AFRP20: New P‐Wavespeed Model for the African Mantle Reveals Two Whole‐Mantle Plumes Below East Africa and Neoproterozoic Modification of the Tanzania Craton

Africa's Cenozoic tectonism is often attributed to mantle plumes, particularly below East Africa, but their morphology, number, location, and impact on the African lithosphere are debated. The broad slow wavespeed African Superplume, ubiquitous in large‐scale tomographic models, originates below South Africa, reaching the surface somewhere below East Africa. However, whether the diverse East African mantle geochemistry is best reconciled with one heterogeneous upwelling, or current tomographic models lack the resolution to image multiple distinct plumes, remains enigmatic. S‐wavespeed tomographic images of Africa are legion, but higher frequency P‐wavespeed whole‐mantle models possessing complementary diagnostic capabilities are comparatively lacking. This hinders attempts to disentangle the effects of Cenozoic hotspot tectonism and Pan African (and older) tectonic events on the East African lithosphere. Here we develop a continental‐scale P‐wave tomographic model capable of resolving structure from upper‐to‐lower mantle depths using a recently developed technique to extract absolute arrival‐times from noisy, temporary African seismograph deployments. Shallow‐mantle wavespeeds are δVP ≈ −4% below Ethiopia, but less anomalous (δVP ≥–2%) below other volcanic provinces. The heterogeneous African Superplume reaches the upper mantle below the Kenyan plateau. Below Ethiopia/Afar we image a second sub‐vertical slow wavespeed anomaly rooted near the core‐mantle boundary outside the African LLVP, meaning multiple disparately sourced whole‐mantle plumes may influence East African magmatism. In contrast to other African cratons, wavespeeds below Tanzania are only fast to 90–135 km depth. When interpreted alongside Lower Eocene on‐craton kimberlites, our results support pervasive metasomatic lithospheric modification caused by subduction during the Neoproterozoic Pan‐African orogeny.

generally reveal a broad (∼500 km-wide), inclined, slow wavespeed anomaly extending to the core-mantle boundary below South Africa: the African Superplume (e.g., Li et al., 2008;Ritsema et al., 1999). However, more recent models (e.g., Chang et al., 2015;Ritsema et al., 2011) provide tentative evidence for multiple narrower plume tails in the upper mantle, centered below Ethiopia and Kenya. Variability in along-rift isotopic signatures also favor a single heterogeneous lower mantle upwelling, common to and tapped by, multiple upper mantle plume stems (e.g., Furman, Kaleta, et al., 2006;Halldórsson et al., 2014;Nelson et al., 2012). A third interpretation posits that two completely distinct upwellings are required to explain the age distribution and some geochemical characteristics of East African basalts (George et al., 1998;Rogers et al., 2000), but seismic evidence supporting this hypothesis is lacking. Elsewhere, the extent to which Cenozoic intra-plate magmatism in Cameroon and Madagascar requires influence from upwellings sourced below the upper mantle is also uncertain (e.g., Cucciniello et al., 2017;French & Romanowicz, 2015;King & Ritsema, 2000;Pratt et al., 2017).
Sutured during the Pan African orogeny along a series of Paleoproterozoic mobile belts, and now interspersed amongst regions of Cenozoic hotspot tectonism, numerous Archean cratons characterize the African plate. It is thus an ideal place to study the modification of cratonic mantle lithosphere (Figure 1). Subduction/orogenesis, rifting, and hotspot tectonism are all capable of modifying Archean lithosphere (e.g., Boyce et al., 2016;Liao et al., 2017;Wang et al., 2015;Wenker & Beaumont, 2018), but isolating evidence for each process is challenging. Temporal constraints on the development of lithospheric mantle composition can be derived from diamond-bearing kimberlites and mantle xenoliths, but such geological samples are spatially limited across Africa (e.g., Begg et al., 2009). Despite their significant temporal separation, the relative importance of Proterozoic "Pan African" orogenesis and Cenozoic hotspot tectonism in modifying Africa's Archean mantle lithosphere is therefore uncertain (e.g., Koornneef et al., 2009), meaning concurrent analysis of seismological and xenolith data is essential. For example, the relationship between the edges of fast wavespeed domains and the spatial distribution of African kimberlite magmatism is debated (e.g., Begg BOYCE ET AL.  (Engdahl et al., 1998). Red triangles: 1994-2019 temporary seismograph station deployments from which absolute arrival-times are obtained via the method of Boyce et al. (2017). Magenta lines: plate boundaries. Light blue dashed circles: The Ethiopian (EP) and Kenyan (KP) topographic plateaus. (b) African crustal geology after Begg et al. (2009). Transparent regions with black dashed outline mark inferred craton extents, presently covered by Phanerozoic sediments. Orange triangles: Quaternary volcanoes. Gray triangles: Carbonatite volcanoes (Woolley & Kjarsgaard, 2008). Black diamonds: kimberlites (Tappe et al., 2018) Celli et al., 2020;Fishwick, 2010;Priestley et al., 2008): some authors suggest that a significant portion of thick African lithosphere has been removed, perhaps by recent mantle plume thermal erosion (Celli et al., 2020). In Tanzania specifically, the cratonic lithosphere displays basal topographic asymmetry and is also significantly thinner (∼160 km on average Adams et al., 2012;Emry et al., 2019;O'Donnell et al., 2013;Weeraratne et al., 2003) compared to other cratons worldwide (≥200 km: Lee & Rudnick, 1999;Priestley et al., 2019). On-craton petrological and seismological evidence for Tanzanian craton evolution toward its present-day state is relatively sparse (Stachel et al., 1998;Wölbern et al., 2012). Hence the areal extent, timing and causal mechanism for post-formation modification (e.g., Pan-African subduction, recent Cenozoic hotspot tectonism and/or rifting) remain unresolved (e.g., Koornneef et al., 2009).
While shear waves are sensitive primarily to temperature over composition (in the absence of melt; Goes et al., 2000), compressional wavespeed constraints are a powerful tool to understand lithospheric modification because they have stronger sensitivity to composition. For example, lithospheric modification through metasomatic addition of silica to olivine leads to enrichment in orthopyroxene by water-rock/ melt-rock interaction during subduction (e.g., Wagner et al., 2008). This can decrease V P at mantle lithospheric depths by >1%, with little accompanying impact on V S and density (Schutt & Lesher, 2010). High resolution African P-wave relative arrival-time studies are abundant (e.g., Bastow et al., 2005;Fouch et al., 2004;Mulibo & Nyblade, 2013b;Reusch et al., 2010), but do not constrain background mean velocity structure; their amplitudes are thus not comparable, or relatable in a consistent way to thermal and chemical mantle properties (e.g., Bastow, 2012). Global-scale P-wave tomographic models (e.g., Hosseini et al., 2019;Montelli et al., 2006;Simmons et al., 2012) sample long wavelength African mantle structure but are typically tuned to resolve mid-to-lower mantle depths. Continent-scale African absolute P-wave tomographic models are scarce (e.g., Hansen & Nyblade, 2013;Hansen et al., 2012), and none tackle the issue of how to extract absolute arrival-times reliably from temporary seismograph networks, whose short station spacing is required for high resolution lithospheric imaging ( Figure 2). The inclusion of data from these networks would also enhance resolution in the lower mantle, pertinent to the aforementioned African plume debates. BOYCE ET AL.   Boyce et al., 2017) with respect to ak135 (Kennett et al., 1995). Negative residuals are early arrivals (fast); positive residuals are late arrivals (slow). Residuals are corrected for station elevation and plotted on an asymmetric color bar from +4 s to −2 s delay time. Magenta lines: plate boundaries. (b) Histogram of residuals binned by year.
Here, we utilize the absolute arrival-time recovery method (AARM) of Boyce et al. (2017) to embed accurately 25 years of data from temporary seismograph networks across Africa ( Figure 1) into a new absolute P-wave mantle tomographic model. Our body wave model facilitates direct comparison of slow wavespeed upper mantle structure underlying the Atlas Mountains, the Cameroon Volcanic Line, Madagascar and East Africa. We explore new seismological evidence for two whole-mantle plumes beneath East Africa, sourced at disparate locations at the core-mantle boundary. We also investigate whether Proterozoic (Pan-African) subduction driven modification below the Tanzanian craton can explain its anti-correlated shear and compressional lithospheric wavespeeds.

