Letter | Published:

Water input into the Mariana subduction zone estimated from ocean-bottom seismic data


The water cycle at subduction zones remains poorly understood, although subduction is the only mechanism for water transport deep into Earth. Previous estimates of water flux1,2,3 exhibit large variations in the amount of water that is subducted deeper than 100 kilometres. The main source of uncertainty in these calculations is the initial water content of the subducting uppermost mantle. Previous active-source seismic studies suggest that the subducting slab may be pervasively hydrated in the plate-bending region near the oceanic trench4,5,6,7. However, these studies do not constrain the depth extent of hydration and most investigate young incoming plates, leaving subduction-zone water budgets for old subducting plates uncertain. Here we present seismic images of the crust and uppermost mantle around the central Mariana trench derived from Rayleigh-wave analysis of broadband ocean-bottom seismic data. These images show that the low mantle velocities that result from mantle hydration extend roughly 24 kilometres beneath the Moho discontinuity. Combined with estimates of subducting crustal water, these results indicate that at least 4.3 times more water subducts than previously calculated for this region3. If other old, cold subducting slabs contain correspondingly thick layers of hydrous mantle, as suggested by the similarity of incoming plate faulting across old, cold subducting slabs, then estimates of the global water flux into the mantle at depths greater than 100 kilometres must be increased by a factor of about three compared to previous estimates3. Because a long-term net influx of water to the deep interior of Earth is inconsistent with the geological record8, estimates of water expelled at volcanic arcs and backarc basins probably also need to be revised upwards9.


The Mariana subduction zone has long been cited as a water-rich system owing to the prevalence of forearc serpentinite mud volcanoes10, a serpentinized mantle wedge11, and hydrous arc and backarc lavas12,13. However, the initial amount of water within the subducting mantle is unknown. The subducting Pacific plate in this region is among the oldest sections of oceanic lithosphere worldwide14 (about 150 Myr). The incoming plate displays widespread plate-bending normal-fault scarps and earthquakes15,16 (Fig. 1b), making it an excellent place to investigate the depth extent of faulting-induced hydration in an old oceanic plate. The potential hydration of old, cold oceanic plates is particularly important for the water cycle because the thermal structure of the plates permits temperature-sensitive hydrous minerals to occur throughout a thicker region2.

Fig. 1: Distribution of seismic stations and bathymetry.

a, Station distribution. Red circles show ocean-bottom seismographs deployed from January 2012 to February 2013. White squares represent the temporary island-based stations. Red squares indicate stations from the US Geological Survey (USGS) Northern Mariana Islands Seismograph Network used in our study. Open triangles show the locations of three large serpentine seamounts within the study area. The dashed white line is the trench axis. Thick solid white lines show the cross-section locations in Fig. 2a–c. The arrow labelled APM indicates the direction of absolute plate motion. Thin solid white lines show magnetic lineations (M22, M23, M24)30. b, High-resolution bathymetry for the outer-rise region of the Mariana trench (indicated by the black rectangle in a) and relocated earthquake locations from the Global Centroid Moment Tensor catalogue15 (black and white focal projections). Magnetic lineations are shown as in a.

We used data collected by broadband ocean-bottom seismographs and island seismographs deployed around the central Mariana trench (Fig. 1a). The deployment covers a sufficiently large region to investigate structure variations in the subducting plate before and after subduction. We derived local Rayleigh-wave group- and phase-velocity dispersion curves using two surface-wave methods (Methods). A Bayesian Monte Carlo algorithm was then used to determine a three-dimensional image of the shear-wave velocity of the crust and uppermost mantle (Methods). Compared to previous active-source seismic studies of other subduction zones4,5,6,7, our study resolves the three-dimensional structure to greater depth and avoids biases caused by possible azimuthal anisotropy. In addition, it images shear velocities that are more sensitive to hydration than are compressive velocities17.

The resulting azimuthally averaged velocity model of vertically polarized S waves shows systematic changes in the incoming plate and subducting slab along the direction normal to the trench (Fig. 2a–c). At distances far from the trench axis (more than 100 km seaward), the subducting plate has a typical structure of old oceanic lithosphere18 at depths of more than 20 km, with velocities greater than 4.5 km s−1. A lower-velocity (about 4.2 km s−1) layer with a thickness of about 10 km is observed immediately beneath the Moho, consistent with active-source seismic results19. This suggests that the very shallow part of the incoming plate starts to be altered far from the trench, as observed in other subduction zones4. A thicker region of low velocities begins about 80 km seaward of the trench axis and deepens towards the trench, with velocities as low as 3.8 km s−1. At the trench axis, the bottom of this low-velocity layer reaches 30 ± 5 km beneath the seafloor, 24 ± 5 km into the upper mantle. (The errors quoted here and elsewhere correspond to the variations in thickness that result from different assumptions about the velocity that bounds the serpentine layer.) This low-velocity layer persists after the Pacific plate is subducted as a (30 ± 5)-km-thick low-velocity layer atop the fast slab mantle between 100 km and 160 km west of the trench axis. Synthetic data tests show that a velocity anomaly of this large magnitude and extent cannot result from the effect of limited resolution of an approximately 6-km-thick layer of low-velocity oceanic crust (Methods, Extended Data Fig. 1).

