INVESTIGATION INTO 3D EARTH STRUCTURE AND SOURCES USING FULL SEISMIC WAVEFORMS

Seismograms are the result of the complex interactions between a seismic source, a propagation medium and the seismograph's response. Through the use of 3-dimensional modeling and full seismic waveform data, we quantify and minimize errors associated with the source and propagation medium within our data sets. We compile a new and unique earthquake catalog for the Middle East that is openly available to the public. We quantify the benefits of using a 3-dimensional model relative to a 1-dimensional model to minimizing error in earthquake moment tensors and identify where in the waveform 3-dimensional models outperform 1-dimensional models. Two new and unique 3-dimensional seismic wave speed models are computed for the Ontong Java plateau and eastern North American margin.Both models are significant improvements to the resolution of wave speed structures in the crust and upper mantle and provide new information for the evaluation of tectonic features.

Geosciences and would like to thank each and every member of the faculty and staff past and present; there are far too many people to name here. You started me on this journey and for that I am grateful. Entering graduate school, I went from sharing the tiniest office on campus with 3 other people, to moving into the only office on campus with a balcony; I guess it pays to stick around.
I would first and foremost like to thank my advisor, Brian Savage. Thank you for your non-stop support, guidance, patience, sense of humor and most importantly belief that I could do this. I'll miss spending days and days writing a unnecessarily long code, only to show you and watch as you condense my weeks worth of work into a single line. It was truly a joy working with you and you have taught me more than I ever thought possible. You have been not only a great advisor but also a great friend.
Thank you to my thesis committee, each one of you has made a significant contribution to not only this thesis but to my graduate career and how I think about the Earth. Yang, you brought me into GSO while I was still an undergraduate and introduced me to seismology. Thank you for all of your years of guidance, advisement and support. Chris, your interpretive dances showed me that sometimes the variables in equations actually do have a physical meaning. Thanks for keeping science fun. I'll especially miss the AGU nights at Lefty's O'Doul's and the more recent parties in the red barn. Karen, thank you for letting me participate in your class and for providing me the opportunity to meet a whole new group of people at Brown. Gopu your unique perspective helped me to think differently about how iii to do inverse problems.
Rob, sometimes I think you're the one keeping the whole campus grounded.
Thanks for all the distractions and your sense of humor, it was more welcome than you know. Katie you showed me that sometimes, albeit in very rare cases, chemistry is important. Thank you especially for the spectacular raspberry pies! To all of my office mates, residents of 2nd floor Horn Bldg. and past and present members of the SeismoLab group. I would especially like to mention Zhigang, Yong, Xiaofeng, Wei and Haiying, without your help and guidance I would not have gotten through the thousands (millions?) of lines of code I've run through the years. Shifra, for listening to me ramble for 4 years, without you listening and occasionally responding back I would have been talking to the wall more than society deems acceptable. To my mom and dad. My whole life you have supported and encouraged me to be the best I can be. Thank you for that.
iv PREFACE The following dissertation examines three distinct geologic settings throughout the Earth. It has been written in manuscript format and is broken into the following three manuscripts: Manuscript one, "A quantitative comparison between 1d and 3d source inversion methodologies: Application to the middle east", investigates the difference between synthetically generated seismograms when using a 1-dimensional or 3-dimensional starting model and how these differences effect the quality of results when inverting for earthquake source mechanisms. Abstracts reflecting the  , was used to estimate the time needed to completely remove a thermal anomaly given a suite of anomaly radius sizes. Two diffusivity, κ, values were used, 0.01 cm 2 s and 0.018 cm 2 s (Gibert et al., 2003). Diffusion time is given in million years (Ma). . A majority of events in the catalog (Table S1)    A gap in between the thick continental lithosphere and the much thinner oceanic lithosphere is aligned with low wave speed features possibly imaging edge-driven convection cells

Introduction
The ability to obtain reliable earthquake source solutions is a useful tool to any tectonic interpretation. Source mechanisms prove invaluable in the assessment of plate motions, accurate characterizations of faults, and defining regional stresses. Holt et al. (1991) showed how moment tensors can be related to the seismically re-leased strain rate, allowing an estimate of the regional tectonic strain rate. Strain rates calculated from earthquake moment tensors can aid in the evaluation of current GPS velocity fields, as well as an evaluation of hypotheses explaining regional tectonic framework. Obtaining accurate source depths and quantifying their uncertainty are an invaluable source of information providing an additional constraint on plate motions, especially in a region of complex geology with multiple emergent subduction zones, complex fold and thrust belts and thickened crust.
Furthermore, accurate source parameters are necessary to the improvement of 3D Earth models in full-waveform inversion methods. The use of full waveforms require well-constrained source parameters to avoid mapping source errors into updated Earth models as well as maximizing the number of measurements in the full-waveform inversions Maggi et al., 2009;.
We solve for source solutions using a full-waveform moment tensor methodology computed using synthetic seismograms from both 1-dimensional (1D) and 3-dimensional (3D) tomographic models at two frequency bands. With each inversion we compute a variance reduction from the initial solution, assess the goodness of fit between the data and synthetic seismograms, and determine the stability of each event solution. A quantitative comparison of each inversion case allows for an assessment of the advantages and limitations of different seismological techniques using similar data sets.

Geologic Setting
Our study region is geographically broad and tectonically diverse, spanning nearly the entire Middle East, parts of western Asia and northern Africa. It can be broken into 7 broad tectonic regions (Figure 1.1 (Vernant et al., 2004;Hatzfeld and Molnar , 2010;Adams et al., 2009). Intracontinental shortening accommodates most of the convergence, especially in Iran; however, large strike-slip faulting occurs along block margins. The interplay between strike-slip and thrusting motion results in compressional structures that strike obliquely relative to the regional convergence direction (Vernant et al., 2004). The transition zone between strike-slip motion in the Zagros (on the Main Recent Fault and North Anatolian Fault) and the Makran subduction zone is marked by large strike slip motion on the Minab-Zendan-Palami Fault (Vernant et al., 2004). The best estimates for the current motion between Arabia and Eurasia is between 18-25 mm/yr, which is slightly slower than the precollision rate of 31 mm/yr (Hatzfeld and Molnar , 2010). Roughly 20% of Arabia-Eurasian convergence is accommodated for in the Zagros (Hatzfeld and Molnar , 2010

Moment Tensor Inversion Methodology Data Set
An original earthquake event list, obtained from the Global Centroid-Moment Tensor (Global CMT) Catalog , was compiled containing events between January 1990 and July 2007; events with a moment magnitude greater than or equal to 5.5 were used. This list contained greater than 200 events within the study region, spanning the Middle East from Turkey to India (30 ¥ E -80 ¥ E) east-west and the Horn of Africa to the Kazakh Platform (10 ¥ N -50 ¥ N) south-north (Figure 1.1). Broadband seismic waveform data was acquired from IRIS DMC from regional and teleseismic stations; a total of 578 stations from 21 receiver networks were used for the moment tensor (MT) inversions. Typical source-to-station distances ranged from a couple hundred kilometers to 90 ¥ .

