Introduction

The Hawaii-Emperor volcanic chain lies in the middle of the North Pacific Ocean, with age-progressive seamounts ranging from ~ 85 Ma on the northern end to 0 Ma at Hawaii (Fig. 1). Since Wilson1 first suggested that the Hawaii chain could be explained by Pacific Plate motion over a fixed spot of melting, the Hawaii-Emperor volcanic chain has become a canonical example of the hotspot hypothesis. The chain has a prominent 60° direction change at the 47 Ma Hawaii-Emperor Bend (HEB) and a subtle change at ~ 80 Ma (Fig. 1). There have been numerous models trying to explain the formation of this hotspot chain, especially the HEB, including a southward drifting Hawaii plume2,3,4, a direction change of Pacific Plate movement from northwest to west at ~ 47 Ma5, and the Hawaii plume attracted by a mid-ocean ridge6.

Fig. 1: Hawaii-Emperor volcanic chain.
figure 1

The bathymetric chart around the Hawaii-Emperor chain. The yellow points denote the seamounts along the chain, labeled by the seamounts’ names and ages (in parentheses) behind.

Full size image

Paleolatitude measurements3,7,8 reveal that the underlying hotspot (source of melt) drifted southward from ~ 80 Ma to ~ 47 Ma during the formation of the Emperor chain. The 78 Ma Detroit Seamount displays the highest paleolatitude of ~ 35°N, while all younger seamounts demonstrate a southward migration trend until 47 Ma with a nearly constant moving speed of about 0.4°/Myr6. Based on the drift rates and direction from geodynamic models, Torsvik et al. 5 showed that the directional change at HEB (~ 47 Ma) cannot be solely due to southward plume motion, which, in the absence of a directional change of the Pacific Plate motion at 47 Ma, requires an unrealistically large migration speed of the Hawaiian plume. Accordingly, they suggest a plate motion history highlighting the 60° Pacific Plate deflection at ~ 47 Ma.

In addition, the southward moving plume model is challenged by the observed paleolatitude of the 86-Ma Meiji Seamount7 (19°N), whose implied latitude location is near that of present-day Hawaii. This data, if reliable, means that the hotspot should have gone rapidly northward at ~ 80 Ma and then moved back to the south. Instead of neglecting this data point as many early studies did, Sun et al. 6 proposed a plume-ridge interaction model where the plume head was temporally captured by the northward-moving Izanagi-Pacific mid-ocean ridge. This model is further supported by trace element analysis9, where basalts from the oldest seamounts, such as Meiji and Detroit, are more MORB(mid-ocean ridge basalt)-like, indicating melting at a spreading center, while basalts from the younger seamounts are more OIB(ocean island basalt)-like, consistent with a plume origin. However, this plume-ridge interaction model is qualitative, with multiple aspects of observation untested, such as the effect of the realistic plate motion history, the amount of plume deflection by the ridge, as well as the resulting shape and age distribution of the simulated Hawaii-Emperor chain.

A recent numerical simulation aiming at quantifying the effect of realistic subduction history on the evolution of the Hawaiian plume10 has been conducted. This study, by showing an LLSVP (large low shear velocity province)-rooted southward migrating plume source and projecting the plume conduit at 350 km directly onto the surface, showed that a systematically older-than-observed hotspot track with all seamounts older than ~ 70 Ma on the Emperor chain should have already been subducted by now. This is inconsistent with the observed age distribution along the hotspot chain, where seamounts as old as 85 Ma are still yet to subduct. Therefore, the mechanism inferred from this numerical study is not satisfactory, and other mechanisms must have been in operation to reproduce the correct seamount age distribution.

Here, we design 3-D data assimilation models to further investigate shallow dynamic processes, especially at asthenospheric depths. These dynamic processes are pertinent to the interaction of the Hawaiian plume and the overlying lithosphere, which may affect melt generation and evolution. Furthermore, by considering both different plate reconstructions11,12,13 and plume migration histories, we attempt to better understand the formation of the Hawaii-Emperor chain through reproducing multiple key geological records, including the shape of the hotspot chain, the age distribution of seamounts, as well as the observed erratic paleolatitude variations.

