Continuous nonthermal slab gap formed by progressive tearing beneath Northeast Asia

Abstract

Slab gaps in subducted plates are widely hypothesized to channel hot mantle upwellings, yet their formation and thermal state remain enigmatic. Here we image a continuous ~1000 km-long arcuate slab gap in the mantle transition zone beneath Northeast Asia with no evidence of anomalous heat. By integrating dense-array teleseismic tomography, 3-D waveform modeling, and a joint inversion of seismic velocities with transition-zone thickness, we constrain the fine-scale geometry and thermal structure of the gap. Sharp lateral boundaries are consistent with a mechanical opening within the stagnant slab, and gap temperatures are comparable to ambient upper‑mantle values and >200 °C lower than plume‑like temperatures. The planform of the gap aligns with the path of a migrating triple junction, suggesting formation through progressive slab tearing during plate-boundary reorganization. These observations demonstrate that a slab gap can persist without hot upwelling, consistent with the diffuse intraplate volcanism above the region.

This study images a gap nearly 1,000 km long that cuts through a subducted slab stalled in the mantle transition zone beneath Northeast Asia. Seismic data suggest progressive mechanical tearing rather than a hot mantle window feeding volcanism.

Introduction

Subduction of oceanic lithosphere transfers surface materials into the deep mantle, drives mantle convection, and governs the long-term evolution of plate motions¹. A significant portion of subducting slabs temporarily stagnates around the mantle transition zone (MTZ)², modulating mantle circulation and geochemical cycles³. The geometry of these stagnant slabs, particularly edges formed by gaps, has a critical impact on both mantle convection and surface tectonics⁴. Such slab-gap structures are often invoked as conduits for mantle upwelling that fuels volcanism, yet whether they represent passive mechanical breaches or active thermal windows remains unresolved. Distinguishing between these scenarios is critical, as each implies different patterns of mass and heat transfer across and above the MTZ and different origins of volcanism near and far from the plate boundaries. Slab morphology also provides important constraints on reconstructions of subduction and plate reorganizations, since sites of slab subduction and stagnation record past configurations of convergent plate boundaries⁵.

Prolonged subduction of the Pacific plate beneath the Eurasian and Philippine Sea plates (Fig. 1) has produced an extensive stagnant slab, imaged as a high-velocity mantle anomaly by seismic tomography⁶. Previous studies have often suggested low-velocity MTZ features extending toward the surface through slab gaps, interpreted as hot mantle upwelling feeding intraplate volcanism^7,8. In contrast, other models have invoked nonthermal processes, highlighting mechanical deformation (e.g., tearing or thinning) and rollback-induced flow, as drivers of mantle circulation^3,9. However, tomographic models have differed greatly on the inferred location, size, geometry, and even existence of the slab gap in this region (Fig. 2a; Supplementary Text 1). These discrepancies arise partly because resolution is limited and strongly influenced by inversion regularization. Furthermore, distinguishing the slab gap from a true thermal upwelling is complicated by uncertainties in MTZ structures, which are sensitive to temperature and composition. These issues have prevented a consistent interpretation of slab morphology and associated mantle conditions beneath Northeast Asia.

Fig. 1. Regional tectonic setting of the northwestern Pacific.

Convergent plate boundaries are shown as red saw-toothed lines, and Cenozoic intraplate volcanoes as red triangles. Colored contours show the depth of the subducting Pacific and Philippine Sea slabs at 50-km intervals¹⁰. The seafloor age is mapped across the oceanic plates, and red arrows indicate plate velocities relative to the Eurasian Plate. Abbreviations: BS Bering Sea ECSb, East China Sea basin, ES East Sea (Sea of Japan), KB Kuril Basin, KP Korean Peninsula, OS Okhotsk Sea, SCSb South China Sea basin, YSb Yellow Sea basin.

Fig. 2. Regional slab structure, seismicity, and seismic stations and events for tomographic imaging.

a Map of the study region showing previously proposed slab anomalies (S1–S6), deep seismicity, and slab-depth contours at 60-km intervals¹⁰. Dots are seismic events (M > 3.0) color-coded by depth. Proposed slab-related features, including stagnant-slab gaps and the tearing of the subducting slabs, comprise S1–S5 in previous studies (see Supplementary Text 1 for references) and S6 from this study (Type 1 in Fig. 3). The solid line is constrained by our tomographic results, while the dashed lines indicate possible extensions of the gap inferred from broader tomographic trends and previous studies suggesting slab tears in nearby regions^6,11. b Seismic stations used in this study (symbols indicate different networks) and teleseismic event distributions for P-wave (red dots) and S-wave (blue dots) tomography. Dashed circles mark 30° distance increments. Seismic networks: F-net/Hi-net, National Research Institute for Earth Science and Disaster Resilience; GSN, Global Seismic Network; JMA, Japan Meteorological Agency; KIGAM, Korea Institute of Geoscience and Mineral Resources; KINS, Korea Institute of Nuclear Safety; KMA, Korea Meteorological Administration; KHNP, Korea Hydro and Nuclear Power.