Fig. 2: Vertical profiles and interpretation.

ac, Cross-sections A–A′ (a), B–B′ (b) and C–C′ (c) showing the azimuthally averaged velocity of vertically polarized S waves (VSV; colour scale). Thick dashed white lines show the location of the forearc Moho31. Thick solid white lines are projected 6-km-thick slab crust (combining an active-source reflection result16 for depths of less than 30 km and the Slab 1.0 model32 for greater depths). Thin white lines are velocity contours of 3.6 km s−1 and 3.9 km s−1, and thin black lines are velocity contours of 4.2 km s−1. Black33, white11 and blue15 circles are relocated earthquakes in the subducting plate around each profile from previous studies. d, Interpretation of seismic structure in b within the area bounded by the two vertical dashed black lines. Dashed black lines are isotherms of 300 °C and 600 °C calculated from ref. 25. Solid grey lines are contours of 0 MPa and 100 MPa extensional stress15. The dashed grey line marks the surface of the slab. The vertical dashed white line indicates the trench axis. Vertical solid blue and black lines mark the locations of the one-dimensional VSV structure shown on the right (colour-coded). The colouring schematically illustrates the velocity, as in ac. UM, unaltered mantle; PS, partial serpentinization; PW, pore water.

We also solved for the azimuthal anisotropy of the phase velocity. These results show trench-parallel fast axes in the incoming Pacific plate for periods of 12–21 s, with a maximum magnitude as large as 9% reached between 14 s and 16 s. By contrast, for periods of more than 27 s, the fast directions rotate to be oblique to the trench strike, close to the palaeo-spreading direction, which is normal to the magnetic lineations (Fig. 3).

Fig. 3: Azimuthal anisotropy results at various periods.

a, b, Results from ambient noise tomography (ANT; red bars) at periods of 12 s (a) and 16 s (b). c, Results from ambient noise tomography and Helmholtz tomography (HT; yellow bars) at a period of 21 s. d, Results from Helmholtz tomography at a period of 27 s. The orientations of the red and yellow bars represent the fast directions and the lengths indicate the magnitude of anisotropy. The trench axis, serpentine seamounts, absolute plate motion and magnetic lineations are shown and labelled as in Fig. 1a. Only nodes with good azimuthal coverage and good azimuthal anisotropy fittings are plotted.

The region of the incoming plate where we observe a reduction in mantle velocity and large azimuthal anisotropy coincides with the plate-bending region, as characterized by substantial normal faulting seismicity15 and large extensional seafloor fault scarps16 (Fig. 1b). There is a clear spatial association between plate-bending-induced faulting and velocity reduction. Velocities within the incoming plate begin to decrease sharply 80 km from the trench, at about the same distance at which intense seismicity and faulting begins on the seafloor16. In this region, the fault scarps and the earthquake fault planes strike approximately north–south, subparallel to the trench axis (Fig. 1b). This is consistent with the observations of the azimuthal anisotropy of the phase velocity at 12–21 s, which primarily sample depths down to about 25 km below the seafloor (Fig. 3). These results reveal trench-parallel fast directions, as would be expected for pervasive trench-parallel water-filled faults or zones of alteration. The flexure model that best fits the bathymetry of the Pacific plate seaward of the trench axis15 predicts a neutral plane at around 30 km (Fig. 2d), which suggests that brittle normal faulting can extend nearly 30 km into the plate. This prediction agrees well with the maximum depth extent (30 ± 5 km) of the low-velocity zone that we observed in the incoming plate near the trench in our study (Fig. 2a–c), which suggests that this zone is related directly to brittle normal faults.

Many previous studies at other subduction zones have attributed low-velocity zones in the upper mantle associated with plate bending to the hydration of mantle peridotite to form low-velocity serpentine minerals. Extensional deformation within the shallow part of the incoming plate produces a pressure gradient that may enable water to penetrate deep into the slab along normal faults20. The serpentinization rate is geologically fast if water delivery to the serpentinization front is efficient21,22. Alternatively, other studies attribute the reductions in mantle velocity to water-filled porosity and cracks23. However, the potential velocity effect of water-filled cracks is usually difficult to estimate directly because it depends critically on the aspect ratio and spatial density of the cracks, which are largely unknown. In this study, we use the increase in velocity as the plate subducts and porosity is reduced to distinguish the effects of water-filled cracks, porosity and serpentinization (Fig. 2d).

The seismic images of the Mariana trench show that the velocity of the low-velocity zone increases as the top of the slab subducts past a depth of about 30 km (Fig. 2a–c). Reductions in porosity due to increased pressure24 reduce or eliminate the velocity effect of water-filled cracks at depth, whereas hydrous minerals remain stable at the cold slab temperatures; this provides a means to separate the complementary velocity effects of porosity and altered minerals and to estimate the concentration of hydrous minerals (PS region in Fig. 2d). Compared to the low-velocity zone within the slab before subduction, the low-velocity zone within the subducted-slab mantle at depths of around 40 km preserves the original thickness (30 ± 5 km) but exhibits a smaller reduction in velocity (about 4.1 km s−1; Fig. 2a–c). We use this smaller velocity reduction to estimate the degree of mantle serpentinization in the downgoing slab. Therefore, we base our estimates of the water content due to serpentinization of the subducting slab on the initial thickness of the low-velocity layer at the trench and on the shear-wave velocities of around 4.1 km s−1 observed at depths of 30–50 km in the subducting plate mantle, after most of the pore water is expelled (PS region in Fig. 2d). The additional velocity reduction (about 0.3 km s−1) within the low-velocity zone in the slab before subduction can be attributed to pore water in cracks (PS + PW region in Fig. 2d).