Results and discussion

Effects of plume-lithosphere interaction

Our numerical models are implemented using the code CitcomS14,15,16 based on data assimilation17,18 that simulates both subduction and plume-lithosphere interaction. The model domain is large enough to capture the subduction processes within the western and northern Pacific Ocean since the Early Cretaceous, and to locate the Hawaiian plume far from the model boundaries. The geographic coordinates of all model inputs are rotated to low latitudes to avoid the polar problem of meshing in a regional model (Supplementary Fig. 1). Besides solving thermochemical convection, we further consider melting and melt advection19,20, so that the model could capture more shallow dynamic effects than recent studies of its kind. The surface hotspot track is defined through two different approaches, with track #A recording the vertical projection of the plume source (at ~ 600 km depth) onto the moving surface, and track #B recording the temporally changing location of simulated melt eruption, where # denotes the model number. More details of the model setup can be found in Methods.

We first test the fixed plume hypothesis (Model 1) based on the plate reconstruction of Müller et al. 11. The plume source is implemented at its present location, which is seismically imaged at ~ 600 km depth slightly northwest of Hawaii island at 157°W, 20°N21,22 (Supplementary Table 1). We adopt a 1° radius of the plume source as imaged21. For most models, we assume a pure thermal (unless otherwise noted) plume whose excess temperature at the asthenospheric depth is 200 °C, consistent with recent estimates22,23. In this case, the location difference between track 1 A (1 denoting model number) and track 1B (Fig. 2) shows the effect of shallow processes due to plume-lithosphere interaction and melt advection. Among them, track 1B provides a closer fit to observation than track 1 A, both in their spatial proximity to the observed hotspot chain and the predicted seamount ages (Fig. 2). Relative to the recent modeling study10 where the > 70 Ma portion of the hotspot track has been subducted by now, our predicted hotspot tracks (1A & 1B) locate all seamounts younger than 90 Ma within the present Pacific Plate, as observed.

Fig. 2: Simulated hotspot tracks using the reconstruction model of Müller et al. 11.
figure 2

The observed Hawaii-Emperor chain is shown in black. The green shades are erupted melt sites predicted by Model 2B. The red dashed lines are plate boundaries from Müller et al. 11. The yellow squares are observed seamounts. Track A shows the modeled trace of the vertical projection of the plume source onto the surface. Track B shows the modeled volcanos based on the erupted melt in our simulation. 1,2 are Model numbers. Tracks A and B are shown as thin and thick lines, respectively. The solid and dashed blue lines represent predicted surface hotspot tracks from Hassan et al. 10. According to this model, white circles along these hotspot tracks mark 10-Myr intervals, where volcanos older than 70 Ma (red dots) should have all subducted by now.

Full size image

As the tracks in Model 1 still deviate from the observed hotspot chain, we further consider a moving plume source at ~ 600 km depth to introduce the effect of plume migration in Model 2. The key reason track 2B differs from 2 A can also be attributed to the effect of plume-lithosphere interaction and lateral melt advection. More specifically, their difference arises from the capture of the rising plume by the northward-moving Izanagi-Pacific ridge during the early part of the Emperor chain. When this ridge is far south of the Hawaiian plume, the plume conduit is clearly bent northward toward shallow depth due to the drag of the fast-moving Izanagi Plate (Fig. 3a). When the ridge approaches the plume from the south, the plume conduit becomes nearly vertical (Fig. 3b). This indicates that the northward drag of the Izanagi plate is compensated by the southward attraction from the ridge. During this period, when the ridge axis overrides the plume (Fig. 3c), the eruption volume notably increases due to enhanced melting by the thinning oceanic lithosphere. The eruption then becomes temporally disrupted when the plume gets captured and advected northward by the ridge (Fig. 3d), characterized by a sudden northward displacement of the top of the plume and the melting region (i.e., hotspot). During the subsequent northward passage of the ridge axis, the southern flank of the ridge arrives above the plume, meaning that the overriding plate switches to the Pacific Plate (Fig. 3d). As the ridge system moves farther away to the north, the ridge attraction effect gradually diminishes to the level of pure Pacific drag, whose influence lasts to as late as 40 Ma when the northward plate motion finally diminishes (Figs. 3e, f and 4a). Therefore, this plume-ridge interaction first decreases the paleolatitude of the melting region due to southward ridge attraction (Fig. 3a–c), then suddenly increases its paleolatitude due to northward ridge advection (Fig. 3c–e), and eventually reduces it again due to the weaker drag of the Pacific plate (Fig. 3e, f). This prediction is reminiscent of the observed erratic paleolatitude variations from Meiji to Detroit and then to Suiko-Koko seamounts (Fig. 1b).