Inversion
We follow a moment tensor inversion methodology by Liu et al. (2004), adapted from a local to regional set of earthquakes, and solve for the 6 independent seismic moment tensor elements (M ij ) plus the event depth. We perform inversions for constrained, zero-trace and double-couple, and unconstrained solutions, azimuthally weighted and unweighted, with and without depth variation. The variety of inversion parameterizations was done to assess the stability of each solution, the robustness of the inversion method, and compare the results using different constraints to determine a robust solution. As in Liu et al. (2004), a comparison of the solutions from different parameterizations showed little difference among the solutions driven by the large number of measurement windows and the stability of the method; our focal mechanism solutions remain consistent for each case.
A zero-trace, azimuthally-weighted solution while solving for depth, is used for error analysis comparisons between wave speed models and frequencies, as this parameterization produced a robust solution with good match between the data and synthetics. Our methodology lends itself to testing the source of non-DC components by comparing solutions between 1D and 3D waves speed models to investigate whether a reduction in misfit reflectes imperfections in the model, as suggested by Liu et al. (2004). If the source of the non-DC component is indeed an effect of poorly constrained Earth structure in the model, we should see a reduction in non-DC going from 1D to 3D parameterizations.
We define the misfit objective function, E, as in Liu et al. (2004) where, E 1 represents the least-square misfit function, (1.2) C 1 pmq is a zero trace moment tensor constraint, C 2 pmq is a double-couple source mechanism constraint, λ is the function weight, µ 1 and µ 2 are Langrange multipliers associated with the constraints; in equation 1.2, A 1 is a normalization factor, w i represents specified weights (eg. azimuthally weighted, w a i ), d i and s i are the data and synthetics respectively, and m is the moment tensor. Synthetics are allowed to shift in time to match data.

Synthetic Seismogram Generation
Full waveform synthetic seismograms are required for the MT inversion. We computed both 1D and 3D synthetics to compare wave speed models independent of the inversion methodology. 1D synthetics were created using mode summation from the PREM (Dziewonski and Anderson, 1981) wave speed model. Synthetics for the 3D reference model case were created using the spectral-element method (SEM) Tromp, 1999, 2002a,b) using the S2.9EA (Kustowski et al., 2008) wave speed model. Benefits and details of using the SEM method-ology over other methodologies are described in Tromp (1999, 2002a,b). The computation cost for computation of the 3D synthetic seismograms was significant, but tractable on a dedicated cluster. The S2.9EA model is a global shear-wave velocity structure model based on the PREM reference and determined from surface wave phase velocities, long-period waveforms, and body-wave travel times (Kustowski et al., 2008). The full 3D wave speed model also uses a Crust 2.0 crustal model (Bassin et al., 2000), attenuation from PREM (Dziewonski and Anderson, 1981), and Etopo5 topography/bathymetry (NOAA, 1988). Compressional wave speed perturbations are scaled from shear-wave speed perturbations by 0.55 as in Kustowski et al. (2008). Initial moment tensor solutions were obtained from the Global CMT Catalog .
Synthetic seismograms and Fréchet derivatives for each component of the moment tensor and depth, were created at all stations within a 90 ¥ by 90 ¥ mesh seen in Figure 1.1, inset. The depth derivative was calculated by the difference between synthetics from the initial solution and synthetics with a depth increased by dh.
Based on synthetic tests, depth perturbations of 1, 10, 15, 20, 25, and 50 km show a less than 1% change in calculated depth derivatives for all perturbations with the exception at 50 km, which is an unreasonable dh value for shallow events. We use a dh value of 1 km.

Data Processing
Data was filtered between 25-125 second (short period) and 60-125 second (long period) to compare the inversion performance and results at different period bands; a maximum period of 125 seconds was used due to band limitations in instrument response. A bootstrap analysis was performed to assess solution robustness (Press et al., 1997). During the bootstrap analysis, we solved for each event solution 200 times using a random selection of components (radial, vertical, and tangential) from the original dataset. The P-axes were then plotted on a focal sphere to 7 quantitatively assess the stability of the plunge and trend, see Figures 1.2.
The Flexwin  algorithm was used to automatically select time windows for input into the MT inversion using a combination of criteria based on phase, amplitude, ratio of short term and long term average, and envelope mismatch. Flexwin allows for a large volume of repeatable measurements to be made on full-waveform data-synthetic pairs that would otherwise be over looked when hand picking only peaks for specific phases or amplitudes. Flexwin has usertunable parameters and the ability to adapt to 1D and 3D models . As per the user-tunable parameters detailed in Maggi et al. (2009), we required a cross correlation value of 0.75 and an amplitude ratio (dlnA) of 1.0 for Flexwin windows to be accepted in the MT inversion. A minimum signal to noise ratio of 3.5 within two measurement windows, and a minimum single window signal to noise ratio of 1.5 was required to use the time series in the MT inversion.

Moment Tensor Inversion Results
We recovered 184 well constrained solutions out off the initial set of events (for the complete earthquake catalog, see Table 1.2); the remaining 11 events had data quality issues that did not produce acceptable results. The average constraint on the trend and plunge is shown as a histogram in Figures 1.2 and 1.3 and in Table 1.1. Standard errors were determined for the trend and plunge of the MT compressional axes, P-axes, using the bootstrap methodology discussed previously.
Standard errors for the trend and plunge of the T-axes were also analyzed; the errors are comparable to the spreads for the P-axes and as such we do not report 8 results for the T-axes.

Comparison
To quantify the effect of wave speed model on the MT inversion, comparisons were made between 1D and 3D moment tensor solutions using an identical data processing scheme and a consistent number of evaluation windows, this allows for direct comparisons of inversion results based on wave speed models and frequency bandwidth without bias to methodology or the number of evaluation windows. The 1D wave speed model does a sufficient job fitting simple body wave signals and large amplitude surface waves at both period bands, example waveforms in Figure   1.4. Complex signals, from body wave propagation and surface wave dispersion due to the continental lithosphere, are not adequately fit by the 1D model at shorter periods. Employing the 3D wave speed model (Kustowski et al., 2008), synthetic seismograms predict a larger portion of the data at all periods, including the late arriving shorter period arrivals due to strong dispersion from the continental lithosphere. Additionally, using an appropriate 3D model improves the amplitude and phase misfits when compared to a 1D model and facilitates the use of more waveform data in the MT inversion.
A quantitative comparison of the variance reduction between 1D and 3D models shows that, within the same frequency band, a reduction in error on the trend and plunge is seen when using the 3D versus the 1D model, Table 1.1. At longer periods, 60-125s, the difference between using the 1D and 3D model is negligable, seen only as an approximate difference of 0.39 ¥ on the trend and plunge. At shorter periods, 25-125s, the improvement is more significant, reducing the error by approximately 4.43 ¥ on the trend and 1.34 ¥ on the plunge, see Figure 1.3.
A metric was created to ease the comparison between inversion runs by defining a variable τ as where λ 1¡5 represents the relative weights ( λ 1 5, λ 2 0.25, λ 3 1, λ 4 0.05, λ 5 0.05),CC is the average cross-correlation value,∆lnA is the average amplitude ratio,ξ is the average misfit, N is the number of windows used, and AZ is the maximum azimuthal gap. Weights were chosen to emphasize the importance of the cross-correlation,CC, and misfit,ξ, to the goodness-of-fit, and downweight the larger numbers associated with the N and AZ variables. Based on this defined metric, where a smaller τ value indicates better goodnes-of-fit, 95% of the events have τ values 10 and 85-90% have τ ranging between 0 and 2. We calculate the mean τ and standard deviation for each inversion set (1D25, 1D60, 3D25, 3D60) to obtain a single number with which to evaluate the entire dataset. Events with metric values, τ ¡ 10 are considered to be very poorly fit and are not included in the inversion set averages. Typically, an event with a large metric, or poor goodness-of-fit has severe data quality or lack of data issues resulting in values of τ in the ¡¡ 100.
Based on trend and plunge standard error improvements, increases in the crosscorrelation coefficient and variance reduction for short periods (25-125s), ( a better fit to the data than does the 1D PREM model (Dziewonski and Anderson, 1981). At shorter periods waveform fit deteriorates slightly, relative to the longer period data (60s vs. 25s), as seen by the decrease in cross-correlation value and increase in the mean τ (Table 1.1); however the benefit is the incorporation of much more seismic data into the source inversion as a result of 3D synthetics predicting a larger portion of the waveform data ( Fig. 1.4) .
Results compare favorably with previously published solutions in both mechanism and depth (Jackson et al., 2002;Tatar et al., 2004;Talebian and Jackson, 2004), as well as Global CMT solutions. Figure 1.5 shows a comparison between Global CMT solutions for 3 events and our calculated solutions, for each event our calculated source mechanism are similar to Global CMT.

