In this study, we model mantle convection in two-dimensional spherical annulus geometry using the finite-volume code StagYY . Free-slip and isothermal boundary conditions are employed at the top and bottom boundaries, with the surface temperature fixed to 300 K and that of the CMB fixed to 4000 K. The numerical experiments are purely bottom heated (i.e., no internal heating). StagYY solves for the conservation equations for mass, momentum, energy, and composition on a staggered grid for a compressible fluid with an infinite Prandtl number. The computational domain is decomposed in 512 × 96 cells, in which around 1.2 million tracers, tracking composition and temperature, are advected (25 tracers per cell). We further impose vertical grid refinement near the boundary layers and near 660 km depth. The size of grid cells therefore varies between 40 and 80 km in the horizontal direction and between 10 and 35 km in the vertical direction. Resolution tests were carried out with up to double the number of grid cells in each direction and up to 50 tracers per cell (Supplement S2). In these tests, we did not observe any significant changes in the dynamics and heterogeneity styles of our models. In fact, lower-mantle heterogeneity is slightly to moderately increased with increasing resolution as in. Our estimates of heterogeneity preservation thereby remain conservative.

In our models, the assumed viscous deformation mechanism is diffusion creep. Viscosity is governed by a simplified temperature-, pressure-, and composition-dependent Arrhenius-type viscosity law (Newtonian rheology): 1η(T,P,c)=η0⋅λc⋅exp⁡Ea+P⋅VaR⋅T-EaR⋅T0, where η0 is the reference viscosity at reference temperature T0 (=1600 K) and zero pressure; Ea and Va are the activation energy and volume, respectively; T is the absolute temperature; P is the pressure; and R is the gas constant (8.314 J mol-1 K-1). The composition dependency of viscosity is considered through prefactor λc, which can be phase dependent (≠1 for primordial and post-perovskite compositions; see Table 1 and Sect. ). All physical and rheological parameters used in this study are listed in Table 1. In cases where a grid cell contains a mixture of compositions, the overall cell viscosity is calculated as a mass-weighted geometrical average of the viscosities of different compositions as in, e.g.,.

The effective rheological properties of lower-mantle materials may be well represented by our relatively low effective activation energy and volume e.g.,. For upper-mantle materials, the applied rheological parameters are significantly smaller than lab measurements . In our Newtonian rheology formulation, low activation energies can account for the effects of grain size or stress-dependent rheology e.g.,. Low values of activation volume may result in reduced mechanical decoupling of the lithosphere and asthenosphere in our models, although our choice of yield stress (Table 1) allows for the development of stiff, mobile plates in our models (see below).

In order to obtain plate-like behavior at the surface, we assume that the material deforms plastically once a critical pressure-dependent yield stress is reached as in. Plate-like behavior is determined using the same diagnostics as in , measuring plateness P (the degree to which surface deformation is localized) and mobility M (the extent to which the lithosphere is able to move). For details regarding these diagnostics, the reader is referred to Supplement S3. Plate-like behavior occurs for P close to 1 and M close to or larger than 1 .

