlen et al. (2000) works strictly speaking only in a continuous velocity model. Hence we need to calculate the kernels from the full waveeld instead.

Computing the rst order perturbation u(t) of a seismic waveform u(t) involves the waveeld from source to every possible scattering location !

u(t,r), and of the scattered waveeld to the receiver. Thanks to source-receiver reciprocity, this second waveeld may instead be replaced by the back-propagating waveeld u(t,r) from receiver to scatterers, which is generally much cheaper computationally (Nissen-Meyer et al., 2007a, eqn. 10). This is conceptually similar to the adjoint method (Tromp et al., 2005).

u =

[integraldisplay]

[circleplustext]

h !vi[notdef] vi+

Eij [notdef]

Ekl Cijkl

we need to store only the displacement eld u(r), rather than the full strain tensor Eij (r). The exact expression of the kernel now depends on the chosen mist criterion. We prefer the cross-correlation traveltime mist measured on one component i (in the case of P-waves usually BHZ), dened as

T = [Rfractur][integraltext]

0 !iui(!) ui(!)d!

[integraltext]

. (11)

With ui(!) = Gij (rs,rr,!) m(!)Mj , we can calculate the

sensitivity of T w.r.t. VP. Using the denition of the kernel in Eq. (5), we arrive at

K(rx) =

0 !2ui(!)ui(!)d!

Xj=12VP(rx)(rx)

[integraltext]

6 0 !2[notdef]Gsr,ij (!) m(!)[notdef]2d! [notdef]

Mj (12)

[notdef][Rfractur]

1 i! Gsr,ij (!)[notdef] m(!)[notdef]2

id3r. (6)

Gj (rx,!)rGj (rx,h!)d!.

The term [notdef] m(!)[notdef], which Dahlen et al. (2000) originally de

noted as the source term, contains the source spectrum, but also the bandpass lters. Gsr(!), the Greens function from source to receiver, introduces any intrinsic attenuation of the reference Earth model into the kernel.

3.3 Spherical Earth kernels

The calculation of a global waveeld at a dominant period of 5 s requires around 104 CPUh when a full 3-D forward solver like SPECFEM (Komatitsch and Tromp, 2002) is used. For a realistic iterative global tomography using > 105 waveforms, the calculation cost would be 109 CPUh, which is com

pletely prohibitive. In a spherically symmetric background model, this cost can be reduced dramatically: 3-D waveelds in a layered Earth can be computed at the cost of the equivalent 2-D waveelds, since the symmetry implies that one dimension may be calculated analytically (Nissen-Meyer et al., 2007a). Forward waveelds are pre-computed for reasonable depth increments (e.g. increments of 1 km for depths from 0 to 100 km, and increments of 10 km from 100 to 700 km, requiring 160 simulations). The backward waveeld needs to be calculated just once, assuming that all receivers are located at the surface or within one P-wavelength of it. The 160 forward calculations need to be done four times (Nissen-Meyer et al., 2007a, p. 1057) for

1. a Mzz monopole source

2. a 12 Mxx + Myy

[parenrightbig]

monopole source

3. a Mxz dipole source

4. a Mxy quadrupole source.

www.solid-earth.net/3/339/2012/ Solid Earth, 3, 339354, 2012

v= @t !u is the velocity eld of the forward propagating wave and !

Eij = 12 [bracketleftbigg]

@rj + @

[bracketrightbigg]

! is the strain tensor. Note that this is

index notation, so summation over repeat indices is implied. The arrows serve as a reminder of the forward and backward nature of the waveelds and strains. For an isotropic medium, the dependence on the Lam parameters is

u =

[integraldisplay]

[circleplustext]

h !vi[notdef] vi+

Eii[notdef]

Ejj + 2!

Eij [notdef]

id3r. (7)

If the inversion is for VP and VS rather than for the Lam parameters , [notdef], we replace them:

[notdef] = V 2S + 2VS VS (8) =

V 2P 2V 2S[parenrightBig]

+ 2VP VP 4VS VS (9)

Replacing [notdef] and in Eq. (7), we have

u = (10)

[integraldisplay]

[circleplustext]

vi[notdef] vi +

V 2P 2V 2S

[notdef]

Eii[notdef]

Ejj + V 2S[notdef]

Eij [notdef]

Eij

[parenrightBig]

+ VP

2VP [notdef]

Eii[notdef]

Ejj

[parenrightBig]

+ VS

4VS [notdef]

Eii[notdef]

Ejj + 2VS [notdef]

Eij [notdef]

Eij

[parenrightBig][bracketrightBig] d3r.

Equation 11 shows that VP-kernels are straightforward to calculate because they only involve the diagonal elements of the strain tensor, which is equivalent to the divergence of the dis

placements:!

Eii = @

@ri = r!u. Hence for P-wave tomography,

346 S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms

8 S. C. Stahler et al.: Finite Frequency Measurements of triplicated waveforms

10 second dominant period

20 second dominant period

Fig. 5. Traveltime sensitivity kernels for vP, calculated by the Axisem spectral element code. Sensitivities are for a cross-correlation traveltime mist, in the time window of 5 s before to 15 s after the estimated arrival. Dominant period T is 10 s for the upper twoand 20 s for the lower two rows. Distances (from top left): 22 , 24 , 26 , 30 , 35 , 40 . Background model: IASP91; source:receiver: z-component. The 410-discontinuity is marked by the dashed line, the 660 by the dashed-dotted line. Note that most kernels feature the famous doughnut hole, which is lled by the sensitivities to the two discontinuities.

Fig. 5. Traveltime sensitivity kernels for [Delta1]VP/VP, calculated by the Axisem spectral element code. Sensitivities are for a cross-correlation traveltime mist, in the time window of 5 s before to 15 s after the estimated arrival. Dominant period is 10 s for the upper two rows, and 20 s for the lower two rows. Distances (from top left): 22 , 24 , 26 , 30 , 35 , 40 . Background model: IASP91; source: explosion; receiver: z-component. The 410-discontinuity is marked by the dashed line, the 660 by the dashed-dotted line. Note that most kernels do not feature the famous doughnut hole, which is lled in by the sensitivities to the two discontinuities.

Fig. 5. Traveltime sensitivity kernels for vP, calculated by the Axisem spectral element code. Sensitivities are for a cross-correlation traveltime mist, in the time window of 5 s before to 15 s after the estimated arrival. Dominant period T is 10 s for the upper two rows, and 20 s for the lower two rows. Distances (from top left): 22 , 24 , 26 , 30 , 35 , 40 . Background model: IASP91; source: explosion; receiver: z-component. The 410-discontinuity is marked by the dashed line, the 660 by the dashed-dotted line. Note that most kernels do not feature the famous doughnut hole, which is lled by the sensitivities to the two discontinuities.