Depth
Comparison of our revised depth versus Global CMT depths are made in Figure   1.6. Differences between 1D and 3D wave speed models are subtle, yet present, especially for shallow events. Agreement between our determined depths and those from Global CMT improves with the inclusion of shorter period data (25-125s), this is especially true at shallow depths. Lack of Global CMT event depths less than 15km is a result of a constraint imposed on their solution, our inversions do not impose this constraint. A cluster of events between 50-150km depth is within agreement with that of Global CMT with the inclusion of shorter periods. For deep events (¡200km) there is a systematic divergence between our solutions and those from Global CMT as the calculated depth is shallower than the initial depth; this is most likely due to S2.9EA's heterogeneous wave speeds at depth. Additionally, Muyzert and Snieder (1996) has shown that these large deviations we see in the depth may possibly be due to unstable initial phase behavior in the long period surface waves.

Discussion
The comparison between Global CMT solutions and the moment tensor solutions presented here show minimal difference in source mechanism, an average of 7.31 ¥ and 7.56 ¥ difference for the P and T axes respectively for the 3D 25-125s case. This is true for all four cases examined (1D vs 3D model, minimum period 25s vs 60s) and when varying constraints applied to the inversion. In all cases, a reduction in variance between data and synthetic is seen between the original solution and our calculated solutions. The greatest reduction in variance is seen in the 3D case when filtered between 25-125s, shown in Table 1.1. A majority of events see a variance reduction of 5-40% which represents a significant improvement in fitting waveforms. Events with the largest variance reduction are a result of a better amplitude agreement between the data and synthetic.
The largest contribution to the misfit in our result can be attributed to poorly constrained shallow Earth structure in our models. The use of shorter periods (25-125s) introduces a potential for contamination from larger errors due to unresolved wave speed structures. At shorter periods, Earth structure will have a greater influence on the propagating wavefield and errors in the Earth model may be mapped into the source solution. There is a small, but significant, increase in standard error on the trend and plunge as well as a decrease in the average cross-correlation value from 60-125s to 25-125s (Table 1.1). We do not feel the degradation in solution stability, as seen by the increase in trend and plunge standard error, and waveform metrics, as seen by the decrease in the cross-correlation value, are justification for the removal of shorter period signal. Alternatively, the addition of shorter period signals, 25-125s, allows for much more seismic data to be incorporated into the inversion, as seen by an increase in the number of evaluation windows from 1D25 to 3D25 (Table 1.1), indicating a robust predictive capability of the 3D wave speed model over a wide period range. An azimuthal gap in seismic stations seen in Kazakhstan and Russia also contributes to the error of the trend of some solutions, but for most events the overall azimuthal coverage is excellent and the trend is well constrained for the entire data set as a whole when using the 3D moment tensor inversion.  . We plot the percentage of double-couple components in our solutions using a 1D and 3D model at 25-125s periods, Figure 1.7, to test this hypothesis that non-double-couple components are a reflection of poorly constrained regional structure in the initial velocity model or complex fault geometry, consistent with suggestions made by Liu et al. (2004) and Henry et al. (2002).
The number of events in our catalog with significant non-double couple components suggests that improvements in the wave speed model are needed. Experiments by Hjörleifsdóttir and Ekström (2010) to evaluate source parameters using synthetic seismograms at periods greater than 40 sec, show small errors in the non-doublecouple component when using a diverse seismic data set, similar to the global catalogs . Experiments here using real data and are consistent with synthetic experiments at longer periods, but the use of a 3D wave speed model and shorter periods, 25 sec, significantly increase the double-couple component. This may suggest a frequency dependance to resolving the non-double-couple components. At longer periods (60-125s), not shown, the difference between 1D and 3D percentage double-couple is minimal, further suggesting that 3D structure at short periods is the cause for increased double-couple component. We would expect an increase in the percentage of double-couple components in the source with improvements made to the wave speed model; however due to the complex nature of real faulting, a completely double-couple source is unlikely. A subset of our deepest events (100 to 260 km) shows a similar trend as for the whole catalog.

Tectonics
Comparisons between our solutions with regional tectonic features show good general agreement with previous geologic interpretations (

Conclusions
Using a full-waveform moment tensor inversion method (Liu et al., 2004), we repeat an identical data processing scheme for four cases using two initial models (1D and 3D) and two frequency bands, allowing for direct comparison between results and the evaluation of model and frequency bandwidth. The number of evaluation windows is consistent for each inversion set permitting an appropriate comparison between models and frequency ranges without bias given to the number of evaluations.
We provide justification for the use of 3D models, in preference to 1D models, by showing a reduction in variance and better constraint on moment tensor solutions, source characteristics, Earth structure and event depth. This is espe-cially true in regions of highly heterogeneous Earth structure, as seen in our study region. The 1D model does not provide an adequate fit to waveforms at shorter periods, especially in regards to fitting complex body wave propagation and surface wave dispersion. Additionally, the 3D model produces a solution with a greater percentage of the source approximated as a double couple, suggesting that the non-double-couple component of our solutions comes from poorly constrained wave speed structure.We achieved an overall agreement in mechanism and depth with regional tectonics across inversion methods, wave speed models, and frequency range confirming the stability and robustness of our methodology and solutions.
Further, the solutions obtained in this study agree with those found by pervious researchers, Global CMT, and also agree with the large scale geologic structures and overall GPS measurements (Adams et al., 2009;Jackson et al., 2002;DeMets et al., 1994;Vernant et al., 2004;Tatar et al., 2004;Hatzfeld and Molnar , 2010).