East African Rift Cenozoic Tectonics and Magmatism
Seismically and magmatically active rifting define the topography of East Africa from Ethiopia and Arabia in the north, to southern Africa ( Figure 1). The southern Red Sea, Main Ethiopian, and East African Rift systems all developed atop broad topographic plateaus. Uplift of the Ethiopian plateau began 20-30 Ma on the basis of (U-Th)/He geothermometry (Pik et al., 2003), with lithospheric foundering in response to extensive heating of the lithosphere perhaps increasing elevation further in Miocene times (Furman et al., 2016;Gani et al., 2007). East African plateau uplift likely began later, at ∼14 Ma (Wichura et al., 2010). The low-lying Turkana Depression was below sea-level in pre-Neogene times (e.g., Sepulchre et al., 2006), and is thought to have risen by ∼600 m concurrently with East African Plateau uplift. Turkana's complex history of multiple rifting phases allows the possibility that the plateaus are one uplifted region, extending from southern Africa to Ethiopia/Arabia. Cenozoic magmatism began in southern Ethiopia and Kenya at 45-30 Ma George et al., 1998), followed by the emplacement of the 31-29 Ma Ethiopian/Arabian flood basalt province. Subsequent low-volume magmatism became pervasive throughout the East African Rift from 10 Ma (e.g., Baker et al., 1996;Rooney, 2020a). There is general consensus in the petrological literature that the Ethiopian and East African mantle is hot (Ferguson et al., 2013;Rooney et al., 2012), with potential temperatures (Tp) elevated by 100°C-140°C at the present day, following a 170°C peak at 30 Ma.

Global and Regional Data Sets
We use a global data set of P, Pn, Pg, pP, PKP, and PKIKP phase absolute arrival-time picks from the "EHB" database (Engdahl et al., 1998) recorded from 1964 to 2007 (Figure 1; Li et al., 2008). Figures S1 and S2 show all earthquakes and stations used in this study. Figure S3 shows mean direct-P absolute arrival-time residuals for Africa and the surrounding regions. The African EHB data set yields 182,853 direct-P phase picks and 48,945 Pn phase picks. Picks from the EHB database are combined with new absolute arrival-time direct-P phase picks from temporary seismograph station deployments from 1994 to 2019 that greatly improve spatial sampling of the African mantle, particularly in the east and south (red triangles in Figure 1). To glean first-breaking (or onset time) information from these stations for the first-arriving phase, we use AARM (Boyce et al., 2017) to accurately extract 87,184 absolute arrival-times from these temporary seismic deployments. For each earthquake, filtered waveforms (0.4-2.0 Hz) are initially aligned on their first visible coherent peak or trough in a narrow 3 s window following the relative arrival-time approach of VanDecar and Crosson (1990) to determine a set of time shifts. These time shifts are then applied to the original, unfiltered seismograms prior to phase weighted stacking, which is further optimized by weighting each trace in a final stack by its cross-correlation coefficient with the initial stack. The first arrival or onset time exists somewhere prior to the first visible coherent peak or trough that was picked during the relative arrival-time analysis. The separation of first arrival-time and subsequent peak or trough (used previously for phase alignment) is unique to the coherent phase arrival from each teleseismic earthquake across a regional seismic network and is primarily controlled by variable background noise level and coherent phase period. The first arrival is revealed on the high SNR final stack and is picked to extract a common time correction for each earthquake. Absolute arrival-times are extracted by applying the common time correction to the individual trace alignment times. To ensure waveform similarity during the alignment process, the temporary seismic station distribution is split into six subregions, namely; Atlas Mountains, Cameroon, Ethiopia, East African Rift, Madagascar and Southern Africa. Data are processed within these subregions (Figures S4-S16) until being combined in the inversion process below.
AARM-derived mean P-wave absolute arrival-time residuals with respect to ak135 (Kennett et al., 1995) are plotted in Figure 2. Because absolute arrival-time residuals record wavespeed variation encountered along the entire source-to-receiver raypath, the first order variations seen in relative arrival-time studies (e.g., Bastow et al., 2008;Boyce et al., 2016) are not expected. For example, much of southern and eastern Africa displays late mean P-wave absolute arrival-times despite the presence of fast wavespeed cratonic lithosphere. However, the main Ethiopian rift shows the lowest mean absolute arrival-time residuals of ≈ 4 s across Africa and a polarity change from low to high mean arrival-time residuals is observed perpendicular to the (Figure 2).

Inversion Technique
We employ the tomographic inversion method of Li et al. (2008) and Burdick et al. (2017). Rays are traced through the 1D reference model ak135 and are clustered by station and event location similarity to form composite rays. Composite rays are weighted by the number of constituent rays (Kárason & Van der Hilst, 2001). The initial model is parameterized on a regular, global grid with cells spaced 45 km in depth, 0.35° in latitude and longitude to account for heterogeneity along the entire raypath. In order to account for incomplete global raypath sampling, adjacent under-sampled grid cells are combined to ensure a minimum of 900 rays per cell ( Figure S17). The iterative, linearized, least squares inversion takes place upon the adaptively parameterized grid and is used to solve for hypocenter mislocation and slowness with respect to ak135. This procedure minimizes the following function: In the first term, the ray-theoretical sensitivity matrix, G, model, m (which includes slowness and hypocenter mislocation parameters) and data, d, are weighted by w. Following previous work, we weight our newly added high-quality absolute arrival-times three times that of the more numerous, yet noisier EHB data set to balance their contribution (e.g., Boyce et al., 2019;Kárason & Van der Hilst, 2001;Li et al., 2008). The inversion is smoothed in the horizontal and vertical directions by the second term, in which L is a first-derivative smoothing operator, and damped by the third (i.e., model norm). These regularization terms are weighted by λ i relative to the first to control model misfit. λ i are determined through trade-off analysis ( Figure S18) which yields a final model with RMS residual reduction of 49.8% after 400 inversion iterations, although convergence is reached much earlier. We plot P-wavespeed anomalies as a percentage with respect to ak135 and focus on the African portion of the global model-space (40°S -40°N and 20°W -60°E).
We account for crustal heterogeneity by making receiver-side corrections to the residuals based on Moho depths within the Crust1.0 model (Laske et al., 2013) supplemented with receiver function constraints for African temporary seismic deployments where available (Akpan et al., 2016;Andriampenomanana et al., 2017;Ebinger et al., 2017;Fadel et al., 2018;Hosny & Nyblade, 2016;Lemnifi et al., 2017;Ogden et al., 2019). Details on generation of this crustal model and its effect on our tomographic inversion are presented in the supplementary material (Figures S19-S21). We propagate rays through the bespoke crustal model, calculate the component of each residual produced by the crustal model and remove this from the data vector (d) prior to the inversion.