Fig. 6. Nominal ray paths of regional P-waves in IASP91. None of the rays bottom in the lower part of the transition zone. Contrast this to the realistic (nite-frequency) sensitivities of 5, which extend broadly across the entire MTZ.

way of processing. The kernel formalism ensures the proper interpretation of the waveforms, as long as the actual measurements use the same window lengths and lters as

515 the waveeld computations.

3.4 Kernel gallery

A selection of 10 s and 20 s VP-kernels at dominant periods of 10 s and 20 s are plotted in Fig. 5, for an explosive source

520 at the surface and the z-component of the seismogram. Especially at 24 epicentral distance, the two frequency bands differ signicantly in their sampling of the MTZ. While the 20 s kernel samples the regions directly below the 410 and the 660, the 10 s kernel samples mainly the region above the

525 660. This is particularly striking when compared to the sensitivity of the rst arriving rays (Fig. 6), which is completely conned to the region directly below the discontinuities. The 22 and 26 kernels in the 10 s band have almost complementary sampling characteristics in the transition zone.

530 The 22 kernel is negative where the 26 one has its largest

which thus should yield good constraints on the MTZ.

The 30 to 40 kernels in the second and forth row look more

535 similar to teleseismic kernels, even though clear imprints of the MTZ remain even for a 40 kernel, resulting in an appearance quite different from the 40 kernel of the original Dahlen method (Dahlen et al., 2000).

Geodynamically interesting regions just above the discon-

540 tinuities, where effects of uid release might be present (Ohtani, 2005), can classically be sampled only by teleseismic waves. Since they traverse the MTZ at a steep incidence angle, their vertical resolution is quite limited in contrast to triplicated phases.

545 Topography on the discontinuities may also inuence the waveforms, an effect for which sensitivities may be computed from the same waveelds (Nissen-Meyer et al., 2007a; Colombi et al., 2012). Ultimately it will be desirable to separate topography effects from those of the volumetric velocity

550 structure, by integrating both kinds of kernels into a joint inversion.

4 The data sets

We identifed two regions that appear particularly suited to tomography of the transition zone using triplicated body-

555 waves: North America and Europe. They are densely instrumented, decently surrounded by earthquake sources, and large enough for wave paths to penetrate the MTZ on their way from source regions to receivers.

4.1 North America

560 Since the advent of USArray, North America is clearly the best-instrumented large landmass on Earth. USArray stations are spaced by 70 km on a regular grid and deliver

superb broadband waveforms.The data from the rst year of installment already brought

565 new insights into the subduction history of the Farallon plate and the formation of the Rocky mountains (Sigloch, 2011) or to uid transport in the Gorda subduction system (Cao and Levander, 2010). The recent move of the deployment into the Great Plains has brought the seismicities of the Guer-

570 rero subduction and along the western Canadian margin into distances of 15 to 30 , creating a large number of crossing paths in triplication range. USArray recordings are supplemented by permanent networks in the U.S., Canada, and Mexico, and by data from temporary experiments. All North

575 American data was downloaded from the IRIS DMC. Currently IRIS delivers about 2000 global stations per event, of which 1400 are located within larger North America, i.e.

between 180 W and 45 W in the northern hemisphere.We found that between 01/01/2001 and 31/12/2011, 92

580 regional earthquakes generated triplicated P-wave record-

Fig. 7. Source and stations in two regional data sets for North America and Europe.The rays connect source-receiver combinations at distances between 15 and 35 degrees. Earthquake locations are shown as beachballs. Station colour codes for the goodness of t between to observed and modelled waveforms at each station. Red denotes median cross-correlation coefcent of CCmax0.8, which we consider a reasonable threshold for accepting any individual measurement, orange and yellow colors indicate a lower median t, which means only some waveforms at this station may be used. Marker size is proportional to the number of events recorded by the station.

S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms 347

length should be chosen with care. A window that is too long will contain more noise than necessary, and the diameter of the kernel will grow, since longer detours from the direct path are allowed. The outer parts of the kernel will contain the higher-order Fresnel zones, which oscillate rapidly and are unlikely to meaningfully contribute in the inversion.Windows too short will contain too little signal and be more prone to cycle skips. The window length should probably be no shorter than twice the dominant period of the bandpass lter. It will not be possible to calculate kernels separately for each of the triplicated phases, since in real data they overlap in time (cf. Fig. 3). However, it is neither necessary or even desirable to revert to this Dirac-type, ray-theoretical way of processing. The kernel formalism ensures the proper interpretation of the waveforms, as long as the actual measurements use the same window lengths and lters as the wave-eld computations.

3.4 Kernel gallery

A selection of VP-kernels at dominant periods of 10 s and 20 s are plotted in Fig. 5, for an explosive source at the surface and the z-component of the seismogram. Especially at 24 epicentral distance, the two frequency bands differ signicantly in their sampling of the MTZ. While the 20 s kernel samples the regions directly below the 410 and the 660, the 10 s kernel samples mainly the region above the 660. This is particularly striking when compared to the sensitivity of the rst arriving rays (Fig. 6), which is completely conned to the region directly below the discontinuities. The 22 and 26 kernels in the 10 s band have nearly complementary sampling characteristics in the transition zone. The 22 kernel is negative where the 26 one has its largest positive values.

With dense arrays like the USArray, we obtain many 22 and 26 recordings for the same earthquake, which thus should yield good constraints on the MTZ. The 30 to 40 kernels in the second and forth row look more similar to teleseismic kernels, even though clear imprints of the MTZ remain even for a 40 kernel, resulting in an appearance quite different from the 40 kernel of the original Dahlen method (Dahlen et al., 2000). Geodynamically interesting regions just above the discontinuities, where effects of uid release might be present (Ohtani, 2005), can classically be sampled only by teleseismic waves. Since they traverse the MTZ at steep incidence angles, their vertical resolution is quite limited, in contrast to triplicated phases. Topography on the discontinuities may also inuence the waveforms, an effect for which sensitivities may be computed from from the same wave-elds (Nissen-Meyer et al., 2007a; Colombi et al., 2012). Ultimately it will be desirable to separate topography effects from those of the volumetric velocity structure, by integrating both kinds of kernels into a joint inversion.

Fig. 7. Source and stations in two regional data sets for North America and Europe. The rays connect source-receiver combinations at distances between 15 and 35 degrees. Earthquake locations are shown as beachballs. Station colour codes for the goodness of t between to observed and modelled waveforms at each station. Red denotes median cross-correlation coefcient of CCmax > 0.8, which we consider a reasonable threshold for accepting any individual measurement. Orange and yellow colours indicate a lower median t, which means that only some waveforms at this station may be used. Marker size is proportional to the number of events recorded by the station.

