Introduction

Soon after the plate tectonics theory was established, subducting slabs were recognised as a fundamental driver of flow in the Earth’s mantle1,2,3. The sinking of oceanic lithosphere indeed induces mantle flow through viscous drag associated with slab down-dip motion4 and through a pressure change from either side of the slab due to lateral slab motion5. As developments and knowledge in seismic imaging, geochemistry and geodynamic modelling progressed, a wealth of studies followed to reveal the geometrical properties of subduction-induced mantle flow. To date, modelling efforts have concentrated on characterising the 3D pattern of the mantle flow associated with slab rollback, since this subduction mode is dominant on Earth6,7,8,9,10. Slab-rollback subduction has been shown to produce a quasi-toroidal mantle flow around the lateral slab edges, with a direction from the sub-slab domain towards the mantle wedge, while poloidal circulation is maximised at the centre of the slab11,12,13,14,15,16,17,18,19,20,21,22,23,24,25. The mantle flow in slab rollback models thus shows strong deviation from plate motion in the vicinity of slabs while it corresponds better to plate motion away from them26, in line with shear-wave splitting observations that suggest a trench sub-parallel flow near the trench27,28,29,30 and a trench sub-perpendicular flow away from it31,32. This change of flow direction near the trench reconciles with the rotating return flow around lateral slab edges observed in the models and suggested for several subduction zones28,29,31,33,34.

Mantle flow produced by slab rollback has been suggested to deflect and interact with mantle plumes21,35,36,37, produce mantle upwellings and promote associated intraplate volcanism17,18,22,38,39,40, transport geochemical signatures laterally over great distances33, deform the overriding plate20,23,41, and curve the trench4,14,42,43. Studying the control that subduction has on mantle flow is thus essential because of its significance in controlling large-scale geodynamic processes. Concerning mantle upwellings and intraplate volcanism, previous research has evidenced that a maximum upward component of slab rollback-induced mantle flow is located near the lateral slab edges17,22. Depending on the rheology of the upper mantle in fast-deforming domains17,39, the temperature of the mantle upwelling and the lithosphere thickness37, such mantle upwelling could trigger decompression melting, thereby leading to intraplate volcanism. Multiple examples of natural volcanism that may be explained with this scenario have been identified, including the Wrangell volcanics near the eastern edge of the Alaska subduction zone17,39, the High Lava Plains east of the Cascadia trench44, Mount Etna and the Iblean Plateau south of the Calabrian slab38, the Shiveluch group of volcanoes and close Aleutian adakites near the northern edge of the Kamtchatka slab45, volcanoes of the Samoan chain near the northern edge of the Tonga subduction zone37, lamproites in the eastern Betic-Rif orogen46 and Cenozoic volcanoes north and south of the Gibraltar arc-Alboran system47.

The quasi-toroidal pattern of the slab rollback-induced mantle flow as well as mantle upwellings near the lateral slab edges have now become a paradigm in subduction dynamics26. However, the pattern of mantle flow induced by slab rollover subduction (trench advance) remains unknown. Does slab rollover subduction produce a mantle flow of comparable toroidal significance? Are prominent mantle upwellings also produced laterally away from the subducting plate? Are there any other mantle upwellings elsewhere (e.g. mantle wedge)? In this paper, we focus on these questions using 4D (3D and time-resolved) buoyancy-driven subduction models, in which we have imaged and quantified the mantle flow velocity field using a tomographic Particle Image Velocimetry (tomoPIV) technique48 to produce 3D views of the flow in a volume of the upper mantle on the side of the subducting plate. We present the results of two experiments developing exclusively slab rollback and two experiments showing rollover subduction after interaction with the upper-lower mantle discontinuity (four experiments in total). The repetition of each experiment for each subduction model is to demonstrate reproducibility and to present the results using a tomoPIV technique and the results using a stereoscopic PIV (stereoPIV) technique. The stereoscopic PIV technique allows for the mapping of the three components of the mantle flow velocity field in a depth section22, thus providing coverage of a larger area, including the mantle wedge. With our rollover subduction models, we image a quasi-toroidal mantle flow directed from the mantle wedge towards the sub-slab domain, opposite to the quasi-toroidal mantle flow produced by slab rollback. We also evidence the occurrence of a large-scale mantle upwelling (with lateral dimensions that are of the same order of magnitude as the depth of the upper mantle) in the broader mantle wedge region throughout the whole experimental duration for the slab rollover model, whereas in the slab rollback model the mantle wedge upwelling lasts only for the period of slab free sinking before the slab interacts with the 660-km discontinuity. Our experimental findings provide the basis for a novel interpretation of the origin of widespread Cenozoic volcanism in Tibet.