Resolution Assessment
Structural and checkerboard tests (Figures 3, 4, and S22-S37) appraise the resolving power of our tomographic model (AFRP20) beneath Africa. Using identical ray paths to the observed data, we calculate arrival-time residuals through synthetic wavespeed models. We invert these data following addition of 0.2 s standard deviation Gaussian noise. Visual defects within input anomaly models arise from imposing our input wavespeed anomalies onto the coarse adaptive grid in poorly sampled regions.
Amplitude recovery of the broad, inclined slow wavespeed structure is ≥ 60% throughout the mid-to-lower mantle ( Figure 3) but is laterally smeared at ≥1,000 km depth where grid cells are coarsely spaced (Figure 3). Slow wavespeed amplitude recovery to the northeast of the Tanzanian craton is muted in the upper mantle likely due to limited overlying station coverage and the δV P = 0% buffer region that surrounds the Tanzanian craton input anomaly. The sub-vertical slow wavespeed anomaly beneath Ethiopia/Afar is recovered at ≥80% to ≥1,200 km depth. Below, amplitude recovery is reduced to ∼30% ( Figure 3) with some lateral smearing. A synthetic upper mantle CVL anomaly is smeared vertically to ∼660 km depth with <50% amplitude ( Figure S22). Slow mid-mantle wavespeeds associated with the synthetic African Super- Figure 3. An inclined, elliptical, input slow wavespeed anomaly (δV P = −2.0%) increases in areal extent from the surface, centered at 10°N, 40°E, to the coremantle boundary, centered at 35°S, 15°E (a and e). We also add a narrower planform, more steeply dipping slow wavespeed anomaly (δV P = −2.0%) extending from the surface, centered at 10°N, 42°E, to the core-mantle boundary, centered at 13°N, 44°E (a and e). We insert δV P = +2% wavespeed anomalies as cratonic cores in the upper 270 km in west Africa, Congo and southern Africa and the upper 180 km in Tanzania. We insert a δV P = −2% wavespeed anomaly below the Cameroon Volcanic Line to 270 km depth (a and e). We ensure that input whole-mantle slow wavespeed anomalies do not overlap upper mantle input anomalies using a δV P = 0% buffer region. Visual defects within the input anomaly model result from a coarse adaptive grid in poorly sampled regions. The cross-section location is indicated in the inset map between (e and f) Recovered anomalies, shown at 100, 1,000, 2,500 km depth and in cross-section (b-d, and f) are displayed on the same color scale as the input.
plume structure to the east are not smeared laterally to below the CVL in our tests ( Figure S22). We expect to identify a discontinuity within slow wavespeed anomalies if amplitudes are reduced to <40% of the over-or-underlying anomaly ( Figures S23 and S24). Further, we expect to identify a ponded slow wavespeed layer at the base of the MTZ with thickness >45 km ( Figure S25).
At broad scales, the west African and Congo cratons are poorly resolved reflecting sparse station density ( Figure 3b). Southern African cratons are recovered with ≥60% amplitude but are vertically smeared to ∼140% of input anomaly thickness. The Tanzanian craton is reasonably well recovered laterally, although the surrounding slow wavespeeds do smear inwards toward Tanzania. The uniform thickness (180 km) Tanzanian input anomaly is recovered with ∼70%-95% of the input thickness. Amplitude recovery is ∼25% for ≥100 km depths, but is ∼50% at shallower depths (Figures 3b and 3f).
Asymmetrical fast cratonic wavespeeds and underlying slow asthenospheric wavespeeds are reasonably well recovered below the Tanzanian craton ( Figure 4). Fast wavespeed amplitude recovery in shallow mantle below the western craton is ∼75% decreasing significantly below ∼90 km depth and to the east. The thicker western region is recovered with ≥80% of the input thickness (Figure 4e), while the thinner eastern region is recovered at approximately the input thickness ( Figure 4h). We can therefore be reasonably confident of the depth extent of imaged fast wavespeeds below Tanzania on the length-scale of our vertical parameterization (45 km). The asymmetry of slow wavespeeds underlying Tanzania is reasonably well recovered with peak amplitude recovery of ∼50%, although some lateral smearing does exist. The upper boundary of slow wavespeed anomalies is constrained to within ±100-150 km depth but the lower boundary is poorly constrained due to significant vertical smearing.
A simple rectangular or elliptical fast wavespeed Tanzanian craton is also resolvable (∼50% amplitude, Figures S32-S34). While lateral smearing is relatively minor, inherent vertical smearing is approximately double the anomaly input thickness in the case of no underlying slow wavespeeds ( Figures S32 and S33). Conversely, when underlain by slow wavespeeds, the depth extent of the fast wavespeed anomaly is recovered at ∼80%-100% of the input anomaly thickness but amplitude recovery is reduced to ∼30% ( Figure S34). Madagascan resolution tests indicate that although upwards and downward vertical smearing is present (±100 km), if slow wavespeed anomalies are isolated to the upper mantle, wavespeed anomaly amplitudes will decrease rapidly below 660 km depth ( Figure S35). Resolution tests of East African upper mantle slow wavespeed anomalies ( Figures S36 and S37), indicate that if ambient mantle is present below the Turkana Depression, wavespeed anomalies will be significantly muted compared to adjacent slow wavespeed regions, despite the lack of overlying stations.