Calibrating the change in seismic velocity to the degree of serpentinization requires knowledge of the seismic velocity of serpentine. We select lizardite, the form of serpentine expected to predominate at lower temperatures22, to interpret the observed velocity reduction. The nominal temperature predicted by plate cooling models25 around 30 km beneath the seafloor is roughly 470 °C (Fig. 2d), possibly higher than the temperature of lizardite breakdown (about 320 °C in ref. 26, although lizardite has been found at temperatures as high as 580 °C in ref. 22). Water circulation in cracks may lower the temperature of the slab mantle in the plate-bending region into the lizardite stability field. In addition, selecting lizardite provides an estimate of the lower bound of water input (Methods). Using the experimental relationship between the change in shear velocity and the serpentine volume fraction for lizardite27, the shear velocity of 4.1 km s−1 observed within the subducted slab corresponds to a change in velocity of about 0.41 km s−1 (Methods), indicating roughly 19 vol% serpentinization (about 2 wt% water). It is possible that the anisotropic effects of serpentine distributed along bending faults could cause additional reductions in velocity, leading to an overestimate of the serpentinization percentage28. However, according to our calculation with a realistic dipping-fault geometry and the frequencies used in this study, the anisotropic effects are not important when estimating the serpentinization percentage (Methods, Extended Data Fig. 2).

We therefore interpret the seismic images as strong evidence for a (24 ± 5)-km-thick, partially serpentinized (2 wt% water) slab-mantle layer. Applying a convergence rate of 50 mm yr−1, the amount of water input into the Mariana subduction zone through mantle serpentinization would be about 79 ± 17 Tg Myr−1 m−1; the total water flux is approximately 94 ± 17 Tg Myr−1 m−1 if water in the sediment and crust is also included from previous estimates3. This estimate of water flux into the Mariana trench is 4.3 ± 0.8 times larger than a previous estimate3, which assumed a 2-km-thick, partially serpentinized slab mantle (2 wt% water). All uncertainties are estimated on the basis of the uncertainty of the thickness of the serpentinized slab mantle.

Our interpretation of serpentinization extending to depths of around 24 km below the Moho in the incoming plate at the Mariana trench has important implications for water flux into subduction zones globally. This depth is greater than the maximum observed depth of large normal-faulting earthquakes and close to the estimated depth of the neutral stress plane15 (Fig. 2d). The maximum depth of serpentinization near trenches has not been well determined for other older incoming plates, because the depth extent is too great to be well constrained by active-source seismic studies4,5,6,7 and surface-wave investigations have not been performed elsewhere. The bending and faulting features of the incoming Pacific plate near the Mariana trench are similar to those observed at other old subduction plates, and the maximum depth of normal faulting and the depth of the neutral plane are generally about the same29. Therefore, it is reasonable that serpentinization extends to similar depths of 20–25 km below the Moho at other sites where old lithosphere subducts. Modifying previous calculations of global water flux into subduction zones to take into account the hydrous alteration of this increased mantle thickness yields an estimated flux of 3.0 × 109 Tg Myr−1 (Methods)—an increase by a factor of about three1,2,3.

This larger estimate of the input water flux at subduction zones is much greater than current estimates of water output from the mantle. Because a large long-term net influx of water to the deep interior is inconsistent with the stability of sea level in the geological record2,8, one possible implication of our result is that the thick layer of serpentinized mantle that we find in the Mariana subduction zone is not characteristic of other old, cold subducting slabs, and that the Mariana slab carries much more water than other subduction zones. However, there is little indication that the incoming plate-bending region of the Mariana subduction zone is substantially different in terms of morphology and intensity of faulting compared to the corresponding regions of other old subduction zones. Thus, the most likely interpretation is that previous estimates of water output from the mantle are also underestimated. Estimates of water output from the mantle at mid-ocean ridges and ocean islands may be relatively well constrained, but estimates for volcanic arcs and backarcs rely on the melt flux and the water content, which are poorly constrained8,9.


Data processing and group- and phase-velocity tomography

The data used in the group- and phase-velocity tomography were collected mainly by 19 ocean-bottom seismographs and seven temporary island-based seismic stations deployed from January 2012 to February 2013 (Fig. 1a). The distribution of ocean-bottom seismographs covers the outer-rise region of the trench and the Mariana forearc region. In addition, we used data from three island stations from the USGS Northern Mariana Islands Seismograph Network that were active over the same time period.

We carried out ambient noise tomography (ANT) following previously described procedures34,35. The daily vertical-component seismograms were corrected for instrument responses and clock errors, and down-sampled to two samples per second. We applied running-average time-domain normalization and spectral whitening to minimize the effects of large earthquakes. Seismograms from all station pairs were then cross-correlated and stacked over the entire time period of the deployment. Frequency–time analysis34,36 was applied to the symmetric components of the stacked cross-correlations to measure Rayleigh-wave group and phase velocities between periods of 8 s and 25 s. For each frequency, only station pairs with distances larger than twice the wavelength were kept. All dispersion curves were screened to exclude those with inconsistent measurements at adjacent periods. For each period, a ray-theory-based tomography method37 was applied to dispersion measurements with signal-to-noise ratios greater than 5 to produce Rayleigh group- and phase-velocity maps on a grid of nodes spaced at 0.2°. The tomographic inversion returns the isotropic and azimuthal anisotropic components of the Rayleigh-wave group and phase velocity (Extended Data Fig. 3).

We applied a Helmholtz tomography (HT) method38 to teleseismic Rayleigh waveforms to determine phase velocities at longer periods. From the International Seismological Centre (ISC) catalogue, we selected seismograms from 380 earthquakes with surface-wave magnitudes (MS) larger than 4.5 and epicentral distances between 25° and 150° that occurred during the time when the stations were operating (Extended Data Fig. 4). The raw seismogram of each event was cut from the time of origin of the earthquake to 12,000 s after it. Before any further analysis, the vertical-component seismograms were down-sampled to one sample per second and instrument responses were corrected. Noise in seismograms at long periods (>50 s) due to ocean swell and associated water-pressure variations, as well as tilt caused by local currents, were removed by correcting the vertical channel using horizontal and pressure channels39,40,41.

