Small-scale thermal upwellings under the Northern East African Rift from S travel-time tomography

There is a long-standing debate over how many and what types of plumes underlie the East African rift and whether they do or do not drive its extension and consequent magmatism and seismicity. Here we present a new tomographic study of relative teleseismic S and SKS residuals that expands the resolution from previous regional studies below the northern East African Rift to image structure from the surface to the base of the transition zone. The images reveal two low-velocity clusters, below Afar and west of the Main Ethiopian Rift, that extend throughout the upper mantle and comprise several smaller-scale (about 100-km diameter) low-velocity features. These structures support those of our recent P tomographic study below the region. The relative magnitude of S to P residuals is around 3.5, which is consistent with a predominantly thermal nature of the anomalies. The S - and P -velocity anomalies in the low-velocity clusters can be explained by similar excess temperatures in the range of 100-200°C, consistent with temperatures inferred from other seismic, geochemical and petrological studies. Somewhat stronger V S anomalies below Afar than west of the MER may include an expression of volatiles and/or melt in this region. These results, together with a comparison with previous larger scale tomographic models, indicate that these structures are likely small-scale upwellings with mild excess temperatures, rising from a regional thermal boundary layer at the base of the upper mantle.


Introduction
East African extension and its accompanying volcanic and tectonic activity (Fig. 1a) are commonly associated with one or more mantle plumes rising below the region [e.g. Burke, 1996;Ebinger and Sleep, 1998]. Indeed, global and continental-scale models have consistently found strong low-velocity regions below the continent, both in the shallow and deep mantle [e.g. Fishwick, 2010;Ritsema et al., 1999]. There is however much debate about whether the deeper structure connects with that of the upper mantle.
Some groups have advocated the existence of a single strongly tilted upwelling rooted in the African Large Low Shear-Velocity Province and rising to fill most of the upper mantle from Tanzania to the Red Sea. This is consistent with the pervasive low velocities in the region [Hansen et al., 2012;Ritsema et al., 1999], broad-scale uplift [Daradich et al., 2003], and a consistent pattern of along-rift directions of fast seismic anisotropy [e.g. Gao et al., 2010;Hammond et al., 2014;Kendall et al., 2005;Montagner et al., 2007]. However, others have proposed the existence of at least two separate branches from the lower into the upper mantle, based on more detailed tomographic imaging [Chang and Van der Lee, 2011;Debayle et al., 2001;Koulakov, 2007;Montelli et al., 2004b], variations in lava chemistry [Furman et al., 2004;George et al., 1998;Pik et al., 2006;Rogers et al., 2000] as well as modelling of the geoid and the evolution of magmatism [Lin et al., 2005]. A recent P tomographic study we performed below the northern East African Rift [Civiero et al., 2015] combined regional seismic datasets across the region to obtain resolution to the base of the transition zone and revealed even smaller structures (scales ~ 100 km) extending across the depth of the upper mantle. We proposed these to be small-scale upwellings rising from a zone of (ponded) hot material at the base of the upper mantle. Such small-scale structures are consistent with those imaged by previous tomographic studies with resolution down to about 400 km depth Bastow et al., 2008;Hammond et al., 2013], and also with proposed interpretations of strongly variable isotopic characteristics of lava along the rift [Furman et al., 2006;Meshesha and Shinjo, 2008].
Here we present a complementary tomographic inversion of relative S and SKS travel times to (a) assess robustness of the structure revealed by our P tomography [Civiero et al., 2015], which we refer to as NEAR-P15, and (b) more importantly, to further constrain the thermal and/or chemical nature of the structures.

Data
This study uses broadband recordings of teleseismic S-and SKS-wave phases from 379 stations across eastern Africa, from Malawi to Eritrea, and the Arabian Peninsula (i.e., all those shown in Fig. 1b). The seismic stations belong to 16 multinational projects and overlap spatially and/or temporally in our region of interest (see supplementary material). There has been some confusion in the literature about the resolving capability of the different generations of tomographic models in the north East-Africa rift (e.g., Reed et al., 2016).