Results
By utilizing AARM (Boyce et al., 2017) to capitalize on African temporary seismograph network data, the AFRP20 tomographic model (Figures 5-9) can resolve short wavelength anomalies in the upper mantle, particularly below densely instrumented eastern and southern Africa, and larger structures within the mid-BOYCE ET AL.
10.1029/2020GC009302 9 of 27 To test the sensitivity of AFRP20 to complex upper mantle wavespeed structure below Tanzania, we insert a rectangular (9-0°S, 30-36°E) fast wavespeed anomaly in the upper mantle below Tanzania (δV P = +2.0%). This anomaly extends to 180 km depth in the west and 135 km in the east. We arrange a series of slow wavespeed anomalies (δV P = −2%) as follows: a shallow (50-300 km depth) anomaly centered at ∼2°N, 35°E to the northeast of the Tanzanian craton, a tabular anomaly below the eastern Tanzanian craton (150-550 km depth) and a tabular anomaly below the western Tanzanian craton at greater depths (410-660 km depth). Visual defects within the input anomaly model result from a coarse adaptive grid in poorly sampled regions. The cross-section locations (A-E) are indicated in (a) Yellow circles and red triangles on map (c) show the locations of "EHB" stations and temporary deployments analyzed in this study. The test input (a, d, g, j, m, and p), output (b, e, h, k, n, and q) and AFRP20 (c, f, i, l, o, and r) are displayed on the same color scale.
to-lower mantle ( Figures S22-S37). This method facilitates significant improvement over relative arrival-time tomographic models that lack resolving power below ∼600-800 km depth (e.g., Bastow et al., 2008;Civiero et al., 2015;Mulibo & Nyblade, 2013b), previous African continental P-wave models (e.g., Hansen et al., 2012) and global P-wave tomographic models (e.g., Hosseini et al., 2019;Li et al., 2008;Montelli et al., 2006;Simmons et al., 2012) less focused on upper mantle resolution (see Figure S38 for comparison with global P-wave models). Lower mantle raypath sampling of our continental data set ( Figure S39) indicates that slow wavespeed anomalies can be resolved beneath the Indian Ocean by significant crossing ray coverage. Over-sampling bias is accounted for during the ray clustering procedure.
Above 300 km depth, the upper mantle within AFRP20 is dominated by slow wavespeeds (δV P ≤ −2%) in the northeast and the fast wavespeed (δV P ≥1%) cores of the Zimbabwe and Kaapvaal cratons in the south as well as some parts of the poorly resolved West African and Congo cratons (Figures 5,8,and 9 (Li et al., 2008). Particularly at these upper mantle depths, resolution offered by the global data set is limited to regions of high earthquake and station density (Figures S1, S2, and S27-S31).
blade, 2007; Priestley et al., 2008) are smeared to ≥410 km depth due to the lack of station coverage (Figures 6  and 8). In the mid-to-lower mantle, fast wavespeed anomalies are limited to below western, northeastern and the very southern tip of Africa, likely reflecting the paucity of subduction below the continent since ∼550 Ma. Instead, the mid-to-lower mantle in AFRP20 is dominated by slow wavespeed anomalies and is reflected in the mean absolute delay time of 1.36 s ( Figure 2) and average wavespeed anomaly within AFRP20 of  P V = −0.216% with respect to ak135. In the following sections we highlight key features of AFRP20 in each region where seismograph stations are present and offer comparisons to previous studies.
10.1029/2020GC009302 11 of 27 Figure 6. AFRP20 African tomographic model at mid-mantle depths of 410, 660, 900, and 1,000 km (a-d) plotted as percentage deviation from ak135 (δV P = ±1%). Gray regions are not constrained by our new AARM-derived African data set (see supporting information), only by the global data set (Li et al., 2008). slow wavespeed anomalies extend below 300 km depth in AFRP20. Although AFRP20 does not include data from seismograph stations in the Turkana Depression of southern Ethiopia and northern Kenya, the slow wavespeed anomalies in the upper mantle (δV P ≈ −1.25%) are not likely an inversion artifact ( Figures  S36 and S37).
East Africa shows a striking difference in the very shallow mantle compared to the rest of the upper mantle (Figures 4 and 5). Fast wavespeeds (δV P ≈ 0.75%-1.5%) are generally limited to the upper ∼135 km below the western Tanzanian craton and the upper ∼90 km below the eastern craton (Figures 4 and 5). Adjacent isolated slow wavespeed anomalies coincide with resolvable portions of the east and west rift branches at these shallow depths. Below, fast wavespeeds are confined to a small, isolated region beneath western Lake Victoria, while the rest of the Tanzanian mantle, extending beyond 200 km depth only displays slow wavespeeds (δV P ≈ −0.9%). Although less well resolved (Figures 3 and 4)  . AFRP20 African tomographic model at lower mantle depths of 1,300, 1,600, 1,900, and 2,300 km (a-d) plotted as percentage deviation from ak135 (δV P = ±1%). Gray regions are not constrained by our new AARM-derived African data set (see supporting information), only by the global data set (Li et al., 2008). Within subplot (d), dashed and dotted dark gray lines mark regions of 5/5 slow-wavespeed votes near the core-mantle boundary in δV S and δV P tomographic models respectively, clustered by Cottaar and Lekić (2016). Anomaly B lies inside these contours while anomaly A lies outside. broadly restricted to beneath the adjacent Congo and southern African cratons below the very shallow mantle.
Depending on the imposed underlying slow wavespeed structure, resolution tests (Figures 3, 4, and S34) indicate that AFRP20 can resolve depth variability in fast wavespeeds below the Tanzanian craton to within 80%-100% of input anomaly thickness on the vertical length-scale of our depth parameterization (45 km). These tests show that thinner input wavespeed anomalies (∼90 km) are recovered with greater reliability in depth than thicker input anomalies (≥180 km), due to the large Pn data set (>1,000 arrivals) at East African EHB stations. Surface wave models (e.g., Adams et al., 2012;Emry et al.,    (a-f) P-wavespeeds at 100 km depth within each region for which spatial sampling is greatly improved by the addition of new AARM-derived absolute arrival-times (Figure 2). Orange triangles: Quaternary volcanoes. Gray triangles: Carbonatite volcanoes; the only active Carbonatite volcano at Ol Doinyo Lengai, Tanzania is outlined pink (Woolley & Kjarsgaard, 2008). Black diamonds: kimberlite locations; those erupted since 500 and 50 Ma are outlined in gray and pink, respectively (Tappe et al., 2018). Yellow regions: Cenozoic volcanic provinces. Black dashed polygons: the craton outlines shown in Figure 1. AF, Afar depression; AFB, Arabian flood basalts; AS, Alboran Sea; CM, Comoros and Mayote; EFB, Ethiopian flood basalts; LV, Lake Victoria. Other acronyms follow Figure 1. Wavespeed plotted as percentage deviation from ak135 on an asymmetric color scale (−4% ≤ δV P ≤ 2%) to illuminate the full extent of wavespeed heterogeneity. west but are perhaps ∼30-50 km shallower in the east. While our results also suggest that fast wavespeeds below the Tanzanian craton are thinner than observed for other cratons elsewhere (typically greater than 200 km), our fast P-wavespeeds appear to be thinner still (∼135-90 km) than the fast shear wavespeeds imaged previously.