Results and Discussion

Slab geometry and surface kinematics

We performed two experiments with contrasting subduction mode, slab geometry and surface kinematics using the tomoPIV technique, namely a rollover subduction model (model rollover_tomo, Figs. 1a and 2) and a slab rollback subduction model (model rollback_tomo, Figs. 1c and 3). Each experiment was reproduced once using the stereoPIV technique (model rollover_stereo and model rollback_stereo, Figs. 1b, d and 4). The experiments are buoyancy-driven, involve a free subducting plate sinking into an upper mantle reservoir, and exclude an overriding plate. The two subduction modes are achieved by varying the length of the subducting plate at the start of the experiment, following Xue et al.6, with a shorter plate resulting in slab rollover and a longer plate resulting in slab rollback. All models follow a multiphase subduction evolution and begin with a free slab sinking phase during which the slab sinks freely into the mantle reservoir and progressively steepens before it interacts with the 660 km discontinuity (Figs. 1, 2a, c, 3a, c, 4a, c). Thereafter, models rollover_tomo and rollover_stereo both produce a slab rollover geometry (a U shape on its side) with trench advance due to the slab rolling forward (Figs. 1a, b, 2b, d, 4b, 5a). Conversely, models rollback_tomo and rollback_stereo both show a slab rollback geometry with backward slab draping and trench retreat after a relatively short period of slab folding at the 660 km discontinuity (Fig. 1c, d; Figs. 3b, d, 4d, 6a). All models show subduction acceleration during the free slab sinking phase with VSP (subducting plate velocity, trenchward is positive) and positive VT (trench velocity, towards the subducting plate side (retreat) is positive) increasing rapidly until shortly before the slab tip touches the 660 km discontinuity (Figs. 5a, b, 6a, b). Noticeably, the rollover subduction models also produce trench retreat during this first phase. Subsequently, the rollover subduction models quickly reach an approximately constant VS (subduction velocity with VS = VSP + VT) of ~6–10 mm/min (~9.6–16.0 cm/yr when experimental values are scaled to nature) while the magnitude of VSP and negative VT (trench advance rate) both increase progressively, in the range ~10–18 mm/min (~16.0–28.8 cm/yr) for VSP and ~1–7 mm/min (~1.6–11.2 cm/yr) for |VT| (Fig. 5a, b). On the other hand, the slab rollback models take longer (~30 min scaling to 15 Myr) to reach an approximately constant VS of ~10–12.5 mm/min (~16.0–20.0 cm/yr) with the magnitude of VSP becoming negligible and VT increasing to ~8.5–12 mm/min (~13.6–19.2 cm/yr) (Fig. 6a, b). Models rollover_tomo and rollover_stereo show very good reproducibility (Figs. 1 and 5). A minor difference in all velocities is observed after the slab free sinking phase where the velocities are reduced in model rollover_tomo for ~5 min (2.5 Myr) because the angle at which the slab tip starts to interact with the 660 km discontinuity is more vertical than in model rollover_stereo, thereby resisting the slab rolling forward motion. Models rollback_tomo and rollback_stereo also show very good reproducibility (Figs. 1 and 6). A minor difference is the longer slab-660 km discontinuity interaction phase in model rollback_tomo, inducing a reduced VSP and VS for a longer time while it takes longer for VT to approach a negligible value. This longer interaction phase is due to a greater slab fold on top of the rigid tank bottom in model rollback_tomo.

Fig. 1: Subduction evolution of the four models.
figure 1

Side view pictures of the four experiments showing the different subduction evolution between (a, b) the rollover subduction models and (c, d) the slab rollback subduction models. The inverted orange triangles indicate initial (empty triangle) and current (filled triangle) trench position.

Full size image
Fig. 2: 3D Mantle flow of the rollover subduction model imaged with the tomographic PIV technique.
figure 2

Two stages of model rollover_tomo showing the 3D mantle flow velocity field in a 10 cm (800 km) wide volume next to the subducting plate. The two subduction stages are presented using two different perspective views. a, b Side perspective view normal to the side of the subducting plate. c, d Oblique perspective view. a, c Present a late stage of the slab free-sinking phase while (b, d) present a mid-stage of the ~steady-state rollover subduction phase. The colour scale represents the vertical component of the 3D mantle flow velocity field (upward is positive) while the direction of the flow is imaged using streamlines overlaid with arrow heads. The image showing the subducting plate is a side view picture of the model that is projected on the perspective view.

Full size image
Fig. 3: 3D Mantle flow of the slab rollback subduction model imaged with the tomographic PIV technique.
figure 3

Two stages of model rollback_tomo showing the 3D mantle flow velocity field in a 10 cm (800 km) wide volume next to the subducting plate. The two subduction stages are presented using two different perspective views. a, b Side perspective view normal to the side of the subducting plate. c, d Oblique perspective view. a, c Present a late stage of the slab free-sinking phase while (b, d) present a mid-stage of the ~steady-state rollback subduction phase. Please see figure legend of Fig. 2 for details.

Full size image
Fig. 4: 3D Mantle flow of the rollover subduction model and slab rollback subduction model imaged with the stereoscopic PIV technique.
figure 4

Top view of the 3 components of the 3D mantle flow velocity field showing 2 different stages of subduction for models (a, b) rollover_stereo and (c, d) rollback_stereo. The depth section of the computed velocity field is at 2.8 cm (224 km) below the model surface. The subduction stages are similar to the ones chosen in Figs. 3 and 4. The in-plane components (VX, VY) are represented with black vectors and the out-of-plane component (VZ, upward is positive) is represented with the colour scale. The subducting plate is represented in grey with the part at the surface in light grey, the slab above the depth section (laser sheet) in dark grey and the slab below the depth section represented by a dashed contour line. The trench is represented with a black line and triangle symbols.

Full size image
Fig. 5: Evolution of velocities in the rollover subduction models.
figure 5

Diagrams illustrating the temporal evolution of velocities for the two rollover subduction models. a Trench velocity (VT in green) and subducting plate velocity (VSP in blue). b Subduction velocity (VS = VSP + VT in black). c Maximum magnitude of mantle upwelling laterally away from the subducting plate (Vup_edge in red) and maximum magnitude of mantle upwelling in the broader mantle wedge domain (Vup_MW in orange). Model rollover_tomo is represented by a continuous line while model rollover_stereo is represented with a dashed line. The black triangles indicate the approximate time at which the slab touches the bottom of the tank. Note that the position of Vup_MW with respect to the trench does not change much, as the distance between them (of the order 500–700 km) remains relatively constant throughout the experiment (cf. Figures 4a, b).