Early studies used stations in and around the Main Ethiopian Rift, resolving P-and S-wave velocities in the top 300-400 km , Benoit et al., 2006, Bastow et al., 2008. The data set was extended into Afar and Eritrea by Hammond et al., 2013 who focussed their interpretation on the top 400 km in relation to rifting in Afar. However, due to the increased aperture of their seismic array resolution, particularly the S-wave model, due to the inclusion of significant SKS data, was good to depths of ~600 km. As a result, the Hammond et al., 2013 model was used to compare against images of transition-zone discontinuity structure in Thompson et al., 2015. The most recent models of Civiero et al., 2015 and this study have extended the data significantly further, including data from Saudi Arabia to Malawi. As shown in Civiero et al., 2015 and Fig. S8 this substantially increases crossing rays in and around the transition zone and thus allows us to confidently interpret seismic velocity models to the base of the transition zone.
We pick arrival times from 590 earthquakes of magnitude m b ≥ 5.5, ranging in epicentral distance from 30° to 130° (30°-90° for S waves, 90°-130° for SKS waves). The azimuthal and distance coverage provided by the selected earthquakes is displayed in Fig. 1c. After applying a 0.04-0.15 Hz band-pass Butterworth filter, S picking was performed on the transverse component, to minimise the effect of P, P-S and S-P conversions, while SKS picking was done on the radial component, as it is dominated by SV energy. Final picks and relative arrival times were determined using the multichannel cross-correlation method (MCCC) of VanDecar and Crosson [1990], in a 12 s window around the initial pick. We deleted all picks with cross-correlation coefficients < 0.80 from our analysis. In total, we retain 16569 traveltime picks, divided into 8730 and 7839 S-and SKS-wave phases, respectively. The mean RMS uncertainty in the delay times, based on the number of picks rather than on crosscorrelation pairs for each event [Tilmann et al., 2001], is 0.38s with a standard deviation of 31 s (S-waves mean RMS uncertainty = 0.34s with standard deviation = 0.28s and SKS-waves mean RMS uncertainty = 0.43s with standard deviation = 0.33s). The range of S-and SKStravel-time residuals is about ± 10 seconds. To reduce the effect of a few higher outliers, residuals larger than 3.0 standard deviations are downweighted iteratively in the inversion [Huber, 1981]; 827 rays end up downweighted, i.e. about 5 % of the data.

Tomographic inversion
The data are inverted using the teleseismic travel-time inversion method of VanDecar et al. [1995], with the same scheme as in our previous P-wave tomographic analysis [Civiero et al., 2015], which has been used and described in numerous previous studies [e.g. Hammond et al., 2013;Ritsema et al., 1998;Schimmel et al., 2003;Wolfe et al., 1997;Wolfe et al., 2009]. We use ray theory and perform a linear inversion; both assumptions should not have a substantial effect on the shape of the anomalies but may somewhat affect amplitudes [e.g. Bastow et al., 2005;Montelli et al., 2004a;Peter et al., 2009]. There is insufficient information on crustal structure below the whole network to correct the data for station statics, so we solve for station (and event) corrections as part of the inversion. We use exactly the same parameterisation as in NEAR-P15, with an inner grid below the stations of 0.5°/0.4° node spacing in latitude/longitude and 50 km in depth, extending down to 1500 km, and an outer grid with 1° node spacing and 100 km depth spacing extending to 28°N-25.40°S in latitude, 25°E-57.20°E in longitude and 2000 km in depth, to avoid mapping outside structure into our region of interest. We regularise the model by suppressing spatial and curvature gradients (smoothing and flattening), as well as, in part of the cases, by damping towards a 3D reference structure. We first investigate the trade-off between the RMS residual reduction and RMS model roughness for different combinations of smoothing and flattening to choose a preferred model that fits the data well, i.e. within a residual reduction that can be justified by the estimated noise level.