In this study, we present fine-scale slab geometry and thermal state by integrating data from dense regional seismic arrays with multi-scale P- and S-wave tomography and forward waveform modeling. The waveform simulations are sensitive to sharp lateral velocity gradients and thus provide an independent test of slab-boundary continuity. We then estimate MTZ potential temperatures (T_p) by jointly analyzing seismic velocities with MTZ thickness (dMTZ), using a conservative framework that includes corrections for tomographic amplitude underestimation and uncertainty quantification across the full parameter space. This approach enables us to robustly quantify the thermal structure around the slab in the MTZ. This comprehensive approach allows us to link present-day slab morphology^10,11 to ~30 Myr of subduction history, to determine whether the observed slab gap results primarily from mechanical tearing of the slab or from thermal upwelling (or both), and to reassess its role in mantle dynamics.

Results

A continuous low-velocity channel in the MTZ with sharp boundaries

Using teleseismic travel-time tomography (see “Methods”), we obtained a seismic velocity model of the MTZ beneath the eastern margin of Northeast Asia that resolves features in greater detail than previous regional or global models (Supplementary Figs. 3 and 4; Supplementary Text 3). The model is dominated by two distinctive domains: a laterally continuous low-velocity channel (Type 1) and surrounding high-velocity regions (Type 2) (Fig. 3). Type 1 is characterized by velocity perturbations of dlnVp <−1% and dlnVs <−2%, along with an elevated dln(Vp/Vs) of approximately +0.5 to +1%. The low-velocity channel extends from the East China Sea, across the southern Korean Peninsula, and then curves eastward into the East Sea (Sea of Japan), spanning roughly 1000 km. Type 2 consists of high-velocity regions (dlnVp > +2.5% and dlnVs > +4%) with low Vp/Vs, which corresponds to the cold, stagnant Pacific slab. These high-velocity domains flank the low-velocity channel, forming a sharp interface between them. The Vs contrast across this interface exceeds 4% over a lateral distance of ~200 km, and this strong gradient persists throughout the MTZ (Supplementary Figs. 6–13). Extensive resolution tests in both Vp and Vs confirm that the location, sharp boundary geometry, ~1000 km continuity, and relative amplitude of the characteristic velocity anomalies at MTZ depths are robustly recovered, with negligible downward smearing from shallow structure, and further show that the slab-gap anomaly is required by the observed data rather than produced by inversion artifacts (see “Methods”; Supplementary Figs. 14–23).

Fig. 3. Seismic velocity structures of the mantle transition zone (MTZ).

a–c Horizontal sections at 510–540 km depth showing dlnVp (a), dlnVs (b), and dln(Vp/Vs) (c). d Map categorizing MTZ structure, highlighting a sharp boundary (black dashed line) that separates the low-velocity slab-gap domain (Type 1) from the surrounding higher-velocity mantle (Type 2). Black dashed outlines mark the Type 1 regions where mantle temperatures within the slab gap were calculated (see Fig. 5c). Brown contours and green dots denote slab-depth contours and deep seismicity, respectively. Dashed straight lines in (a) are locations of vertical sections shown in Supplementary Fig. 7.

The sharpness of this boundary is further corroborated by abrupt, systematic variations in teleseismic travel-time residuals observed across the region (Supplementary Fig. 24). As the backazimuth of incoming waves sweeps across the inferred Type 1 boundary, the residuals shift abruptly by up to ~1.0 s for P-waves and ~2.5 s for S-waves. Forward calculations reproduce these azimuth-dependent residual patterns only when a sharp lateral velocity contrast is included in the MTZ.

3-D waveform modeling to confirm the properties of the MTZ channel

To independently test the existence and sharpness of the imaged velocity anomalies, we analyzed teleseismic S-wave (SH) amplitudes and waveforms. The SH-wave amplitudes show roughly a twofold peak-to-peak difference between wave paths sampling the Type 1 channel versus those sampling the Type 2 slab (Δlog10 amplitude ≈ 0.3). Moreover, 3-D simulations demonstrate that the observed back-azimuthal migration of high-amplitude lobes is replicated only when the sharp MTZ anomalies from our model are included (Fig. 4; see “Methods”).

Fig. 4. Synthetic waveform simulations based on three-dimensional (3-D) tomography models.

a Map of three teleseismic events (see Supplementary Table 1) and their great-circle ray paths to the seismic array, plotted on the Vs model at 560 km depth. b Transverse-component S waves (SH) records from Event 1, with observed waveforms (black) compared to synthetics (red). The inset highlights examples of a defocused waveform (I) vs. a focused waveform (II) relative to a 1-D reference model (green). c S-wave amplitude variation map for Event 1, with observed amplitudes (circles) overlain on predicted amplitudes (background color). Wavefield focusing through the Type 1 low-velocity slab gap produces amplified amplitude lobes (blue dashed circles), whereas defocusing by the high-velocity Type 2 slab yields reduced amplitudes (red dashed circles). Pink lines mark the MTZ boundary between lower-Vs domain and high Vs surroundings (from Fig. 3b). Thick dashed lines indicate ray trajectories. d Predicted S-wave amplitude map for a control model without the MTZ slab gap (i.e., the 3-D velocity structure is applied only above 400 km depth). e–g Same as (b–d) for Event 2. h–j Same as (b–d) for Event 3.