Full size image
Fig. 6: Evolution of velocities in the slab rollback subduction models.
figure 6

Diagrams illustrating the temporal evolution of velocities for the two slab rollback subduction models. Please see figure legend of Fig. 5 for details, including explanations on panels a–c. Model rollback_tomo is represented by a continuous line while model rollback_stereo is represented with a dashed line. Note that the position of Vup_MW with respect to the trench changes during the experiment and that the distance between them progressively increases from ~600 km during the free sinking phase to ~2200 km in an advanced stage of the steady-state phase (Fig. 4c, d). Further note that the position of Vup_edge changes with respect to the trench from next to the sub-slab domain during the free sinking phase to next to the mantle wedge domain during the steady-state phase (cf. Fig. 4c, d).

Full size image

3D mantle flow and upwellings

During the free slab sinking phase, all models induce a 3D mantle flow with strong toroidal component going from the sub-slab domain around the lateral slab edges to the mantle wedge (Figs. 2a, c, 3a, c, 4a, c). Associated with this return mantle flow, an upwelling component is observed with maximum magnitude located to the side of, but close to, the subducting plate (next to the sub-slab domain) some 18–50 mm (144–400 km) away (trench-normal distance) from the trench position at the lateral slab edge, and another upwelling is located in the mantle wedge ~55–125 mm (~440–1000 km) from the trench. During the steady-state phase, the rollover subduction models generate a return flow in which flow in the toroidal cells circulates in the opposite direction compared to the slab rollback models. Indeed, the rollover subduction-induced mantle flow goes from the mantle wedge around the lateral slab edges to the sub-slab domain, whereas the slab rollback-induced mantle flow keeps a similar pattern as during the slab free sinking phase. All models continue to produce an upwelling component laterally away from the subducting plate, but the location of the maximum magnitude away from the trench is different depending on subduction style. The rollover subduction models produce a lateral upwelling near the trailing edge corners of the subducting plate and the upwelling close to the lateral slab edges has disappeared (Figs. 2b, d, 4b), whereas the lateral upwelling occurs closer to the trench, some 0–48 mm (0–384 km) from it, and now also occurs more prominently next to the mantle wedge region during the steady-state slab rollback phase (Figs. 3b, d, 4d). Furthermore, a prominent upwelling (which we define as an upwelling with a maximum vertical velocity VZ > 0.5 mm/min) remains in the broader mantle wedge region during rollover subduction ~55–125 mm (~440–1000 km) from the trench, whereas it disappears during steady-state slab rollback while a moderate and more localized one appears above the slab leading edge that rests on top of the 660 km discontinuity. The depth at which the maximum magnitude of the lateral mantle upwellings is observed depends on the subduction style and phase. During the slab free sinking phase, it is in the mid-upper mantle for model rollover_tomo (~20–40 mm scaling to 160–320 km, Fig. 2a, c) and it encompasses the whole upper mantle for model rollback_tomo (Fig. 3a, c). During the steady-state subduction phase, the depth is 30–45 mm (240–360 km) in model rollover_tomo (Fig. 2b, d), whereas it is close to the surface (~10–15 mm scaling to 80–120 km) in model rollback_tomo (Fig. 3b, d). Furthermore, a minor downward velocity component occurs in the mantle next to the sub-slab, lateral slab edge and mantle wedge during the rollover subduction phase (light blue colours in Fig. 2b, d).

In terms of their magnitude, the upwelling in the mantle wedge due to rollover subduction occurs at a faster rate and is broader (~890 km in trench normal direction) than the one produced by slab rollback (~560 km in trench normal direction) (Figs. 4b, d, 5c, 6c). Contrastingly, the mantle upwelling produced laterally away from the subducting plate is faster in the slab rollback models than in the rollover subduction models. Furthermore, the rate of these mantle upwellings follows an episodic evolution as the one observed in the surface kinematics. For all models, the free slab sinking phase shows an acceleration of the lateral slab edge maximum upwelling rate (Vup_edge) and mantle wedge maximum upwelling rate (Vup_MW) followed by a rapid slowdown during the slab-660 km discontinuity interaction (Figs. 5c, 6c). In the rollover subduction models, Vup_edge reaches a maximum of ~0.8–1.3 mm/min (1.28–2.08 cm/yr) and then drastically reduces to become negligible during the ~steady-state rollover subduction phase, while the broad mantle wedge reaches a Vup_MW of 1.15 mm/min (1.84 cm/yr) during the slab free sinking phase and maintains a rate of 0.60–0.95 mm/min (0.96–1.52 cm/yr) during ~steady-state rollover subduction. In the slab rollback models, a comparable Vup_edge of 0.95–1.05 mm/min (1.52–1.68 cm/yr) is observed for the lateral slab edge upwelling followed by a noticeable slowdown during the slab-660 km discontinuity interaction. During the following steady-state slab rollback evolution, Vup_edge attains a near steady-state velocity of 0.90–1.20 mm/min (1.44–1.92 cm/yr) while Vup_MW decreases to reach values of ~0.4–0.5 mm/min (0.64–0.8 cm/yr). As also observed in the surface kinematics, models with the same subduction style show a good reproducibility on the evolution of the magnitude of their mantle upwellings, despite a few minor differences pertaining to a somewhat different slab-660 km discontinuity interaction.

Surface kinematics and slab geometry of the two contrasting subduction styles

The two subduction styles simulated in this study, as obtained by varying the length of the subducting plate, are characterised by different surface kinematics and slab geometries, similar to analogue subduction models by Xue et al.6. The trench is advancing during the steady-state subduction phase of the two rollover subduction models, indicating that subduction is primarily accommodated by VSP, whereas VSP is close to zero for the two slab rollback models, indicating that slab rollback and associated trench retreat mostly accommodate subduction. The variation in partitioning of subduction as found for the two subduction styles is comparable to that described by Xue et al.6, who explain that short plates provide less resistance to trenchward subducting plate motion through a lower viscous basal drag compared to longer plates, thus favouring a high VSP. In turn, a high VSP promotes a relatively higher slab dip angle, facilitating the occurrence of a slab rollover geometry after interaction with the 660 km discontinuity. Our results agree well with previous subduction modelling studies showing the influence of VSP/VS on controlling subduction style8,13,49,50, and with statistics on natural subduction zones showing that trench advance is associated with a higher VSP compared to slab rollback subduction51.