We obtain the same preferred regularisation parameters for our final model (which we will call NEAR-S16), as for NEAR-P15, although they were determined independently. Preferred flattening and smoothing factors are 4800 and 153600, respectively (Fig. 2).
In most previous travel-time inversions for the region, a 1-D starting model was used (i.e., zero anomaly, as the relative travel-time method has no sensitivity to 1-D structure) and no explicit damping to it was performed Bastow et al., 2008;Hammond et al., 2013]. However, teleseismic travel-time tomography provides relatively poor resolution of the crustal and uppermost mantle structures above ~200 km depth because of the paucity of crossing ray paths in that depth range. Regional surface-wave data can resolve this shallow part of the upper mantle better. Therefore, as for our NEAR-P15 model, we damp the inversion using the regional surface-wave model of Fishwick [2010] [F2010] as our starting model above 350 km depth (shown in supplementary Fig. S1). The surface wave model is estimated to reach a vertical resolution of ~50 km and lateral resolution between 150 and 500 km. In the area of interest, given the large number of stations and path coverage, we are using one of the best-resolved parts of the model. Where required by the travel-time data, our inversion will add smaller-scale structure to this long-wavelength starting model. Below 350 km, we damp to a 1D (zero-anomaly) reference. The benefits and effects of such damping are discussed in detail in Civiero et al. [2015].
We test a range of damping factors from 0 (no damping) to a fairly strong factor of 70 (where the shallow part of the model is largely the same as F2010). As was found for the Pwave models, the RMS residual reduction does not change significantly with changes in damping (for our preferred flattening and smoothing factors of 4800 and 153600, the residual reduction is 85.8, 86.1, 85.8%, for a damping of 0, 35 and 70, respectively). However, the heavily damped case (damping = 70) has stronger shallow anomalies and weaker (but still required) deeper structure while the undamped case (damping = 0), which is fully determined by the travel times, shows the same spatial distribution of anomalies, but has weaker shallower and stronger deeper anomalies (Suppl. Fig. S3, S4).

Resolution
We performed a range of resolution tests to assess the quality of our retrieved tomographic models. Standard checkerboard tests are included in the supplementary material ( Fig. S9 and S10), as well as several tests for how much smearing occurs from structures above and below the transition zone into the transition zone ( Fig. S11). In addition, we run a set of tests for the three alternative plume hypotheses ( Fig. 3): (A) a large-scale plume head filling all of the upper mantle, (B) a single plume (with assigned width of 380 km, defined as the distance where the anomaly amplitude is 20% of that in the centre of the synthetic plume structure), (C) two narrow vertical plumes (with widths of 150 km). For case B and C a hot layer is added at the top of the input structures to mimic plume material spread in the rift zone below the plate. As a starting model for the damped plume resolution tests, we used the corresponding synthetic input models, filtered to mimic the effect of the surface wave resolution. The filter applies a variable vertical averaging (lower vertical resolution as absolute depth increases) and a variable Gaussian spatial filter (lower spatial resolution as depth increases).
The checkerboard tests illustrate that we achieve quite good resolution for anomalies of 125 km and 250 km diameter through most of the transition zone (Suppl. Fig. S9, S10). At shallow depths (around 100 km), the resolved area is confined to directly below the stations, and from about 300 km depth covers much of the region of interest shown in Fig. 1. At the edges of the model, the checkers become more smeared with depth. For 125-km sized features, resolution becomes poor below about 600 km depth for most of the area (Fig. S10).
Larger features, with diameters of at least 250 km, are resolved over a wider region and through the transition zone, down to 700-800 km depth (Suppl. Fig. S9).
The plume tests in Fig. 3 illustrate the importance of using a-priori constraints for the shallow structure in the starting model and damping towards these structures. Without such constraints, the expected laterally widespread low velocity structures above 200 km depth can not be retrieved. While undamped inversions emphasise lateral variations in velocity structures, which in the case of models B and C means the plume tails, in the damped inversions the geometry of the imaged structures is much closer to that of the input structures.