The composition in our modeled mantle has three lithological end-member components: harzburgite, basalt, and primordial material. Each tracer carries either a primordial-material composition or a mechanical mixture of harzburgite and basalt (hz and bs). Harzburgitic and basaltic mantle materials are treated as a mixture of olivine and pyroxene–garnet systems for details, see, following a parametrization based on mineral physics data . Primordial material is not defined in terms of a specific mineral composition but solely through its material properties. It corresponds to a bridgmanite-enriched material with a (Mg + Fe) / Si ratio of ≈ 1.0 and is thought to represent cumulates resulting from the fractional crystallization of a magma ocean in the early Earth . Phase transition parameters and physical properties for each mineral system and for primordial material are given in Table 2. The density profiles of the mantle materials are shown and discussed in Supplement S1.2. Our reference primordial material roughly corresponds to Mg0.88Fe0.12SiO2 (Mg number of 0.88) or any other material with the same density profile (Supplement S1.2). The modeled mantles are initialized with a bottom primordial layer below a ≈ pyrolitic layer (see Sect.  below). The initial chemical density difference between such a primordial and pyrolitic layer stabilizes chemical stratification, while the thermal density difference destabilizes chemical stratification. The competition between the two is expressed as the non-dimensional buoyancy ratio B : 2B=ΔρCΔρT=ΔρCραΔT, where ΔρT and ΔρC are the appropriate thermal and compositional density contrasts, α is the thermal expansivity, ρ is the lower-layer density, and ΔT is the super-adiabatic temperature contrast between surface and CMB. B is calculated for relevant lower-mantle depths, taking depth-dependent parameters ρ (density of primordial material), ΔρC, and α at 1500 km depth. Accordingly, our reference model with a Mg number of 0.88 corresponds to a buoyancy number of 0.28.

Changes in composition, carried on tracers, arise from melt-induced differentiation. Tracers in the basalt–harzburgite space undergo partial melting as a function of composition, pressure, and temperature, which generates basaltic oceanic crust and a complementary depleted residue (for details, see ). Tracers of primordial composition undergo “melting” once they cross the relevant solidus of pyroxenite melting . This solidus is used as pyroxenes are the low-pressure polymorphs of bridgmanite. During such a “melting” episode, the primordial tracer is converted into a tracer with 40 % basalt and 60 % harzburgite, and its primordial flag is removed. Any melting and related degassing would likely reduce, or even destroy, the ancient isotopic fingerprint of the previously primordial material . The 40:60 basalt / harzburgite ratio is used as it corresponds to a (Mg + Fe) / Si ratio of ≈ 1.0, such as that in bridgmanite. Due to this tracer conversion, the final composition of the non-primordial mantle (from here on termed “ambient mantle”) fbs,ambfinal evolves over time as primordial material is processed (melts) in the upper mantle: 3fbs,ambfinal=fbs,ambini⋅Vambini+cbs,prim⋅Vprim,moltenfinal/Vambini+Vprim,moltenfinal, where fbs,ambini is the initial bulk composition of the ambient mantle with volume Vambini, cbs,prim is the conversion factor of primordial material into bs–hz space (=0.4) and Vprim,moltenfinal is the total volume of primordial material that has melted during model evolution (=1-χprimpres⋅Vprimini with a primordial preservation factor of χprimpres=VprimfinalVprimini).

The initial mantle temperatures are calculated from an adiabat with a potential temperature of 1900 K together with the top and bottom boundary layers and small random perturbations. The resulting temperature profile is ≈ 300 K warmer than that Earth's present-day geotherm. Such high initial mantle temperatures are applied to roughly mimic the thermal evolution of the mantle, with high vigor of mantle convection and near-surface melting that likely occurred in the early Earth e.g.,.

In terms of composition, our models are initialized as simplified two-layered (and for selected cases, three-layered) profiles, motivated by a fractional-crystallization sequence of the magma ocean . Similar to previous work , we impose a bridgmanitic “primordial” material layer in the lower mantle with layer thickness Dprim (1568–2230 km). By including randomly distributed 5 % pyrolitic “noise” in this primordial layer, the layer does not represent a pristine fractional-crystallization end-member cumulate. The upper mantle is initially a mechanical mixture of 15 % basalt and 85 % harzburgite, i.e., close to pyrolitic composition (Supplement S1.1). For some models, a third layer is added just above the CMB (150, 200, or 250 km thick) for which we impose the same physical properties as for basalt. This intrinsically dense layer is thought to represent an iron-enriched material, e.g., originating from the last cumulates from a crystallizing basal magma ocean e.g., or from ancient crust that settled at the CMB .