Our study further concurs with Xue et al.6 and the above modelling studies that long-term trench advance correlates with the rollover slab geometry, while long-term trench retreat is related to slab rollback as classically observed in subduction modelling and natural settings6,8,13,52,53,54,55,56. In our study the two contrasting subduction styles were obtained varying the plate length for a specific plate to mantle viscosity ratio. Nevertheless, it should be kept in mind that other parameters have a strong influence on subduction style, including subducting plate to mantle viscosity ratio, mantle to subducting plate thickness ratio, overriding plate properties, slab width and slab age13,19,42,57,58,59.

Control of subduction style on 3D mantle flow and upwellings

Our subduction modelling study demonstrates that the return mantle flow is tied to the kinematics of subduction and slab sinking, hence it depends greatly on subduction style. A quasi-toroidal flow around the lateral slab edges is observed during slab rollback, starting in the sub-slab domain and displacing mantle material towards the mantle wedge domain. This general pattern of slab rollback-induced mantle flow agrees well with previous subduction modelling11,12,13,14,15,16,17,18,19,20,22,23,24,25,39,40,60, including the existence of a mantle upwelling that is laterally offset from the subducting plate17,18,22,37,38,40,60. In contrast, roll forward slab motion and associated trench advance generate a quasi-toroidal flow in the opposite direction, starting from the mantle wedge and going towards the sub-slab domain. The quasi-toroidal flow laterally away from the subducting plate in the rollover subduction models is generally more horizontal, with a slightly downward velocity component laterally away from the slab (Figs. 2–4). On the other hand, the upwelling in the mantle wedge is more prominent and broader in the rollover subduction models than in the slab rollback models (Fig. 3). An upwelling is observed close to the subducting plate trailing edge but it occurs as a result of the trenchward motion of this trailing edge.

Because slab sinking drives mantle flow through viscous drag due to slab down-dip motion4 as well as through pressure changes due to lateral (trench-normal) slab motion5, mantle flow velocities will be controlled by these two driving mechanisms. Nevertheless, the geometry of the subduction-induced flow, i.e. the geometry of the streamlines, will be controlled by the partitioning of subduction between VSP and VT (expressed as VSP/ VS, with VS = VSP + VT), which controls the relative rate of down-dip and lateral slab motion, as well as by the slab dip angle. Thus, the rate of mantle upwelling/downwelling next to lateral slab edges is controlled by |VT| , VSP/VS, and the slab dip angle. In the slab rollback models, |VT| is generally high during the steady-state phase (8–12 mm/min) and the slab dip angle and VSP/VS are relatively low, causing upward-inclined streamlines (Fig. 3b) and relatively high upwelling rates next to the slab edge (continuous and dashed red lines in Fig. 6c). In the slab rollover models, |VT| is lower during rollover (2–8 mm/min), while the slab has an approximately vertical geometry and VSP/VS is relatively high, causing sub-horizontal streamlines (Fig. 2b) and negligible upwelling rates during rollover (continuous and dashed red lines in Fig. 5c). Strak and Schellart40 demonstrated that the quasi-toroidal mantle flow geometry and associated upwelling magnitude are controlled by the slab dip angle during slab rollback, which is consistent with the modelling results presented here. In contrast to upwelling rates next to lateral slab edges, upwelling in the mantle wedge is much faster and occurs over a broader region during slab rollover than during steady-state slab rollback. The faster rate is because during slab rollover the upwelling rate in the mantle wedge is determined by the rate of trench and slab advance, as the advancing slab pushes the mantle forward and upward61, and the trench advance is fast. In contrast, during slab rollback the upwelling rate in the mantle wedge is determined by the subducting plate velocity and downdip slab sinking rate, which are low and thereby cause the poloidal return flow in the mantle wedge to be only weakly developed, thus resulting in low upwelling rates in the mantle wedge.

Earlier studies have also presented geodynamic models of slab rollover6,8,13,20,49,57,59,61,62, from which it can be concluded that slab rollover only results in a setting of long-term trench advance. Considering that there are only a few present-day proposed examples of slab rollover in nature61, including the India-Eurasia continental subduction/collision zone63, the Makran subduction zone61 and a slab below the Caroline microplate64, it follows that long-term trench advance in nature is rare. This is consistent with the general observation that trench retreat is more common (some 62-78%, the exact number depending on the choice of absolute reference frame) than trench advance (remaining 22–38%) for present-day subduction zones on Earth7. Several modelling studies simulating slab rollover have investigated the mantle flow patterns associated with such rollover20,61,62. In Giuseppe et al.62, one numerical subduction model is presented showing slab rollover, but the trench and slab are advancing at a slow rate (up to 3 mm/yr), and so the toroidal return flow is not very well developed in their model. In the models presented here, the trench advance and slab advance rates are much larger (of the order 5–10 cm/yr, Fig. 5a), resulting in a well-developed toroidal return flow and a fast and broad upwelling in the mantle wedge (Figs. 2b, d, 4b, 5c). In laboratory experiments of slab rollover presented in Xue et al.61, mantle flow is only imaged in cross-sections through the centre of the subduction zone, and the results are consistent with those presented here. In Schellart and Moresi20, there is only one numerical subduction model that shows a phase of trench advance and rollover, and this is a model with a two-dimensional spatial set-up. This model shows a short period (~6 Myr) of slab rollover with a maximum trench advance rate of 1.2 cm/yr. As this is a 2D model, it can obviously not show any toroidal flow, and so a detailed comparison with our analogue models is not useful.