This implementation of the HT method recovers frequency-dependent phase and amplitude information via the narrow-band filtering of the broadband cross-correlations between the vertical-component seismogram from a given station and the time-windowed seismograms from all other nearby stations. The phase delays and amplitude information were determined by fitting the narrow-band-filtered cross-correlations with a Gaussian wavelet38. To eliminate the influence of poor-quality records, we estimated the coherence between waveforms from nearby stations for a series of periods from 21 s to 53 s, and only included those measurements with coherence larger than 0.5. For each earthquake and each period, we inverted the phase delays for spatial variations in dynamic phase velocity via the Eikonal equation42. We then further corrected the propagation effect via HT43, producing maps of structure phase velocity with a spacing of 0.2°. This tomographic method returns the azimuthal isotropic phase velocity and the azimuthal anisotropic component at each node simultaneously.

Bayesian Monte Carlo inversion

We combined the ANT and HT results to provide more complete measurements of phase velocity for the vertically polarized S wave (SV wave) velocity inversion (Extended Data Fig. 5). The two sets of dispersion curves were combined in the geographic region that was well resolved by both methods. Phase velocities were interpolated onto a uniform grid of nodes with a spacing of 0.2° before being combined at each node. For phase velocities from ANT (8–25 s), the uncertainties were normalized at each period so that the uncertainty of the best-resolved node is 0.075 km s−1. The uncertainties of the group velocity (8–21 s) were normalized so that the best-resolved node has an uncertainty of 0.188 km s−1, 2.5 times the value for phase velocities44. For phase velocities from HT (21–53 s), the uncertainties were normalized at each period so that the best-resolved node has an uncertainty equal to the standard deviation of velocity differences between HT results and results from a two-plane-wave tomography method45 (Extended Data Fig. 6). We used a linear weighting average method to combine phase velocity measurements and uncertainty estimates for overlapping periods (22–25 s). A running average was then applied to make the resulting dispersion curve smoother. Group velocity results from ANT (8–21 s) were also included for the SV-wave velocity inversion to better fit water thickness and to improve resolution for shallower structure.

We use a Bayesian Monte Carlo algorithm46 to invert the azimuthally averaged SV-wave velocity at each node. This approach allows us to apply prior constraints on crustal thickness and other parameters in a systematic way, to avoid any potential bias of the starting model47 and to derive formal estimates of velocity uncertainty.

The Bayesian Monte Carlo method constructs an a priori distribution of SV-wave velocity models at each node, defined by perturbations relative to the starting model and model constraints. Each model consists of four layers on top of a half-space: (1) water with starting thickness from bathymetry48 that has been smoothed with a Gaussian filter (at a length of 125 km) and an allowed perturbation of ±1.5 km; (2) sediments; (3) crust; and (4) upper mantle from the Moho to a depth of 180 km. The sedimentary layer is described by two parameters: a layer thickness of 0.5 km with an allowed perturbation of ±0.5 km and a constant VSV of 2.0 km s−1 with a perturbation of ±1.0 km s−1. The crust is assumed to have linearly increasing velocity with depth and is described by three parameters: a layer thickness, and VSV at the top and bottom of the layer. For the incoming plate east of the trench, the crustal thickness is allowed to vary by ±1.5 km around the starting value of 6.5 km. The forearc crustal thickness perturbs within 3 km, with starting values from a previous seismic refraction survey31 at the southern edge of the study region ranging from 19 km to 6.5 km. The top and bottom crustal VSV are set at 3.0 km s−1 and 3.2 km s−1, respectively, with a perturbation of ±1.0 km s−1. The upper-mantle VSV is parameterized by a B-spline, which is defined by seven nodes, with a perturbation of ±30% for the first five and of ±20% for the last two. We impose the constraint that the jumps in VSV from the sediment to the crust and from the crust to the mantle are positive.

We also apply a physical dispersion correction with a reference period of 1 s (ref. 49) using a one-dimensional attenuation (Q) model simplified from a seismic attenuation study in the same region50. Compared to the Preliminary reference Earth model, our one-dimensional Q model for the forearc has a high-attenuation layer in the uppermost mantle: QS = 60 from the Moho to a depth of 100 km. For the incoming plate region, the uppermost mantle is set to have a typical lithospheric attenuation: QS = 300 from the Moho to a depth of 100 km.

For each grid node, the best-fitting model is identified and models are accepted if their χ2 misfit is less than 50% higher than that of the best-fitting model44,46. We also exclude models with mantle velocity higher than 4.9 km s−1. The posterior distribution thus provides statistical information on all possible SV-wave velocity models that satisfy the Rayleigh-wave dispersion curves within tolerances depending on data uncertainties. An average model is then calculated from all accepted models and used for plotting and interpretations46,51. Examples of the SV-wave velocity inversion at four representative nodes are shown in Extended Data Fig. 5. The results of the Bayesian Monte Carlo inversion fit the measured group- and phase-velocity dispersion curves well.

Robustness of the thick low-velocity layer

The application of the Bayesian Monte Carlo algorithm46 helps to avoid the potential bias of the starting models and provides better prior constraints on crustal thickness and other parameters. However, it uses a B-spline method to parameterize the upper-mantle VSV, and so may smooth a thin low-velocity layer over a wider depth range. Here we run simulations to test whether the thick low-velocity region observed above and below the surface of the subducting slab can be caused by smoothing of the 6-km-thick lower-velocity crust as a result of the inversion parameterization and the lack of precise depth resolution.