Relative to synthetics from a 1-D reference model, the observed SH waveforms that traverse the Type 1 low-velocity channel exhibit narrower pulse widths and slightly delayed onset times (Fig. 4b, e, h). When MTZ velocity anomalies are omitted from the model (i.e., the 3-D structure is applied only above 400 km depth), the simulations fail to reproduce the observed amplitude variations. The predicted amplitude perturbations are weak (Δlog10 amplitude <0.1), and the positively and negatively amplified regions remain nearly stationary in and around the southern Korean Peninsula across event back-azimuths, varying only modestly in the degree of amplification (Fig. 4d, g, j). This stationary pattern is incompatible with the systematic back-azimuth-dependent migration of the amplified lobes seen in the observations, indicating that the sharp MTZ anomalies are required to reproduce the observed amplitude pattern. Complementary synthetic resolution tests with input anomalies confined above 400 km further confirm that vertical leakage of upper-mantle structure into the MTZ is quantitatively limited (see “Methods”), so its contribution to the observed amplitude pattern is correspondingly small.

These observations agree with theoretical expectations for wavefield focusing and defocusing caused by strong lateral velocity gradients¹². In essence, seismic waves propagating through a low-velocity body (Type 1) are focused, whereas propagation through a high-velocity body (Type 2) defocuses the waves. Consequently, waveforms traversing the Type 1 channel have larger amplitudes than those predicted by a 1-D reference model, while waveforms crossing the Type 2 region have systematically smaller amplitudes than predicted (Fig. 4b,e, h).

Not anomalously hot MTZ channel: joint analysis with MTZ thicknesses

We next estimated the MTZ temperature field to test whether the low-velocity slab channel is a thermal anomaly or instead reflects a mechanically torn slab gap. The temperature structure was inferred by a joint analysis of our Vp and Vs models together with dMTZ variations (see “Methods”), which is sensitive to temperature and thus provides an independent constraint on mantle thermal structure¹³. The dMTZ exhibits lateral variations exceeding 50 km across the study area (Fig. 5a). A thinned MTZ (5–15 km thinner than the reference dMTZ of ~255 km) underlies the southern Korean Peninsula and extends westward into the East China Sea, spatially coincident with the Type 1 low-velocity channel. Surrounding this thin corridor are thickened MTZ regions (10–35 km greater than the reference) beneath the Yellow Sea, the East Sea, and southwest Japan, consistent with colder slab material in those areas (Type 2). MTZ thinning shows a close spatial correspondence with low velocities and thickening with high velocities, with overlaid vertical cross-sections and along-profile comparisons confirming the co-location of these features (Fig. 6; Supplementary Figs. 31 and 32). The observed polarity, with low velocities associated with a thinned MTZ, is diagnostic of a temperature-dominated control, because non-thermal factors would produce different trends and hydration in particular would yield the opposite relationship¹⁴. Applying 3-D velocity corrections to our composite dMTZ map (see “Methods”) sharpens the boundaries between these thin and thick MTZ domains (Fig. 5a and Supplementary Fig. 33). The characteristic dMTZ pattern, therefore temperature, is robust with respect to the choice of reference model (Supplementary Text 5): using a simple 1-D model yields smaller-amplitude differences (<30 km), whereas using a 3-D model introduces a systematic bias of up to 20 km in dMTZ. The 3-D velocity correction to dMTZ slightly raises the estimated T_p for the Type 1 anomaly; this is a more conservative approach in the context of interpreting the Type 1 anomaly as a nonthermal structure.

Fig. 5. MTZ thickness and thermal structure.

a Map of MTZ thickness (dMTZ) from receiver-function studies (see “Methods”). Unsampled or low-resolution regions are masked out (Supplementary Fig. 26). b MTZ mantle temperature anomalies relative to a 1350 °C adiabat, derived from the seismic velocity models (Fig. 3) and the dMTZ map. Brown contours and green dots indicate slab-depth contours and deep seismicity (>400 km), respectively. c Violin plot comparison of mantle potential temperature (T_p). The first six violins (leftmost) show T_p anomalies for the slab gap (Type 1) and stagnant slab (Type 2) in Northeast Asia, including Monte Carlo–based ensemble estimates (labeled “Ensemble”) that incorporate uncertainties across all parameters affecting MTZ temperature, together with results derived from amplitude-scaled and original (unscaled) tomographic models. Immediately to the right, a yellow violin depicts T_p estimates for the source mantle of Cenozoic intraplate volcanoes in East Asia¹⁸ (Fig. 1), followed by four violins for hotspot provinces (purple, green, cyan) versus mid-ocean ridges (orange)¹⁷. Vertical hollow bars on the far right denote MTZ T_p anomalies from other regions (Hainan plume³⁹, Middle East⁴⁰, and Pacific Coast Range³⁸). The width of each violin indicates probability density; horizontal lines mark the maximum, mean, and minimum values; white circles denote medians; and black/white vertical bars span the 68, 95, and 99.7% percentile ranges. Blue and green shaded bands highlight T_p ranges that are insufficient to drive significant mantle upwelling or to produce active plume-fed ascent¹⁷.

Fig. 6. Joint analysis of MTZ seismic velocity and thickness.

a Map view of S-wave velocity perturbations (dlnVs) and MTZ thickness (dMTZ) within the overlapping sampling region. dMTZ is corrected for upper-mantle 3-D velocity structure (Methods). Red symbols denote low dlnVs (<0%) with thin dMTZ (<255 km); blue symbols denote high dlnVs (>0%) with thick dMTZ (>255 km); gray symbols indicate other cases. The reference thickness (~255 km) corresponds to a pyrolitic mantle at T_p ~ 1300 °C. Black lines outline the inferred slab-gap region, and the three profiles A–A′, B–B′, C–C′ for panels c and d are marked. b Scatter plot of dlnVs versus dMTZ. Background color indicates data density, error bars show uncertainties, and red lines mark the reference values. The Pearson correlation coefficient is shown in the lower right (r = 0.48, p < 0.001). c Vertical cross-sections of the S-wave tomographic model along A–A′, B–B′, and C–C′ overlain with the 410- and 660-km discontinuities from receiver-function studies (purple lines). Gray dashed lines mark reference depths, black ellipses highlight the slab-gap region, and gray shading indicates reduced resolution. d Along-profile comparison of dMTZ (purple) and MTZ-averaged dlnVs (black, or gray where dMTZ is unavailable) for the three profiles. Yellow shading marks slab-gap intervals with dlnVs <0% and thinned dMTZ. Corresponding results for P-wave velocity are provided in Supplementary Figs. 31 and 32.