Generic implications for rollover subduction-induced mantle flow in nature

While slab rollback-induced mantle flow has been extensively studied in previous work (see section 4.2), this study presents a type of mantle flow produced by the rollover subduction style that has not been imaged before in experiments in three dimensions. The analogue rollover subduction models presented in this study are in particular comparable with the India-Eurasia continental subduction zone. Indeed, tomography models63,65 image a rollover slab geometry (see e.g. Figure 4b in Van der Voo et al.63). In this study we observe a contrasting evolution of the mantle flow field and upwellings in the rollover subduction model, with prominent upwellings in the mantle wedge, next to the slab and laterally away from the sub-slab domain during the free-falling phase, whereas the ~steady-state rollover subduction phase drives an opposite toroidal flow with broad upwellings remaining significant in the mantle wedge but vanishing next to the slab and sub-slab. This contrasting evolution means that if decompression melting occurs because of these subduction-induced upwellings, intraplate volcanism can be predicted at different locations depending on the subduction phase.

The capacity to induce intraplate volcanism may depend on the depth at which the fast upwelling occurs, because a fast upwelling occurring close to the peridotite solidus near the lithosphere-asthenosphere boundary would facilitate decompression melting. Thus, the slab rollback subduction style should promote decompression melting laterally away from the subducting plate compared to the rollover subduction style (compare Figs. 2 and 3), while rollover subduction should promote decompression melting in a large area of the mantle wedge (Fig. 4b). However, the generation of decompression melting is also controlled, amongst several factors, by the peridotite solidus and its relation to hydration, as well as the depth of the lithosphere-asthenosphere boundary66,67, which makes decompression melting by slab rollback-induced mantle upwelling a non-ubiquitous mechanism causing intraplate volcanism. Indeed, intraplate volcanism near lateral slab edges of retreating slabs is observed at several natural locations but not close to all subduction zone terminations. We thus propose that other factors contribute to trigger decompression melting, including the interaction between the subduction-induced mantle upwelling and hydrated sediments on the lateral slab edge, which produces intraplate volcanism of Adakitic signature45,68, and/or the interaction between the mantle upwelling and a mantle plume37,47.

Implications for Tethyan-Indian subduction at the India-Eurasia collision zone and Cenozoic Tibetan volcanism

The origin of Tibetan Cenozoic volcanism can be discussed in light of our modelling results (Fig. 7). Cenozoic volcanism has been documented in several provinces throughout a broad area of Tibet, spanning the Lhasa, Qiangtang and Songpan-Ganzi terranes ~230–1000 km north of the Main Boundary Thrust (MBT, Fig. 7b) and extending across an E–W distance of ~1400 km69. The geochemistry and age of the volcanic rocks vary from south to north. (1) In the Lhasa terrane, Linzizong volcanic rocks aged 65–37 Ma are geochemically characterised by an arc signature and they spread over a distance between ~230 and ~470 km from the MBT (Fig. 7b). (2) Strikingly, widespread volcanism with a potassic signature is observed throughout the Lhasa, Qiangtang and Songpan-Ganzi terranes. Three groups of potassic volcanics with different ages can be distinguished at different locations. Potassic volcanic rocks aged 25–8 Ma are found ~280–380 km from the MBT, while another group aged 50–29 Ma is found ~440–850 km from the MBT and a third group aged 17–1 Ma is observed ~850–1000 km from the MBT. (3) Furthermore, alkali basaltic rocks aged 60–45 Ma have been documented in the Qiangtang terrane, ~570–760 km away from the MBT.

Fig. 7: Geodynamic reconstruction of subduction, mantle flow and Cenozoic volcanism in Tibet.
figure 7

Conceptual geodynamic cross-sectional reconstruction to explain Cenozoic volcanism in Tibet based on this subduction modelling study compared with the location of Tibetan Cenozoic volcanism. a Pre-collisional stage before ~50 Ma during which subduction of Tethyan oceanic lithosphere drives large-scale poloidal flow in the mantle wedge, inducing a deep-seated mantle upwelling far away (570–760 km) from the trench. This deep-seated mantle upwelling triggers sub-lithospheric mantle melting below Asian lithosphere to produce the Qiantang Alkali basaltic volcanism dated 65–45 Ma. During this stage, the Linzizong arc volcanism is very active due to fast subduction. b Collisional stage (~50–0 Ma) during which long-lived rollover subduction of Tethyan lithosphere followed by Indian continental lithosphere drives large-scale mantle upwelling in the broader mantle wedge region. This broad mantle upwelling triggers mantle melting at the lithosphere-asthenosphere boundary and involves the thickened lithospheric mantle over a large area of Tibet, producing potassic volcanism in the Lhasa, Qiangtang and Songpan-Ganzi provinces in a domain located ~440–1000 km north of the Main Boundary Thrust (MBT). During this stage, the Linzizong arc volcanism becomes progressively extinct due to continental subduction and indentation, continuously thickening the Himalayan lithosphere and due to the slowdown of subduction, while potassic volcanism takes place instead slightly to the north from 25 Ma, possibly as a result of an upwelling following Tethyan slab breakoff. Position of volcanism in (b) is based on present-day location of Cenozoic volcanics from Ding et al. (2003). LM: lithospheric mantle. The low subduction angle of the Indian slab above 150 km and deeper angle below in (b) approximate the geometry found in seismic images84 and in a geodynamic numerical subduction-collision study85.