In this study, we systematically investigate the styles of chemical heterogeneity in the mantle as a function of the physical properties of primordial material. The physical properties explored are the intrinsic density (FeO enrichment in (Mg, Fe)SiO3) and viscosity contrast (silica enrichment) relative to pyrolite as in. The density profile of primordial material is shifted throughout the mantle, which intrinsically changes the initial buoyancy ratio B between the models (varying between 0.07 and ≈ 0.57, as per Eq. ). The approximate Mg number of the primordial material thereby ranges from 0.9 to 0.86 (Supplement S1.2). We further impose a viscosity contrast λprim (λc in Eq. ) between primordial and pyrolitic mantle material in the lower mantle that is motivated by the high intrinsic strength of bridgmanite relative to ferropericlase .

Additionally, we varied the initial conditions of chemical layering in the models (Supplement S1.1). The initial layer thickness of primordial material Dprim was changed such that the final composition of the ambient (i.e., convecting) mantle is similar for all cases, with a target value of a 25 % basalt + 75 % harzburgite mechanical mixture (fbs,ambfinal=0.25). This value is appropriate for the Earth, particularly if considering that the ambient lower mantle may be moderately enriched in ROC compared to the upper mantle in the present day e.g.,. As fbs,ambfinal for any given model depends on the extent of primordial-material processing over time, which creates basaltic materials through near-surface melting (Eq. ), Dprim is set to a different value for a given combination of parameters B and λprim. The applied Dprim ranges from 1569 to 1844 km (i.e., 60 vol %–75 vol % of the lower mantle) between the models. These choices for the initial conditions of our models lead to final ambient-mantle bulk compositions of fbs,ambfinal=0.24–0.26, i.e., very close to our target value for all cases. In order to calculate these appropriate values for Dprim in the first place, we ran an additional suite of test models with a constant initial volume of primordial material. From the predictions of these test models and Eq. (), the appropriate values for Dprim for each case of the main model suite were computed. For more details on this approach and the results of the test suite, see Supplement S4.

Finally, we compare model outcomes of the selected cases which include a thin, dense FeO-rich layer just above the CMB (three-layered models) with those of corresponding models in the main model suite (two-layered model).

Our results reveal different styles of long-term convection and mixing in the mantle. We label these regimes in line with previous work on variable styles of primordial heterogeneity preservation as regimes I–III. For all main models, primordial and ROC heterogeneity can coexist in a fully convecting mantle (partial heterogeneity preservation, regime III), with several distinct preservation styles (Fig. 2). A subset of the test experiments predicted various styles of stable chemical layering (regime II), which are summarized in Sect.  and discussed in more detail in Supplement S4. In contrast to previous work , we do not observe any models with pervasive convective mixing of ROC and primordial mantle materials (regime I) in this study.

The regime diagram shown in Fig. 3 pinpoints the parameter space where the various heterogeneity styles discussed above are identified. The viscous blobs heterogeneity style (III.B) occurs for the majority of the parameter space (≈ λprim>20 and variable B), while the total volume of the preserved blobs strongly depends on the intrinsic viscosity contrast λprim. High λprim results in a delayed overturn, more localized deformation, and slow entrainment of primordial material by the convecting mantle (Figs. 3, 2a). In contrast, low λprim causes an earlier overturn, increases convective vigor and mixing, and prevents large primordial domains from being sustained in the convecting mantle: a marble cake mantle (III.M) occurs for low B, while primordial material is diffusely distributed throughout the mantle (III.D) for moderate to large B. In a narrow parameter space (high B and λprim of approximately >0.45 and >300, respectively), metastable, primordial “piles” settle due to their intrinsic negative buoyancy (regime III.P).