For the target node, we set up a one-dimensional shear-velocity model without a low-velocity serpentinized slab mantle layer, based on our prior knowledge of the geometry of the Mariana subduction zone (Extended Data Fig. 1a), and calculate the synthetic phase and group dispersion curves52. We then apply the Bayesian Monte Carlo inversion with the same parameterizations as in our study to the synthetic phase and group data and obtain a one-dimensional reconstructed shear-velocity structure. Examples for two nodes are shown in Extended Data Fig. 1b, c, and suggest that the thick low-velocity layer observed in our study cannot be caused purely by smearing of the subducting oceanic crust.

Serpentine and water-filled cracks

Previous studies at various convergent margins generally attribute the observed upper-mantle slow velocity anomalies to the presence of serpentine6,53,54,55,56,57. Although the three main serpentine minerals—lizardite, chrysotile and antigorite—have the same water content (about 13 wt%), their physical properties, including seismic velocity17,27 and stability field58,59, are different, owing to the different crystal structure. Lizardite and chrysotile are the more abundant serpentine minerals in hydrated mantle rocks formed at low temperatures and are stable up to about 320 °C at 1 GPa. When the temperature reaches between 320 °C and 390 °C, lizardite is progressively replaced by antigorite at the grain boundaries and in the core of the lizardite meshes26. Antigorite is the main stable serpentine mineral at higher temperature (up to about 620 °C at 1 GPa)26,59,60,61,62. Lizardite and chrysotile have much lower shear velocities (roughly 2.3 km s−1) compared to antigorite (roughly 3.7 km s−1) at 600 MPa (ref. 27). At 600 MPa, the experimental relationship between shear-velocity Vs and serpentine volume fraction (Φ) is Vs = 4.51 − 2.19Φ for lizardite and chrysotile, and Vs = 4.51 − 0.84Φ for antigorite27. For the same reduction in shear velocity, the serpentinization percentage (and thus the water content) estimated assuming an antigorite component will be roughly 2.5 times that assuming a lizardite and chrysotile component27. This feature makes it imperative to decide which serpentine minerals are present before making any further interpretations of velocity reductions. To estimate a lower bound for the serpentinization percentage, we use 4.51 km s−1 as the reference velocity for the unaltered mantle instead of the 4.7 km s−1 that we observe (Fig. 2a–c).

It has been argued23 that the same reduction in velocity can also be caused by water-filled porosity, without involving substantial bulk hydration. This argument presumes a non-fractured isotropic media, which is incompatible with the field observations in the Mariana subduction zone, where numerous normal faults and normal-fault earthquakes have been detected15,16. Instead, a porous media with aligned cracks is the more appropriate assumption63,64. On the other hand, this argument can be applied to the slab only before subduction. When the slab starts to subduct, the confining pressure increases with increasing depth, causing closure of cracks and expulsion of free water within these cracks and/or porosity24. Thus, this hypothesis is not applicable to the velocity reduction observed within the slab at greater depth after subduction.

Estimation of global subduction-zone water flux

We recalculated the global water flux into the subduction zone on the basis of a previous estimate3, by re-evaluating the water content in the slab mantle. Because serpentine minerals are stable up to 620 °C, young and warm subducting plates have less potential to be serpentinized to great depth. According to thermal models for oceanic plates25, only plates older than about 40 Myr have their 600 °C isotherm deeper than about 30 km beneath the seafloor; we therefore set 40 Myr as an age threshold for the subducting plate to be affected by deeper serpentinization. For subduction zones with subducting plates younger than 40 Myr, we take the water-flux estimates from ref. 3. For subduction zones with incoming plates older than 40 Myr, we assume that the slab mantle is partially serpentinized (2 wt% water) to 20 km below the Moho and keep the water volume in the sediment and crust as in ref. 3. This rough estimate suggests that the global subduction-zone water flux should increase to 3.0 × 109 Tg Myr−1.

Anisotropy effect of serpentine along bending faults

The anisotropic effects of serpentine distributed along bending faults could lead to an overestimate of the serpentinization percentage28. We calculated the long-wavelength azimuthal anisotropy produced by evenly distributed serpentine layers65. According to our estimate of the serpentinization percentage (roughly 19 vol%), we assume that pure serpentine layers (450 m thick) were evenly distributed within isotropic peridotite with a spacing of 2 km. We show results for two layering geometries: vertical layering and 45° dipping layering (Extended Data Fig. 2). Serpentine layers were set to have the following properties: isotropic compression-wave velocity VP = 5.1 km s−1, isotropic shear-wave velocity VS = 2.32 km s−1 and density ρ = 2.52 g cm−3. Peridotite layers were set to have the following properties: VP = 8.1 km s−1, VS = 4.51 km s−1 and ρ = 3.32 g cm−3. The azimuthally averaged quasi-SV-wave velocity for the case of 45° dipping layering is 4.08 km s−1, very close to the Voigt average SV-wave velocity (4.1 km s−1) that we use to estimate the serpentinization percentage directly. This result suggests that the anisotropy effect of serpentine distributed along bending faults may be less important when estimating serpentinization percentage28, especially when the bending faults are dipping.

Data availability

Raw seismic data are available at the Data Management Center of the Incorporated Research Institutions for Seismology (http://www.iris.edu/dms/nodes/dmc) under network IDs MI and XF. Network and station information can be found at the IRIS website (http://www.ds.iris.edu/mda).

Additional information

Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.


  1. 1.

    Hacker, B. R. H2O subduction beyond arcs. Geochem. Geophys. Geosyst. 9, Q03001 (2008).

  2. 2.

    Rüpke, L. H., Morgan, J. P., Hort, M. & Connolly, J. A. D. Serpentine and the subduction zone water cycle. Earth Planet. Sci. Lett. 223, 17–34 (2004).