The inferred MTZ temperatures beneath our study area range from −225 °C to +75 °C relative to a reference T_p of 1350 °C, consistent with normal MORB-source mantle^15–17 (Fig. 5b, c). The high-velocity flanks of the channel (the stagnant Pacific slab) are on average ~80 °C colder than this reference, with local minima near −225 °C (Fig. 5c), the coldest anomaly being south of Japan (near the deeply subducted Izu-Bonin slab segment). By contrast, the Type 1 channel is slightly colder than ambient mantle (mean and median ≈ –25 °C), with modest local positive anomalies at its center (up to +75 °C). These values match petrologically inferred T_p estimates for Cenozoic intraplate volcanism in the region¹⁸ and indicate that the mantle in the gap is not particularly hot in absolute terms (Fig. 5c). Notably, the regional mean T_p (~1300 °C) is comparable to MTZ temperature estimates near other subduction zones^15,16 and is lower than the global average beneath mid-ocean ridges¹⁷ (Fig. 5).

To test the extreme limits, we additionally applied a scaling correction to the velocity anomalies to estimate the maximum possible T_p within the Type 1 region. In this case, we directly amplified the tomographic anomalies by factors derived from fitting the back-azimuthal travel-time residuals to account for potential underestimation of true amplitudes by regularized tomography (see “Methods”). To robustly quantify uncertainties, we further sampled the physically plausible ranges of all temperature-controlling parameters through a Monte Carlo ensemble of ~10⁵ realizations, yielding the full distribution of inferred T_p (see “Methods”).

After the scaling correction, the first-order temperature pattern and regional mean T_p remain essentially unchanged (Supplementary Fig. 35), although local extrema become ~1.1–1.6 times more pronounced. Importantly, even with this amplitude boosting, the mean T_p of the Type 1 channel increases only to ~1350 °C, with an extreme of 1475 °C (Fig. 5c). The ensemble yields approximately normal slab-gap temperature distributions with a small standard deviation of ~70 °C. Even accounting for the full extreme range, the distribution overlaps predominantly with those of cold hotspots and mid-ocean ridges rather than with warm or hot hotspot provinces (Fig. 5c). This is difficult to reconcile with thermally driven upwellings¹⁷ and effectively rules out an active plume conduit in the upper mantle (Fig. 5c).

Discussion

A nonthermal continuous gap in the stagnant Pacific slab

Our tomographic model reveals a laterally continuous Type 1 channel, characterized by slightly reduced seismic velocities, a modestly elevated Vp/Vs, a slightly thinned dMTZ of ~240 km, and a sharp boundary (Figs. 3 and 5). This channel is flanked by Type 2 domains of high velocities and low Vp/Vs. The temperature estimates indicate that the Type 1 region has essentially no excess thermal anomaly on average (1325 °C–1350 °C), with even the maximum values remaining modest (1430 °C–1475 °C), whereas Type 2 is significantly colder (Fig. 5). Such cold MTZ anomalies (Type 2) are commonly linked to subducted oceanic lithosphere or detached fragments of continental lithosphere⁶. We interpret these high-velocity, cold anomalies primarily as the stagnant Pacific slab (and possibly with some portions of the Philippine Sea plate). The measured temperature values and the contrasts between Types 1 and 2 match expectations for slab versus ambient upper mantle based on seismic observations¹⁹ and numerical models²⁰. This identification is also supported by geodynamic models that relate the young subducting age of the Pacific slab (~10–20 Ma) and its buoyancy to slab stagnation in the MTZ²¹, and slab geometries inferred from plate kinematics²². These observations consistently indicate that the intervening Type 1 domain is a structurally distinct gap within the otherwise cold stagnant slab.

Notably, our results suggest that previously reported slab tearing and gaps in this region (features S3 to S5 and possibly the southern part of S2 in Fig. 2) could be partial observations of one continuous slab gap. Although previous tomography models are broadly similar in their first-order slab geometry, they differ in detail. Statistical analysis (Supplementary Text 3) shows two coherent low-velocity trends that approach within ~50 km beneath the southern Korean Peninsula. One trend strikes northeast from the southwestern end of the slab gap, whereas the other extends westward from the triple junction where the subducting Pacific slab bends (Supplementary Fig. 5). Their connection is unclear and varies among previous models, likely reflecting limited local resolution. By focusing on a localized tomography domain and incorporating dense regional arrays directly above the gap, our tomography improves the recoverable scale beneath the southern Korean Peninsula to ~100–150 km, as demonstrated by comprehensive resolution tests (see “Methods”). This scale is comparable to the expected finite-frequency Fresnel scale of teleseismic P waves in the MTZ, so we cannot completely exclude the possibility that the gap contains unresolved slab remnants or is divided by sub-100-km slab fragments. However, this alternative is unlikely for two reasons. First, 3-D thermomechanical models show that mature slab tearing requires stress localization controlled by trench offset geometry and plate rheology, with representative thresholds of ~150–200 km rather than tens of kilometers^23,24. Second, if slab material with a size of a few tens of kilometers existed within the gap, its dlnVs should still differ from the surrounding low-velocity gap by > 5%. Our sub-resolution spike tests show that such a strong small-scale anomaly would not be completely hidden, because it retains about half of its amplitude within ~50 km even after spatial smearing (Supplementary Figs. 21 and 22). It should appear either as a localized faster anomaly within the low-velocity gap or as an intermediate velocity zone after tomographic mixing and smearing with the gap anomaly²⁵. Instead, our model shows a uniformly low-velocity anomaly across the proposed gap (Fig. 3), indicating that one continuous slab gap is more likely than multiple smaller gaps or slab fragments below the resolution limit.