Finally, we analyzed the potential relationship between dense piles, viscous blobs, and mantle dynamics for a model that displays both piles and blobs (model M300dD, regime III.B, illustrated in Fig. 2a). Figure 8 shows 2D histograms relating selected model output quantities with one another, including pile characteristics (pile height, formation age of pile material) and mid-mantle characteristics (primordial-material fraction, radial velocity, and temperature anomaly over a 1000–2000 km depth range). Several relations between these parameters can be inferred: first, the presence of primordial material in the mid-mantle is related with little radial dynamics (near-zero vz, Fig. 8a). Similarly, most piles are stable at near-zero radial velocity, although thick (>100 km) piles are also stabilized in regions of upwellings. Second, the absence of primordial domains (fprim=0, bottom rows in Fig. 8e, f) and piles (Dpile=0, bottom rows in Fig. 8e, f) is associated with columns of cool downwellings (vz<0 and dT<0). Third, the presence of a large amount of primordial material in the mid-mantle is associated with positive temperature anomalies (Fig. 8b), and thus large primordial blobs (high fprim) are typically slightly warmer than average (also see Fig. 2a, b). There is also a slight positive correlation between pile height and mid-mantle temperature anomaly: large piles are often related to overlying warm primordial blobs (see below). Although the histograms remain scattered, there is a slight positive trend in formation age of pile material versus mid-mantle primordial fraction and formation age versus pile height (Fig. 8c, g), indicating that piles containing older material are more often overlain by primordial domains than younger piles and that old ROC tends to be preserved near to top of the piles (see above). Finally, Fig. 8d and h indicate a spatial relationship (in radial direction) between the presence of primordial material in the mid-mantle and the height of the underlying ROC material. Even though this relationship is not strict, it indicates that the highest ROC piles tend to be centered just below the largest primordial blobs. This configuration away from upwellings and downwellings (Fig. 8b, f) appears to promote the coupled preservation of both these thermochemical domains.

Many of the identified dynamic styles of primordial-material preservation in our prior study are here established in the presence of plate-like behavior and recycling of oceanic crust into the deep mantle. This indicates that primordial-material preservation robustly occurs for significant rheological contrasts, regardless of tectonic behavior or the presence of dense, recycled materials in the lower mantle. This study even shows the robustness of primordial-material preservation with respect to initial layering set-up. Despite no detection of models in which both primordial and recycled materials are efficiently mixed throughout the mantle (regime I in ), we expect such a behavior to occur for small rheological contrasts of primordial material in combination with lower-density anomalies of ROC than modeled here . Indeed, different ROC properties may affect the segregation, accumulation and survival of ROC in the deep mantle e.g.,.

All of our numerical experiments are run in 2D geometry. Admittedly, the planform of mantle flow is different for 3D than for 2D geometry . An additional toroidal component of mantle flow in 3D can enhance mixing relative to mantle flow in 2D with its rather simple geometry of convection. In terms of the preservation of intrinsically strong heterogeneity, however, we expect competing effects to occur: on one hand, coherent blobs may be difficult to preserve in the presence of toroidal and poloidal flow components ; on the other hand, mantle flow may be more efficiently guided around viscous blobs in 3D than in 2D geometry . In addition, sheet-like upwellings in 2D can enhance mixing. Indeed, several studies showed that for high Rayleigh number convection models, the different geometry of boundary layer instabilities between 2D and 3D does not significantly modify the timescale or the regime of mixing e.g.,, although it is uncertain whether this would be the case in the presence of composition-dependent rheology as modeled here.

The relevance of each regime for terrestrial planet evolution depends on the actual initial compositional profile of the mantle at the onset of long-term solid-state convection. The initial conditions of our models are mainly motivated by an enrichment of a thick, basal layer in (Mg, Fe)SiO3 bridgmanite (and hence in silica with respect to the overlying mantle). This may occur due to fractionation during magma ocean or basal magma ocean (BMO) crystallization e.g.,. Indeed, bridgmanite is the liquidus phase in a magma ocean over a wide compositional and pressure range e.g.,. Alternatively, incomplete core–mantle equilibration during the last stages of planetary accretion or disproportionation of FeO during core formation may have promoted the formation of (bridgmanite-enriched) heterogeneity in the lowermost mantle. Moreover, chemical transfer between the outer core and the (basal) magma ocean in a hot early Earth may have caused delivery of SiO2 to the BMO and extraction of iron from the BMO into the core, stabilizing bridgmanite as the primary liquidus phase during prolonged periods of BMO crystallization e.g.,. Indeed, various studies have proposed SiO2 exsolution from the core during secular cooling of the Earth or Earth-like planets e.g.,.