Data and Resources
The facilities of the IRIS Data Management System, and specifically the IRIS  are indicated by (latitude, longitude) beneath the event date. A better fit to the data is seen for complex body waves between 150-450 seconds for II.KURK and 600-900s for station KIEV as well as surface wave dispersion at greater than 800 seconds when using the 3D derived synthetic seismograms (highlighted by the blue box). Our solutions compare well with the initial solution, but variance between the data and synthetic is reduced while fitting more of the data using a 3D model. Percentage represents the variance reduction between our solutions and the Global CMT solution. A majority of events in the catalog (Table S1) see a variance reduction between 5-40% compared to the Global CMT solution. This result provides evidence that the % of non-DC component in the solutions is a result of imperfect and poorly resolved Earth structure within the initial velocity models.  Zhao, L., T. Jordan, K. Olsen, and P. Chen, Fréchet kernels for imaging regional earth structure based on three-dimensional reference models, Bulletin of the Seismological Society of America, 95 (6), 2066, 2005.

Introduction
The Ontong Java Plateau (OJP) represents the largest preserved Large Igneous Province (LIP) by volume on the Earth . At the surface, the OJP's area of 2 M km 2 also makes it the largest oceanic plateau . Taylor (2006) showed that the OJP, Manihiki Plateau (MP) and Hikurangi Plateau (HP) were once part of the same feature and were subsequently separated by seafloor spread-ing during the Cretaceous. Including the MP and HP as originating from the same edifice, greater than 4 M km 2 of ocean floor has been effected by the massive outpouring of material associated with the three plateau's formation . Connections with the Louisville Hotspot Chain have also been made, suggesting that the chain represents a plume tail   Current understanding of the OJP comes from a broad range of research. Sampling of the OJP's surface has been done using geochemistry and petrology on recovered rock samples from DSDP and ODP cruise legs as well as sampling on nearby islands in the Solomon chain Tejada et al., 1996Tejada et al., , 2002Tejada et al., , 2004. Crust and upper mantle structure has been investigated using gravity and magnetic surveys  as well as active-source seismic profiling .
Deeper seismic structures have been imaged using Rayleigh-wave seismic tomography , seismic attenuation  and anisotropy .
The result of these studies paint a complex geologic history spanning over 120 million years (Ma) and despite previous work, a consensus regarding the genesis of the OJP is lacking and several hypotheses have developed. Two main hypotheses on the OJP's origin invoke either 1) the surfacing of a buoyant plume head or 2) vigorous passive mantle upwelling at or near a spreading ridge, as responsible for the plateau's emplacement.

Plume Source
The prevailing mechanism for the origin of LIPs has been the decompression melting of a surfacing mantle plume head Campbell , 1998).
A Rayleigh-Taylor instability originating from the core-mantle boundary or the 660 km transition zone can be positively buoyant due to either a thermal or compositional anomaly compared to the ambient mantle Campbell , 2005). The OJP, and other LIPs, would be a product of high degrees of melting requiring high mantle temperature anomalies that rise quickly and adiabatically through the mantle , resulting in widespread melting, drying and depletion of the mantle beneath a forming plateau .  suggested this excess of heat, originating from the core-mantle boundary, could alter normal mantle convection, changing the magnetic reversal frequency and leading to the observed mid-Cretaceous magnetic quite zone following the formation of the OJP.
The most compelling evidence for a plume source to the OJP is the volume and rate of erupted material. Erupted volume estimates range from 44 to 57 M km 3 over 6 ¡ 14 M yrs Tejada et al., 2002). Tejada et al. (1996Tejada et al. ( , 2002 determine the main plateau forming event occurred around 120 M a, with a smaller, but significant, volume of material emplaced around 90 M a. Geochemically, samples represent high degrees of melting to a relatively homogeneous and well-buffered OIB-like source (Tejada et al., 1996;. Enrichment in siderophile elements, such as molybdenum (Mo), members of the platinum group, and gold (Au), may suggest a core-mantle boundary source, consistent with a plume hypothesis .
Despite the evidence of a plume source for the OJP, complications arise when examining the emplacement depth and isostatic topography of the plateau. The vesicularity of the OJP lavas and presence of microfossils suggest that plateau emplacement was entirely submarine, at depths greater than 800m below sea level .  suggested that based on a realistic geotherm for a mantle hot enough to induce melting, the plateau should have been emplaced at or above sea level based solely on the isostatic topography; the addition of a buoyant plume head would dynamically raise the plateau further.
Using numerical models, , suggested uplift of approximately 5 km above abyssal sea floor when lithospheric extension is allowed in their model, similar to the pre-emplacement tectonic setting near the OJP.  suggested the formation of a viscous "plug" due to significant melt extraction and dehydration. Flow around the plug could result in high melt extraction rates and limited uplift; further the viscous plug would be more resistant to mantle flow and able to persistent for ¡ 120 Ma. In addition to a lack of uplift, post-emplacement subsidence has been retarded relative to normal seafloor and seafloor adjacent to the plateau , suggesting a remnant positive buoyancy within the mantle beneath the OJP.

Passive rift driven upwelling
As an alternative to the plume-driven hypothesis,  proposed that entrainment of dense eclogite, by vigorous plate-driven mantle flow due to fast plate spreading rates, could explain both the topography and geochemistry of the OJP.  suggested the dense eclogite comes from recycled subducted crust. Initial formation of the combined plateaus occurred in the vicinity of the Tongareva triple junction (Pacific-Phoenix-Farallon); the Osbourn Trough separates the MP and HP, while spreading in the Ellice Basin seperated the OJP and MP Billen and Stock , 2000;Viso et al., 2005;Taylor , 2006).
Nearby magnetic lineations (M0-M7) imply a half spreading rate of 7.7 cm{yr between 120 ¡ 129M a .  suggested that this rapid spreading rate alone should be large enough to entrain material denser than nominal mantle, and would only be enhanced by the presence of a nearby triple junction.
What follows is a description of our tomography methodology, using a unique data set combining ambient noise and earthquake waveforms. This allows us to obtain the resolution at depths in the crust and upper mantle necessary for interpretation of the OJP's wave speed structure and arrive at a hypothesis regarding its formation.

Methodology
To determine the 3-dimensional wave speed structure beneath the OJP we employed a two phase, iterative, tomography using full-waveform ambient noise and earthquake data. Due to the sparse coverage of seismic stations and earthquakes in the Pacific ocean and the relative isolation of the OJP, a two step process was used to image the seismic wave speed structure beneath the plateau. The first phase used Green's functions derived from ambient noise data at periods up to 200 seconds. This ambient noise-only model provided an improved base model for subsequent iterations using joint ambient noise and earthquake data. Use of Green's functions from ambient noise as a starting point effectively exploits the 20 years of continuous, broadband seismic data, and is not reliant on earthquake distributions and solving for earthquake source mechanisms. This eliminates a source of error, the earthquake mechanism and location, during the initial iterations and allows the model converge on large scale features. Ambient noise further exploits small temporary seismic station deployments that may not gather sufficient earthquake data during their deployment window.
We added earthquake waveform data after changes in the wave speed structure 41 converged using only ambient noise Green's functions. The use of both ambient noise and earthquake data is complementary as each data set samples different portions of the 3-dimensional model space. The addition of earthquakes tripled the number of total measurements used in the inversion and provided higher quality measurements than those strictly from ambient noise. The large magnitude sources associated with earthquake events produced more distinct surface wave signals with higher signal-to-noise ratios. The surface waves measured with both data sets have path coverage sensitive to crust and upper mantle ( 500 km) Earth structure, and is key to the understanding of the OJP's wave speed structure and formation.

