The principle of seismic tomography is similar to that of medical imaging: travel times or attenuation of certain types of waves are measured on crossing paths from many sources to many receivers allowing to locate and determine the shape and amplitude of three-dimensional (3D) ‘anomalies’ with respect to a standard reference model. Global seismic models are obtained using natural seismicity as sources of elastic waves and data collected at seismographic stations all over the world. A regrettable limitation is the uneven distribution of sources (mostly along tectonic plate boundaries) and stations (mostly on continents) which introduces severe limitations in the resolution of deep 3D structure. As there are two types of elastic waves, compressional (P) and shear (S), there are also presently two main types of tomographic models. ‘P-models’ are generally constructed using bulletin data, collected by the International Seismological Centre (ISC) in the UK (e.g., [3,4,25]). They rely heavily on first arrival travel time data, which are the most consistently reported, and therefore have best resolution around subduction zones, where both the density of stations and sources is highest. ‘S-models’ are constructed using travel times and waveforms of S waves collected over the last 15 years at broadband stations of the international global seismic network, of which the French led GEOSCOPE [41,50] is a significant component. While P velocity (V_p) tomographic models are able to resolve fine details of structure in the vicinity of subduction zones (Fig. 1), the long wavelength features of lower mantle structure are best constrained in global tomographic models of S velocity (V_s), because of the more uniform global sampling they attain, in particular under the oceans, since there are more data types available. Global 3D V_s models consistently indicate an increase in the strength of lateral heterogeneity in the lowermost 500 km of the mantle, mirroring, to some extent, what is observed at lithospheric depths, i.e. at the Earth's top boundary layer (Fig. 2). There is also a distinct change in the character of lateral heterogeneity, as expressed in its spectrum, which is relatively rich in short wavelengths (i.e. ‘white’) throughout the bulk of the lower mantle, but becomes progressively dominated by low degrees as depth increases. In the lowermost 400–500 km of the mantle, degree 2 clearly stands out (Fig. 3). The striking long wavelength features that give rise to this strong ‘degree 2’ are two large regions of low velocity, centered in the Pacific and under Africa, and surrounded by a ring of high velocities (Fig. 4). The latter is thought to be related to remnant slab material that has piled up at the bottom of the mantle over geological times, as seen in geodynamic simulations (e.g., [45]). However, the connection between obliquely plunging downgoing lithospheric slabs at the top of the lower mantle, and the high velocity structure in the circum Pacific ring at the base of the mantle, is relatively clear only for some very specific profiles in P velocity models (e.g., central America, [25]) and only in some S velocity models (Fig. 5). In fact, in most regions where they do not lie flat at the upper-mantle/lower-mantle boundary, slabs can be followed down no further than about 1500 km. This has led Fukao et al. [13] to conjecture that a major ‘flushing’ event may have occurred about 45 Myr ago and caused a gap between older, remnant slabs piled up at the CMB, and the currently observable, continuous subducting lithosphere.

The two prominent regions of low velocities at the base of the mantle have been identified as regions of major thermal upwelling and are referred to as ‘superplumes’. Most tomographic S models indicate that these upwellings rise high above the CMB, but become significantly narrower at depths greater than 500 km above the CMB (Fig. 6), and some models distinctly indicate that they extend at least to the top of the lower mantle (e.g., SAW24B16, [40]). Also, the tilt observed on some SW–NE sections through the African plume [48] may be an artifact due to smearing effects associated with preferential directions of sampling in some models. Recently, in a tomographic study of upper mantle attenuation, we have shown evidence for a vertical continuation into the upper mantle of the high temperature anomalies associated with the Pacific and African superplumes [52].

There is increasing evidence that these ‘superplumes’ also involve a component of chemical heterogeneity, at least at their base. This evidence comes from the joint tomographic inversion of seismic data sensitive to V_p and V_s, which allows the inspection of depth variations in the ratio R=dlnV_s/dlnV_p, as well as the comparison of structure in V_s and in bulk sound velocity (V_Φ), which both give insight into the chemical/thermal nature of heterogeneity. The ratio R, which is of the order of R∼1.7 in the upper mantle, has been shown to increase in the lower mantle, reaching values between 3 and 4 at depths of 2000 km and larger (e.g., [36,49,51]). When confronted with the results of mineral physics experiments, such high values, well in excess of R=2.5, cannot be explained by thermal effects alone, even when including pressure and anharmonic effects [1,26]. At depths greater than 2000 km, the correlation between large scale V_s and V_p distributions becomes increasingly poorer (e.g., [49]) and the long wavelength V_Φ distribution is anti-correlated with that of V_s [23,55]. All these results are difficult to explain without invoking a compositional component to the heterogeneity in the deepest mantle.