4 The data sets

We identied two regions that appear particularly suited to tomography of the transition zone using triplicated body-waves: North America and Europe. They are densely instrumented, decently surrounded by earthquake sources, and large enough for wave paths to penetrate the MTZ on their way from source regions to receivers.

www.solid-earth.net/3/339/2012/ Solid Earth, 3, 339354, 2012

348 S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms

S. C. Stahler et al.: Finite Frequency Measurements of triplicated waveforms 7

2000 global stations per event, of which 1400 are located

in larger North America, i.e. between 180 W and 45 W in the Northern Hemisphere.

We found that between 01/01/2001 and 31/12/2011, 92 regional earthquakes generated triplicated P-wave recordings of acceptable signal-to-noise ratio. This yielded 26 016 unique, acceptable wave paths and broadband waveforms in total. We applied Gabor bandpass lters (bandwidth one octave) at centre periods of 20 s, 10 s, 5 s, 2.5 s, 1.25 s.Since we make cross-correlation measurements, the correlation coefcient between observed (ltered) waveforms and their synthetics using IASP91 serves as the primary measure of goodness of t. We denote by T the time shift that maximizes the cross-correlation between observed seismogram and matched lter as in Eq. (4), and by CCmax the correlation coefcient at this optimum time shift. Hence CCmax acts as a quality measure, and T is the actual nite-frequency observable to be interpreted by tomography. Previous experience (Sigloch and Nolet, 2006; Sigloch, 2008) lets us assume that CClim 0.8 is a good threshold for acceptance. If

we only accept ltered waveforms with CCmax 0.8, 42 949

out of 130 080 remain (Table 1). Fitting waveforms in the ve separate frequency bands thus increases the number of usable data more than fourfold, compared to tting only the broadband seismograms. In particular, we can often accept at least one passband measurement even when the CCmax of the broadband measurement is clearly too low this salvages many important wave paths. For teleseismic waves, another frequency band below 20 s might reasonably be added, but for regional P-waves this is less interesting, since the 20 s kernels already ll up the entire depth range of the transition zone (see Fig. 5).

Figure 8 shows a comparison between matched lters and real waveforms for the seismograph station at Mount St. Helens crater, which is by no means an ideal station. Still three out of ve bands have CCmax > 0.8 and could contribute to tomography. These seismograms were calculated using IASP91 velocity model. For an actual tomography of Northern America it may be benecial to calculate matched lters and kernels in a regional 1-D model; for a recent model and a comparison between previous results see Chu et al. (2012).

Figure 9 shows that the average absolute traveltime anomaly T is similar in all bands, with a standard deviation of 2.4 s. From our experience with teleseismic P-waves, we

know that a large part of this T signal is due to mislocated sources rather than to mantle heterogeneity; both contributions are jointly inverted for by our tomography software.

These T measurements on triplicated data will be embedded into a larger data set of teleseismic P-wave data (Sigloch, 2011). The latter have better azimuthal coverage, since they include earthquakes from the east (mid-Atlantic ridge and southern Europe). Unfortunately, very few seismic sources east of North America are located at regional distances.

Solid Earth, 3, 339354, 2012 www.solid-earth.net/3/339/2012/

465

Fig. 4. The t between a triplicated P-wave and its synthetic, ltered into frequency passbands. The epicentral distance is 18.6 , so

that the waveform contains triplications from both the 410 and 660 discontinuities. This example uses the same broadband waveforms as seen in Fig. 3; the seismic station is VALT.

440 3.3 Spherical Earth kernels

The calculation of sensitivity kernels at a dominant period of 5 s requires around 105 CPUh when a full 3-D forward solver like SPECFEM (Komatitsch and Tromp, 2002) is used (Colombi et al., 2012). For a realistic global tomography

445 using > 105 waveforms, the calculation cost for individual source-receiver kernels would be 1010 CPUh, which is

completely prohibitive. Even if the adjoint method is invoked such that only the number of sources contributes to the computational cost, each iteration would require 108 CPUh

450 which is beyond currently available resources on state-ofthe-art supercomputer systems.

In a spherically symmetric background model, this cost can be reduced dramatically: 3-D waveelds in a layered Earth can be computed at the cost of the equivalent 2-D waveelds,

455 since the symmetry implies that the azimuthal direction may be calculated analytically (Nissen-Meyer et al., 2007a). Forward waveelds are pre-computed for reasonable depth increments (e.g. increments of 1 km for depths from 0 to 100 km, and increments of 10 km from 100 to 700 km, re-

cated at the surface or within one P-wavelength of it.

The 160 forward calculations need to be done four times (Nissen-Meyer et al., 2007a, p. 1057) for

1. a Mzz monopole source,

Fig. 8. The t between a triplicated P-wave and its synthetic, ltered into frequency pass bands. The epicentral distance is 18.6 , so that the waveform contains triplications from both the 410 and 660 discontinuities. This example uses the same broadband waveforms as seen in Fig. 3; the seismic station is VALT (Mount St. Helens crater rim).

4.1 North America

Since the advent of USArray, North America is clearly the best-instrumented large landmass on Earth. USArray stations are spaced by 70 km on a regular grid and deliver superb

broadband waveforms.

The data from the rst year of instalment already brought new insights into the subduction history of the Farallon plate and the formation of the Rocky mountains (Sigloch et al., 2008) or to uid transport in the Gorda subduction system (Cao and Levander, 2010). The recent move of the deployment into the Great Plains has brought the seismicities of the Guerrero subduction and along the western Canadian margin into distances of 15 to 30 , creating large numbers of crossing paths in triplication range. USArray recordings are supplemented by permanent networks in the US, Canada, and Mexico, and by data from temporary experiments. All North American data was downloaded from the IRIS data management center (DMC). Currently IRIS delivers about

2. a 12 (Mxx +Myy) monopole source,

3. a Mxz or Myz dipole source,

4. a Mxy or Mxx Myy quadrupole source.

Since the response of a station at azimuth to a Mxz-source

470 equals that of a station at /2 to a Myz-source, and re

sponse of a (Mxx Myy)-source at equals that of a Mxy-

source at /4, we can reconstruct the response of a spher

ically symmetric Earth to an arbitrary moment tensor from these four calculations.

475 The backward propagation needs to be done twice:

1. pz monopole source for the Z-component of the seismo-gram

2. px or py dipole source for the E- or N-component of the seismogram

480 Again, px or py are equivalent, if the receiver location is rotated by /2.

Using the 3-D to 2-D reduction strategy applied in the SEM-code Axisem (Nissen-Meyer et al. (2007b) and Nissen-Meyer et al. (2008)), the computation cost for a 3-D global

485 waveeld of dominant period T = 5 s is around 16 CPUh on an i7 machine, or 104 CPUh for the entire waveeld

library. Since the calculation grows at O(1/T 3), a library

of dominant period 1.25 s would still require less than 106 CPUh, which is easily feasible nowadays.