Mid-to-Lower Mantle
Slow wavespeeds (δV P ≤ −1%) characterize the mantle transition zone beneath northeast Africa (Figures 6  and 8). The slow wavespeed region separates into two distinct anomalies of δV P ≤ −0.8% below 660 km depth, centered beneath the Ethiopian and Kenyan plateaus (A: ∼8°N, 42°E and B: ∼7°S, 34°E, respectively). Between these slow wavespeed regions, below the Turkana depression, a positive, low amplitude wavespeed anomaly at 900-1,200 km depth is also imaged (Figures 6 and 8, anomaly C). Two distinct wavespeed anomalies below the uplifted Ethiopian and Kenyan plateaus have been imaged to ∼800 km depth previously in global tomographic models (Chang et al., 2015;Li et al., 2008;Montelli et al., 2006;Ritsema et al., 2011), but their extension into the lower mantle is unclear. Slow wavespeed anomalies penetrating the lower mantle beneath the EAR and/or Arabia have also been imaged by regional studies (e.g., Bastow et al., 2008;Chang & Van der Lee, 2011;Civiero et al., 2015;Mulibo & Nyblade, 2013b), however such studies typically suffer a significant resolution decrease below the upper mantle.
Over the depth range 1,300-2,300 km, the two slow wavespeed anomalies (δV P ≤ −0.6%) broaden and separate with depth. The northern anomaly (A), shifts approximately eastwards with depth (Figures 7 and 8) and appears to bifurcate at ∼1,600 km, reaching the lower mantle below the Indian Ocean, northeast of Madagascar (Figures 5-8). Hosseini et al. (2019) show evidence for distinct slow wavespeeds in the upper and lower mantle below Ethiopia and the Indian Ocean respectively but do not establish a connection between these depths. The southern anomaly (B), often referred to as the African Superplume, extends southwest with increasing depth over a broader spatial footprint (Figures 7 and 8) and has been imaged frequently (e.g., Chang et al., 2015;Hansen et al., 2012;Hosseini et al., 2019;Li et al., 2008;Ritsema et al., 1999;Ritsema et al., 2011). Because these anomalies appear to dip out of plane the reader is encouraged to study the cross-sections presented in Figure 8 and the three-dimensional rendering of the AFRP20 tomographic model within the supplementary material ( Figures S40-S43). Resolution tests indicate two relatively broad subvertical features are well resolved in the mid-to-lower mantle ( Figure 3) and horizontal gaps in these slow wavespeed anomalies would be resolved if present (Figures S23 and S24).

Southern Africa
Southern Africa is dominated by two fast wavespeed regions centered on ∼26°S, 28°E and ∼17°S, 30°E of δV P ≤ 2% (Figures 5 and 9). These anomalies are collocated with the Archean Kaapvaal and Zimbabwe cratons (e.g., Begg et al., 2009), in general agreement with previous P-wavespeed results (e.g., Fouch et al., 2004;Youssof et al., 2015). The lateral extent of these fast wavespeed regions narrows significantly between 100 and 200 km depth ( Figure 5). They are separated and surrounded by less anomalous wavespeed domains of δV P ≈ 1%, such as the Proterozoic Limpopo belt. Southwestern Africa hosts a weak but distinct slow wavespeed anomaly (δV P ≈ −0.3%) at ∼32°S, 21°E, beneath the Namaqua-Natal mobile belt. Similar to the resolution tests, fast wavespeed anomalies in southern Africa are somewhat smeared below 300 km depth (Figures 3 and 6) but their amplitudes are greatly reduced (δV P ≤ 0.6%).

Madagascar
Madagascar's upper mantle is largely characterized by slow wavespeeds of up to δV P ≈ −0.8%, primarily underlying northern and central Cenozoic magmatic provinces and the southwestern region (Figures 5 and 9). Adjacent to the central Madagascan volcanic province, a circular fast wavespeed anomaly exists (δV P ≥1.5%), centered at ∼19°S, 48°E. This feature extends to the northeast with depth and is located beneath thicker lithosphere (∼120 km) than is present below adjacent magmatic provinces (Rajaonarison et al., 2020). Resolution tests ( Figure S35) demonstrate that this feature is not likely an inversion artifact, although anomaly amplitude appears to be strongly influenced by the crustal correction ( Figure S21). Resolution tests also indicate that the extension of slow wavespeeds beneath central Madagascar to below upper mantle depths is not entirely the result of vertical smearing ( Figure S35). Regions of slow upper mantle wavespeeds below central and northern Madagascan magmatic provinces have been imaged previously in the depth range ∼80-165 km (Barruol et al., 2019;Celli et al., 2020;Emry et al., 2019;Mazzullo et al., 2017;Pratt et al., 2017). Perhaps due to decreased surface wave resolution beyond the upper mantle, these studies debate whether slow upper mantle wavespeeds extend pervasively to the northeast (e.g., Barruol et al., 2019;Mazzullo et al., 2017) or directly below central Madagascar with depth (e.g., Emry et al., 2019;Pratt et al., 2017). The distinctly slow upper mantle wavespeeds below Madagascar in AFRP20 (δV P ≤ −0.6%) extend southwest through the mantle transition zone but terminate at ∼1,000 km depth ( Figure 6).

Cameroon
The onshore component of the CVL is underlain at 100 km depth by a slow wavespeed anomaly (δV P ≤ 1%: Figures 5 and 9) around 7°N, 11°E that extends to the northeast as imaged previously (Adams et al., 2015;Celli et al., 2020;Emry et al., 2019;Reusch et al., 2010). An offshore wavespeed anomaly would not be resolved since station coverage there is limited. To the southeast a fast wavespeed anomaly (δV P ≥2%) underlies the Congo Craton. Previous studies (Adams et al., 2015;Reusch et al., 2010Reusch et al., , 2011 have suggested that both the slow and adjacent fast wavespeeds beneath the CVL and Congo Craton respectively are limited to the upper 300 km. We observe slow wavespeeds extending below 660 km depth ( Figures 5 and 6), likely resulting from significant vertical smearing ( Figure S22). However, resolution tests indicate that the imaged slow wavespeed connection at mid-mantle depths extending from below the CVL to the lower mantle below southern/eastern Africa is not entirely attributable to further lateral/vertical smearing ( Figure S22).

Discussion
By embedding ≥87,000 high quality P-wave teleseismic absolute arrival-times from seismograph stations across Africa within a global P-wave arrival-time data base and tomographic inversion, AFRP20 offers fresh scope to investigate the African mantle across the entire depth range. We begin with an investigation into broad-scale mantle heterogeneity below Africa, before focusing in on upper mantle wavespeed structure below regions of Cenozoic magmatism and cratonic lithosphere that possess good seismograph station coverage (Figures 1 and 2).