A provocative suggestion by Ishii and Tromp [23] has led to the controversial issue, whether the large scale density structure at the base of the mantle is correlated or anticorrelated with the velocity structure. In their recent normal mode study, Ishii and Tromp [23] suggested that, in D″, the low-velocity, ‘superplume’ regions in the central Pacific and under Africa may correspond to higher than average density. While normal mode spectra from very large deep earthquakes have some sensitivity to the longest wavelengths of lateral variations in density, this sensitivity is weak, and other studies have shown that the density structure, and in particular, the nature of its correlation with velocity structure, depends strongly on initial and regularization constraints on the inversions (e.g., [29,44,51]). Whether the density structure at the base of the mantle is such as advocated by Ishii and Tromp [23] therefore remains unresolved.

Although the anticorrelation of V_s and V_Φ at the base of the mantle can be reconciled with dynamical models of thermal and chemical heterogeneity that imply buoyant material associated with the two major African and Pacific upwellings [12], the possibility of dense material residing at the base of these superplumes has intrigued geodynamicists and has motivated recent numerical (e.g., [38]) as well as laboratory (e.g., [11]) experiments. Such experiments also relate to renewed interest in the possible existence of a chemically distinct reservoir in the lower mantle [27]. In order to satisfy the global geochemical and thermal data, which seem to imply a considerable ‘missing’ volume of material enriched in radioactive elements, somewhere in the Earth, this layer would have to be much thicker than D″ in the regions of upwellings, and present an upper boundary with significant topography. No seismological evidence for such a boundary in the lower mantle has been found until now, and, as we have seen, the morphology of the superplumes tends to indicate that they are wide-based in D″, but tend to become narrower as they rise into the lower mantle (Fig. 6). This implies that the volume available within the superplumes for the geochemical reservoir may not be large enough.

A number of studies have used differential travel times of phases sensitive to the base of the mantle (core or diffracted phases), referenced to core phases in order to minimize contributions from the upper mantle, in order to map lateral variations of structure in D″. These studies have confirmed the large scale variations observed from whole mantle tomography, and in particular some regional discrepancies between S and P models (e.g., [56,57,67]).

While seismic tomography presently gives consistent information on the large scale structure in the deep mantle, uneven spatial coverage, particularly at large depths in the mantle, combined with errors in the data and imperfections in the theory impose damping conditions in tomographic inversions that prevent accurate recovery of the amplitudes and gradients of lateral heterogeneity. Insight into both the strength of heterogeneity as well as the true length scales can be obtained through the use of forward modelling, presently limited to specific well sampled ‘corridors’.

One such well sampled region is the central Pacific ocean, which lies in a favorable geometry for paths between the highly seismic Fiji–Tonga subduction zone and north America, populated by numerous broadband stations at appropriate epicentral distances for the observation of diffracted waves interacting with the Pacific superplume. Forward modelling of S, Sdiff and ScS travel times (relative to reference phases such as SKS to minimize contributions from upper mantle structure) has shown that existing tomographic models are able to successfully reproduce the observed trends of travel times (as referenced to a spherically symmetric model) with distance, but they fail to provide good estimates of the amplitudes of the variations observed (e.g., [6]). In fact, tomographic models underestimate these variations by a factor of two or three. By starting with a global tomographic model and keeping the shape of the anomalies approximately fixed, it is possible to obtain a much better fit to the observed travel times by simply modifying the amplitudes in parts of and immediately above D″ (e.g., [5]). The resulting ‘modified’ models exhibit lateral variations in S velocity at the base of the mantle that reach on the order of 8–10% over distances of several hundred kilometers, implying strong lateral gradients of structure on the borders of the Pacific superplume. Such studies also show that the strongly reduced velocities associated with the superplume continue to large heights above the CMB (possibly 1000 km, Fig. 7), and confirms that this superplume is characterized by a broad base, which becomes narrower as it rises into the lower mantle [7].

Similar results have been found under Africa, in the region bordering the African superplume, along profiles made possible by the deployment of the temporary Tanzania PASSCAL array (e.g., [47,64]). In particular, detailed waveform modeling of the African data reveals extremely sharp boundaries on both sides of the superplume [43]. Such strong lateral gradients cannot be explained by thermal effects alone and therefore imply the existence of chemical heterogeneity at the base of the mantle. A contributing factor, especially when comparing P versus S sensitive data, could also be anisotropy, as suggested by Wysession et al. [68] to explain rapid decrease of the ratio of V_p to V_s towards Alaska, in the northern Pacific.

Seismic anisotropy, indicative of strong past or present shear associated with convective motions, has been shown to be strong in the upper mantle. On the other hand, the bulk of the lower mantle is, within the limits of the present-day resolution, devoid of any indication of anisotropy (e.g., [39]). In contrast, there is now ample evidence for the presence of anisotropy in D″. Vinnik et al. [60] first suggested that the observation of elliptically polarized diffracted S waves (S_diff) was due to anisotropy in D″, along paths sampling the central Pacific Ocean. The presence of anisotropy was later confirmed by other studies of S_diff or ScS in the same region (e.g., [46,53,61,62]) as well as in the circum Pacific region [16,28,37]. In most cases, SH leads SV, implying horizontal layering or flow, however, short-scale variations have been found and interpreted as evidence of complex shear flows, possibly associated with partial melt, at the root of the Pacific superplume [53].