Data Preparation
To extract usable Rayleigh wave signals from continuous raw seismic data, we use an ambient noise processing procedure outlined in Shen et al. are then removed from the time series. Using the one day long records, a cross correlation between station pairs was then calculated with one station acting as a "virtual" source. The cross correlated records were then summed into monthly stacks; the total sum of these records represents our empirical Green's Function (EGF) following a time derivative (Figure 2.10). Monthly stacks are computed to quantify the error of the EGF.
Earthquake data was added to the inversion problem after iteration four. Seismic waveform data was collected from IRIS DMC for earthquakes between 1990 and 2012 with moment magnitudes (M w ) between 5.0 and 6.0.

Synthetic Waveform Generation
Synthetic seismograms were calculated by propagating seismic waves from a virtual source to each receiver using a nonstaggered-grid finite-difference method . The initial model is a combination of the global surface wave diffraction model, CUB  and AK135  for depths greater than 396 km. Wavefield simulations were carried out on a multinode Linux cluster with 24 core per node; each simulation took approximately 8 hours using a single node.
With the addition of earthquake data, inaccurate representations of the earthquake source mechanism and location needed to be addressed. As such, earthquake moment tensor solutions were gathered from the Global Centroid Moment Tensor (GCMT) catalog . GCMT solutions were applied in finite difference forward simulations using a bell-shaped source time function (STF) of 4s duration for numerical stability purposes. To directly compare data and synthetics, the synthetic waveforms were convolved with a STF scaled in duration by the earthquake magnitude. Moreover, the STF used in the finite difference simulation was convolved with the observed data. An appropriate earthquake STF was calculated based on the event's magnitude, using where L is the surface rupture length, a is 5.08, b is 1.16 from Table 2A in , T R and V R is the rupture time and velocity, 2.86 km/s. This source equalization process is summarized in the equations below where S d is the observed data seismogram, S s is the synthetic seismogram, G is the Green's Function, and Λ d and Λ s are the STFs of the data and synethic respectively, The above pair of convolutions align the data and synthetics in time by accounting for phase-shifts and allowing for a straight-forward measurement to be made between the two. Filters used in the measurement are much longer in duration than either of the STFs, reducing these convolutions to simple time-shifts.

Phase Delay Measurement and Inversion
Empirical green functions (EGFs), earthquake data and synthetics were filtered, two-pass butterworth, at five overlapping finite frequency bands, 200 ¡100s, 150 ¡ 75s, 100 ¡ 50s, 75 ¡ 30s, 50 ¡ 25s. Phase delays, dT , were measured between the data and synthetics by cross-correlation for each frequency band. Low quality signals were removed from the measurements using a minimum signal-to-noise ratio and a minimum cross correlation coefficient criteria (  Chen et al., 2007b,a; is used to calculate perturbations in V p and V s using a finite frequency full-waveform tomography process (see Appendix for details).
Finite frequency techniques have been used effectively to iteratively improve three-dimensional models of southern California (Tape et al., 2007, the northwestern US , Europe (Zhu et al., 2012b,a;, Tibet  and the Middle East . Finite frequency sensitivity kernels better recover perturbation amplitudes and wave speed geom-etry than ray based methods , reducing smearing, leading to higher amplitude and more constrained anomalies (Becker , 2012

Resolution
The sparse receiver distribution in the Pacific has previously limited the ability to obtain high resolution images of the OJP region. Using our unique data set of combined ambient noise and earthquake waveform measurements, we are able to significantly improve resolution of the wave speed structure beneath the plateau from previous work   Malaita (Tejada et al., 1996) suggest that we are possibly resolving the source of this recent volcanism. To the west of the OJP, slow wave speeds roughly trace the plate boundary between the North Bismark/Manus and the South Bismark plates (Bird , 2003). The dense, subducting Solomon Sea plate is also neatly outlined by a seismically fast region in our model (between approximately 150 ¥ ¡ 154 ¥ ). The agreement of these features in our model with known geologic structures further provides confidence in our model's resolving capabilities.

Discussion
Our resulting model has a region of fast shear wave speeds beneath the OJP that differs significantly from previous Rayleigh wave tomography  and is faster than 120 Ma oceanic lithosphere (Maggi et al., 2006).
We can rule out melt and/or volatiles as a source of the fast anomaly, as their presence would decrease shear wave speed . In the following, we discuss three possible explanations, 1) the observed data do not require such high wave speeds, 2) there is significant V s v ¡ V s h anisotropy in the region or 3) the wave speed structure is a result of a mantle compositionally different than 120 Ma oceanic mantle.
Are the very high wave speeds required by the observed data?
To address question 1 we forward simulate select event-station pairs that cross the high wave speed anomaly using two model cases, limiting V s values to a maximum of a) 4.5 and b) 4.75 km/s above 400 km depth. Shown in Figure 2 Kustowski et al., 2008; and faster than the expected shear wave speeds of 120 Ma oceanic lithosphere (Maggi et al., 2006;Beghein et al., 2014).

Anisotropy
The second possibility for the anomalously fast wave speeds could be related to In case 2 we increase our V s values by the percentage of V s h /V s v anisotropy in anisotropic PREM. This is analogous to adding PREM anisotropy to our calculated values, V s h ¡ V s v . For all three event-station pairs, the anisotropic case fits the data better than the purely isotropic model, seen as a decrease in the dT for nearly all frequency bands (Figure 2.9). The ability to fit both vertical and transverse component surface waves when applying simple V s h ¡ V s v anisotropy, suggests that the fast anomaly is probably anisotropic.