Broad-Scale African Mantle Structure: Evidence for Multiple Whole-Mantle Upwellings?
AFRP20 illuminates two distinct, diverging, slow wavespeed anomalies in the mid-to-lower mantle (anomalies A and B: Figures 6-8) and so motivates discussion of competing models of plume related magmatism in East Africa. High 3 He/ 4 He ratios (R/Ra ≈ 20) in the Ethiopia-Afar volcanic province and the lack of a depleted MORB mantle source of pre-rift magmas provide compelling support for a deep mantle source for East African Cenozoic magmatism (e.g., Marty et al., 1996;Nelson et al., 2012;Pik et al., 2006). Geochemical measurements and geodynamic models have frequently been interpreted in light of global seismic tomographic models that possess modest African resolution reflecting limited data availability. In a recent review, Rooney (2017) summarized the three most popular mantle plume models invoked to explain East African magmatism: (1) a single thermochemical upwelling feeding all magmatism (e.g., Ebinger & Sleep, 1998;Forte et al., 2010;Hansen et al., 2012;Simmons et al., 2007), (2) two branches of the same thermochemical upwelling feeding northern and southern rift volcanism (e.g., Furman, Kaleta, et al., 2006;Nelson et al., 2012) or (3) two distinct thermochemical upwellings below the northern and southern re-gions, perhaps sourced at different depths or disparate locations at the core-mantle boundary (e.g., George et al., 1998;Rogers et al., 2000Rogers et al., , 2010. A single thermochemical upwelling feeding East African magmatism was championed by modeling work (Ebinger & Sleep, 1998;Forte et al., 2010) and previous large-scale tomographic models (e.g., Hansen et al., 2012;Ritsema et al., 1999). These studies revealed a broad (>500 km wide) inclined slow wavespeed anomaly extending from the African LLVP in the lower mantle below southern Africa (e.g., Cottaar & Lekić, 2016) to the upper mantle below the EAR in Kenya: the African Superplume. The Superplume in the lower mantle is routinely interpreted to be chemically distinct from "normal" thermal plume material on account of its inclination (Simmons et al., 2007) and sharp sides (Ni et al., 2002). Apparent plume inclination may result from whole-mantle convection processes (Steinberger & Torsvik, 2012) or the tomographic manifestation of vertically rising plumes at sequentially advanced stages of ascent (Davaille et al., 2005).
A developing body of geochemical literature, documenting variability in isotopic ratios of along-rift magmas but with no clear north-south trend, proposes that erupted basalts are heterogeneous samples of one broad thermochemical upwelling (e.g., Furman, Kaleta, et al., 2006;Nelson et al., 2012;Halldórsson et al., 2014). In this scenario, multiple upper mantle plume stems rise from a common, deep mantle source, the most likely candidate being the African LLVP and overlying Superplume. Heterogeneity may arise from variability in thermochemical entrainment within upper mantle plume branches, lithospheric contamination and geochemical sampling strategies (e.g., Barry et al., 2013;Williams et al., 2015). Below the Ethiopian and Kenyan plateaus, two narrow cylindrical slow wavespeed anomalies, confined to the upper mantle, underlain by the Superplume structure have been imaged previously (Chang et al., 2015;Ritsema et al., 2011) providing tomographic support for the multiple upper mantle plume stem model. However, these tomographic models utilize fewer African seismic stations than AFRP20. Regional seismic studies have also interpreted a number of small slow wavespeed conduits beneath the EAR and/or Arabia but have lacked resolution into the lower mantle (Chang & Van der Lee, 2011;Civiero et al., 2015;Emry et al., 2019).
Lastly, two distinct thermochemical upwellings below the northern and southern regions, perhaps sourced at different depths or from different regions at the core-mantle boundary are supported by dating (George et al., 1998) and isotopic ratios (Rogers et al., 2000(Rogers et al., , 2010 of Kenyan and Afar basalts. Geodynamic modeling of African mantle plumes shows that the timings of volcanism are hard to reconcile with a single plume model (Lin et al., 2005), and different plume types/morphologies can coexist in relative proximity (Farnetani & Samuel, 2005). AFRP20 shows evidence for two slow wavespeed anomalies that extend throughout the mid-to-lower mantle (anomalies A and B: Figures 6-8). Although not readily distinguished in the upper mantle, these anomalies do not appear to merge at greater depth or sample a similar portion of the lower mantle (Figures 3, and S23-S25). The elongate heterogeneous wavespeed structure within anomaly B is not easily reconciled with the model of multiple vertically rising plumes at sequentially advanced stages of ascent (Davaille et al., 2005) meaning an alternative mechanism to explain plume tilt is favored below Africa (e.g., Farnetani & Samuel, 2005;Simmons et al., 2007;Steinberger & Torsvik, 2012). Further, supported by specific resolution testing ( Figure S25), AFRP20 shows little evidence for a ponded layer of slow wavespeed material at lower mantle transition zone depths, above which multiple upper mantle plume stems could emerge. However, we cannot rule out the presence of a ponded slow wavespeed layer of thickness less than our vertical parameterization of 45 km. Our resolution testing shows that a significant horizontal interruption in slow wavespeeds would also be visible if present (Figures S23 and S24). While numerous tomographic models image the African Superplume, tentative evidence for a second anomaly exists within previous models but a clear connection from two anomalies within the mid-mantle to the core-mantle boundary (Li et al., 2008;Montelli et al., 2006;Simmons et al., 2012) or between the upper and lower mantle (Hosseini et al., 2019) is not conclusive ( Figure S38). AFRP20 may therefore be the first large-scale tomographic model to convincingly image two distinct whole-mantle plumes underlying East African magmatism.
While AFRP20 cannot constrain uniquely the thermochemical state of the imaged mantle upwellings, their location within the lower mantle is notable. The African Superplume (Figures 6-8, anomaly B) appears to be anchored in the lower mantle (>2,300 km depth) west of southern Africa at ∼25°S, 2°E. This cor-relates well with the African LLVP beneath southern Africa and the Atlantic (Cottaar & Lekić, 2016, see Figure 7d). Although anomaly A appears to bifurcate below ∼1,600 km depth, its extent into the lower mantle below the Indian Ocean, northeast of Madagascar, places its source outside the canonical African LLVP (Figure 7d). However, recent multi-frequency P-wave tomography that includes Pdiff phases (Hosseini et al., 2019), shows slow P-wavespeeds collocated with the lower mantle extent of anomaly A. An upwelling rooted outside the conventional LLVP implies a potentially different composition from the upwelling sourced from inside the African LLVP.