To explain the origin of the Type 1 anomaly, a deep‑seated hot and dry plume could produce coincident velocity and dMTZ changes. However, this would require unrealistically high temperatures compared to the low average temperature in this region⁸. Alternatively, a non-gap, warm‑over‑cold structure (i.e., a cold stagnant slab near 660 km depth overlain by an anomalously hot mantle domain near 410 km) could potentially produce the near-neutral or slightly elevated temperatures and near-normal or slightly thinned dMTZ we observe (Fig. 5). In such a configuration, the cold slab would thicken the MTZ while the overlying hot mantle would thin it, yielding an apparently balanced overall dMTZ. However, this scenario is unlikely. First, reproducing the observed ~20–30 km of lateral MTZ thinning (Fig. 5a) purely via heating would require a temperature excess on the order of 200 °C–300 °C at ~410 km depth, along with large velocity reductions (dlnVs ≈ −2.5 to −4%) at the top of the MTZ. These extreme conditions are inconsistent with our seismic velocity and temperature models and with previous observations⁷. Second, a warm‑over‑cold stack would tend to mute the observed seismic focusing and defocusing signal: fast (cold) and slow (hot) velocity segments along the same wave paths would partially cancel each other's effects, yielding amplitude contrasts much smaller than what we observed.

Among non‑thermal mechanisms, geochemical signatures in intraplate volcanism suggest possible compositional heterogeneity, linked to subduction of fast‑moving Pacific‑plate sediments. However, the expected range of basaltic compositions in the MTZ yields only minor changes in seismic velocity (and Vp/Vs), and it is difficult to concentrate such compositional anomalies into a continuous, focused corridor, such as Type 1²⁶. A hydrous MTZ is also conceivable in areas of slab stagnation³, yet water alone tends to increase dMTZ and to be distributed broadly above the slab rather than spatially focused^27,28. Crucially, our thermometry explicitly includes plausible ranges of mantle hydration (see “Methods” and Supplementary Text 4), so a purely hydrous effect can be quantitatively ruled out¹⁴. Lastly, hydrous partial melting could, in principle, produce a gap-like feature via water‑induced localized melt lenses²⁹. But melt is generally considered unlikely within the MTZ³⁰. A melt layer atop the MTZ that mimics a warm‑over‑cold effect would strongly affect seismic velocities and thus does not match our observations. While a sub‑resolution thin melt layer could be averaged into the tomographic model (appearing weak despite lowering Vs and raising Vp/Vs), such layers are generally expected to be broadly distributed above a stagnating slab due to negative buoyancy³⁰. We considered potential contributions from various factors (see “Methods”), but they are all less likely compared to the pronounced velocity contrast required by the data and its straightforward thermal interpretation.

In the northeastern segment of the slab gap beneath the western East Sea (~130–132°E, 35–37°N), our tomography reveals a low-velocity anomaly (Fig. 3) that still corresponds to a cold mantle region in our thermal model (Fig. 5). Our synthetic recovery tests show that, as in other sections of the gap, a strong Type 1–Type 2 velocity contrast is required here. This segment lies near mapped shallow tearing of slabs (S3, S4 in Fig. 2) and coincides with a stagnant-slab gap inferred in prior studies (S4 in Fig. 2). A plausible interpretation, therefore, is that the gap in this area is not confined entirely to the MTZ. Instead, the gap likely represents a through-going tearing linking the deep stagnant-slab gap (S5) to the shallower disruptions of the subducting slab closer to the trench (S3)¹¹ (Fig. 2), forming one extensive gap. The shallower tearing appears as a narrow low-velocity anomaly adjacent to the subducting slab, and synthetic tests show that its presence at ~500 km depth is resolvable when its width is comparable to our effective resolution, with the real tomography reproducing a clearer low-velocity anomaly than expected for a continuous slab (Supplementary Fig. 40). Together with the localized absence of deep seismicity in this segment (Fig. 2), this supports a physical continuity between the deep stagnant-slab gap and the shallower tearing. Tearing narrows toward its initiation depths in the shallower upper mantle, where it falls below the resolution of teleseismic tomography. Meanwhile, the pronounced high‑temperature anomaly farther east beneath Honshu is unrelated to a slab gap and occurs where no subducting slab is present²⁹.