Perhaps the most intriguing recent discovery regarding the lowermost mantle, is that of patches of extremely reduced P velocity near the CMB, called ‘ultra low velocity zones’ (ULVZ), indicative of the existence, albeit intermittent, of a thin basal mantle layer of thickness 10–40 km with corresponding reductions in P velocity of the order of 10%. First detected in the central Pacific from the analysis of travel times of the phase SP_dKS, whose P_d leg travels a short distance along the CMB [15], they have now been found in different locations (e.g., [22]), although most of the surface of the CMB where such studies have been possible so far do not show any evidence of ULVZ's (e.g., [66]). A possible correlation of the locations of ULVZ's with the hot spot distribution has been proposed (Fig. 8). The existence of a thin basal layer is also suggested from the study of travel times of short-period diffracted P (P_diff), which indicates an average reduction in P velocity of at least 4% at the very base of the mantle, compared to reference global models (e.g., [58]). Some authors propose a related reduction in V_s of up to 30% (e.g., [42,65]), which favors a partial melting interpretation. However, other studies find no evidence for an equivalent reduction in S velocity (e.g., [10,54]).

While the absence of an ULVZ in V_s associated with the one in V_p is puzzling, various interpretations for this basal layer of dramatically lower V_p velocities have been proposed. These include partial melting (of the order of 5–30% in volume, [63,65]) and/or infiltration of iron from the core [34]. The observed effects could also originate in a thin zone of finite rigidity at the top of the outer core (e.g., [8,17]). Whatever their origin, the strong lateral variations in the vicinity of the CMB are indicative of processes that involve core–mantle interactions, and this should provide important implications on mantle dynamics.

As we have seen, the bulk of the lower mantle is, according to tomographic results, characterized by relatively short-wavelength heterogeneity, the details of which are not very well resolved. On the other hand, except in regions of subduction, there is little evidence for coherent energy originating from sub-horizontal reflectors deeper than the 670-km discontinuity: slowness/frequency stacks generally appear very quiet in front of major upper-mantle discontinuity reflected phases, all the way back to D″ depths, except for anecdotal evidence such as that for a deeper reflector under southern Africa (e.g., [32]) or sub-vertical reflectors suggestive of ancient subducted lithosphere near the Mariana arc (e.g., [24]). In contrast, studies of precursors to core phases (PKP) indicate strong scattering in D″ (e.g., [2,21]), although some questions remain as to the vertical extent of the scattering region, which may reach high above D″ [20].

In addition, numerous studies have documented the existence of a sub-horizontal reflector at depths of about 250±100 km above the CMB, more often observed for S waves than for P waves (e.g., [69]). This reflector does not seem to correspond to any plausible mineralogical phase change and some have questioned its classification as a ‘discontinuity’, arguing that it could be an artifact of the effects of strong 3D heterogeneity in D″ on wave propagation (e.g., [59]).

While the bulk of the lower mantle 3D structure remains poorly resolved, so that agreement has not yet been reached on whether subducted slabs can be traced down to the core–mantle boundary and whether major upwellings reach into the upper mantle, many seismological studies converge to indicate that strong heterogeneity is present in D″, and, in some regions, extends to at least a thousand kilometers above the CMB. There is now unequivocal evidence for strong gradients of structure in D″ that cannot be explained by thermal effects alone, as well as extensive evidence for anisotropy, the precise origin of which is however not well resolved, due to limitations in the sampling afforded by the present-day dataset. It is thus now practically established that the nature of heterogeneity in the lowermost mantle is both thermal and compositional, and the presence of ULVZ's implies some degree of partial melt and/or mixing with iron from the core. Because the strongest heterogeneity appears to be largely confined to the lowermost few hundred kilometers of the mantle, it is not clear whether a distinct lower mantle geochemical reservoir of sufficient volume is compatible with the seismic data, as much of it would have to be undetected by seismic methods. On the other hand, the African and Pacific superplumes appear to be associated not only with the upwelling of hotter than average material, but also may comprise a chemically distinct core. An emerging view, from the combination of velocity and attenuation tomography, is that the thermal anomaly persists well into the upper mantle (through the transition zone) while the chemical ‘root’ may remain confined to the lowermost mantle, which may explain why the superplumes are so much more prominently visible in D″.

Future progress in mapping lower-mantle structure relies on increased seismic resolution, and in particular, better sampling under the oceans, which will require sustained efforts to establish long-term ocean floor observatories.