Fig. 3: Snapshots of the simulated Hawaiian plume evolution (Model 2B).
figure 3

The 3D contour representing the + 100 °C temperature anomaly is shown in red. The 3D contour of the 1% melt fraction is shown in white. The core-mantle boundary is shown in light blue. Plate boundaries are shown as green lines. Plate velocities are shown as black arrows. The length of the arrows qualitatively shows the magnitude of the plate velocity. The background shows the log-normalized (relative to 1021 Pa s) mantle viscosity profile. The snapshots of before (a, b), during (c), and after (d–f) the ridge arrival are shown. Both the ridges and the Hawaiian plume produce melt. The top right corner of each snapshot shows the temperature contour (black contour) and melt area (outlined by the red line) of the plume, and the profile of the temperature contour in each snapshot is denoted by a black dotted line on the top of the plume.

Full size image
Fig. 4: Paleolatitude predictions and plume source motion.
figure 4

a Simulated paleolatitudes of the Hawaii-Emperor seamounts. Orange dots with error bars (95% confidence limits) are data by Woodworth and Gordon38. Blue dots are data (95% confidence limits) from the Ocean Drilling Program (ODP)3 and Deep Sea Drilling Project (DSDP) boreholes7. Purple points are paleolatitudes estimated by correcting plate motion effects from the present Hawaii chain using the reconstruction of Müller et al. 11. The red line is the paleolatitude calculated from Model 1. Paleolatitudes of subducted seamounts are shown as the red dashed line (see “Method”). The green line is the paleolatitude calculated from Model 2. The northward velocity of the Pacific Plate is shown in the cyan dashed line11. The gray dashed line is the latitude of the plume source trajectory from Model 2. b The best-fitting lateral trajectory of the plume root is sourced at ~ 600 km depth (Model 2). The trajectory contains two linear motions during 80–47 Ma and 47-0 Ma. The cumulative southward plume source movement is about 7° since 80 Ma. 120 Ma is the starting time of our simulated plume source, and the plume is stagnant before 80 Ma.

Full size image

To further investigate the effect of ridge location on seamount evolution, we ran several additional models with various ridge locations. In the model (Model 2 C) that is identical to Model 2B but without a ridge (thus no ridge attraction and advection), the track deviates from Model 2B significantly before ~ 70–65 Ma (Supplementary Fig. 10). This is because in Model 2C, as there is no ridge between the Izanagi plate and Pacific plate, both plates move at the same Pacific plate velocity, which delays the arrival of the Pacific plate at the plume compared with Model 2B (Fig. 3d). At the same time, as the Izanagi plate changes its motion to the NW direction instead of the nearly northward direction as in Model 2B (Fig. 3e), the northwestward deflection of the track develops. The ridge effect diminishes after ~ 70–65 Ma because the ridge is far away from the plume (Supplementary Fig. 10 and Fig. 3f). In other models that are modified from Model 2B by varying the initial ridge location, the predicted tracks during 0–55 Ma are similar to those in Model 2B, but clearly drift off before Detroit formation (~ 75–90 Ma), especially at Detroit (Supplementary Fig. 11). Since during this period of these models, the plume moves southward while the ridge moves northward, shifting the ridge further south will delay the timing of plume-ridge interaction, and vice versa, resulting in the plume either meeting the ridge too early (blue, purple) or too late (red). In addition, except for the track in the model with a much southern ridge (Supplementary Fig. 11, model a, − 10°), others deviate from the reference model (2B) toward the northwest side (Supplementary Fig. 11), which may result from the change of the Pacific plate motion from NW to nearly northward during ~ 75–90 Ma (Fig. 3d, e). It is worth noting that all tracks are more (purple track) or less (yellow track) shifted eastwards compared to the observed seamounts during ~ 75–55 Ma. However, the mechanism is quite different from the previous one, as during this period, the Pacific plate movement was relatively stable. Based on the result of Model 2B, we propose that it may be the plume-plate boundary interaction that influences the tracks. As shown in Model 2B (Fig. 3d, e), during ~ 75–55 Ma, the plate boundary between the Pacific plate and the Izanagi plate sits very close to the plume. While the velocities of the adjacent plates are quite different, there should be a pressure difference, which may lead to a Couette flow beneath that could affect the drag on the plume and result in a track deviation at the surface. In conclusion, the ridge location determines the timing of plume-ridge interaction, plume-plate boundary interaction, as well as the relative motion between plates, thus shaping the geometry and age distribution of the seamount chain.