Tectonic origin of the slab gap: progressive tearing of subducting slabs

East Asia has undergone protracted subduction since the Mesozoic, and its present configuration features a trench-trench-trench triple junction where the Pacific, Eurasian, and Philippine Sea plates meet (Fig. 1). Following cessation of Izanagi Plate subduction, the Izanagi-Pacific spreading ridge approached the margin at ~55 Ma, reducing slab pull on the Pacific plate and triggering regional plate-boundary reorganization³¹. At that time, the newly formed Philippine Sea plate became the overriding plate above the subducting Pacific plate. Extension and thermal rejuvenation of a relic arc initiated Izu-Bonin subduction at ~52 Ma³². By ~35 Ma, continued northward motion of the Philippine Sea plate led to the establishment of the triple junction; this was followed by rollback of the Philippine Sea plate and clockwise rotation of southwest Japan (Fig. 7a). These processes facilitated back-arc opening and culminated in a major mid-Miocene reorganization at ~20 Ma³³ (Fig. 7b).

Fig. 7. Tectonic origin and reconstructed history of tearing and the gap in the western Pacific slab.

a Map at 600 km depth combining our P-wave model (black square) with the regional model⁶. The thick dashed curve with an arrowhead shows the reconstructed path of the Pacific–Philippine Sea–Eurasian triple junction (numbers indicating ages in Ma)²². A green rectangle marks a previously proposed shallow slab tearing site¹¹. Dashed black lines outline regions interpreted as stagnant slab gaps from our tomography (Type 1). b Schematic 3-D diagrams of the Pacific slab gap and associated tearing (feature S6 in Fig. 2). The subducted Philippine Sea plate is omitted for clarity. Green and pink stars mark the locations of high-Mg andesite (12–14 Ma) and Abukuma adakite (20–14 Ma) volcanism, respectively⁴⁹. c–f Reconstructed slab configurations at key stages of triple-junction migration. (c, 35 Ma): Triple junction established; slab tearing initiates along the plate margin, (d, 20 Ma): Major plate reorganization and triple-junction migration induce a pronounced bend in the tearing path (dashed arrow in a), (e, 16 Ma): Triple junction migrates further; tearing propagates eastward, and (f, 0 Ma): Present setting. Black arrows indicate motions of oceanic plates relative to the Eurasian plate.

The arcuate MTZ gap that we image (Type 1) strikes NE–SW at its southwestern end and rotates to E–W in the east, closely following the reconstructed trajectory of the migrating triple junction between ~35 and 16 Ma²² (Fig. 7). Geodynamic models indicate that progressive slab tearing is localized at stress concentrations that develop during trench retreat, slab bending, and mechanical decoupling between adjacent overriding plates²³. Numerical simulations incorporating plate-boundary kinematics further predict that tearing of the Pacific slab would track the triple-junction path, producing a gap within an otherwise stagnant slab system^21,22. Our high-resolution imaging provides direct observational support for this scenario that slab tearing progressed as plate boundaries evolved. A pronounced bend in the strike of the gap (Fig. 7, dashed black arrow) likely records a rapid reorientation of the triple junction. This reorganization coincides with a shift from northeastward triple-junction motion (pre-20 Ma) to dominantly eastward migration during the main phase of the back-arc extension (~20–15 Ma), supporting a tectonic link between slab tearing and back-arc opening³³. Geodynamic models have shown that slab tearing redistributes slab pull and produces along-strike variations in subduction rate and trench retreat^23,34,35, in agreement with the spatially variable rollback recorded by the staged opening of back-arc basins along this margin²². Within this kinematic framework, it is plausible that the slab gap connects to the present-day shallow slab disruption beneath Honshu¹¹ (Fig. 7, green square) along the plate-junction path, forming an extensive mechanically segmented slab system from East China to southwest Japan.

Comparably progressive slab tearing is observed in other mature oceanic slabs, where tearing tends to localize at geometric irregularities (e.g., trench offsets, sharp bends in slab curvature)²⁴. In the northwest Pacific, the tearing along the triple-junction trajectory was likely facilitated by the contrast in the properties between the Eurasian and Philippine Sea overriding plates, as well as by variations in slab dip. Such inherited segmentation is consistent with multi-level tearing documented in other subduction regions, including the Aegean, Calabrian, and Lesser Antilles arcs²⁴. This suggests that the process we have resolved in East Asia represents a fundamental geodynamic mechanism for slab segmentation in long-lived subduction systems. Importantly, this mechanism can operate independently of induced upper-mantle upwellings, underscoring that purely mechanical slab tearing should be considered an essential factor when interpreting surface tectonic structures and volcanism in subduction zones worldwide with slab stagnation.

We evaluated several alternative plate-reconstruction models, but none can reproduce the continuous, arcuate, sharply bounded geometry and nonthermal structure of the MTZ gap. First, a scenario in which a large marginal basin was consumed during Philippine Sea plate expansion with double-dipping subduction³⁶ predicts wide slab gaps and symmetric remnants, rather than a sharply bounded gap confined to the MTZ. Second, reconstructions that position the Izu-Bonin trench near its present position by ~35 Ma³⁷ imply a near-stationary trench, which would yield a broad slab stagnation instead of an isolated gap. Third, attributing the primary control to the arrival of the Izanagi-Pacific ridge at ~55 Ma³¹ would produce broad, hot slab gaps.