Note that due to the decrease in thermal sensitivity of V S with depth (from about 3%/100K at 100 km depth to <1%/100K at 600 km depth, Suppl. Fig. S12) and the expectation that temperatures of adiabatic upwellings only change by a few 10s of degrees over the depth of the upper mantle, we would expect plume anomalies to decrease with depth. Thus tomographic images of upper mantle plumes are not expected to have simple cylindrical morphologies.

Our preferred S-SKS model (with intermediate damping) is shown in Fig 4. Undamped
and strongly damped models are included as Suppl. Fig. S3 and S4. The colour scale used is the same in all panels to illustrate how anomaly amplitudes vary with depth. The scale best illustrates the transition zone anomalies and saturates at depths from 300 km upwards for which the structures have been analysed in detail in previous studies Bastow et al., 2008;Hammond et al., 2013]. The damped models in particular emphasise that the strongest low velocities are shallow, where they are widespread. Deep structures persist at all degrees of damping and are hence required by the data. Also, the synthetic test for a large anomaly confined to the top 400 km shows limited smearing to deeper depths (Suppl. The mantle structure shows the same first-order features as NEAR-P15 [Civiero et al., 2015]. The shallow upper mantle, down to ~100km, reveals a broad low-velocity layer with the strongest anomalies following the rift morphology (-4.5% < δV S < -3.0%), similar to other previous tomographic studies for this region [e.g. Bastow et al., 2008;Fishwick, 2010;Hammond et al., 2013].
At larger depths, the low-velocity structure splits into several separate anomalies, that persist down to at least ~650 km depth. In NEAR-P15, these deeper anomalies clearly form two clusters, each about ~400 km in diameter, beneath Afar and west of the MER (Fig. S5), correlating with active volcanism in the Afar region and volcanism along and off-axis of the MER. In the S-SKS model, the same two clusters are also clear from about 200 to 400 km depth. In the transition zone, the Afar cluster is more strongly defined than the one below the MER, which breaks up into several smaller-scale features. Note that, in NEAR-P15, the clusters also contained additional small-scale structure in the transition zone (Fig. S5, S6).
The amplitude of the low-velocity features in the intermediately damped case decreases with depth from ~ -4% in the uppermost mantle to ~ -0.5% below the transition zone. The deepest resolution is too limited to determine how the transition zone structure relates with that below 600-700 km depth.
Complexity in the structure may be to some extent influenced by anisotropy. However, we have quite good azimuthal coverage (as shown in supplementary Fig. S8), where the core of the region is traversed by rays spanning an azimuthal range of 270-360 degrees, which should aid in averaging out azimuthal anisotropic effects. And although we combine SH and SV data, we get images that are quite consistent with those of the P data, so if radial anisotropy exists it does not appear to map significantly into the S structures we resolve.
Nonetheless, this would be an aspect that warrants future work.
Vertical cross sections through the models (Fig. 5) further illustrate the vertical continuity of the two clusters from the shallow mantle through the transition zone, again consistent with the P-wave structure. However, different from the P-velocity structure (Suppl. Fig. S5, S6), the anomalies below Afar are stronger in amplitude and more continuous than the anomalies west of the MER (see profiles E-F and G-H in Fig. 5).

Ratio of P and S velocity anomalies
Several studies have used the ratio of relative changes in shear-and compressional-wave velocities, defined as R S,P (= dlnV S /dlnV P = [dV S /V S ]/[dV P /V P ]), as diagnostic of the cause of the seismic anomalies [e.g. Karato and Karki, 2001;Masters et al., 2000;Robertson and Woodhouse, 1996;Simmons et al., 2009]. The expected R S,P for purely thermal variations spans a range from 1.3 (for low values of temperature, with a weak anelasticity effect) to 2.2 (at high temperatures, where the anelasticity effect is strong) [Cammarano et al., 2003;Goes et al., 2000]. Where water is present, it may enhance the anelastic sensitivity to temperature, by lowering the melting temperature [Karato and Jung, 1998] and/or form hydrated minerals that usually have lower velocity than average mantle minerals [e.g. Abers and Hacker, 2016;Angel et al., 2001]. The anelastic effects of fluids can increase R S,P to up to about 2.3; hydrous minerals, if present in large enough quantities possibly somewhat more [e.g Goes et al., 2000;Hacker and Abers, 2012].