Full size image

The cause for widespread Cenozoic potassic volcanism in Tibet remains debated. Proposed mechanisms to explain this volcanism are a mantle upwelling resulting from slab breakoff of Tethyan lithosphere70,71, mantle upwellings due to Asian lithospheric mantle delamination70, and intracontinental subduction driving melting of hydrated mantle69. The problem with slab breakoff models and intracontinental subduction models is that they predict local volcanism, and so cannot explain the widespread nature of the potassic volcanism. Another problem with one of these conceptual models is that it involves a very large amount of continental flat slab subduction (> 600 km), which is geodynamically unfeasible72. Here, we propose a mechanism derived from our geodynamic model, namely rollover subduction-induced mantle wedge upwelling, as responsible for the Tibetan Cenozoic potassic volcanism. The broad orogen-perpendicular extent of the upwelling (up to ~1000 km from the trench, Fig. 4b) makes it a good candidate to explain the widespread nature of the Cenozoic potassic volcanism (up to ~1000 km from the MBT, Fig. 7b). In addition, considering the northward motion of the India-Eurasia collisional boundary through time, and the ~north-south oriented shortening of the Eurasian lithosphere, this type of broad mantle wedge upwelling resulting from rollover of the subducted slab can also logically explain the northward progression of the potassic volcanism from the group located 440–850 km (50–29 Ma) to the group located 850–1000 km from the MBT (17–1 Ma). However, the upwelling location coincides less well with the group located in the southern Lhasa province 280–380 km from the MBT (25–8 Ma). We thus propose that the potassic volcanism in the southern Lhasa province group is due to a different mechanism that was partly coincidental with the mechanism of widespread slab rollover-induced upwelling in the mantle wedge domain. This different mechanism is possibly an upwelling following Tethyan oceanic slab breakoff from the more buoyant Indian continental slab that would have occurred before 25 Ma, which would generate more localized volcanism closer to the convergent plate boundary, as proposed previously70,71. We propose that the two above types of mantle upwelling drive partial melting at the lithosphere-asthenosphere boundary and in the overthickened Tibetan lithospheric mantle and crust, thereby generating the potassic signature in the volcanic rocks73. Thus, slab rollover is responsible for the volcanisms at 440–850 km and 850–1000 km from the MBT, slab breakoff is responsible for the volcanism at 280–380 km from the MBT, while Tethyan subduction caused formation of the Linzizong volcanic arc located at ~230–470 km from the MBT.

The origin of far backarc alkali basaltic volcanism in the Qiangtang terrane still requires an explanation. We discuss this question considering an Andean-type margin along the southern margin of Asia prior to collision73. Previous modelling showed that a very wide, Andean-type subduction zone drives large-scale poloidal flow in the mantle wedge due to the downdip slab sinking component20,74. For example, in Schellart and Moresi20, a model is presented that involves Andean-style overriding plate shortening during trench push-back, mostly downdip slab sinking and poloidal flow in the mantle wedge, with mantle upwelling occurring at 600–1000 km from the trench (see e.g., Figure 8b in their paper). We propose that such type of subduction and mantle flow applies to the pre-collisional stage before 50 Ma, inducing a deep-seated mantle upwelling far away from the trench. This deep-seated mantle upwelling triggered sub-lithospheric mantle melting due to an Asian lithosphere with a normal thickness since little deformation occurred at this distance from the subduction margin74, thereby producing the Qiangtang alkali basaltic volcanism dated 65–45 Ma. Intraplate alkali basaltic volcanism occurring far away from subduction margins has been reported elsewhere, for example in the Changbai and Fukue volcanic provinces75,76, but it has been attributed to slab sinking and its interaction with the mantle transition zone76,77. Because the depth and geometry of the sinking Tethyan slab in the mantle at that period of time are unknown, it remains difficult to constrain whether a similar interaction could have occurred to produce the Qiangtang alkali basaltic volcanism.

Fig. 8: Model set-ups.
figure 8

3-D sketchs of the laboratory set-ups with the subducting plate at an early stage of the subduction process. a Tomographic Particle Image Velocimetry (PIV) setup used to compute the 3D mantle flow in a 10 cm (800 km) wide by 100 cm (8000 km) long by 8.25 cm (660 km) deep volume next to the subducting plate. b Stereoscopic PIV setup used to image the 3 components of the mantle flow in a horizontal section located at 2.8 cm (224 km) below the model surface.

Full size image

An Andean-type margin prior to collision would have also produced overriding-plate driven trench and slab push-back and slow trench retreat, where the overriding plate trenchward motion is driven by the large-scale poloidal mantle flow74. The slab push-back would have resulted in a low slab dip angle, which is consistent with the large distance of ~230–470 km between the MBT and the Linzizong arc volcanism prior to collision. This is similar to the central Andean setting where a trench-arc distance of ~300–400 km is observed. The Linzizong arc volcanism was active in the period 65–37 Ma before becoming extinct. Since the onset of collision and continental subduction, the plate boundary motion changed to advance (northward) and the continental subduction setting was characterized by slab rollover. We propose that the Linzizong arc volcanism became progressively extinct due to the continuously thickened Himalayan lithosphere and progressive slowdown of subduction, and due to the change from normal to continental subduction, while potassic volcanism took place instead slightly to the north from 25 Ma, possibly as a result of an upwelling following Tethyan slab breakoff as previously suggested70,71.