  3. 3.

    van Keken, P. E., Hacker, B. R., Syracuse, E. M. & Abers, G. A. Subduction factory: 4. Depth-dependent flux of H2O from subducting slabs worldwide. J. Geophys. Res. 116, B01401 (2011).

  4. 4.

    Fujie, G. et al. Systematic changes in the incoming plate structure at the Kuril trench. Geophys. Res. Lett. 40, 88–93 (2013).

  5. 5.

    Ranero, C. R., Phipps Morgan, J., McIntosh, K. & Reichert, C. Bending-related faulting and mantle serpentinization at the Middle America trench. Nature 425, 367–373 (2003).

  6. 6.

    Shillington, D. J. et al. Link between plate fabric, hydration and subduction zone seismicity in Alaska. Nat. Geosci. 8, 961–964 (2015).

  7. 7.

    Van Avendonk, H. J. A., Holbrook, W. S., Lizarralde, D. & Denyer, P. Structure and serpentinization of the subducting Cocos plate offshore Nicaragua and Costa Rica. Geochem. Geophys. Geosyst. 12, Q06009 (2011).

  8. 8.

    Parai, R. & Mukhopadhyay, S. How large is the subducted water flux? New constraints on mantle regassing rates. Earth Planet. Sci. Lett. 317–318, 396–406 (2012).

  9. 9.

    Grove, T. L., Till, C. B. & Krawczynski, M. J. The role of H2O in subduction zone magmatism. Annu. Rev. Earth Planet. Sci. 40, 413–439 (2012).

  10. 10.

    Fryer, P. Serpentinite mud volcanism: observations, processes, and implications. Annu. Rev. Mar. Sci. 4, 345–373 (2012).

  11. 11.

    Barklage, M. et al. P and S velocity tomography of the Mariana subduction system from a combined land-sea seismic deployment. Geochem. Geophys. Geosyst. 16, 681–704 (2015).

  12. 12.

    Kelley, K. A. et al. Mantle melting as a function of water content beneath the Mariana arc. J. Petrol. 51, 1711–1738 (2010).

  13. 13.

    Shaw, A. M., Hauri, E. H., Fischer, T. P., Hilton, D. R. & Kelley, K. A. Hydrogen isotopes in Mariana arc melt inclusions: implications for subduction dehydration and the deep-Earth water cycle. Earth Planet. Sci. Lett. 275, 138–145 (2008).

  14. 14.

    Müller, R. D., Sdrolias, M., Gaina, C. & Roest, W. R. Age, spreading rates, and spreading asymmetry of the world’s ocean crust. Geochem. Geophys. Geosyst. 9, Q04006 (2008).

  15. 15.

    Emry, E. L., Wiens, D. A. & Garcia-Castellanos, D. Faulting within the Pacific plate at the Mariana trench: implications for plate interface coupling and subduction of hydrous minerals. J. Geophys. Res. 119, 3076–3095 (2014).

  16. 16.

    Oakley, A. J., Taylor, B. & Moore, G. F. Pacific plate subduction beneath the central Mariana and Izu-Bonin fore arcs: new insights from an old margin. Geochem. Geophys. Geosyst. 9, Q06003 (2008).

  17. 17.

    Christensen, N. I. Serpentinites, peridotites, and seismology. Int. Geol. Rev. 46, 795–816 (2004).

  18. 18.

    Nishimura, C. E. & Forsyth, D. W. The anisotropic structure of the upper mantle in the Pacific. Geophys. J. Int. 96, 203–229 (1989).

  19. 19.

    Feng, H. S.-H. Seismic Constraints on the Processes and Consequences of Secondary Igneous Evolution of Pacific Oceanic Lithosphere. PhD thesis, Massachusetts Institute of Technology and Woods Hole Oceanographic Institution (2016).

  20. 20.

    Faccenda, M., Gerya, T. V., Mancktelow, N. S. & Moresi, L. Fluid flow during slab unbending and dehydration: implications for intermediate-depth seismicity, slab weakening and deep water recycling. Geochem. Geophys. Geosyst. 13, Q01010 (2012).

  21. 21.

    Reynard, B. Serpentine in active subduction zones. Lithos 178, 171–185 (2013).

  22. 22.

    Nakatani, T. & Nakamura, M. Experimental constraints on the serpentinization rate of fore-arc peridotites: implications for the upwelling condition of the slab-derived fluid. Geochem. Geophys. Geosyst. 17, 3393–3419 (2016).

  23. 23.

    Korenaga, J. On the extent of mantle hydration caused by plate bending. Earth Planet. Sci. Lett. 457, 1–9 (2017).

  24. 24.

    David, C., Wong, T.-F., Zhu, W. & Zhang, J. Laboratory measurement of compaction-induced permeability change in porous rocks: implications for the generation and maintenance of pore pressure excess in the crust. Pure Appl. Geophys. 143, 425–456 (1994).

  25. 25.

    Stein, C. A. & Stein, S. A model for the global variation in oceanic depth and heat flow with lithospheric age. Nature 359, 123–129 (1992).

  26. 26.

    Schwartz, S. et al. Pressure–temperature estimates of the lizardite/antigorite transition in high pressure serpentinites. Lithos 178, 197–210 (2013).

  27. 27.

    Ji, S. et al. Seismic velocities, anisotropy, and shear-wave splitting of antigorite serpentinites and tectonic implications for subduction zones. J. Geophys. Res. 118, 1015–1037 (2013).

  28. 28.

    Miller, N. C. & Lizarralde, D. Finite-frequency wave propagation through outer rise fault zones and seismic measurements of upper mantle hydration. Geophys. Res. Lett. 43, 7982–7990 (2016).

  29. 29.