Composition
Finally we are left with a compositional source for the fast wave speeds in the region. Historically, eclogite has been used to account for such high wave speeds in other anomalous regions throughout the world. Ultra high pressure (UHP) eclogites from Sulu and Dabie region in China are weakly anisotropic ( 3%) and have fast axis V s and slow axis V s velocities ranging from 4.89 ¡ 5.05 km/s and 4.84 ¡ 5.01 km/s respectively (Bascou et al., 2001; . Eclogite samples from the Western Gneiss Region in Norway show Vs wave speeds ranging from 4.5 ¡4.99 km/s (Bascou et al., 2001;. Several studies of the Slave Craton in northern Canada have shear wave speeds similarly fast as our results for OJP (Cammarano and Romanowicz , 2007; and eclogite has been interpreted as the source for these abnormally fast wave speeds .
To discriminate between a garnet-rich fertile peridotite and various types of eclogite, we ran sample cases using the Excel macro of .
We added garnet to a fertile peridotite with modal abundances of 55% forsterite, 25% enstatite, 18 % diopside. Garnet was added in increments of 2%, 4%, 6%, 8%, 10% and 20% as the modal abundances of the other minerals was adjusted proportionally, see Table 2.2 for modal abundances. Mineral wave speeds were simulated at 2.5 GPa and a lithosphere temperature of 1125 ¥ C at that depth, which is an approximate geotherm of 15 ¥ C per km (Stein and Stein, 1992). An 1125 ¥ C estimated lithosphere temperature falls within the range of possible values for 120 Ma lithosphere at 70 km (Stein and Stein, 1992). Calculated shear wave speed is less than 4.65 km/s for all garnet-rich peridotite cases (Table 2.2) . We then calculated a suite of eclogite compositions, again incrementally increasing the % of garnet (20%, 30%, 40%, 50%, 60%). Modal abundances of zoisite amphibole eclogite, amphibole eclogite, zoisite eclogite, lawsonite amphibole eclogite, coesite eclogite, and diamond eclogite were also used . Calculated shear wave speed was less than 4.50 km/s for all eclogite cases. Finally wave speeds were calculated using approximate modal abundances for UHP eclogites from  that contain large amounts of garnet and jadeite, see Table   2.2 for compositions. We were able to reproduce the measurements of V s to those measured in the literature (Bascou et al., 2001; for UHP eclogite compositions.
Our resulting wave speeds and the above sensitivity tests support a possible compositional anomaly for the fast wave speeds in the region. This does not discriminate, however, between the potential plateau formation theories, as both a vigorous upwelling  or a plume scenario (Bercovici and Mahoney, 1994;Tejada et al., 1996; could support entrainment of eclogite.  suggested the entrainment of dense eclogite fragments non-uniformally distributed throughout the mantle to explain the intriguing buoyancy characteristics of the plateau.  showed evidence for significant amounts of ancient recycled subducted crust in their xenolith samples from mantle beneath the OJP. The authors suggest a chemically heterogeneous plume  as the source of the OJP.
Models of thermochemical plume heads, with as much as 15 wt% eclogite, have also been shown to retard surface uplift, as well as cause extensive delamination and thinning of the lithosphere, compared to a purely thermal plume head case . Further,  suggested the presence of excess heat at the base of the lithosphere representing remanent portions of plumes unable to penetrate a cratonic lithosphere. This hypothesis could explain the slow wave speeds seen directly beneath the shallower seismically fast and likely strong anomaly on the plateau. An eclogite composition could also reconcile the anomalous subsidence history of the plateau. Foundering or delamination of the eclogite could allow the plateau to remain relatively buoyant over the last 120 Ma.
In addition to the plume or vigorous upwelling hypotheses,  suggest two models to explain a Bouguer anomaly greater than predicted Airy isostasy beneath the plateau. The first involved late stage magmatic underplating beneath an already thickened plate. It is possible that this underplating could force oceanic crust within the stability range for eclogite formation.
Additional evidence from geochemistry is needed to further discriminate between source hypotheses. Geochemically, samples represent high degrees of melting to a well-buffered OIB-like source (Tejada et al., 1996;. Enrichment in siderophile elements, such as molybdenum (Mo), members of the platinum group, and gold (Au), suggest a core-mantle boundary origin that is consistent with a plume hypothesis .
A compositionally heterogeneous mantle beneath the OJP is required to generate the observed wave speed structure seen in our model. Regions of fast wave speed, ¡ 4.75 km/s, can be explained with a garnet rich composition that is likely eclogite. The fast seismic wave speeds suggest a feature that is strong, able to persist 120 Ma and is related to the plateau's formation. The feature may represents remnants of a larger structure that has undergone erosion due to 120 Ma of plate motion. Laboratory tank modeling suggests that the plate motion history is extremely important to the interpretation of the modern tectonic features Kincaid et al., 2013; . Mantle fabrics present 120 Ma ago are likely to be significantly deformed as a result of complex interactions with plate tectonic process. While our seismic model can not discriminate between the two main plateau formation hypotheses, the fast wave speeds in combination with previous geochemical observations is consistent with hypotheses for a compositionally heterogeneous plume with entrained eclogite.

Conclusions
A unique dataset using a combination of ambient noise and earthquake waveforms was used to determine the seismic wave speed structure of the Ontong Java Plateau. Our model's resolution represents a significant improvement over previous research and the highest wave speed resolution to date for the region. We have shown a significant improvement in our model relative to the starting model, CUB , seen as a decrease in the phase delay, dT , through iterations, Figure 2.3, and in the synthetic waveform fit to the data, Figure 2.4.
Beneath the plateau we image a region of shear wave speeds ¡ 4.75 km/s, possibly up to 5.00 km/s. These wave speeds are faster than normal oceanic lithosphere 120 Ma in age (approximately 4.5 km/s) (Maggi et al., 2006;Beghein et al., 2014) and are similar to as observed in cratonic environments Kustowski et al., 2008;. Tests for V s v ¡ V s h anisotropy through the fast anomaly are ruled out by the data. However, the addition of simple V s h ¡ V s v anisotropy slightly improves synthetic fit to the data relative to our isotropic model.
Our observed wave speeds beneath the plateau are consistent with a compositional anomaly and likely a result of UHP eclogite compositions. While our seismic model cannot conclusively discriminate between the two major formation hypotheses, previous studies of rock samples from the Solomon Islands suggest a compositionally heterogenous plume source for the OJP (Tejada et al., 1996;. We suggest that the surfacing plume head entrained eclogite from the deep mantle, resulting in a denser than normal eruption that retarded surface uplift and accounts for the 54 observed fast wave speeds beneath the plateau.   that is late compared to the data for all frequency bands (a-c). The applied PREM anisotropy case fits the data slightly better than our isotropic model, suggesting that the fast anomaly is weakly anisotropic.
Included in the appendix is a description of the scattering integral (SI) methodology Chen et al., 2007b,a; and additional figures and tables.  Jadeite (jd), the sodium bearing clinopyroxene was used in place of Omphacite.
Coesite (coe) was used in place of Rutile because it was the closest mineral structure substitute. Phengite and Opaques were combined into Muscovite (mu). References HA04 refers to  and Ji03 refers to . The

76
The scattering integral (SI) methodology constructs the station strain green tensors (SGTs) from a 3D reference model, here using a finite difference simulation, of the response to a force at a "source" location. Travel time anomalies are then measured from the observed and synthetic waveforms at each station. Station SGTs are used to calculate finite-frequency sensitivity kernels to perturbations in V p and V s . The travel-time measurements and sensitivity kernels are used to invert for Earth structure which are then added to the 3D reference model and can later be repeated and iterated on. The methodology varies slightly whether using ambient noise data or earthquake data. For ambient noise data, the SGT is calculated from a Gaussian pulse located at the coordinates of another "virtual source" station. The approach using earthquake data applies the earthquake's moment tensor acting at its source location to the finite difference calculation.
From Chen et al. (2007b,a), a forward problem for displacement can be written as δd Aδm where d is the data represented as functionals of an Earth model, m and calculated for a starting Earth model,m. Data sensitivity kernels, K d , are derivatives of the data with respect to the model parameters within them volume, V , at every point within the model, x. Generalized seismological data functionals (GSDFs) are used to map synthetic waveforms (u i pωq) into observed waveforms (ū i pωq) in the frequency domain, using two frequency-depandet quantities, the phase delay time (δτ p pωq) and the amplitude reduction time (δτ q pωq) and equation 2.8 below, where δτ p,q pωq are measured at a set of discrete frequencies, ω n Chen et al., 2007b,a). We use only the phase delay time, δτ p pωq.