On the basis of this result, we further analyze the paleolatitude of modeled hotspot tracks (Fig. 4). The calculated paleolatitudes based on the reconstruction of Müller et al. 11 (red line in Fig. 4) are lower for Detroit and higher for Meiji than their respective observational counterparts. But the result that Meiji’s paleolatitude is lower than that of Detroit matches observation. The calculated paleolatitudes in models (Models 3 & 5) based on other plate reconstructions are shown in Supplementary Fig. 2. In the reference case (Model 1, Supplementary Fig. 12), the predicted paleolatitude offset (north of the present Hawaii island) increases from the initial 0.8° at 105 Ma, when the ridge first encounters the plume, to 4.1° at 97 Ma, and to the highest value of 5.8° at 82 Ma, after which the value gradually diminishes. This predicted maximum offset (5.8°) at 82 Ma accounts for half of the total offset by further including plume migration (green line in Fig. 4), and its timing is slightly earlier than that observed (~ 78 Ma).

Another major shallow process that affects the hotspot track is plate drag. This process is long-lasting and beyond the duration of the plume-lithosphere interaction. According to the model results, as the plate moves, it excites a Couette flow (the amplitude of horizontal velocity decreases with increasing depth) in the underlying asthenosphere, which will drag the plume in the same direction. For example, the fast-moving Izanagi Plate bends the plume northward (Fig. 3a and Supplementary Fig. 12a), but this effect is almost entirely balanced by the southward attraction of the ridge when it approaches the plume (Fig. 3b and Supplementary Fig. 12b). Similarly, the slower-moving Pacific plate entrains the plume as well, although with reduced strength. Furthermore, the varying speed of the Pacific Plate explains the two local maxima of calculated paleolatitudes (Model 1) at 82 and 97 Ma during plume-ridge interaction: The drop in predicted paleolatitude at 97 Ma is in response to the temporary southward Pacific movement; the calculated paleolatitude rises again when the Pacific switches to moving northward after 95 Ma (Fig. 4a). On the basis of these observations, we can estimate the maximum amount of the calculated paleolatitude offset caused by direct ridge attraction to be about 3.3° (from 0.8° at 105 Ma to 4.1° at 97 Ma when the Pacific motion is small) and that by Pacific drag in response to the ~ 10 cm/yr northward Pacific velocity increase to be about 2.7° (95–80 Ma) (Fig. 4a). Note that these two effects occur at slightly different times, their combined effect could start at ~ 109 Ma when the ridge is close enough to attract the plume from the south and end at ~ 80 Ma when the ridge is far away, resulting in a maximum of 5.8° offset (Supplementary Fig. 12 and Fig. 4a).

Plume buoyancy can also potentially influence the model results. Studies suggest that the Hawaii plume may include some eclogites20,24, which may represent recycled oceanic crust. Therefore, we evaluate the effect of eclogite density in our models, by testing how variations in net plume buoyancy can influence the predicted paleolatitudes. We collect test results when the plume with reduced buoyancy (Models 10, 11, 12, 13) can rise to the surface and form a continuous volcanic track as in Model 1. As shown in Supplementary Fig. 3, reducing plume buoyancy has little effect on the extent of plate drag during plume-ridge interaction, thus, the resulting paleolatitude predictions. Among these tests, the maximum modeled paleolatitude offset always occurs around 85 Ma by no more than 6°.

We also test the effect of asthenospheric viscosity by increasing and reducing its value by a factor of 2 (Model 13, 14). The choice of this value range is based on multiple of our previous studies with data assimilation that matches both past subduction history and the present mantle structures17,18. As shown in Supplementary Fig. 9, the resulting paleolatitude offsets are not sensitive to the tested asthenospheric viscosity. This is likely due to the tradeoff between the vertical and lateral motions of the plume: while reducing asthenospheric viscosity allows the plume to rise faster, the plume also migrates faster laterally due to pressure-driven flow with reduced viscous resistance, thus resulting in a similar eruption site and hotspot location.

Contribution of plume migration

In order to further improve the model to fit with the observation, we also consider the movement of the plume source. We treat this exercise as an inverse problem where we explore different trajectories of simple plume motion (assuming a two-stage linear movement history unless otherwise noted, Fig. 4b) by minimizing model mismatches with multiple aspects (shape, location, age, and paleolatitude) of the observed Hawaii chain (Figs. 1, 2, 4a). In practice, we focus on varying the two-stage plume source trajectory only to fit the locations of the two kinks (at HEB and Detroit) within the hotspot track. The optimal Model 2 (with track 2B), compared to Model 1 (track 1B) that assumes a fixed plume source, not only significantly improves the overall shape of the observed hotspot track, but also reproduces all seamount ages within a 5-Myr error range (Fig. 2) and paleolatitudes of nearly all seamounts (green line in Fig. 4a). This result is more reasonable than that in Hassan et al. 10, whose predicted seamounts older than 70 Myr are all north of the Bering trench. Importantly, the predicted paleolatitudes with the moving plume source closely match the observed values of all seamounts, with Meiji being the only exception (Fig. 4a).