For most plausible large-scale compositions (i.e., more or less melt-depleted forms of peridotite, or alternative chondritic compositions), the effect of composition on V P and V S above the transition zone is small compared to that of temperature [Afonso et al., 2010;Cammarano et al., 2003;Goes et al., 2000;Worthington et al., 2013]. In the transition zone, the sensitivity of velocity to changes in composition is complicated by the different depths, Clapeyron slopes and widths of the phase transitions expected for different bulk compositions [Cammarano et al., 2003;Xu et al., 2008]. Overall, for most upper-mantle compositions and depths, the sensitivity of R S,P to temperature and composition is likely not different enough to distinguish the two factors.
The largest effects on R S,P are expected from partial melt [e.g. Faul et al., 2004;Hammond and Kendall, 2016;Hammond and Humphreys, 2000;Schmeling, 1985] with values that can range from around 1.6 for ellipsoidal melt inclusions, to 2.2 for films in a geometry taken from natural samples [Hammond and Humphreys, 2000], to up to 4.0 for aligned film and layer geometries [Takei, 2002], which would also lead to strong anisotropy.
Several studies use the ratio of P-and S-wave relative arrival-time residuals δ S,P , i.e., using the data rather than the tomographically imaged anomalies [e.g. Bastow et al., 2005;Gao et al., 2004;Rocha et al., 2011]. δ S,P should be proportional to the ratio of absolute dV S and dV P along the chosen station-event pairs. Thus, the slope δ S,P equals (V P /V S ) R S,P or about √3 times R S,P . Bastow et al. [2005] perform a δ S,P analysis below the Main Ethiopian Rift and find a gradient between ~5 and ~10, which they attribute to significant fractions of shallow melt ponding beneath the rift, consistent with other geochemical and geophysical results [e.g. Bastow et al., 2005;Ebinger and Casey, 2001;Gao et al., 1997;Hammond et al., 2014;Kendall et al., 2005;Rooney et al., 2007]. Other studies have used δ S,P to argue for chemical heterogeneity. For example, Gao et al. [2004] conclude a δ S,P of ~2.9 for fast anomalies below the Colorado Plateau is on the high side to invoke temperature perturbations. Similarly, Rocha et al. [2011] find δ S,P > 2.9 below the São Francisco Craton and interpret it as due to compositional effects. However, from our preceding discussion of the expected range of R S,P values including the effect of anelasticity, δ S,P up to 3.6-3.8 correspond to R S,P which fall within the thermal range.
To get further insight into the possible physical causes of the seismic anomalies we imaged below the Northern East African Rift, we perform a comparison of the P-and S-SKSwave relative arrival-time residuals for common events (Fig. 6). Prior to the analysis, we remove all residuals downweighted by the Huber iterations. A straight line fit through our measurements (least-squares, accounting for the estimated picking errors) yields a δ S,P of 3.69 ± 0.01 which corresponds to an R S,P of ~2.1. This value is towards the higher end of estimates of purely thermal R S,P between ~1.3 to 2.2 [Cammarano et al., 2003;Goes et al., 2000], and is as expected, given the likely elevated mantle temperatures (and thus expected higher anelasticity effect) below the region indicated by previous work [Armitage et al., 2015;Ferguson et al., 2013;Rooney et al., 2012].
This result seems contradictory to the previous δ S,P , between ~5 to ~10, for stations within the MER . However, when we calculate the δ S,P only for the subset of stations centred around MER (i.e., more or less corresponding to the area covered by the EAGLE experiment that Bastow et al. [2005] analysed), we find a higher slope of 5.60 ± 0.03 corresponding to an R S,P ~3.2. This indicates that what Bastow et al. [2005] interpreted as the signature of melt, is likely local and shallow. Our new travel-time data set covers a significantly longer time span and a larger region of the rift, and thus the effects of local shallow melt signatures are mixed in with deeper mantle thermal signatures.