The East African Rift
In Ethiopia, the Miocene-Recent East African Rift exposes subaerially the transition from continental rifting to oceanic spreading within a ∼30 Ma old continental flood basalt province (e.g., Chorowicz, 2005;Rooney, 2017). P-wave absolute arrival-times recorded in the Ethiopian rift are generally delayed by > 3 s ( Figure 2). The largest station-mean delay time (minimum 10 arrivals) is 4.7 s at 9°N, 39.1°E, on the western Ethiopian rift flank. Consequently the most anomalous wavespeeds of δV P ≤ −4% occur below the western rift flank and adjacent plateau (Figures 5, 8, and 9). This corroborates earlier studies in Ethiopia that cite observations of markedly low wavespeeds and delayed phase arrival-times as evidence that the Ethiopian mantle is amongst the slowest worldwide (e.g., Bastow et al., 2005;Bastow et al., 2008). Markedly slow shear wavespeeds (δV S ≤ −11%) have been also observed in the upper 150 km (Gallacher et al., 2016). AFRP20 also confirms that the slowest upper mantle P-wavespeeds exist below the MER and adjacent Ethiopian plateau, not below Afar (Bastow et al., 2008).
Petrologically determined mantle potential temperatures (Tp) are elevated by ∼140°C in Ethiopia, toward the low end of the temperature range of global hotspots and large igneous provinces (Rooney et al., 2012). The MER is currently extending at ∼6 mm/yr, slower than the adjacent Red Sea rift at ∼15-20 mm/yr (e.g., McClusky et al., 2010;Vigny et al., 2006) and is doing so without significant crustal thinning (e.g., Mackenzie et al., 2005). Therefore, the anomalously slow wavespeeds cannot be easily attributed to the traditional hotspot model of a purely thermal anomaly or decompression melting due to rapid plate thinning above a mantle plume (White & McKenzie, 1989). Instead, melt likely contributes considerably to our observations, both within and below the lithosphere.
Evidence for melt throughout the Ethiopian lithospheric mantle comes from studies of seismic anisotropy (e.g., Bastow et al., 2010;Kendall et al., 2005), while asthenospheric melt is also required to explain the region's depressed surface wave wavespeeds (Bastow et al., 2010;Gallacher et al., 2016). Below depths at which decompression melting of normal mantle peridotite can reasonably be invoked to explain slow wavespeeds, a hypothesis of CO 2 -assisted mantle melting can explain anomalously depressed seismic wavespeeds, beyond those readily interpreted thermally (Rooney et al., 2012). Carbonatite volcanism in East Africa, such as is found at Ol Doinyo Lengai, Tanzania (see Figure 9), does not provide direct evidence for anomalous mantle enrichment of CO 2 (Fischer et al., 2009). However, on the strength of xenolith evidence and CO 2 degassing measurements, carbonatite melts and CO 2 enriched fluids are likely to be present throughout the East African mantle (Rudnick et al., 1993;Frezzotti et al., 2010;Muirhead et al., 2020). Uneven CO 2 availability offers an appealing explanation for the apparent internal wavespeed architecture within a thermochemically heterogeneous upper mantle (≤410 km depth: Figures 5,  8, and 9) because AFRP20 is less easily reconciled with ponding of deep sourced plume material below the transition zone rising as plumelets into the upper mantle (e.g., Civiero et al., 2015;Farnetani & Samuel, 2005).
Further south in the EAR, in Kenya, where rifting is at a less mature stage than in Ethiopia (Ebinger et al., 2017), arrival-times are less delayed and consequently upper mantle wavespeeds are less depressed than below Ethiopia (Figures 2, 5, 8, and 9). However, petrologically determined Tp estimates do not differ markedly below Ethiopia and Kenya (Rooney et al., 2012). At <100 km depth, the relative paucity of melt below the EAR in Kenya can explain this observation (Rooney, 2020b). At greater depth, reduced CO 2 -driven mantle melting and/or other thermochemical effects may better account for discrepant wavespeeds between the Ethiopian and Kenyan upper mantle.

The Confluence of Multiple Upwellings Below the East African Rift
AFRP20 lacks abrupt separation between slow wavespeeds centered below the Ethiopian and Kenyan plateaus in the upper mantle (<660 km depth), yet below, evidence for two distinct whole-mantle upwellings becomes apparent (Figures 5-8). Our thorough resolution testing (specifically Figures S36 and S37) strongly indicates that distinct upper mantle slow wavespeed anomalies below the Ethiopian and Kenyan plateaus separated by ambient mantle below Turkana would be evidenced by significant reduction in negative wavespeed amplitudes, contrary to what is observed in AFRP20. Furthermore, separation of slow wavespeed anomalies below the mantle transition zone, as observed in AFRP20 is also robust. The persistently depressed wavespeeds below East Africa suggest mixing of heterogeneous material may be taking place in the upper mantle. Mixing of heterogeneous upwellings could be influenced by competing, yet largely unconstrained factors including the thermochemical nature of each upwelling, the motion of Africa prior to docking with Eurasia at ∼30 Ma, the surrounding mantle flow field and the variability in the plume impingement time below the African plate.
Primitive helium signatures are likely sourced from within LLVPs in the lower mantle (e.g., Williams et al., 2019), yet are found along the entire continental rift from Afar to southern Tanzania (Hilton et al., 2011;Halldórsson et al., 2014). This implies that material derived from the African LLVP is present throughout the East African upper mantle (Rooney, 2020b). Incomplete or uneven mixing of heterogeneous material from the deep mantle, largely dominated by that transported from the LLVP in the African Superplume, may explain mantle geochemical signatures of magmatism that vary between the Ethiopian and Kenyan rift branches and so do not easily fit within the single thermochemical upwelling model (e.g., Rogers et al., 2000). Additionally, although the Ethiopian rift is more mature and extending faster than the EAR further south (Saria et al., 2014;Ebinger et al., 2017), we cannot preclude the possibility that additional heating provided by a second plume can, at least in part account for the more anomalous upper mantle wavespeeds (Figures 5, 8, and 9) and modestly higher Tp estimates (Rooney et al., 2012) below Ethiopia/ Afar.

The Turkana Depression: Rifting or Dynamic Topography?
The cause of the two distinct uplifted topographic plateaus in Ethiopia and Kenya (typically over 1,000 m elevation), separated by the Turkana Depression (<500 m elevation) is often attributed, at least in part, to the influence of mantle plumes (e.g., Pik et al., 2008). Some global tomographic models show two distinct slow wavespeed anomalies beneath Ethiopia and Kenya in the upper mantle (e.g., Chang et al., 2015;Ritsema et al., 2011). Higher resolution, continental shear-wave models show some parts of the Turkana Depression to be underlain by slow wavespeeds (Celli et al., 2020;Emry et al., 2019). Presently, it is unclear whether these anomalies are coherent throughout the entire upper mantle depth range or represent active upwellings.
On the strength of resolution tests ( Figures S36 and S37), AFRP20 suggests slow upper mantle wavespeeds exist below the Turkana Depression meaning it remains challenging to explain the relatively subdued topography of Turkana by present-day upper mantle processes alone. Although we resolve two midto-lower mantle upwellings below the uplifted plateaus and a low amplitude, fast wavespeed anomaly at 900-1,200 km depth below the Turkana depression between them (Figures 6 and 8, anomaly C), surface dynamic topography is broadly controlled by thermal anomalies in the upper mantle (Hager et al., 1985). Alternatively, Turkana is host to a failed, NW-trending, Mesozoic rift system (e.g., Chorowicz, 2005), which may explain Turkana's low elevations. If the Turkana lithosphere had not been thinned by a combination of Mesozoic and more recent Cenozoic rifting events, high elevations may have been continuous across the Ethiopian and Kenyan plateaus. Additional constraints on shallow mantle wavespeeds from body waves will be provided by ongoing work in the Turkana depression.