490 From these 642 pre-computed waveelds, kernels for any arbitrary time window in the seismogram can be calculated on the y in the frequency domain, which is memory intensive but computationally cheap.

Time window choices for these kernels are not limited to

495 any specic phase arrivals or wave types, although it is certainly judicious to chose parts of the seismogram that contain signicant seismic energy rather than noise. The window length should be chosen with care. A window that is too long will contain more noise than necessary, and the

500 diameter of the kernel will grow, since longer detours from the direct path are allowed. The outer parts of the kernel will contain the higher-order Fresnel zones, which oscillate rapidly and are unlikely to meaningfully contribute in the inversion. Too short time windows will contain too little

505 signal and be more prone to cycle skips. The window length should be no shorter than twice the dominant period of the bandpass lter.

It will not be possible to calculate the separate kernels for each of the triplicated phases, since for real data they overlap

From our experience with teleseismic P-waves, we know that

610 a large part of this T signal is due to mislocated sources rather than to mantle heterogeneity; both contributions are jointly inverted for by our tomography software. For the tomography, these T measurements on triplicated data will be embedded into a larger data set of teleseismic P-wave data

615 (Sigloch, 2011). The latter have better azimuthal coverage, since they include earthquakes from the east (mid-Atlantic ridge and southern Europe). Unfortunately, very few seismic sources to the east of North America are located at regional distances.

620 4.2 Europe

Compared to Northern America, the European seismic networks are less homogeneous, even though the NERIES initiative brought a great step forward. We downloaded our waveforms from the ORFEUS DMC. This archive currently

625 contains around 400 seismic stations, mainly in Europe, but also several temporary installations by European agencies, e.g. in Indonesia. The data coverage is quite uneven. Several large countries such as France and Poland are sparsely instrumented, other networks do not share their data yet. There

630 have been several large temporary installations in Central Europe and Scandinavia, data which would be very interesting but are being released only slowly. The IRIS DMC does not maintain data from any European broadband stations be-

Fig. 8. Measured traveltime anomalies T for all paths in the North American data set, in all ve frequency bands (distinguished by color). Only data whose t exceeds CCmax > 0.7 are plotted. The

histograms are normalized so that at T = 0, each bar has value 1.

S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms 349

Table 1. Number of acceptable traveltime measurements in the various frequency bands, as a function of the chosen rejection threshold CClim. In total 26016 (NA) and 4916 (Europe) broadband seismo-grams were available.

Northern American data set

CClim 0.70 0.75 0.80 0.85 0.90 0.95

20 s 19 897 18 563 16 717 14 225 10 553 5018 10 s 17 646 15 440 12 995 10 124 6317 2143 5 s 12 691 10 331 7900 5452 2861 7292.5 s 8180 5852 3856 2235 1056 2201.25 s 3561 2368 1481 773 258 13

total 61 975 52 554 42 949 32 809 21 045 8123

broadband 14 552 12 177 9949 7141 3715 1080

European data set

CClim 0.70 0.75 0.80 0.85 0.90 0.95

20 s 3450 3161 2840 2423 1891 775 10 s 2953 2477 1922 1261 577 84 5 s 1556 1037 598 264 83 52.5 s 882 483 233 91 19 01.25 s 277 164 72 29 5 0

total 11 627 9311 7042 4840 2876 886

broadband 2509 1989 1377 772 301 22

4.2 Europe

Compared to North America, the European seismic networks are less homogeneous, even though the NERIES initiative brought a great advance. We downloaded our waveforms from the ORFEUS DMC. This archive currently contains around 400 seismic stations, mainly in Europe, but also several temporary installations by European agencies, e.g. in Indonesia. Figure 7 shows that data coverage is quite uneven. Several large countries like France and Poland are sparsely instrumented, other networks do not share their data yet. There have been several large temporary installations in Central Europe and Scandinavia, data which would be very interesting but are being released only slowly. The IRIS DMC holds hardly any broadband stations above and beyond the ORFEUS DMC. On the other hand, the geometry and seismicity of Europe is more favourable than for North America. Hellenic seismicity is recorded over a large azimuthal range from Spain to Russia and produces many crossing paths, together with Western Mediterranean seismicity recorded in Central Europe, and seismicity on the North Atlantic ridge between Iceland and Svalbard. One of the most interesting features in the transition zone are the slab remnants below the Northern Balkans and the Carpathians, which cannot be easily connected to the established subduction regimes. Sampling of the MTZ by triplicated waves is very good in this region, and should help in further illuminating its mantle processes. A clear difference to the North American data

Fig. 9. Measured traveltime anomalies T for all paths in the North American data set, in all ve frequency bands (distinguished by colour). Only data whose t exceeds CCmax > 0.8 are plotted. The histograms are normalized so that at T = 0, each bar has value 1.

is that few earthquakes above magnitude 6 are recorded in Europe. Since we obtain best correlations between data and synthetics at around magnitude 6.5, the data quality is generally lower than in North America. In total, the number of available stations in larger Europe (dened as the region between 45 W and 75 E in the northern hemisphere) is 390.

Our dataset contains 66 earthquakes, which recorded 4916 broadband seismograms (unique wave paths). 7042 out of 24 580 passband measurements have a data-synthetic t that exceeds CCmax > 0.8. Compared to the North America, we obtain about ve times fewer waveforms, and seven times fewer acceptable T measurements. This lower average signal quality seems to be mainly due to the weaker seismicity around Europe.

4.3 Information content of triplicated P-waveforms

Before embarking on tomographic inversions, we want to convince ourselves that triplicated waveforms do indeed contain coherent and usable structural information. The broad footprint of USArray may be sliced up into dense seismic proles at various back-azimuths, each featuring dozens of stations. Here we consider one such quasi linear section, generated by an earthquake from Guerrero, Mexico. Figure 10 shows the P-waveforms of 84 displacement seismograms in a range between 22 and 35 distance. The traces are color coded, where green means zero displacement, blue is negative displacement, and red is positive. Traces are time-aligned on the IASP91-predicted arrivals of the rst P-phase.

The rst arrival in the real seismograms occurs around +2 s for all traces, a systematic bias w.r.t. to IASP91 that is most likely due to source mislocation. The blue triangular move-outs (e.g. to +7 s at 21 distance) are the triplicated phases of

www.solid-earth.net/3/339/2012/ Solid Earth, 3, 339354, 2012

350 S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms

P. Starting at +10 s, the whole P-arrival sequence is echoed (also in blue) this is the depth phase pP, in itself triplicated, for this shallow event (12 km deep).