For Earth, geodynamic styles that predict the coupled preservation of primordial material and ROC, such as regimes III.D (diffuse primordial domains), III.M (marble cake mantle), and III.B (viscous blobs), can potentially explain a wide range of geophysical, geochemical, and geological observations. First, these geodynamic styles are consistent with whole-mantle convection operating on Earth, which is motivated by seismic tomographic images of recently (<200 Myr) subducted lithosphere in the lowermost mantle e.g., and deep-rooted plumes that rise through the entire mantle . Second, they all include the existence of thermochemical piles in the lowermost mantle, which are robustly featured in seismic tomographic studies as two antipodal, anomalously slow domains (LLSVPs) just above the CMB. e.g.,. Third, coupled preservation of primordial and recycled materials in the models agrees with the geochemical record of igneous rocks carrying the footprint of both (old) recycled crust e.g., and primordial mantle sources, such as 182W / 184W and 142Nd / 144Nd isotopic anomalies and/or high 3He / 4He ratios e.g.,. The specific isotopic fingerprints of the different mantle reservoirs predicted by our models (e.g., primordial and recycled) and the subsequent sampling thereof by mantle plumes remain key topics for future research. Finally, the plate-like behavior displayed by the numerical models agree with various geological indicators of plate tectonics operating on Earth for at least several billions of years before present day and references therein.

Out of the likely geodynamic styles (diffuse domains, marble cake mantle, and viscous blobs), the primordial blobs style (III.B) is most commonly predicted (Fig. 3). Model predictions for this sub-regime are also consistent with the stagnation of some slabs in a depth range that is similar to primordial-domain roofs, while other slabs readily sink into the deep mantle , as seen in Figs. 2, 7, and videos in the Supplement . Furthermore, the presence of viscous domains in the mid-mantle may explain the observation of sharp seismic velocity contrasts in the uppermost lower mantle (850–1100 km depth), usually occurring away from major upwellings and downwellings . Moreover, such bridgmanite-rich regions would suppress the average effects of the pressure-induced spin transition of iron in ferropericlase on seismic properties . Indeed, a recent study noted the absence of the expected signal of this spin transition for pyrolitic compositions at mid-mantle depths in global seismic profiles , corroborating the presence of bridgmanite-rich regions in a depth range as predicted here. Finally, in terms of seismic tomography, the lack of clear evidence for primordial domains in the mid-mantle may be related to a trade-off between the seismic effects of their compositional (intrinsically fast bridgmanite, ) and thermal (slightly warmer than the ambient mantle) anomalies.

Although various alternative hypotheses have been put forward to explain each of the observations above, a hybrid style of present-day Earth mantle heterogeneity as predicted here can provide a unified explanation for these various geochemical and geophysical observations. This hybrid style is illustrated in Fig. 9 and includes the coexistence of viscous primordial material and dense ROC material in a convecting mantle, with several primordial streaks and/or blobs forming a larger composite bridgmanitic structure (100 s–1000 km) around which convective flows are organized. Except for very large λprim, this prediction is somewhat different from that in , who predict large coherent blobs. ROC material is present as marble-cake streaks throughout the mantle with local enhancement in the mantle transition zone (MTZ) and at the CMB (i.e., forming thermochemical piles) as in. Piles in this regime are mostly of recycled origin but may include ancient FeO-enriched material that is preserved near their top if a FeO-rich basal layer ever existed in the early Earth. In this scenario for mantle composition and structure, the upper mantle is on average pyrolitic, while the lower mantle is significantly more enriched in silica, in line with previous studies e.g.,.