In our best-fit model (Model 2B, Fig. 2), the required amount of meridional migration of the plume during the Emperor chain formation (86–47 Ma) is much smaller than recent inferences3,4,10. Our optimal plume source model has a constant southward speed of 0.13°/Myr, or a total of 4.4° southward migration between 80 and 47 Ma (Fig. 4), about one-third of that implied by paleolatitude measurements (Fig. 1b). Constraining the longitudinal motion of the plume source from paleomagnetic records remains a major challenge. Our exercise, by forcing the modeled hotspot track to match observation, provides a means to infer this motion. Although the result depends on the adopted plate motion history, models based on all three plate reconstructions (Models 2, 4, 6, Supplementary Table 1) considered here produce a consistent SSW-directed motion of the plume source prior to 47 Ma, after which the plume turns to ESE (Fig. 4b and Supplementary Figs. 2, 4). We interpret this E-W flipping motion at 47 Ma as reflecting a return-flow response of the lower mantle to the increased westward motion of the Pacific around this time.

Another unexpected improvement in the moving plume models (Models 2, 4, 6) is in the timing of the Meiji-Detroit bend (Figs. 1a, 2) and associated paleolatitude change (Figs. 1b, 4a and Supplementary Fig. 4). In model 2, because the Hawaiian plume source is about 6.4° more north around 80 Ma than in the fixed-plume model (Model 1), it encounters the Izanagi-Pacific ridge much later. Consequently, the timing of the simulated peak paleolatitude offset is delayed to around the formation of Detroit (green line in Fig. 4a), more consistent with observation than in Model 1 (red line in Fig. 4a). In addition, the maximum simulated paleolatitude offset in Model 2 is increased to 11.4° around the time of the Detroit formation. This amount nearly equals the combined offsets due to plume source migration (~ 6°) and to the plume-lithosphere interaction (~ 5.8°), thus verifying the respective contributions from the deep and shallow mantle processes.

We note that the simulated paleolatitude offset of Meiji is always higher than observed, with the misfit being 5−10° from models based on all three reconstructions (Fig. 4 and Supplementary Figs. 2, 4). The reason may lie in the inaccuracies of (1) the paleolatitude measurements at Meiji, (2) the adopted plate reconstruction model, or (3) the assumed plume migration history. Examination of the first possibility is beyond our capability. Regarding the second one, we first choose a plate reconstruction model11, combined with the observed positions and ages of the present Hawaii-Emperor chain, reversing the chain back to the time when it was generated, and then obtaining the Meiji paleolatitude. However, the resulted paleolatitude offset of Meiji is still much higher than that observed (purple dots in Fig. 4, Supplementary Fig. 4 for other reconstruction models). This means that plate motion history cannot account for the abnormally low Meiji paleolatitude. As the plume’s precise moving history is unknown, here we perform another model (Model 7) to test the third possibility and obtain a three-stage plume movement history (Supplementary Fig. 5). It differs from Model 2 only in that the plume source moves northward at 90 Ma from the current location of Hawaii to the 80-Ma location in Model 2. In this case, the paleolatitude of Meiji is indeed improved (Supplementary Fig. 5b), but its present location is far more south, thus still violating certain observations (Supplementary Fig. 5c). Accordingly, the above presumed three-stage plume movement is not a reasonable scenario.

On the basis of these additional tests regarding Meiji’s abnormally low paleolatitude, we propose that one way to reconcile this dilemma is to suggest that the Pacific plate rotated clockwise during the formation of Meiji but stopped before Detroit formed. Chandler et al. 25. find that the Ontong Java Plateau, which was formed at ~ 123 Ma and is now located in the southwest Pacific, may have experienced ~ 40° of clockwise rotation. If this plateau was part of the Pacific Plate back then, the region where Meiji formed should have also rotated by a similar amount. This scenario could naturally account for the difference between the simulated and observed Meiji-Detroit chain orientations, which also differ by ~ 40° (Supplementary Fig. 5c).