If the Earth were truly spherically symmetric, the P-arrivals would be aligned smoothly if not along a vertical line of t = 0, then along some other smooth line of steep

move-out. This is clearly not the case, signalling the kind of lateral mantle heterogeneity that seismic tomography targets.Strong unevenness is for example observed between 25 and 31 distance, where the P-onset varies between 0 s and +3 s.

The largest delays are present between 29 and 31 distance. This might be explained by a low-velocity anomaly directly below the 660, since only waves of this distance range bottom there (see Fig. 6). Recordings at more distant stations would be less inuenced by this hypothetical anomaly, since they traverse it at a steeper angle, with correspondingly lower sensitivity. However, a second observation leads to another explanation: The second triplication (here arriving at 5 s at

28 ), represents the phase bottoming directly above the 660.

According to IASP91, it should be recorded only to a distance of 27.8 , but here we observe it to distances exceeding 29 . This could be explained by a depression of the 660-km discontinuity to a depth of 680 km, which would also explain the delay of the rst phase at this distance. A depressed 660 in this region of subduction is plausible, since the lower temperatures of the slab would shift the phase transition to higher pressures and deeper depths. Indeed, studies from SS precursors Houser et al. (2008) showed a 660 depressed to 668 671 km in Northern Mexico.

The second subplot of Fig. 10 shows several stations at 28 to 29 distance, where the second triplication should not be recorded. We do observe the phase, but only at some of the stations (at +4 to +5 s), which suggests that the depression of the 660 is quite localized. From the azimuthal spread of the stations, we can estimate their bottoming points to be spaced only by tens of kilometres, and yet the imprint of the 660 in the seismograms is quite variable. Since the maximum frequencies in these waveforms are less than 1 Hz, the scale of the undulations on the 660 must be close to the resolution limit of these waves. The aforementioned precursor study lacked the resolution to resolve such small features.

5 Discussion

Using triplicated P-waves lls an important gap in waveform tomography. Teleseismic waves have relatively poor depth resolution in the MTZ, and little sensitivity to the discontinuities themselves. Since regional waves have this resolution, we hope to greatly increase the resolution of future P-wave models in the transition zone. So far, triplicated body waves have hardly ever been used for tomography, since they are not well modelled by classical ray theory. The appearance of the actually measured waveform has a clear nite-frequency character rather than resembling a sequence of Dirac pulses.

Fig. 10. Evidence of 3-D structural information contained in trip-licated P-waves. Top: we extract a quasi 1-D prole of 84 USArray stations, recording an earthquake in Mexico (Event: 2009/04/27 16:46:28, Mag: 5.8, Lat: 17.03, Lon: 99.45, Depth: 35 km). Middle: Section of broadband seismograms at triplication distances (between 22 and 35 , green in the map plot). Z-component, colour coded, green is zero displacement, blue negative, red positive. The traces are normalized separately, and aligned to the rst arriving P-wave as predicted by IASP91. See text for discussion. Bottom: Seismograms from the ensemble of stations at quasi-constant distance of 28 to 29 (red triangles in map plot). The second P-triplication is observed on some of the traces (at t = 45 s) but not on others, pointing to variations on the 660-km discontinuity.

Our choice of mist criterion is the cross-correlation traveltime effectively a phase shift measurement, given our relatively narrow pass-bands. Cross-correlation is the optimal strategy for the detection and estimation of a known signal (the synthetic waveform or matched lter) in a noisy version of the same signal (the measured waveform), where the noise is assumed to be white additive Gaussian noise in each

Solid Earth, 3, 339354, 2012 www.solid-earth.net/3/339/2012/

S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms 351

passband. Intuitively it might seem that cross-correlation may not be a suitable mist for triplicated waves, since for all but the highest frequencies, the time window will necessarily encompass all three triplicated phases, each of which has different spatial sensitivity. Hence calculating one traveltime delay on all three overlapping phases may seem unphysical.We believe that this understanding of the cross-correlation mist is misguided by (ray-theoretical) intuition. While it is true that one cross-correlation applied to a broadband signal only calculates one delay time for the whole broadband time series, this is actually not what we propose to do. When splitting the signal into multiple frequency bands k and calculating separate Tk, several, ideally independent measurements are done on the waveform. The corresponding sensitivity kernels describe the sensitivity of the model towards each of these measurements. It does not describe the sensitivity towards the traveltime of one particular (ray-theoretical) phase. The kernel computation formalism ensures that the proper sensitivity for each window and lter is computed, whatever they may be, meaning a time window containing triplicated waves can be used just as any other time window.One should just not expect a ray-theoretically intuitive meaning corresponding to each measurement.

The calculation of full waveelds in a laterally heterogeneous 3-D Earth model is still prohibitively expensive for frequencies above 0.05 Hz. Hence waveform tomography must currently choose between one of two compromises:

1. In a laterally heterogeneous Earth model, calculate low frequency waveelds (< 0.05 Hz), since high frequencies are not affordable. This means focusing on surface waves and the low-frequency part of body waves.Due to the large wavelengths of body waves, their low-frequency part offers little resolution in the mantle transition zone (already at a dominant period of 20 s, the kernel lls the entire transition zone). The high cost of waveeld calculations mandates that they need not be done too often. The adjoint method proposed by Tarantola (1984), and applied in continental-scale seismology by Tromp et al. (2005), Fichtner et al. (2009) and Zhu et al. (2012), offers an efcient solution by calculating waveelds only once per source, resulting in composite, so-called event kernels, which indicate the descent direction for the (linearised) gradient search. Several iterations are possible and customary, since waveelds can be computed in arbitrary Earth models, in particular also the updated ones.

2. In a spherically symmetric Earth model, calculate kernels that span any or all parts of the seismically relevant frequency spectrum. This broadband capacity has dened nite-frequency tomography since Dahlen et al. (2000). Nissen-Meyer et al. (2007b) showed that full waveeld kernel computations are feasible for frequencies up to 1 Hz, due to the extreme computational savings that result from smart exploitation of the spherical

symmetry. The use of high-frequency waves promises accordingly higher image resolution. The method is so efcient that all source-to-receiver kernels are explicitly calculated, turning the problem into one of matrix inversion, for which powerful analysis concepts and computational tools are well known. The disadvantage is that several iterations are not possible, since the Earth model would lose its spherical symmetry after the rst update.This might be a serious drawback in media with heterogeneity of strong magnitude, where the single-scattering (Born) approximation starts breaking down, but where several linearised iterations might still lead signicantly closer to the global mist minimum than a single one.