Surface influences of the continuous mechanical slab gap

Slab gaps in the MTZ can reorganize mantle circulation and modulate surface tectono-magmatism by opening lateral windows through which deeper mantle or sub-slab asthenospheric material infiltrates overlying regions. Where such slab-gap pathways are hot and buoyant, they can channel focused, long‑lived volcanism associated with hot upper-mantle plumes⁸. MTZ slab gaps in NE Asia have often been interpreted as hot slab-induced upwellings^7,8, with the velocity contrast between the gap and the surrounding cold slab (up to ~8% in dlnVs) taken to indicate thermal excesses of ~120 °C–240 °C. In contrast, a mechanically maintained stagnant-slab gap that is not anomalously hot will produce relatively weak and spatially diffuse surface expressions of volcanism and other associated tectonic processes. A compilation of regional and global tomography models for this area shows that the actual gap-to-slab contrast is smaller than previously inferred for hot upwellings. It does not exceed 4% in dlnVp or 5.5% in dlnVs across the full range of compiled models. Our model falls within this range (Supplementary Fig. 5 and Supplementary Text 3) and implies at most ~75 °C of excess temperature relative to ambient mantle.

The inferred T_p range overlaps typical mid-ocean-ridge mantle temperatures and is far below the ~200 °C excess in T_p¹⁷ characteristic of regions of anomalously hot upper mantle³⁸, active lower-mantle upwellings³⁹, or plume-heated slab gaps⁴⁰. A gap with a slightly cold (at most modestly warm) T_p yields limited buoyancy, implying slow ascent rates on the order of ~1–2 cm/yr, insufficient to form a focused plume by thermal buoyancy¹⁷. Such weak upwellings are readily dissipated by shallower mantle circulation such as edge-driven convection⁴¹ and rollback-induced toroidal flow³⁴. Consistently, recent full‑waveform tomography⁶ and geochemical observations⁴² show no clear sign of elevated heat in this region. Regional Cenozoic intraplate basalts have geochemical signatures indicative of derivation from ambient-to-modestly-elevated asthenosphere¹⁸. Notably, magmas coeval with the triple-junction migration (~22–13 Ma) around the southern Korean Peninsula record source potential temperatures of ~1320 °C^43,44, comparable to ambient mantle.

At shallower asthenospheric depths, ongoing tearing can channel the sub-slab material toward the overriding plate. Numerical and analog models show that tearing drives toroidal flow around tear edges, focuses deformation in the overriding plate, localizes trench retreat, and promotes back-arc extension^35,45. Regional seismic anisotropy observations support this picture, showing trench-normal return flow with mantle fabrics locally deflected by slab morphology and focused toward tearing regions^46,47. In southwest Japan, fore‑arc volcanism spanning OIB‑like basalts (17–15 Ma), high‑Mg andesites (14–12 Ma), and voluminous felsic eruptions (Fig. 7b) has been linked to Philippine Sea Plate subduction⁴³. The high‑Mg andesites require interaction between slab‑derived melts and hot peridotite. Likewise, pulses of adakite volcanism in northeast Japan (20–14 Ma) require transient heating or access to fusible portions of the old Pacific slab⁴⁸. Progressive tearing along the migrating triple‑junction path between 20 and 12 Ma would have opened short‑lived conduits for asthenospheric influx into the back‑arc, facilitating interaction with hot asthenosphere and oceanic sediment components. This provides the necessary material delivery mechanism without invoking input from upper-mantle plumes⁴⁹.

It is worth noting that hot, low-viscosity mantle can still be entrained beneath subducting slabs by dynamic processes²⁹ and escape through slab gaps or shallower tearing zones to drive localized upwellings at MTZ depths^3,8. However, our observation of the nonthermal present-day slab-gap in the MTZ indicates that any such upwelling events were transient⁵⁰. Dynamic processes at shallower depths likely consumed much of the available heat early in the tearing history³⁵. Any remaining thermal anomalies could then be exhausted in the MTZ via melt extraction and diffusion into surrounding colder slab material during the early stage of slab stagnation³⁰, with endothermic phase transitions.

Methods

Seismic data and relative travel-time measurements

We compiled waveforms from 565 seismic stations across eight networks in the southern Korean Peninsula and southwest Japan (Fig. 2b). From the International Seismological Center (ISC) catalog, we selected 1,175 teleseismic events (2013–2018) with mb ≥ 5.4 and epicentral distances of 30–95°. P‑ and S‑phase arrivals were picked on the vertical and transverse (SH) components, respectively, and band‑pass filtered (0.1–5.0 Hz for P, 0.1–2.0 Hz for S). Conservative Fresnel‑zone estimates (Supplementary Text 2) indicate these passbands can resolve features down to ~140 km at MTZ depths. Relative travel‑time residuals were measured using an adaptive‑stacking approach based on inter‑station coherency⁵¹. All teleseismic waveforms were visually inspected, and noisy or incoherent traces were discarded before measurement. The final datasets comprise 281,920 P‑wave rays from 725 events and 92,838 S‑wave rays from 258 events (Fig. 2b), with ray coverage shown in Supplementary Fig. 1. Measurement uncertainties, estimated from waveform similarity, were used as weights in the inversion.