77
Once phase anomalies are measured, perturbation kernels (J sr in ) can be constructed for the nth misfit measurement made on the ith component of the seismogram, generated by source, s, and recorded at receiver, r, by where the seismogram perturbations are related to density, ρ, and strain, c jklm , Jadeite (jd), the sodium bearing clinopyroxene was used in place of Omphacite. Coesite (coe) was used in place of Rutile because it was the closest mineral structure substitute. Phengite and Opaques were combined into Muscovite (mu). References HA04 refers to  and Ji03 refers to . The m prefix signifies the modal abundances have been modified from the published values.

Introduction
The eastern North American margin (ENAM) is presently a passive continental margin and the result of multiple episodes of continental collision and rifting dating back greater than a billion years ago (1 Ga) .   Ma in present day southeastern North America .
The initial rifting and breakup of Pangea and formation of the Atlantic basin 83 are associated with multiple magmatic provinces scattered on both margins of the present day Atlantic basin. Two in particular, the Central Atlantic Magmatic Province (CAMP) and the East Coast Margin igneous province (ECMIP) have been suggested as a result of either 1) a mantle plume or 2) continental rifting in the form of 2a) reactivation of Paleozoic structures or 2b) upwelling convection cells at the edges of cratons

Methodology
To determine the 3-dimensional seismic wave speed structure beneath the Eastern North American Margin (ENAM) we use a iterative, finite-frequency tomography approach using full-waveform ambient noise data. Green's functions are derived from continuously recorded broadband seismic data at periods up to 200 seconds. Use of Green's functions from ambient noise exploits 20 plus years of broadband seismic data recorded at stations throughout North and Central America and the Caribbean, not reliant on earthquake distributions and solving for the earthquake source mechanism. The use of data from ambient noise reduces a source of error, from an earthquake's location and source mechanism, and is able to exploit smaller temporary seismic station deployments unable to gather sufficient earthquake data during their deployment window. We measure surface waves sensitive to crust and upper mantle ( 500 km) Earth structure, key to understanding the nature of the margin. A detailed description of the methodology can be found in Chapter 2 of this volume. Here we will summarize the most important steps.
Continuous, vertical component seismic data recorded between 1990 and 2014, was gathered from IRIS DMC for 203 stations located in the eastern United States, Caribbean, Central and South America (Figure 3.1). To extract usable Rayleigh wave signals from the raw data, an ambient noise processing procedure outlined in  and  was used. After removing the instrument response a frequency time normalization (FTN)  was used to normalize the data. Earthquake signals are removed and a cross correlation between station pairs is calculated with one station acting as the "virtual" source.
The cross correlated records are stacked and following a derivative, represent our empirical Green's functions (EGFs).
Synthetic seismograms are calculated by propagating seismic waves from a virtual source to each receiver using a nonstaggered-grid finite-difference method . As in Chapter 2, the initial model is a combination of the global surface wave diffraction model, CUB , and AK135  at depths greater than 396 km.
EGFs and synthetics are filtered using a two-pass butterworth filter at five overlapping finite frequency bands, 200 ¡ 100s, 150 ¡ 75s, 100 ¡ 50s, 75 ¡ 30s, 50 ¡ 25s. Phase delays, dT , were measured between the data and synthetics by cross-correlation for each frequency band. Low quality signals are removed from the measurement using a minimum signal-to-noise ratio and a minimum cross correlation coefficient criteria (

Resolution and data fit
Along the ENAM an extensive network of seismic stations is ideal for use in ambient noise tomography, Figure 3.1. For the given station distribution, within the continental interior our resolving ability is excellent. However, stations off of the continent are limited to Bermuda (IU.BBSR) and the Caribbean, resulting in a reduction of resolving power of offshore features. The inclusion of stations in Bermuda and throughout the Caribbean increases coverage offshore into the Atlantic ocean but a lack of crossing paths hinders our ability to resolve fine scale features. To demonstrate the resolving capabilities of our dataset, Figure 3.2 shows the computed model domain perturbed with a 1 ¥ , 3 ¥ , 5 ¥ , and 7 ¥ sized harmonic pattern of positive and negative 5% wave speed anomalies.
Resolution is excellent on the North American continent and we are able to recover anomalies of 3 ¥ , 5 ¥ , and 7 ¥ in size with little to no smearing or lose in amplitude to depths ¡ 300 km. We are also able to resolve features at 1 ¥ , but at a significant loss of amplitude. Off of the continent, anomalies 5 ¥ or larger can be interpreted approximately 500 km offshore into the Atlantic ocean and 7 ¥ resolution is obtained throughout the Caribbean.
These resolution tests show that wave speed structure is best resolved between depths of approximately 30 and 300 km, similar to as seen in Chapter 2 of this volume and in Gao and Shen (2014) using a similar methodology and period range.
Histograms showing the range of dT and number of measurements for each iteration is shown in Figure 3.3. After a single iteration, an overall reduction in traveltime and increase in the number of measurements indicates the model is improving and is representing large and small scale Earth structures.

Shear wave speed structure
Seismic wave speed structure is plotted in Figure 3.4 as absolute wave speed in km/s and as perturbation relative to the initial model, CUB , in Figure 3.5. Our initial iterations show the data requires a larger contrast between the fast cratonic lithosphere of the Grenville Province and the slower A small-scale, but prominent anomaly appears beneath southern Maryland and Virginia at depths greater than 50 km, this feature is also observed in Van der  and . This anomaly gets stronger and grows in scale at deeper depths. To the west, the mantle beneath the Appalachian Mountains have low wave speeds from Pennsylvania through Tennessee and throughout the crust. In the upper mantle (¡ 50 km) a continuous low wave speed anomaly in West Virginia extends to depths greater than 250 km. A low wave speed region is present just north of the New Madrid Seismic Zone, only between 60 ¡ 90 km.
In southern North America a low wave speed anomaly along the coast of South Carolina and Georgia extends from 30¡200 km into the mantle; this feature is only observed in Van der  and  model at depths greater than 100 km. Similar to  and , a low wave speed region is also observed up Florida's gulf coast, with the highest amplitudes located to the southeast of the Florida panhandle.
Between approximately 120 ¡ 190 km depth the southern coastal anomalies converge into a single anomaly aligned along the North American continent margin as nearly continuous low wave speed structure from Florida to Maryland. This is in contrast to global shear wave tomography models TX2000 and TX2011 (Grand , 2002) that show a much simpler structure between the continental and oceanic lithosphere, absent the low wave speeds observed in our model.