Model limitations and reproducibility

Our models are simplifications of natural subduction zones because side plates and an overriding plate are not included, the 660 km mantle discontinuity is simulated as an impenetrable boundary, and Newtonian rheologies are used. As discussed in previous modelling studies13,49,78,79, these simplifications do not affect significantly the geometric and kinematic evolution of subduction for narrow slabs. Natural subduction zones are indeed characterised by a weak subduction zone interface and weak transform plate boundaries along the sides of the plate, which can be represented with a low-viscosity Newtonian syrup. In addition, a rigid bottom boundary represents an end-member case compared to the natural lower to upper mantle viscosity ratio. Nevertheless, this approximation remains reasonable considering comparable subduction evolution in our models and nature. Furthermore, the Newtonian rheology employed in our models can affect mantle flow velocities but, to a lesser extent, mantle flow geometries, as shown in subduction models using a non-linear upper mantle rheology that produce more localised and faster mantle upwellings and toroidal return flow cells, but a similar mantle flow pattern as the slab rollback models of our study24,39. Another simplification of the generic subduction experiments concerns the application to the Himalaya-Tibet continental subduction setting, in particular when it concerns the slab width (trench-parallel extent), which is narrow for the models (representing 800 km) but intermediate for the Himalaya-Tibet setting (~2400 km). The choice for a narrow experimental slab was mostly practical, as it allows for a better and more accurate visualisation of flow in the broader mantle wedge region, as there is less chance that slab deformation will block the view of the mantle wedge. This simplification does not have a major impact on the model outcomes when it concerns the style of mantle flow (upwelling or downwelling and landward or trenchward mantle flow) in the broader mantle wedge region, as this is simply determined by the style of subduction (rollover or rollback). It can, however, have an effect on the rate of upwelling during slab rollover, as the rate is determined by the rate of trench migration, which depends on the slab width.

Conclusions

Our geodynamic subduction experiments quantifying the 3D subduction-induced mantle flow field for advancing and retreating narrow slabs (800 km) allow us to draw the following main conclusions.

- Subduction evolution and induced mantle flow show a multiphase evolution as controlled by the slab-660 km discontinuity interaction in both the rollover and slab rollback subduction models (Figs. 5 and 6).

- The first subduction phase corresponding to the slab sinking freely into the upper mantle generates a mantle flow of similar pattern between the rollover and slab rollback subduction models, with quasi-toroidal flow driving mantle material from the sub-slab domain around the lateral slab edges towards the mantle wedge (Figs. 2a, c, 3a, c, 4a, c). Prominent mantle upwellings during this phase are produced in the mantle wedge 440–1000 km from the trench and laterally away from the subducting plate close to the lateral slab edges.

- The rollover subduction models then evolve into a rollover subduction phase with an advancing trench (and slab), which produces a reversed quasi-toroidal mantle flow, with mantle material flowing from the mantle wedge domain around the lateral slab edges to the sub-slab domain. A prominent mantle upwelling during this phase is still present in the broader mantle wedge region, while the mantle upwelling close to the lateral slab edges disappears (Figs. 2b, d, 4b).

- In contrast to the rollover subduction mode, the slab rollback models evolve into a slab rollback phase, during which the quasi-toroidal mantle flow preserves the same direction. In this phase, the mantle wedge upwelling vanishes while the lateral slab edge upwellings remain significant after a slow-down phase due to the slab-660 km discontinuity interaction, and they extend slightly closer to the mantle wedge side due to a decrease in slab dip angle (Figs. 3b, d, 4d).

- The specific kinematics of rollover subduction with its fast downdip slab motion promotes mantle upwelling in the broader mantle wedge region (Fig. 4b) while the high slab dip angle and moderate lateral slab motion produce negligible upwellings laterally away from the subducting plate (Figs. 2b, d, 4b). In contrast, slab rollback promotes mantle upwellings around the lateral slab edges due to a more inclined slab dip angle (Figs. 3b, d, 4d) and relatively fast trench retreat, while the mantle wedge upwelling is more localized and much weaker because of a much slower downdip slab motion (Fig. 4d).

- We propose a driving mechanism to explain widespread Cenozoic potassic volcanism in Tibet during the Cenozoic (Fig. 7), namely rollover subduction-induced broad mantle wedge upwelling starting from ~50 Ma, while potassic volcanism in the southern Lhasa province has occurred due to another type of upwelling, possibly resulting from Tethyan slab breakoff that occurred between ~37 and ~25 Ma. Alkali basaltic volcanism in the Qiangtang terrane is likely to have resulted from large-scale poloidal mantle flow-induced upwelling during Tethyan slab subduction before ~50 Ma.

Methods

Our experimental approach builds upon previous modelling work that considers the slab’s negative buoyancy force as the main driver for subduction, of which the energy is dissipated in the advection of the mantle and the progressive viscous bending of the slab over geological timescales4,12,40,49,54,79. In this paper, an overriding plate is not included to allow for the calculation of the mantle flow velocity field in the broader mantle wedge domain using the stereoPIV technique. We perform four fluid dynamic subduction models involving two linearly viscous (Newtonian) layers contained in a rectangular Plexiglas tank of 100 cm by 60 cm (Fig. 8). The top layer, representing the subducting oceanic lithosphere, is made of a homogeneous mixture of Korasilon silicone and fine (< 150 µm) iron powder. This modelled subducting plate has a thickness of 1.2 cm (scaling to 96 km) in all models, while its length is 30 cm (2400 km) in the rollover subduction models and 50 cm (4000 km) in the rollback subduction models. The different lengths for the subducting plate are used as a means to control the subduction mode6. The lower layer, representing the sub-lithospheric upper mantle, is made of transparent glucose syrup and continues to a depth of 8.25 cm (660 km). The width of the subducting plate is 10 cm (800 km) in all models, simulating a narrow subduction system, similar to the Scotia, Hellenic and Makran subduction zones, which have produced trench retreat (Scotia and Hellenic) and trench advance (Makran). A space of 25 cm is present between either side of the subducting plate and the sidewalls, thereby minimising sidewall effects. Boundary conditions are no slip at the bottom and sidewalls, and free surface at the top. The bottom of the tank represents an impenetrable 660 km discontinuity. The lateral edges and the trailing edge of the subducting plate are free so they represent strike-slip faults and a mid-oceanic ridge, respectively, offering negligible resistance to subducting plate motion.