Implications on deep mantle dynamics and hotspot track formation

In this study, we quantitatively evaluate the effects of shallow dynamic processes on the formation of the Hawaii-Emperor chain, using several recent plate reconstruction models11,12,13. On the base of this result, we further infer the necessary latitudinal migration of the plume source by best reproducing the observed hotspot chain properties. This workflow is motivated by the fact that our knowledge of surface plate kinematics is better than that of the deepest mantle. Indeed, current models of the Hawaii-Emperor chain formation, even by acknowledging the observed paleolatitude offsets26, involve drastically different views on lower mantle dynamics, including a stationary plume6, a moving plume toward a stationary LLSVP27, and a migrating plume originated from the retreating edge of LLSVP10.

To circumvent the uncertain lower mantle dynamics, we attempt to infer the effective plume motion by correcting the contribution from the better-constrained shallow mantle processes through quantitative numerical calculations. In practice, we place the present-day plume source at ~ 600 km depth where seismic tomography detects the plume conduit21. Our result agrees with that from recent convection studies10,28 in that the plume source moves southward during most of the time when the Emperor seamounts formed. But our inferred amount of southward plume motion is only half of that from these recent results, with the shallow processes accounting for the other half. This conclusion is based on the above presented systematic study through which the preferred model could match more data (track shape, seamount age, and paleolatitude) than all previous models. Therefore, the lower mantle flow that dislocates the plume and/or the LLSVP should be less vigorous than previously thought. Our result also quantifies the contribution of shallow processes as being equally important in forming the Hawaii-Emperor chain. So, future research concerning mantle evolution and plate reconstruction should take such shallow effects into account.

Our preferred model with two-stage plume movement (Model 2) predicts 11.4° of maximum paleolatitude offset around the Detroit seamount, which is within the uncertainty range (8–18°) of the observed mean offset (12°). Although allowing more southward plume migration may slightly increase the maximum offset10, such exercise fails to reproduce seamount ages with a large portion of the observed Emperor chain being subducted by now (Fig. 2 and Supplementary Fig. 6). Therefore, the 11.4° offset should represent a realistic value, which implies that some very large offsets (> 20°) near Detroit may represent inaccurate records of the true paleolatitudes resulting from limited sampling data and analytical error29, possibly due to post-eruption deformation by the underlying plume-lithosphere interaction. Similar reasoning can be applied to the very low paleolatitude offset (~ 0°) of Meiji28 if future research rules out the scenario of rapid clockwise rotation of the Pacific plate prior to 80 Ma as implied in our three-stage plume motion model (Supplementary Fig. 5).

An even broader implication is that this identified plume-lithosphere interaction with a nearby ridge could have affected other hotspot tracks as well. Paleolatitude data computed from GMHRF-3 (global moving hotspot reference frame)30 shows that, besides Hawaii, New England, Reunion, and Tristan all experienced prominent paleolatitude shifts by different amounts (5° to 10°) (Supplementary Movie 1). The paleolatitude of the New England hotspot decreased rapidly starting at 120 Ma, as could reflect a similar mechanism to that for Hawaii: the northward velocity (thus drag) of the North American plate that the hotspot interacted with prior to 70 Ma suddenly decreased after 120 Ma, followed by ridge attraction of the approaching North America-Africa ridge from the south. When the ridge passes over the plume, the ridge attraction is simply compensated by the plate drag, resulting in no change in the paleolatitude trend of the volcanoes. Eventually, the plate drag is further reduced as the plume migrates below the slow-moving African plate (Supplementary Fig. 7a and Movie 1). Similarly, the sudden increase of Reunion’s paleolatitude around 70 ~ 60 Ma is consistent with it traversing the northward moving Indian-Africa ridge, and the subsequent rapid paleolatitude decrease reflects reduced plate drag (due to slower plate velocity) south of this ridge, a mechanism that has also been suggested by Bredow et al. 31,32. (Supplementary Fig. 7b and Movie 1). It is worth noting that during ~ 53–39 Ma, plate motion increases while paleolatitude decreases, which indicates that the fading ridge effect may still play an important role during this stage. Tristan’s paleolatitude gradually decreases after ~ 120 Ma, which could also be related to both the diminishing northeastward plate drag and the long-lasting attraction from the Africa-South America ridge, where the hotspot has been hovering around since the Late Cretaceous (Supplementary Fig. 7c and Movie 1).

In conclusion, we demonstrate that the plume-ridge interaction can temporally modulate the long-lasting plume-lithosphere interaction, leading to significant hotspot track deviations and paleolatitude offsets in multiple hotspot chains. Consequently, the observed abnormal hotspot motions require much less contribution from the deep plume source than previously thought. For the Hawaii chain, plume-lithosphere interaction could account for around 50% of paleolatitude change, and lower mantle dynamics or plume source motion may account for the other 50%. These inferences should enlighten further research in both mantle dynamics and plate reconstruction.