We believe that the nite-frequency approach is currently better suited to tomography of the upper mantle and the transition zone, since (a) being able to use the highly resolving body waves across their entire frequency range should be a big advantage, and (b) all previous studies let us expect only relatively weak perturbations from a layered background model like IASP91, on the order of a few percent, so that the single-scattering approximation should be completely adequate.

A global reference model may not be best suited for a regional study. Rather, careful selection of a suitable regional reference model is a crucial step of a tomography.

We note that the mixed use of forward modelling codes (WKBJ, reectivity method) has historical reasons and reects an unnished (though functional) stage of development. Ultimately all steps will be carried out with the most complete method, Axisem. For inversion of source time functions from teleseismic data, WKBJ has been an efcient tool since (Sigloch and Nolet, 2006). It is completely adequate for teleseismic waves, and this STF inversion step is treated as independent from the modelling of triplicated waveforms. Reectivity is used for efcient forward modelling of the triplicated broadband seismograms (WKBJ cannot easily compute triplicated phases, and Axisem is not yet set up to do true broadband computations efciently for the number of computations needed). In order to compute sensitivity kernels, the full waveeld is needed, and we obtain it from full numerical forward modelling using Axisem (but not yet routinely up to frequencies of 1 Hz).

Analysis of triplicated waveforms has so far been mostly applied to deep events (Tajima and Grand, 1995; Tajima et al., 2009), in order to separate the inuence of the discontinuities from the depth phases. We have shown that shallow events can be modelled when source parameters are carefully estimated. Routine catalogue estimates (Global CMT (Dziewoski et al., 1981) or NEIC) do not deliver all the parameter we need (source time function), or not to the required accuracy (e.g., source depth).

However, using our own source inversion results from tele-seismic P-waves, we can model the sources sufciently well for our purposes. The ability to use shallow earthquakes

www.solid-earth.net/3/339/2012/ Solid Earth, 3, 339354, 2012

352 S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms

enlarges the data base enormously, since most earthquakes occur shallower than 30 km depth. In particular, good tomographic resolution requires good azimuthal coverage by sources, but few regions on Earth are surrounded by subduction zones to generate deep earthquakes from all directions.

The regional data sets are promising, both for North America and for Europe. They will be seamlessly embedded into a global inversion that also contains teleseismic P-wave measurements Sigloch (2011). A regional tomography of Europe can thus still benet from events in eastern Asia, i.e. beyond the triplicated range. Hence the fact that the regional seismocity around Europe is weaker than in North America, is compensated by the superiour azimuthal coverage at teleseismic distances. Nevertheless, the European network is nowhere as dense as the USArray, which will make any American mantle structure far better resolved in the foreseeable future. Average waveform cross-correlation is poorer in the European data set than in the North American one. Three possible explanations come into mind:

1. Station quality might be lower. We do not believe this is generally the case, even though the average distance to the nearest coast is smaller in Europe than in North America.

2. European mantle structure could be more complicated than under Northern America, generating a larger mismatch between observed and modelled waveforms. The tectonic history of Europe is often thought to be more complex, but tomographic studies since the advent of USArray have also revealed very heterogeneous mantle structure under North America (Pavlis et al., 2012;Becker, 2012). Hence it is unlikely that the true seismic velocity structure in the mantle under Europe deviates signicantly more from a layered model than under North America.

3. Earthquake sources around Europe may be less suited.

We think this is the main issue. In North America, we can use rather strong strike-slip events along the west coast, and numerous deep events along the Guerrero subduction zone, many of them exceeding magnitude 6.European seismicity is weaker in magnitude and tends to consist of complicated events in the Aegean subduction and along the Anatolian fault; strike-slip events along the Atlantic ridge are rather weak. Since some of these events are not contained in the IRIS WILBER archive and thus only the ORFEUS stations were available to us, our source inversion for them might be more error prone as well.

The seismic section from the Mexican earthquake demonstrates the high signal quality of USArray recordings (Fig. 10). The discussed strong imprint of a depressed 660 in this signal shows that we may need to be careful to properly parametrize the inversion such that depressed or elevated

discontinuities are detected and become part of the tomography result, rather than smearing into bulk velocity structure. In order to separate the two effects, we are considering the use of boundary topography kernels (Colombi et al., 2012), in addition to the volumetric velocity kernels shown in Fig. 5.

6 Conclusions

We conclude that regional body waves should be usable and useful for waveform tomography. We plan to invert these data using nite-frequency tomography and kernels from full numerical waveeld computations (Nissen-Meyer et al., 2007b), but our conclusions about the nature of the data have broader validity.

After careful deconvolution of the source time function and other source parameters from teleseismic waveforms, we obtain good cross-correlations between observed and modelled triplicated seismograms, across the broadband range and even for shallow sources. Due to the much more complex nature of the waveforms, these ts are lower than what we obtain for teleseismic P-waves, but are still sufcient to assemble two decently sized tomography data sets for North America and Europe. The inclusion of these data greatly increase and complement the sensitivity to transition zone structure, and in particular to the discontinuities at 410 km and 660 km, which so far must be investigated using waves of far lower signal-to-noise ratio. The abundance and high quality of data from USArray make the transition zone under North America the natural target for a rst waveform inversion using triplicated P-waves.

Acknowledgements. We thank F. Tajima and B. L. N. Kennett for sharing their experience on triplicated waves. G. Nolet, M. v. Driel and an anonymous reviewer helped improving the manuscript in the Solid Earth Discussion stage. All waveform data came from the IRIS and ORFEUS data management centres. S. C. S. was funded by the Munich Center for Advanced Computing (MAC/IGSSE).

Edited by: R. Carbonell

References

Becker, T.: On recent seismic tomography for the western United States, Geochem. Geophys. Geosyst., 13, 111, http://dx.doi.org/10.1029/2011GC003977

Web End =doi:10.1029/2011GC003977 , 2012.

Cao, A. and Levander, A.: High-resolution transition zone structures of the Gorda Slab beneath the western United States: Implication for deep water subduction, J. Geophys. Res., 115, B07301, http://dx.doi.org/10.1029/2009JB006876

Web End =doi:10.1029/2009JB006876 , 2010.

Chaljub, E. and Tarantola, A.: Sensitivity of SS precursors to topography on the upper-mantle 660-km discontinuity, Geophys. Res.Lett., 24, 26132616, http://dx.doi.org/10.1029/97GL52693

Web End =doi:10.1029/97GL52693 , 1997.

Chapman, C.: A new method for computing synthetic seismograms,

Geophys. J. Roy. Astr. S., 54, 481518, 1978.