We also computed R S,P values after the tomographic inversion. These ratios are strongly scattered, ranging from about 0 to 4 (Fig. S13), as they are affected by any differences in spatial resolution of the two models. For this reason, care must be taken interpreting the R S,P distribution. Broadly, the R S,P distribution is peaked in the thermal range (median 1.7 for damping=35). The R S,P in the damped models is more strongly peaked in the thermal range, as it conditioned by our thermally-scaled starting model for P. This does however confirm that a thermal interpretation is compatible with the data. Also, as expected for thermal mechanisms (but also from melt or hydration), the peak in R S,P shifts to higher values in low velocity regions (median 1.9 for damping=35; Fig. S14). This trend is seen in both damped and undamped models.

Temperature estimates
As, to first order, the anomalies appear to be thermal in nature, we convert the lowvelocity anomalies in the two main clusters (see boxes in Fig. S7) to temperature anomalies, using the same method as in Civiero et al. [2015]. To account for the fact that the tomography underresolves anomaly amplitudes and the recovered amplitudes depend on regularisation, we first scale (Fig. 7a) the S velocities according to the amplitude resolution with depth estimated from an idealised plume model (Fig. 3, Fig S15). The effect, and uncertainties of the amplitude resolution are discussed in detail in Civiero et al. [2015], where the results from differently damped inversions were in a similar range. Next, the scaled dV S anomalies are converted to dT using a smooth dV S /dT derivative for a pyrolitic composition along a 1300°C adiabat (taken from Styles et al. [2011], Fig. S12). Uncertainties in this conversion to temperature anomalies are a few tens of degrees including consideration of uncertainties in the reference profile, and would largely work in the same direction for the conversion from P and S velocity anomalies, as both depend on shear modulus and density. Figure 7b shows that the inferred estimates of excess temperatures for both wave speeds agree well within their uncertainties and scatter, and fall between 100 and 200 °C.
Note that although the temperature estimates from dV P and dV S for both the Afar and MER clusters agree, the significant uncertainties and scatter would allow for some influence of other mechanisms. dT S estimates tend to be slightly higher than dT P above 500 km depth, while at larger depths dT S tends to be somewhat lower than dT P . The latter is probably due to the stronger decrease of the resolution with depth of V S than V P . The former may be a hint of a signature of deep melts. Thompson et al. [2015] used receiver functions to image a distinct low velocity zone above the transition zone beneath Afar. They interpreted this to be a melt layer caused by the release of volatiles from an upwelling below Afar. Rooney et al. [2012] also proposed a contribution of deep CO 2 -assisted melting to the very slow velocities below Afar. Our tomographic model does not have the resolution to distinguish such local features [Civiero et al., 2015], but does leave room for such an interpretation.
Our temperature estimates are consistent with those inferred from geochemistry and petrology [Ferguson et al., 2013;Rooney et al., 2012], from seismic studies of the shallow mantle [Armitage et al., 2015;Rychert et al., 2012] and transition zone [Thompson et al., 2015], and from estimates of melt production [Armitage et al., 2015]. The estimates are lower than inferred below some other hotspots, e.g., Hawaii, North Atlantic and Deccan [Armitage et al., 2010;Watson and McKenzie, 1991]. They are also somewhat lower than those from Mulibo and Nyblade [2013], who in a recent tomographic study conclude that a thermal perturbation of ~150-300K reconciles P-and S-wave velocities beneath Tanzania and northwestern Zambia. The temperature estimates below the Northern EAR are consistent with the interpretation that the low velocities through the transition zone are due to warm thermal upwellings.