Methods

Model setup

In order to reproduce the history of the Hawaii-Emperor hotspot chain while considering possible dynamic effects from all nearby plate boundaries, our regional models manage to include the subduction of the Izanagi, Kula, and Northern Pacific plates since the middle Mesozoic. Because these features are too close to the North Pole, the default mesh of a regional model defined by latitudes and longitudes in CitcomS will not work. Therefore, we conducted a coordinate transformation, and the transformed computation domain is shown in Supplementary Fig. 1. Similarly, all the boundary data from the adopted plate reconstruction models are transformed as well.

We model the mantle as an incompressible viscous fluid under the Boussinesq approximation. Besides, we use a data assimilation method17,18 to reproduce past subduction and mantle flow. In our model, all model boundaries, except the surface, are free to slip. The model’s surface assimilates kinematic boundary conditions from the plate reconstruction models, which also provide information about seafloor age and plate boundary geometries at a 1-Myr time interval. The input seafloor age and plate motion are smoothly interpolated between the two adjacent snapshots. To get robust results, we test three different reconstruction models from Müller et al. 11, Müller et al. 12, and Torsvik et al. 13.

In our model, we prescribe a thoroughly weakened vertical zone along all sidewalls to minimize the artificial flow due to the impermeable vertical boundaries17,18. In addition, the study region is near the center of the model domain, and the plume does not move much laterally. Hence, the model’s side boundaries are always far from the plume, and the plume evolution is not affected by the side boundaries.

We solve for thermochemical convection, where compositional tracers at 20 per element are used to track key tectonic features, including the crust, the mantle, and the plume. Tracers with different physical properties (density and viscosity) are used to facilitate realistic subduction17,18. Additional model components, like eclogite inside the plume and simplified melt migration, are also realized through tracers.

The main difference between our model and previous studies10 is that, instead of simply projecting the location of the plume conduit from its source to the surface, we further consider the lateral advection of the plume within the asthenosphere due to interaction with the moving plate at the surface. In addition, we also incorporate a simplified melting function into the evolution of the thermal plume, such that the melt could accumulate over time and advect with the ambient mantle flow. The plume conduit could then be bent by plate drag and mid-ocean ridge attraction, while the melt region could move laterally within the plume head, as shown in Fig. 3.

Simulating partial melting

In order to track the formation of the Hawaii-Emperor chain, we consider the melting process of the mantle. In this study, we only consider the melt of peridotite and adopt a linear melting function for the volumetric degree of melting (before considering extraction and depletion) \({\varphi }_{0}\)19.:

$${\varphi }_{0}=\min \left\{1,\max \left\{0,\frac{{T-T}_{{{{\rm{solidus}}}}}}{{T}_{{{{\rm{liquidus}}}}}-{T}_{{{{\rm{solidus}}}}}}\right\}\right\}$$
(1)

where \({T}_{{{{\rm{solidus}}}}}\) and \({T}_{{{{\rm{liquidus}}}}}\) are pressure-dependent solidus and liquidus constrained by experiments on dry peridotite33,34:

$${T}_{{{{\rm{solidus}}}}}=1080+134.2P-6.581{P}^{2}+0.1054{P}^{3}$$
(2)
$${T}_{{{{\rm{liquidus}}}}}=1762+57.46P-3.487{P}^{2}+0.077{P}^{3}$$
(3)

where P is pressure in GPa, and T is temperature in °C. The pressure is calculated using a 1D reference density model35.

We assume melt materials would erupt at a rate of 10%/Myr when the local melt percentage reaches a threshold of \({M}_{\max }=4\%\)19. When an eruption happens, the erupted melt is put on the surface instantaneously. This simplified modeling may result in that the depth at which the melt reaches the eruption threshold is very shallow, usually right below the LAB. This means melt percolation happens mostly within the high-viscosity lithosphere, where the mantle wind is absent. Thus, such simplification would not cause a notable mismatch at the surface eruption site. Melt depletion \({\varphi }_{{extracted}}\) is also calculated to record the amount of erupted melt. The in situ melt fraction is \(\varphi={\varphi }_{0}-{\varphi }_{{extracted}}\).