Solid Earth, 3, 339354, 2012 www.solid-earth.net/3/339/2012/

S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms 353

Chu, R., Schmandt, B., and Helmberger, D. V.: Upper mantle P velocity structure beneath the Midwestern United States derived from triplicated waveforms, Geochem. Geophys. Geosyst., 13, 121, http://dx.doi.org/10.1029/2011GC003818

Web End =doi:10.1029/2011GC003818 , 2012.

Colombi, A., Nissen-Meyer, T., Boschi, L., and Guardini, D.: Seismic sensitivity to global boundary topography, Geoph. J. Int., 191(2), 832848, http://dx.doi.org/10.1111/j.1365-246X.2012.05660.x

Web End =doi:10.1111/j.1365-246X.2012.05660.x , 2012.

Dahlen, F. A., Hung, S. H., and Nolet, G.: Frchet kernels for nite-frequency traveltimes I. Theory, Geophys. J. Int., 141, 157 174, http://dx.doi.org/10.1046/j.1365-246X.2000.00070.x

Web End =doi:10.1046/j.1365-246X.2000.00070.x , 2000.

Deuss, A.: Global observations of mantle discontinuities using SS and PP precursors, Surveys in geophysics, 30, 301326, http://dx.doi.org/10.1007/s10712-009-9078-y

Web End =doi:10.1007/s10712-009-9078-y , 2009.

Deuss, A., Redfern, S. A. T., Chambers, K., and Woodhouse, J. H.:

The Nature of the 660-Kilometer Discontinuity in Earths Mantle from Global Seismic Observations of PP Precursors, Science, 311, 198201, http://dx.doi.org/10.1126/science.1120020

Web End =doi:10.1126/science.1120020 , 2006.Dziewoski, A. M.: Preliminary reference Earth model, Phys. Earth

Planet. In., 25, 297356, http://dx.doi.org/10.1016/0031-9201(81)90046-7

Web End =doi:10.1016/0031-9201(81)90046-7 , 1981.

Dziewoski, A. M., Chou, T.-A., and Woodhouse, J. H.: Determination of Earthquake Source Parameters From Waveform Data for Studies of Global and Regional Seismicity, J. Geophys. Res., 86, 28252852, http://dx.doi.org/10.1029/JB086iB04p02825

Web End =doi:10.1029/JB086iB04p02825 , 1981.

Fichtner, A., Kennett, B. L. N., Igel, H., and Bunge, H.-P.: Full seismic waveform tomography for upper-mantle structure in the Australasian region using adjoint methods, Geophys. J. Int., 179, 17031725, http://dx.doi.org/10.1111/j.1365-246X.2009.04368.x

Web End =doi:10.1111/j.1365-246X.2009.04368.x , 2009.Fuchs, K. and Mller, G.: Computation of Synthetic Seismograms with the Reectivity Method and Comparison with Observations, Geophys. J. Roy. Astr. S., 23, 417433, 1971.

Grand, S. P.: Upper mantle shear structure of North America, Geophys. J. Roy. Astr. S., 76, 399438, http://dx.doi.org/10.1111/j.1365-246X.1984.tb05053.x

Web End =doi:10.1111/j.1365- http://dx.doi.org/10.1111/j.1365-246X.1984.tb05053.x

Web End =246X.1984.tb05053.x , 1984.

Grand, S. P.: Mantle shear structure beneath the Americas and surrounding oceans, J. Geophys. Res., 99, 1159111621, 1994.Grand, S. P.: Mantle shearwave tomography and the fate of subducted slabs, Philos. T. R. Soc. A, 360, 2475, 2002.

Houser, C., Masters, G., Flanagan, M. and Shearer, P. M.: Determination and analysis of long-wavelength transition zone structure using SS precursors, Geophys. J. Int., 174, 178194, http://dx.doi.org/10.1111/j.1365-246X.2008.03719.x

Web End =doi:10.1111/j.1365-246X.2008.03719.x , 2008.

Kennett, B. L. N.: The Seismic Waveeld: Introduction and theoretical development, Cambridge University Press, 2001.

Kennett, B. L. N. and Engdahl, E. R.: Traveltimes for global earthquake location and phase identication, Geophys. J. Int., 105, 429465, http://dx.doi.org/10.1111/j.1365-246X.1991.tb06724.x

Web End =doi:10.1111/j.1365-246X.1991.tb06724.x , 1991.Komatitsch, D. and Tromp, J.: Spectral-element simulations of global seismic wave propagation I. Validation, Geophys. J. Int., 149, 390412, 2002.

Marquering, H., Dahlen, F. A. and Nolet, G.: Three-dimensional sensitivity kernels for nite-frequency traveltimes the banana-doughnut paradox, Geophys. J. Int., 137, 805815, http://dx.doi.org/10.1046/j.1365-246x.1999.00837.x

Web End =doi:10.1046/j.1365-246x.1999.00837.x , 1999.

Melbourne, T. I. and Helmberger, D. V.: Whole mantle shear structure beneath the East Pacic Rise, J. Geophys. Res., 107, 2204, 2002.

Montelli, R., Nolet, G., Dahlen, F. A., and Masters, G.: Finite-frequency tomography reveals a variety of plumes in the mantle,

Science, 303, 33843, http://dx.doi.org/10.1126/science.1092485

Web End =doi:10.1126/science.1092485 , 2004. Nissen-Meyer, T., Dahlen, F. A., and Fournier, A.: Spherical-earth

Frchet sensitivity kernels, Geophys. J. Int., 168, 10511066, http://dx.doi.org/10.1111/j.1365-246X.2006.03123.x

Web End =doi:10.1111/j.1365-246X.2006.03123.x , 2007a. Nissen-Meyer, T., Fournier, A., and Dahlen, F. A.: A two-dimensional spectral-element method for computing spherical-earth seismogramsI. Moment-tensor source, Geophys. J. Int., 168, 10671092, http://dx.doi.org/10.1111/j.1365-246X.2006.03121.x

Web End =doi:10.1111/j.1365-246X.2006.03121.x , 2007b.

Nissen-Meyer, T., Fournier, A., and Dahlen, F. A.: A 2-D spectral-element method for computing spherical-earth seismograms-II. Waves in solid-uid media, Geophys. J. Int., 174, 873888, http://dx.doi.org/10.1111/j.1365-246X.2008.03813.x

Web End =doi:10.1111/j.1365-246X.2008.03813.x , 2008.

Nolet, G.: A Breviary of Seismic Tomography: Imaging the Interior of the Earth and Sun, Cambridge University Press, 2008. Ohtani, E.: Water in the mantle, Elements, 1, 25, http://dx.doi.org/10.2113/gselements.1.1.25

Web End =doi:10.2113/gselements.1.1.25 , 2005.