Nonrift Related Magmatism: A Lower Mantle Contribution to the Cameroon Volcanic Line?
Away from the uplifted Ethiopian and Kenyan plateaus, in Madagascar, Cameroon and the Atlas Mountains, mean absolute arrival-times are typically delayed (≈1-2 s: Figure 2). Accordingly, when compared to Ethiopia, these volcanic centers are characterized by less anomalous upper mantle wavespeed anomalies in AFRP20 (typically −2% ≤ δV P ≤ −1%: Figures 5 and 9). The cause of magmatism along the CVL has long been debated because the chain lacks the linear age progression expected for a stationary plume underlying a moving plate (e.g., Fitton & Dunlop, 1985). Lavas along the CVL are typically alkaline basaltic in composition (e.g., Suh et al., 2003) which, when reviewed in light of the region's relatively felsic crust (Gallacher & Bastow, 2012), has led some studies to suggest that low melt volume hypotheses for CVL development such as shear zone reactivation or small-scale convection are most appropriate (e.g., De Plaen et al., 2014;Fourel et al., 2013;Gallacher & Bastow, 2012;Milelli et al., 2012). Our observations of only subtly anomalous absolute arrival-time residuals (Figure 2) and modestly slow wavespeeds below Cameroon compared to the markedly delayed arrivals and δV P ≈ −4% anomalies below Ethiopia, corroborate this view. Numerous regional studies have also concluded that slow wavespeeds below the CVL are confined to the upper mantle (<410 km depth; Reusch et al., 2010Reusch et al., , 2011Adams et al., 2015).
Synthetic resolution tests ( Figure S22) show vertical smearing of the CVL slow wavespeed anomaly to ∼660 km depths. Based on our specific tests, the imaged extension of slow wavespeeds below the CVL upper mantle, southwest to greater depth in AFRP20 ( Figures S22e, and S22f) is not entirely the result of smearing. Therefore, we cannot preclude the possibility that lower mantle material contributes to magmatism somewhere along the CVL. This agrees with some global tomographic models documenting whole mantle plumes (e.g., French & Romanowicz, 2015) and the continental-scale study of Emry et al. (2019), who suggest that the northern CVL may be fed by deeper material whilst the southern CVL is more consistent with a shallower causal mechanism.
A striking observation in Figures 4, 5, and 9 is that fast P-wavespeeds below the Tanzanian craton are largely restricted to shallow mantle depths and perhaps display basal topography, extending only to 135-90 km depth west-to-east. Fast shear wavespeeds are also thinner than expected for typical cratonic cores (≥200 km) and are asymmetric (Adams et al., 2012;Celli et al., 2020;Emry et al., 2019;O'Donnell et al., 2013;Weeraratne et al., 2003); thicker on the western side (∼175 km), thinner on the eastern side (∼135 km). Our fast P-wavespeeds are significantly thinner than the imaged fast shear wavespeeds that on average, provide a better match to petrologically defined lithospheric thickness estimates (∼160 km; Lee & Rudnick, 1999).
Thermal erosion by underlying warm asthenospheric material has been proposed to explain the anomalously thin Tanzanian mantle fast shear wavespeeds (e.g., Celli et al., 2020;Weeraratne et al., 2003). While thermal erosion can explain the anomalously thin Tanzanian lithosphere, it cannot account for the discrepancy between our P-wavespeed images and previous shear wavespeed models because lithospheric removal or increased lithospheric temperatures would decrease both δV P and δV S (e.g., Goes et al., 2000;Schutt & Lesher, 2006). Excess water or mantle melting at the plate base would also dramatically reduce δV S and δV P (Wagner et al., 2008). We therefore likely require a steady-state compositional explanation for the observed decrease in P-wavespeeds in the lower portion of the anomalously thinned Tanzanian lithosphere.
Tanzanian xenoliths show the mantle lithosphere is depleted, particularly above 120 km depth (Lee & Rudnick, 1999), yet V P sensitivity to lithospheric depletion is low (Schutt & Lesher, 2006). However, from 120 to 150 km depth, the initially refractory lithosphere may have undergone refertilization (Lee & Rudnick, 1999), primarily involving silica enrichment of olivine by fluids and/or melt, resulting in orthopyroxene formation (e.g., Gibson et al., 2013;Lee et al., 2011), as observed in on-craton xenolith samples (Begg et al., 2009;Stachel et al., 1998). This process may lower V P by >1% without affecting V S or density (Schutt & Lesher, 2010;Wagner et al., 2008) and hence provides an appealing mechanism to reconcile the anticorrelation between P-and S-wavespeed anomalies within the thinned Tanzanian cratonic lithosphere. Wölbern et al. (2012) also suggest the observed mid-lithospheric discontinuity beneath the northwest Tanzanian craton may result from the upper limit of fluid or melt infiltration.
Because on-craton kimberlite samples from 47 to 53 Ma capture evidence for earlier lithospheric orthopyroxene enrichment (Stachel et al., 1998), prior to the onset of flood basalt related magmatism in East Africa (∼31 Ma Baker et al., 1996;George et al., 1998;Rooney, 2020a), P-wavespeed evidence for metasomatic modification of the Tanzanian craton is best explained by a widespread tectonic process, preceding both kimberlite magmatism and younger, perhaps ongoing, rift or plume related mechanisms (Celli et al., 2020;Koornneef et al., 2009). A clear candidate is therefore Proterozoic-age, "Pan-African" subduction at ∼550 Ma (Möller et al., 1998). Plume and rift related craton modification may instead favor lithospheric thinning (e.g., Celli et al., 2020;Weeraratne et al., 2003) and marginal processes (Koornneef et al., 2009), respectively. By considering both the impact of prior metasomatic modification, lowering V P only (Figures 4, 5, and 9) and more recent thermal erosion, evidenced by lowering of both V S (Adams et al., 2012;Celli et al., 2020;Emry et al., 2019;O'Donnell et al., 2013;Weeraratne et al., 2003) and V P , we can reconcile seismological and petrological evidence pertinent to the evolution of the Tanzanian cratonic lithosphere.

Conclusions
AFRP20 is a new P-wave whole-mantle tomographic model for Africa. It combines data from all temporary broadband seismograph deployments from 1994 to 2019 with an existing global database of absolute arrival-time measurements. The western flank of the main Ethiopian rift exhibits the largest station-mean absolute arrival-times across Africa (≈4.7 s). Consequently, the upper mantle in this region and adjacent plateau hosts the most anomalous slow P-wavespeeds (δV P ≤ −4%). We resolve the African Superplume as an inclined, heterogeneous, whole-mantle slow wavespeed structure extending from the African LLVP to the upper mantle below the Kenyan plateau. However, we also image a second whole-mantle slow wavespeed anomaly extending upwards from the core-mantle boundary below the Indian Ocean, distal from the African LLVP, to the upper mantle beneath the Ethiopian plateau. Our tomographic images indicate two, disparately sourced whole-mantle plumes may influence magmatism in East Africa. Other sites of African Cenozoic volcanism such as Cameroon and Madagascar, exhibit much lower amplitude upper mantle wavespeed anomalies than below northeast Africa, but may extend beyond upper mantle depths.
Below the Tanzanian craton, fast P-wavespeeds extend to ∼90-135 km depth. This captures the previously observed asymmetry in depth but is even thinner than corresponding surface wavespeed images. While we cannot constrain whether the Tanzanian craton is presently undergoing thermal erosion (Celli et al., 2020), this provides an appealing explanation for the thinner than global average cratonic thickness of fast P-and S-wavespeeds beneath Tanzania. Our observation of even thinner P-wavespeeds compared to shear wavespeed images is explained best in light of petrological evidence for modification of the lower lithosphere via subduction-driven metasomatism during the ∼550 Ma Pan-African orogeny.