Model rheology

Our models adopt a depth-, temperature-, and composition-dependent viscosity law for the whole model domain before considering melting effects:

$${\eta }_{{{{\rm{no}}}}\,{{{\rm{melt}}}}}={N}_{0}(r)\cdot C\cdot \exp \left(\frac{{E}_{0}(r)}{T+{T}_{0}(r)}-\frac{{E}_{0}(r)}{{T}_{m}+{T}_{0}(r)}\right)$$
(4)

where ηno melt is the effective viscosity before considering melting effects, \({N}_{0}(r)\) is the depth-dependent background viscosity, C is the viscosity multiplier due to compositional effects, \({E}_{0}(r)\) is the activation energy, \({T}_{0}(r)\) is the temperature offset, and r is the radial distance from the earth’s core. T is the temperature and Tm is the background temperature. These parameters and their values are listed in Supplementary Table 2. Most parameters have been used and tested in other studies17,18,36.

The background mantle includes a four-layer viscosity profile: 0–44, 44–410, 410–660, and 660–2867 km. The background viscosity in each depth range is 1020 Pa·s, 1020 Pa·s, 1021 Pa·s and 1023 Pa·s, respectively. The temperature dependence leads to four orders of magnitude in lateral viscosity variation. This viscosity structure is largely consistent with our previous inferences based on data assimilation models for simulating past subduction and mantle flow17,18,36, which provide a tight range on the values of viscosity at each mantle depth. Besides, a viscosity cut-off (1019 – 1023 Pa·s) was employed in our models to avoid too large variation due to compositional and melting effects. These model parameters have been extensively tested in recent studies, as shown to be appropriate for simulating subduction and mantle flow dynamics17,18,36.

We consider two kinds of melt weakening mechanisms to account for melt effects on viscosity. The first concerns melt extraction, during which the viscosity of the whole rock column between the melt source and the surface would be weakened37:

$$\eta={\zeta }_{1}\cdot {\eta }_{{{{\rm{no}}}}\,{{{\rm{melt}}}}}$$
(5)

where \({\zeta }_{1}\) is the melt weakening pre-factor. This mimics the process of melt eruption. The whole eruption channel from the node where melt is extracted to the surface would get weakened, following a recent study37.

Another weakening mechanism arises from the fact that rheologically weak melt can be retained inside the mantle20. We assume the effective viscosity depends on the in situ melt fraction following:

$$\eta={\eta }_{{{{\rm{no\; melt}}}}}\times \exp (-{\zeta }_{2}\times \varphi )$$
(6)

Where ηno melt and η are the effective viscosity before and after considering melting effects, respectively. \({\zeta }_{2}\) is another melt weakening pre-factor20, and φ is the in situ melt fraction. Since the maximum in situ melt fraction in our model is limited to 4%, this weakening mechanism leads to a maximum viscosity reduction of one order of magnitude.

Volcano tracking

We develop surface hotspot tracks through two different approaches. Track A records the vertical projection of the plume source at 600 km depth onto the moving surface, and track B records the temporally changing location of the simulated melt eruption.

For track A, because the location of the plume source is known a priori, we simply advect the corresponding tracers at the Earth’s surface following plate motions from the reconstruction models.

For track B, after the erupted melt is moved to the surface instantly, this erupted material would be attached to an existing tracer near the surface, representing the volcano. Thus, it could move with the plates and even subduct under another plate. Since this melt has been moved to the surface and solidified, it has no effect on viscosity anymore. In practice, this volcano material is removed from the model when it subducts to a depth greater than 1000 km.

Then, we use a postprocessing script to collect volcanic materials to generate a 2D map of the hotspot track. We first choose several control points, and then use a moving rectangle window to get the weighted mean location and age of volcanoes inside the window. By moving the window from one control point to another, we can obtain a modeled hotspot track (Supplementary Fig. 8). Simulated paleolatitude offsets are calculated using the latitude difference of the seamounts at the same age in the two tracks between the predicted track #A and track #B.

Robust plume source migration results

To test the robustness of our results, we further tested the reconstruction models of Müller et al. 12 and Torsvik et al. 13 in addition to the model of Müller et al. 11, as shown in the main text. We tested both the fixed and moving plume source trajectories for each reconstruction. All the results are shown in Supplementary Fig. 2.

For models with a moving plume source (Supplementary Fig. 2c), all three models share the same north-south motion. However, different east-west plume movements are required for different reconstructions to match the locations of the HEB and Meiji-Detroit bends. Based on these reconstructions, all three models produce a consistent SSW-directed motion of the plume source prior to 47 Ma, after which the plume turns to ESE. The results of the robustness tests on the motion of the Hawaii hotspot are shown in Supplementary Fig. 4. The paleolatitudes inferred from these additional tests are similar to those based on Müller et al. 11. Specifically, the paleolatitude of Detroit is always larger than that of Meiji.