    Emry, E. L. & Wiens, D. A. Incoming plate faulting in the northern and western Pacific and implications for subduction zone water budgets. Earth Planet. Sci. Lett. 414, 176–186 (2015).

  30. 30.

    Nakanishi, M., Tamaki, K. & Kobayashi, K. Magnetic anomaly lineations from Late Jurassic to Early Cretaceous in the west-central Pacific Ocean. Geophys. J. Int. 109, 701–719 (1992).

  31. 31.

    Takahashi, N., Kodaira, S., Tatsumi, Y., Kaneda, Y. & Suyehiro, K. Structure and growth of the Izu-Bonin-Mariana arc crust: 1. Seismic constraint on crust and mantle structure of the Mariana arc–back-arc system. J. Geophys. Res. 113, B01104 (2008).

  32. 32.

    Hayes, G. P., Wald, D. J. & Johnson, R. L. Slab1.0: a three-dimensional model of global subduction zone geometries. J. Geophys. Res. 117, B01302 (2012).

  33. 33.

    Emry, E. L., Wiens, D. A., Shiobara, H. & Sugioka, H. Seismogenic characteristics of the northern Mariana shallow thrust zone from local array data. Geochem. Geophys. Geosyst. 12, Q12008 (2011).

  34. 34.

    Bensen, G. D. et al. Processing seismic ambient noise data to obtain reliable broad-band surface wave dispersion measurements. Geophys. J. Int. 169, 1239–1260 (2007).

  35. 35.

    Lin, F.-C., Moschetti, M. P. & Ritzwoller, M. H. Surface wave tomography of the western United States from ambient seismic noise: Rayleigh and Love wave phase velocity maps. Geophys. J. Int. 173, 281–298 (2008).

  36. 36.

    Levshin, A. L. & Ritzwoller, M. H. Automated detection, extraction, and measurement of regional surface waves. Pure Appl. Geophys. 158, 1531–1545 (2001).

  37. 37.

    Barmin, M. P., Ritzwoller, M. H. & Levshin, A. L. A fast and reliable method for surface wave tomography. Pure Appl. Geophys. 158, 1351–1375 (2001).

  38. 38.

    Jin, G. & Gaherty, J. B. Surface wave phase-velocity tomography based on multichannel cross-correlation. Geophys. J. Int. 201, 1383–1398 (2015).

  39. 39.

    Bell, S. W., Forsyth, D. W. & Ruan, Y. Removing noise from the vertical component records of ocean-bottom seismometers: results from year one of the Cascadia Initiative. Bull. Seismol. Soc. Am. 105, 300–313 (2015).

  40. 40.

    Crawford, W. C. & Webb, S. Identifying and removing tilt noise from low-frequency (<0.1 Hz) seafloor vertical seismic data. Bull. Seismol. Soc. Am. 90, 952–963 (2000).

  41. 41.

    Webb, S. C. & Crawford, W. C. Long-period seafloor seismology and deformation under ocean waves. Bull. Seismol. Soc. Am. 89, 1535–1542 (1999).

  42. 42.

    Lin, F.-C., Ritzwoller, M. H. & Snieder, R. Eikonal tomography: surface wave tomography by phase front tracking across a regional broad-band seismic array. Geophys. J. Int. 177, 1091–1110 (2009).

  43. 43.

    Lin, F.-C. & Ritzwoller, M. H. Helmholtz surface wave tomography for isotropic and azimuthally anisotropic structure. Geophys. J. Int. 186, 1104–1120 (2011).

  44. 44.

    Shen, W. et al. A seismic reference model for the crust and uppermost mantle beneath China from surface wave dispersion. Geophys. J. Int. 206, 954–979 (2016).

  45. 45.

    Yang, Y. & Forsyth, D. W. Regional tomographic inversion of the amplitude and phase of Rayleigh waves with 2-D sensitivity kernels. Geophys. J. Int. 166, 1148–1160 (2006).

  46. 46.

    Shen, W., Ritzwoller, M. H., Schulte-Pelkum, V. & Lin, F.-C. Joint inversion of surface wave dispersion and receiver functions: a Bayesian Monte-Carlo approach. Geophys. J. Int. 192, 807–836 (2013).

  47. 47.

    Wei, S. S. et al. Seismic evidence of effects of water on melt transport in the Lau back-arc mantle. Nature 518, 395–398 (2015).

  48. 48.

    Lindquist, K. G., Engle, K., Stahlke, D. & Price, E. Global topography and bathymetry grid improves research efforts. Eos 85, 186 (2004).

  49. 49.

    Kanamori, H. & Anderson, D. L. Importance of physical dispersion in surface wave and free oscillation problems: review. Rev. Geophys. 15, 105–112 (1977).

  50. 50.

    Pozgay, S. H., Wiens, D. A., Conder, J. A., Shiobara, H. & Sugioka, H. Seismic attenuation tomography of the Mariana subduction system: implications for thermal structure, volatile distribution, and slow spreading dynamics. Geochem. Geophys. Geosyst. 10, Q04X05 (2009).

  51. 51.

    Wei, S. S. et al. Upper mantle structure of the Tonga-Lau-Fiji region from Rayleigh wave tomography. Geochem. Geophys. Geosyst. 17, 4705–4724 (2016).

  52. 52.

    Herrmann, R. B. Computer programs in seismology: an evolving tool for instruction and research. Seismol. Res. Lett. 84, 1081–1088 (2013).

  53. 53.

    Contreras-Reyes, E., Grevemeyer, I., Flueh, E. R., Scherwath, M. & Heesemann, M. Alteration of the subducting oceanic lithosphere at the southern central Chile trench-outer rise. Geochem. Geophys. Geosyst. 8, Q07003 (2007).

  54. 54.