Discussion
Our resulting wave speed model shows features broadly similar to other regional   .
At depths in the model less than 50 km the wave speed structure is generally a reflection of crustal thickness across the ocean continent margin. Within the upper mantle, Figure 3.4 (e) at 126 km, thinner lithosphere and seismically slower wave speeds are observed along the Atlantic coast and thicker crust with faster wave speeds are observed in the continental interior including the Appalachians and Grenville Province . A region of low wave speed beneath the southern Appalachians is consistent with thickened crust, greater than 50 km, in the region observed by Hawman (2008) and . Ekström (2014) observed very slow phase velocities at short periods, 5 ¡20 seconds, along the Gulf Coast likely the result of a thick sedimentary layer. Our resolution is limited at such shallow depths and throughout the Gulf Coast, but we do not observe this broad slow feature in our model. As observed in Van der  and Van der , south and east of the craton is dominated by a series of isolated low wave speed anomalies that will be discussed below.
The most discernible low wave speed anomaly is beneath New England and continuing offshore in the direction of the New England Seamounts. Li et al.  (2002) calculated a mantle potential temperature beneath New England of around 1200 ¥ C at 100 km depth, significantly warmer than the average Eastern North America geotherm. The calculated geotherm and lower wave speeds are consistent with a thinner continental lithosphere as imaged by .
In our wave speed model the New England anomaly is present between about 50 ¡ 250 km. The location and extent of the anomaly is consistent with Van der Lee and Nolet (1997), , , , and Van der  and its association with the New England Seamount chain offshore is tough to discount. In cross section, Figure   3.6, the low wave speed anomaly shallows seaward. The slowest wave speeds are 90 seen adjacent to the craton to the west and at approximately 200 km depth. As suggested in , this low wave speed anomaly may be a result of thermal erosion associated with the Monteregian hotspot or asthenospheric melt or volatiles . To test whether a thermal anomaly could persist for 100 Ma, we calculate a diffusion time, t, using a scaled version of the diffusion equation, where t radius 2 κ . Two values for the diffusivity (Gibert et al., 2003), κ, were used and a suite of sizes for the radius, see Table 3.2. The time needed to diffuse a thermal anomaly is longer than the 100 ¡ 120 Ma since the Monteregian hotspot for all cases except the 50 km radius anomaly, which significantly underestimates the size of our observed low wave speed feature.
The observed slow wave speeds beneath South Carolina, Georgia and continuing offshore roughly align with the locations of the Brunswick and East Coast magnetic anomalies . We observe a low wave speed feature spanning ¡81 ¥ to ¡76 ¥ longitude, dipping seaward and to depths extending from the crust to greater than 200 km. We do not have the shallow resolution, 35km, to compare our observations directly with  and  but the seaward dipping nature of the feature may imply the shallow structure is related to the deeper low wave speed anomalies we observe in our model.  and  described this region offshore as being characterized by seaward-dipping reflections, high seismic velocities (V p 6.5 ¡ 7.5 km/s) and high densities (2870 ¡ 3090 kg m 3 ) indicating a margin that is highly volcanic. They describe a transitional crust, in between the rifted continent crust and new oceanic crust, that is 24 km thick and accreted to the margin during rifting. Any connection between our observed low wave speed anomalies and the volcanic transitional crust would have to be persistent in the mantle since rift initiation 230 Ma.
Another possible explanation for the observed low wave speed anomalies in the south is that they are related to small scale, edge-driven convection on the edges of continents , reflecting either a temperature anomaly or melt.
Abrupt lateral changes in lithosphere thickness, as observed on the ENAM, in combination with normal plate motion causing a "mantle wind", has been shown to drive convection and asthenospheric upwelling . "Hot" cells have also been hypothesized to be a result of the insulating effects of continents and the absence of subduction related cooling causing localized upwelling . Seen in Figure 3.4 at depths between 126 ¡ 187 km and in cross section (Figure 3.7), is a continuous low wave speed feature that follows the edge of the North American continent.
The low wave speeds are located in a "transitional" zone, or gap, between the thicker continental lithosphere and 230 Ma oceanic lithosphere. Further north, this gap between the oceanic and continental lithospheres decreases in size and the transition becomes smoother. Higher amplitude anomalies of low wave speeds adjacent to South Carolina, Georgia, Delaware and Florida may be exploiting weakness in the lithosphere from past episodes of volcanism. Lizarralde et al. (2007) observed short-length scale variability in the style of rifting in the Gulf of California and attributed these changes to inherited mantle fertility and hydration; where wider, magma-poor rifts were from depleted mantle and narrower, magma-rich rifts from fertile mantle. Moreover, the crustal structure observed across the Carolina trough  and the Guaymas segment in the Gulf of California (Lizarralde et al., 2007) were both shown to have thickened, high velocity crust from magma-rich rifting. Following Lizarralde et al. (2007), this may suggest the Carolina segment was more fertile and/or hotter than surrounding areas during rifting. This interpretation is consistent with the low wave speed anomalies reflecting weaker, rifted lithosphere that is more susceptible to plate tectonics forces and/or temperature variations.

92
Further,  described abrupt changes in the velocity gradient in the mantle and crustal thickness changes as a result of changes in the spreading rate. The difference in transition style between the continental and oceanic lithosphere from south to north up the coast of North America, may be reflecting changes in the spreading rate or style of spreading as the margin began rifting. While this is appropriate for shallow structure, the deeper observed anomalies in our model are likely due to a different, present day deformation in the mantle.

Conclusions
We image the seismic wave speed structure of the ENAM using data from EarthScope's transportable array (TA) deployment in addition to other permanent and temporary networks throughout the United States, Canada, Central America and the Caribbean. Our model has excellent resolution of less than 3 ¥ for the ENAM region and resolution greater than 7 ¥ in the Atlantic ocean and Caribbean.
Our imaged wave speed structure is consistent with previous continent and regional scale tomography models for depths of 30 ¡ 300 km. We observe a distinct transition from the fast wave speed Grenville Province to a low wave speed feature beneath New England that is likely related to the Monteregian hot spot.
Nearly continuous low wave speeds on the edge of the continent, between 126 ¡187 km, are consistent with numerical modeling by , which describes edge-driven asthenosphere upwelling due to abrupt lateral changes in lithosphere thickness. Cells of higher amplitude anomalies, reflecting lower wave speed, are observed off of the coast of South Carolina and Delaware that may represent centers of upwelling convection cells. The transition from continental to oceanic lithosphere varies with latitude, this may be a result of changes in the spreading rate from the southern to the northeastern portion of the margin.

93
Our wave speed model represents a significant improvement in resolution from previous work and provides a good starting model for future studies that plan to fully exploit the entire deployment window the EarthScope's TA.   , was used to estimate the time needed to completely remove a thermal anomaly given a suite of anomaly radius sizes. Two diffusivity, κ, values were used, 0.01 cm 2 s and 0.018 igure 3.2: 1 ¥ ,3 ¥ ,5 ¥ and 7 ¥ sized harmonic pattern of positive and negative 5% wave speed anomalies. We are able to recover the shape of the 1 ¥ sized anomalies for depths shallower than 100 km however there is a large decrease in amplitude.
Anomalies 3 ¥ and larger are very well resolved in both shape and amplitude for the all of the eastern North American margin and approximately 500 km offshore. Anomalies of 7 ¥ or larger can be interpreted for much of the Atlantic ocean and  Figure 3.3: (a) Histogram of the phase delay, dT , measured between the data and synthetic waveforms. In grey is iteration 01, outlined in black is iteration 02. (b) Shows the number of measurements plotted with iteration. An increase in the number of measurements indicates that the model is better able to fit the data and more measurements fall within the minimum acceptance criteria.