Conclusions
In this study, we combine S-and SKS-wave relative travel times to achieve a highresolution tomographic shear-wave model, NEAR-S16, that extends throughout the whole upper mantle below of the Northern East-Africa. This S model complements a recent P travel time model, NEAR-P15, for the same region [Civiero et al., 2015]. The patterns of P-and Swave velocities strongly resemble each other in terms of geometry and scale. The tomographic images reveal widespread low velocities in the upper 200 km of the mantle below the region. Below this depth, two clusters of low velocities emerge, situated below Afar and west of the MER. These extend through the transition zone. Both the P and S low velocity anomalies yield excess temperature estimates in the range of 100-200°C in these two clusters. These results give us more confidence in interpreting the structure imaged in both Pand S-SKS-wave models below the Northern East-African rift as being mainly the signature of multiple small-scale upper-mantle thermal upwellings.
A comparison of our imaged structure with that of previous global tomographic studies, reveals the consistency of our structures with these large-scale models, as shown in Fig. 8 for the global anisotropic S velocity model SGLOBE-RANI [Chang et al., 2015]. As more data have been added to the global models, the complexity of the low-velocity structure below the East African Rift has increased, e.g., compare S20RTS [Ritsema and van Heijst, 2000] which imaged a single large-scale feature that appeared to be continuous from the lower mantle below southern Africa to the upper mantle from Kenya/Tanzania to Afar, with S40RTS  and SGLOBE-RANI [Chang et al., 2015], which find two low-velocity anomalies in the upper mantle below Kenya/Tanzania and Afar, as well as variability within the lower-mantle low-velocity anomalies (Fig. 8) . The comparison between our travel-time model and the global model SRANI clearly shows that what appeared like one large-scale feature below Afar in several previous studies [e.g., Hansen et al., 2012], has in fact internal small-scale structure, as was previously suggested based on surface wave tomography by Debayle et al. [2001]. Models S40RTS and SGLOBE-RANI even hint at a low-velocity anomaly at the base of the transition zone directly below the strong low velocities we image.
As discussed fully in Civiero et al. [2015], this might be a ponding of warm material that forms the source of the small-scale upwellings we infer, a feature that is seen in analogue and numerical models of plume transition-zone interaction [e.g., Kumagai et al. [2007]; Tosi and Yuen [2011]]. Evaluation of the global models along different cross sections indicates that this low velocity region at the base of the transition zone may connect with one of the lowvelocity features that appear to rise off the LLSVP below Africa. Further work to improve resolution of lower-mantle structure below the rift system, with more stations and/or by making use of a larger part of the waveforms will be necessary to identify the lower mantle source of the small-scale shallow mantle upwellings.   (panel b,f,h) (c,g,k) Recovered models using the same parameters as in the data inversion, but with no damping. (d,h,l) Fig. S5 to see the first-order similarities in structure between the independently inverted S/SKS-and P-velocity models. Compared to the undamped model (Suppl. Fig. S3), the shallow mantle anomalies are enhanced and more extensive.
Otherwise, similar features, including the low velocities below Afar and west of the MER, appear.
© 2016 American Geophysical Union. All rights reserved. The features in S model are similar to those seen in the P model (j,k).
© 2016 American Geophysical Union. All rights reserved.  Civiero et al. 2015) versus S-wave (this study) relative arrival-time residuals for all common earthquakes. The solid red line is a least-squares fit including picking errors for all data (blue and white circles) and has a gradient, δ S,P , of 3.69 ± 0.01 for all data corresponding to a ratio R S,P =lnV S /dlnV P of ~2.1, consistent with a dominantly thermal origin of the anomalies. The dashed red line is a least-squares fit including only data close to the MER (white circles) and has a gradient of 5.60 ± 0.03 (R S,P = ~3.2), previously interpreted as implying the presence of melt in a region around the MER ].
© 2016 American Geophysical Union. All rights reserved.  SGLOBE-RANI. NEAR-S16 is shaded gray where there are less than 5 rays per node. Note the consistency between the two models, with NEAR-S16 resolving more small-scale structure inside the large low velocity region in SGLOBE-RANI.