    Contreras-Reyes, E. et al. Deep seismic structure of the Tonga subduction zone: implications for mantle hydration, tectonic erosion, and arc magmatism. J. Geophys. Res. 116, B10103 (2011).

  55. 55.

    DeShon, H. R. & Schwartz, S. Y. Evidence for serpentinization of the forearc mantle wedge along the Nicoya Peninsula, Costa Rica. Geophys. Res. Lett. 31, L21611 (2004).

  56. 56.

    Garth, T. & Rietbrock, A. Constraining the hydration of the subducting Nazca plate beneath Northern Chile using subduction zone guided waves. Earth Planet. Sci. Lett. 474, 237–247 (2017).

  57. 57.

    Savage, B. Seismic constraints on the water flux delivered to the deep Earth by subduction. Geology 40, 235–238 (2012).

  58. 58.

    Evans, B. W. The serpentinite multisystem revisited: chrysotile is metastable. Int. Geol. Rev. 46, 479–506 (2004).

  59. 59.

    Guillot, S., Schwartz, S., Reynard, B., Agard, P. & Prigent, C. Tectonic significance of serpentinites. Tectonophysics 646, 1–19 (2015).

  60. 60.

    Perrillat, J.-P. et al. Kinetics of antigorite dehydration: a real-time X-ray diffraction study. Earth Planet. Sci. Lett. 236, 899–913 (2005).

  61. 61.

    Ulmer, P. & Trommsdorff, V. Serpentine stability to mantle depths and subduction-related magmatism. Science 268, 858–861 (1995).

  62. 62.

    Wunder, B. & Schreyer, W. Antigorite: high-pressure stability in the system MgO-SiO2-H2O (MSH). Lithos 41, 213–227 (1997).

  63. 63.

    Gurevich, B. Elastic properties of saturated porous rocks with aligned fractures. J. Appl. Geophys. 54, 203–218 (2003).

  64. 64.

    Hudson, J. A., Liu, E. & Crampin, S. The mechanical properties of materials with interconnected cracks and pores. Geophys. J. Int. 124, 105–112 (1996).

  65. 65.

    Backus, G. E. Long-wave elastic anisotropy produced by horizontal layering. J. Geophys. Res. 67, 4427–4440 (1962).

Download references


We thank P. J. Shore, H. Jian and the captains, crew and science parties of the RVs R. Revelle and Melville for data collection; S. Wei and M. Pratt for helping with data processing; R. Parai and M. J. Krawczynski for discussions; and X. Wang for support. IRIS PASSCAL and OBSIP provided land-based seismic instrumentation and ocean-bottom seismographs, respectively. This work was supported by the GeoPRISMS Program under NSF grant OCE-0841074 (D.A.W.).

Reviewer information

Nature thanks C. Rodríguez Ranero & D. Shillington for their contribution to the peer review of this work.

Author information

C.C. and M.E., advised by D.A.W., analysed the seismic data. W.S. developed and modified the Monte Carlo inversion code. C.C. and D.A.W. took the lead in writing the manuscript, and all authors discussed the results and edited the manuscript.

Competing interests

The authors declare no competing interests.

Correspondence to Chen Cai.

Extended data figures and tables

  1. Extended Data Fig. 1 Robustness test of the low-velocity zone.

    a, The assumed geometry of the subduction zone according to our prior knowledge. b, c, Simulation results for nodes 80 km (b) and 110 km (c) landward from the trench. The black dashed lines are the input one-dimensional models; blue dashed and solid lines are the best-fitting and average models from the Monte Carlo inversion of the synthetic dispersion curves, respectively; red dashed and solid lines are the best-fitting and average models from the Monte Carlo inversion of the real data.

  2. Extended Data Fig. 2 Azimuthal anisotropy from evenly distributed serpentine layers (of thickness 450 m and with a spacing of 2 km).

    a, Result for vertical layering. b, Result for 45° dipping layering. Numbers in the parenthesis are the mean velocity for quasi-P, quasi-SV or quasi-SH. The incidence angle is defined relative to the strike of the layer: 0° is parallel and 90° is normal to the strike.

  3. Extended Data Fig. 3 Maps of azimuthally averaged group and phase velocity.

    a, b, Group velocity (colour scale) at periods of 10 s (a) and 21 s (b) inverted by ANT. c, d, Phase velocity (colour scale) at periods of 10 s (c) and 21 s (d) from ANT. e, f, Phase velocity (colour scale) for periods of 25 s (e) and 40 s (f) inverted by HT. 3-km, 4-km and 5-km bathymetry contours are shown as thin grey lines. The trench axis and serpentine seamounts are shown as in Fig. 1a.

  4. Extended Data Fig. 4 Earthquakes used in this study.

    Blue dots represent ISC earthquake locations. The red star shows the location of the Mariana trench.

  5. Extended Data Fig. 5 Examples of Monte Carlo inversion and phase-velocity sensitivity kernel.

    ad, The joint Rayleigh phase and group dispersion data (error bars, one standard deviation) and computed phase (red solid lines) and group (blue solid lines) dispersion curves from the Bayesian Monte Carlo averaged model, for four locations: a, inner forearc; b, outer forearc; c, trench high; d, Pacific plate. eh, Shear-velocity model from the Bayesian Monte Carlo inversion for the four example locations. i, Phase-velocity sensitivity kernels at example periods, calculated using the average velocity model in g.

  6. Extended Data Fig. 6 Comparison between Rayleigh-wave isotropic phase velocities determined from teleseismic tomography using HT and a two-plane-wave method.

    a, At 27 s. b, At 36 s.

Rights and permissions

To obtain permission to re-use content from this article visit RightsLink.

About this article

Further reading


By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.