Thermochemical piles formed in our models comprise of ROC with varying formation ages. For models including an ancient FeO-rich basal layer, piles consist of both recycled and ancient materials (e.g., a “basal melange”, ). Towards the roof of the piles, the formation age of ROC increases, while the overall ROC content decreases (Fig. 7). A radial distinction into regions of a high- and low-ROC content within piles was also found by . For a high buoyancy ratio of basaltic material, the thermochemical piles consist of a high-density basal layer covering nearly the entire CMB, overlain by high-topography piles with a much lower fraction of ROC , as is the case for the thermochemical piles in our models. Moreover, in our experiments, ancient FeO-rich material survives for several billion years at the roof of some piles (Fig. 7). Compositional layering of ancient versus recycled domains within LLSVPs has been previously proposed , albeit with an inverted maturity gradient as found here.

Chemical layering strongly depends on the density profile and rheological properties of the material that makes up the thermochemical piles, which remain poorly constrained. We modeled the ancient FeO-rich basal layer to have the same physical properties as ROC. Final preservation of an ancient basal layer would be advanced, and the maturity gradient potentially inverted, if the layer were enriched in FeO or SiO2 relative to ROC. Such an enrichment may or may not occur depending on the formation scenario of the ancient basal layer, e.g., the style of magma ocean crystallization e.g., or core–mantle interaction e.g.,. Moreover, ancient crustal rocks may have different chemistry and hence density at lower mantle depths than young basalts e.g.,, which may also affect the layering sequence as seen in our models.

The slight positive correlation between pile height and overlying primordial material (Fig. 8) differs from the previously conceptional model proposed in , in which the authors suggested that dense mantle material ideally piles up in zones of convergence atop the CMB, free from overlying viscous blobs (BEAMS). As we find piles to comprise of older material when overlain by primordial blobs (Fig. 8d, h), we infer that primordial domains can shield piles from any incoming subducted material and aid their longevity. The presence of primordial domains in the mid-mantle may therefore constrain the shapes and spatial fixity of LLSVP piles .

Our results further contribute to the ongoing debate about whether thermochemical piles are intrinsically stable features in the deep mantle, which spatially determine mantle convective patterns or are instead pushed around by subduction zones . In our models, piles and blobs both mostly reside well away from lower-mantle downwellings, as they are pushed away from them (videos in the Supplement, ). Indeed, as noted in previous studies e.g.,, we observe that downgoing slabs are mainly controlling the spatial distribution of piles (as well as viscous blobs) and not the other way around.

An interesting prediction of our models involves the breakdown of ancient compositional layering (“overturn”), marking the beginning of whole-mantle convection (see Fig. 6, Supplement S3). The timing of the overturn varies between 0.5 and 2.5 Gyr of model evolution, mainly depending on the intrinsic strength of the primordial layer. This overturn triggers a shift in surface-tectonic style, as plate-tectonic behavior (P and M) robustly occurs in the models following the event. For Earth, a major change in geodynamic style is commonly proposed to mark the onset of “modern-style” plate tectonics e.g.,and references therein, shifting from either an (episodic) stagnant lid e.g., or a sluggish lid with mostly vertical tectonics e.g., to a mobile-lid style with horizontal plate-like tectonics. Estimates for the timing of this event vary between studies (e.g., 2–3.2 Ga, , and references therein). According to our model predictions, the breakdown of ancient compositional layering may be related to such a tectonic shift in early Earth. The related stirring of the mantle during the overturn may have drastically changed upper-mantle composition e.g.,, which is likely followed by the oxygenation of the atmosphere, with implications for the evolution of life .