Pavlis, G. L., Sigloch, K., Burdick, S., Fouch, M. J., and Vernon,F.: Unraveling the Geometry of the Farallon Plate: Synthesis of Three-dimensional Imaging Results from USArray, Tectono-physics, 532535, 82102, http://dx.doi.org/10.1016/j.tecto.2012.02.008

Web End =doi:10.1016/j.tecto.2012.02.008 , 2012.

Shearer, P. M. and Masters, G.: Global mapping of topography on the 660km discontinuity, Nature, 355, 791796, 1992.

Sigloch, K.: Multiple-frequency body-wave tomography, Ph.D. thesis, Princeton, 2008.

Sigloch, K.: Mantle provinces under North America from multifrequency P wave tomography, Geochem. Geophys. Geosyst., 12, 27 pp., http://dx.doi.org/10.1029/2010GC003421

Web End =doi:10.1029/2010GC003421 , 2011.

Sigloch, K. and Nolet, G.: Measuring nite-frequency body-wave amplitudes and traveltimes, Geophys. J. Int., 167, 271287, http://dx.doi.org/10.1111/j.1365-246X.2006.03116.x

Web End =doi:10.1111/j.1365-246X.2006.03116.x , 2006.

Sigloch, K., McQuarrie, N., and Nolet, G.: Two-stage subduction history under North America inferred from multiple-frequency tomography, Nat. Geosci., 1, 458462, http://dx.doi.org/10.1038/ngeo231

Web End =doi:10.1038/ngeo231 , 2008.

Tajima, F. and Grand, S.: Evidence of high velocity anomalies in the transition zone associated with southern Kurile subduction zone, Geophys. Res. Lett., 22, 31393142, http://dx.doi.org/10.1029/95GL03314

Web End =doi:10.1029/95GL03314 , 1995.

Tajima, F., Katayama, I., and Nakagawa, T.: Variable seismic structure near the 660 km discontinuity associated with stagnant slabs and geochemical implications, Phys. Earth Planet. Int., 172, 183 198, http://dx.doi.org/10.1016/j.pepi.2008.09.013

Web End =doi:10.1016/j.pepi.2008.09.013 , 2009.

Tape, C., Liu, Q., Maggi, A., and Tromp, J.: Adjoint tomography of the southern California crust., Science, 325, 988992, http://dx.doi.org/10.1126/science.1175298

Web End =doi:10.1126/science.1175298 , 2009.

Tarantola, A.: Inversion of seismic reection data in the acoustic approximation, Geophysics, 49, 12591266, http://dx.doi.org/10.1190/1.1441754

Web End =doi:10.1190/1.1441754 , 1984.

Thomas, C. and Billen, M. I.: Mantle transition zone structure along a prole in the SW Pacic: thermal and compositional variations, Geophys. J. Int., 176, 113125, 2009.

Tian, Y., Zhou, Y., Sigloch, K., Nolet, G., and Laske, G.: Structure of North American mantle constrained by simultaneous inversion of multiple-frequency SH, SS, and Love waves, J. Geophys. Res., 116, http://dx.doi.org/10.1029/2010JB007704

Web End =doi:10.1029/2010JB007704 , 2011.

Tromp, J., Tape, C., and Liu, Q.: Seismic tomography, adjoint methods, time reversal and banana-doughnut kernels, Geophys. J. Int.,

www.solid-earth.net/3/339/2012/ Solid Earth, 3, 339354, 2012

354 S. C. Stahler et al.: Finite-frequency measurements of triplicated waveforms

160, 195216, http://dx.doi.org/10.1111/j.1365-246X.2004.02453.x

Web End =doi:10.1111/j.1365-246X.2004.02453.x , 2005.

VanDecar, J. C. and Crosson, R. S.: Determination of teleseismic relative phase arrival times using multi-channel cross-correlation and least squares, Bull. Seismolog. Soc. Am., 80, 150169, 1990.Zhou, Y., Dahlen, F. A., and Nolet, G.: Three-dimensional sensitivity kernels for surface wave observables, Geophys. J. Int., 158, 142168, http://dx.doi.org/10.1111/j.1365-246X.2004.02324.x

Web End =doi:10.1111/j.1365-246X.2004.02324.x , 2004.

Zhou, Y., Nolet, G., Dahlen, F. A., and Laske, G.: Global upper-mantle structure from nite-frequency surface-wave tomography, J. Geophys. Res., 111, B04304, http://dx.doi.org/10.1029/2005JB003677

Web End =doi:10.1029/2005JB003677 , 2006.

Zhu, H., Bodag, E., Peter, D., and Tromp, J.: Structure of the European upper mantle revealed by adjoint tomography, Nat. Geosci., 5, 493498, http://dx.doi.org/10.1038/ngeo1501

Web End =doi:10.1038/ngeo1501 , 2012.

Solid Earth, 3, 339354, 2012 www.solid-earth.net/3/339/2012/

Word count: 14269

Show less

Abstract

Translate

Triplicated body waves sample the mantle transition zone more extensively than any other wave type, and interact strongly with the discontinuities at 410 km and 660 km. Since the seismograms bear a strong imprint of these geodynamically interesting features, it is highly desirable to invert them for structure of the transition zone. This has rarely been attempted, due to a mismatch between the complex and band-limited data and the (ray-theoretical) modelling methods. Here we present a data processing and modelling strategy to harness such broadband seismograms for finite-frequency tomography. We include triplicated P-waves (epicentral distance range between 14 and 30°) across their entire broadband frequency range, for both deep and shallow sources. We show that is it possible to predict the complex sequence of arrivals in these seismograms, but only after a careful effort to estimate source time functions and other source parameters from data, variables that strongly influence the waveforms. Modelled and observed waveforms then yield decent cross-correlation fits, from which we measure finite-frequency traveltime anomalies. We discuss two such data sets, for North America and Europe, and conclude that their signal quality and azimuthal coverage should be adequate for tomographic inversion. In order to compute sensitivity kernels at the pertinent high body wave frequencies, we use fully numerical forward modelling of the seismic wavefield through a spherically symmetric Earth.

Details

Title

Triplicated P-wave measurements for waveform tomography of the mantle transition zone

Author

Stähler, S C; Sigloch, K; Nissen-Meyer, T

Pages

339-354

Publication year

2012

Publication date

2012

Publisher

Copernicus GmbH

ISSN

18699510

e-ISSN

18699529

Source type

Scholarly Journal

Language of publication

English

DOI

https://doi.org/10.5194/se-3-339-2012

ProQuest document ID

1757024129

Triplicated P-wave measurements for waveform tomography of the mantle transition zone

Jump to:

Full text

Abstract

Details

Suggested sources