The subducting plate has a density of 1488 kg m−3 and a dynamic viscosity of ~2.8 × 104 Pa s while the upper mantle has a density of 1408 kg m−3 and a dynamic viscosity of ~40 Pa s at a lab temperature of 20.5 °C. The subducting plate-upper mantle density contrast of 80 kg m−3 makes the subducting plate negatively buoyant, and reflects natural conditions of a 100 km-thick, 80-Ma-old oceanic lithosphere80, although it may also represent a slightly younger lithosphere because of the presence of surface tension effects that resist subduction in the models, whereas these effects are negligible in nature13. Subduction is initiated manually by forcing downward a sinking instability of 2 cm long with an ~30˚ angle at the tip of the subducting plate’s leading edge. The models are thereafter left to evolve without any external control. The subducting plate to upper mantle viscosity ratio is ~700, which is in the range of natural estimates13,57,81 and allows for the generation of the two different subduction modes (rollback and rollover) when varying the plate length6.

To ensure the dominance of viscous forces over inertial forces that should be negligible at a million-year timescale in nature, we conduct the experiments at very low Reynolds number (Re, estimated between 2.3 and 5.8 × 10−4) :

$${Re}={\rho }_{{UM}}{VL}/{\eta }_{{UM}}$$
(1)

where ρUM is the upper mantle density, V is the characteristic flow velocity (V = 8–20 × 10−5 m s−1), L is the characteristic length scale (L = 0.0825 m) and ηUM is the upper mantle viscosity. Such low Re guarantees dynamic similarity between the models and nature, since Re needs to be significantly smaller (more than an order of magnitude) than 1 for the models to have a similar flow regime as natural subduction systems13. The models are scaled following earlier fluid dynamic subduction experiments4,40,79, using a length scale ratio \({L}_{m}/{L}_{n}\) of 1.25 × 10-7 (1 cm in the models represents 80 km in nature) and a time scale ratio \({t}_{m}/{t}_{n}\) of 3.81 × 10-12 (1 min in the models represents 0.5 Myr).

The imaging system consists of six high-resolution CCD cameras (2752 × 2200 pixels), a laser (wavelength of 532 nm) equipped with volume and sheet illumination optics and software for continuous recording and calculation of the velocity fields (Fig. 8). We set up two hardware configurations to employ the tomoPIV and stereoPIV techniques, so that by reproducing the model for each subduction mode we could compute the 3D mantle flow velocity field in a 100-cm-long (trench-normal) by 10 cm-wide (trench-parallel) by 8.25 cm-deep volume of the mantle next to the subducting plate, as well as the 3 components of the mantle flow velocity field in a horizontal section at 2.8 cm (224 km) depth covering the whole tank. In the tomoPIV setup, four cameras positioned obliquely to a sidewall record pictures of fluorescent polymer particles (PMMA-RhB, with a size of 20–50 µm) illuminated by a laser volume and homogeneously mixed within the glucose syrup. The tomoPIV technique is an extension of the stereoPIV technique (see details in Strak and Schellart22). It uses an iterative tomographic reconstruction method to reconstruct the 3D intensity distribution of the passive particles in a volume, creating a volume composed of voxels with various intensities. 3D cross correlation is then applied to detect the particle (voxel) shift iteratively using an interrogation volume. Pictures of an embossed calibration board with marks at regular intervals are used to correct the raw images of the individual cameras, by producing a calibration map where all marks of the different cameras overlap in the central vertical plane of the volume in the same Cartesian coordinate system. The marks that are not in this central vertical plane thus appear offset on different cameras, allowing for calibration of the out-of-plane spatial component and for volumetric calculation. For accurate tomographic volume reconstruction, volume self-calibration is an additional calibration procedure that improves accuracy to better than 0.1 pixel (voxel). The tomoPIV technique has emerged some fifteen years ago48,82 and is mostly used in research that investigates turbulent flow83. However, the technique has never been used in the Solid Earth Sciences. In addition to the cameras visualising the fluorescent particles, two cameras record images of the subducting plate from the side and top to analyse slab geometry and surface kinematics (subducting plate velocity VSP with positive values indicating trenchward motion, trench velocity VT with positive values indicating trench retreat, and subduction velocity VS = VSP + VT). The stereoPIV configuration uses a total of four cameras, two oblique cameras from the top to visualise the fluorescent particles and compute the 3 components of the mantle flow in a horizontal plane, and two cameras to provide normal views of the model as in the tomoPIV setup. The recording rate is 0.2 Hz in the tomoPIV models and 0.1 Hz in the stereoPIV models.

The 3D tomoPIV-derived velocity field is computed on images with a spatial resolution of ~10.8 pixels/mm, an interrogation cube of 416 × 416 × 416 voxels with an overlap of 75%, and a time lapse of 30 s, that are adjusted to attain an optimal signal to noise ratio, resulting in a 3D−3C (three-dimensional with three components) velocity field of 37 × 11 × 14 vectors. The stereoPIV-derived velocity field uses images with a spatial resolution of ~3.3 pixels/mm, an interrogation window of 48×48 pixels with an overlap of 25%, and a time lapse of 20 s, resulting in a 2D−3C velocity field of 77 × 62 vectors. A seeding density of ~40 fluorescent particles/cm2 is used across all models (particle concentration from side view images). No filtering process nor any smoothing algorithm are used to not alter the pick velocities in domains showing maximum upwelling velocity.