Another implication of this geodynamic shift is that deep-rooted mantle plumes can reach the upper mantle for the first time. Accordingly, the geochemical signatures of mafic rocks of this age should carry stronger lower-mantle signatures compared to older mafic rocks. Many geochemical studies indicate a rapid, widespread change in isotopic signatures in basalts around ≈ 3 Ga e.g.,and references therein. For example, μ182W anomalies in basalts show a steady trend of mostly positive values for rocks of before ≈ 3 Ga but display negative values in the modern mantle. This change has previously been attributed to inner-core segregation or the onset of deep slab subduction and recycling of crustal material and sediments e.g.,. Here, we propose that it may (additionally) be related to the onset of primordial-material entrainment by mantle plumes via whole-mantle convection. The primordial material could have obtained a negative μ182W signal through SiO2 exsolution from the core, which has a strongly negative μ182W signature e.g.,, to the (basal) magma ocean during rapid initial cooling of the planet e.g.,.

It must be noted that the modeled thermal evolution of the mantle and related early convective vigor somewhat depends on our parameter choices. Although our models do not take into account heating from radioactive decay or core cooling , the modeled mantles cool over time with internal potential temperatures from 1900 to ≈1600–1650 K, consistent with petrological studies on Earth's mantle temperatures over time . A prior study found that early model dynamics differ as a different initial mantle temperature or core temperature is applied, although the ultimate thermal and chemical structure after 4.5 Gyr does not depend much on these parameters . In particular, the scale of early mantle flow and the geometry of boundary layer instabilities are different when evolving core temperatures and internal heating sources are incorporated . With our constant core temperature of 4000 K, early Earth's CMB heat flow is likely underestimated and present-day heat flux slightly overestimated. The presence of internal heating from the decay of radioactive nuclides (or heat-producing elements, HPEs) increases convective vigor and decreases the length scales of mantle flow. On the other hand, it suppresses active hot instabilities and mantle plumes, which instead become diffuse and passive return flows e.g.,. Keep in mind that erosion of viscous blobs and the entrainment of primordial material is primarily accommodated by plumes. Nevertheless, considering uniform internal heating is likely to impede preservation, as blobs should heat up and rise into the upper mantle, where they would be efficiently processed. However, HPEs are unlikely to be uniformly distributed. In fact, they are unlikely to be incorporated into bridgmanitic material for a wide range of formation scenarios. For example, bridgmanitic magma ocean cumulates will not incorporate any significant levels of highly incompatible elements, including HPEs . Instead, if most HPEs are partitioned into the dense, final MO cumulates (i.e., such as the ancient FeO-rich basal layer discussed in Sect. 3.3.2), they would be “trapped” in this insulating layer.

As in any geodynamic study, the numerical models remain a simplified approximation of nature. For example, our model resolution only allows us to simulate relatively large heterogeneities (on scales of several kilometers or longer). In addition, we do not include any felsic material in our models, which may contribute to lower mantle heterogeneity by recycling of continental material . Furthermore, future investigations are needed to clarify how the heterogeneity styles operate in 3D geometry or in the presence of internal heating sources and core evolution.

The work presented here is primarily a numerical study that establishes various feasible regimes of mantle dynamics and heterogeneity mixing and preservation as a function of initial conditions and primordial-material properties. In the future, a thorough integration with inter-disciplinary observations on mantle heterogeneity is needed to further establish the applicability of some of our detected regimes for the Earth's (lower) mantle. For example, model results should be quantitatively compared with seismic constraints: piles are predicted to be anomalously hot and dense, which would leave a clear seismic signature e.g.,. On the other hand, bridgmanitic blobs in our models are usually only slightly warmer than typical ambient mantle temperatures (Fig. 8). Therefore, the related seismic anomalies may be much less evident in tomography, particularly when considering that bridgmanite is slightly faster than ferropericlase . Indeed, seismic tomographic images do not show obvious large-scale seismic anomalies in the mid-mantle, and future studies should aim at establishing the presence or absence of these domains in the present-day Earth, e.g., focusing on seismic anisotropy or out-of-plane reflections.