Tomographic inversion

We conducted teleseismic travel-time tomography to resolve the 3-D velocity structure of the upper mantle. This method exploits travel-time residuals from common events (rays that share nearly identical source-side and far-field paths) and is suited to imaging lateral velocity variations at depth while minimizing bias from event mislocation and heterogeneities outside the model volume. This approach has been applied to delineate slab geometries, identify segmentation, and characterize surrounding mantle structure in other studies. Our tomographic inversion employed fast-marching ray tracing combined with the subspace inversion method⁵². The model space extends from the surface to ~800 km depth, with grid spacings of ~10 km in the crust and ~30 km in the mantle. The starting model consisted of a 3-D crustal structure with a fixed Moho interface overlying the ak135 1-D mantle model⁵³. Crust and upper‑mantle velocities were inverted simultaneously while holding the Moho geometry fixed. For Vp and Vs (expressed as dlnVp and dlnVs, respectively), damping and smoothing were determined from trade-off analyses between data misfit, model roughness, and variance (Supplementary Fig. 2)⁵². Using the same earthquake dataset, we also inverted for Vp/Vs structure to investigate first-order compositional heterogeneity. Perturbations in Vp/Vs (expressed as dln(Vp/Vs)) are sensitive to partial melt, hydration, and the depletion state. The dense arrays enable recovery of short-wavelength patterns in Vp/Vs within and around the MTZ that were unresolved in previous studies. The total objective function combined P- and S-wave residuals, weighted by their norms and uncertainties. To reduce systematic biases from unequal ray coverage, both datasets were restricted to common source–receiver pairs, yielding 86,918 rays from 516 teleseismic events. Additional regularization entailed a second-derivative smoothing penalty on dln(Vp/Vs)⁵². Perturbations are reported relative to ak135. Tests with alternative 1-D references show negligible impact on the recovered pattern. The Vp/Vs smoothing parameter was selected using the same variance–roughness trade-off approach applied to Vp and Vs (Supplementary Fig. 2).

Resolution tests

We assessed model resolution through extensive synthetic recovery experiments that mirror the real acquisition geometry and inversion setup. Synthetic datasets were generated using the same source–receiver configurations as the observed data, with Gaussian random noise added to match estimated residual uncertainties (58 ms for P waves, 137 ms for S waves). To demonstrate that our data coverage and inversion regularization are sufficient to resolve the sharply defined, laterally continuous MTZ gap and its flanking high‑velocity domains, we evaluated four tests.

1. Checkerboard tests. Anomalies were applied with cell sizes of ~100 km (Supplementary Fig. 14a, b) and ~150 km (Supplementary Fig. 15a, b) and amplitudes of ±4% for Vp and ±7% for Vs. For Vp/Vs, we used cell sizes of ~120 km (Supplementary Fig. 14c) and ~150 km (Supplementary Fig. 15c) with an amplitude of ±3%. To test depth‑dependent recoverability, we applied smaller anomalies (~150 km) from the surface to ~400 km and larger ones (~450 km) below ~400 km (Supplementary Fig. 16), reflecting the wavelength variations in the final models. Results show highest sensitivity beneath the southern Korean Peninsula and southwestern Japan: ~100 km anomalies are recovered to ~600 km depth for Vp and Vs, and ~150 km anomalies to ~800 km (Supplementary Fig. 15a, b). Vp/Vs resolves ~120 km features to ~400 km and ~150 km features to ~600 km (Supplementary Figs. 14a and 15c). Depth‑dependent tests confirm that multi‑scale heterogeneity can be recovered simultaneously (Supplementary Fig. 16).

2. Structure-based tests. We created forward models of high‑velocity bodies representing a stagnant slab, both with and without an arcuate gap (Supplementary Fig.17a–d). We also examined the influence of shallow upper‑mantle anomalies (0–400 km) on deeper recovery (Supplementary Fig. 17e, f). The inversions clearly delineate sharp gap boundaries with velocity reductions exceeding −2%. Where no gap was included in the input model, the inversion recovered a continuous slab with only weak, patchy artifacts (<−0.5%). Shallow volumetric anomalies have minimal impact on the reconstruction of deeper stagnant‑slab structure (Supplementary Fig. 18).

3. Azimuthal distribution tests. Because the actual event distribution is dominated by sources from the southeast, we repeated inversions after down‑weighting and removing a subset of those events to create a more uniform azimuthal distribution. The key features of the final models, including the stagnant slab and the associated gap, remain robust; only minor reductions in anomaly amplitudes are observed, and geometry and continuity are essentially unchanged (Supplementary Fig.19).

4. Vertical smearing and slab-gap robustness tests. Potential vertical smearing from shallow upper-mantle structure and the robustness of the MTZ slab-gap anomaly were assessed with three experiments (Supplementary Figs. 20–23). First, shallow heterogeneity leakage tests imposed alternating anomalies (~250 km lateral scale; ±4% dlnVp, ±6% dlnVs) only at 100–350 km depth, with no perturbations below 400 km (Supplementary Fig. 20). The recovered anomalies remain confined to the input depths, with vertical leakage of <0.5% in dlnVs and <0.1% in dlnVp. Second, point-anomaly recovery tests placed isolated spike anomalies (10% dlnVs, 8% dlnVp) at individual grid nodes across the upper mantle and MTZ, with each spike inverted independently to quantify the smearing kernel at each location (Supplementary Figs. 21 and 22). Recovered anomalies are well localized, with ~60–80% amplitude recovery within ±50 km of the input depth in well-sampled regions and vertical smearing <50 km in the slab-gap region. Third, we tested whether the slab-gap low-velocity anomaly is required by the observed data using a two-step subtraction test (Supplementary Fig. 23). We first constructed a gap-free reference model by replacing the slab-gap low-velocity anomaly in our final model with an ambient high-velocity perturbation (+0.6% dlnVs, +0.3% dlnVp), and verified that inverting its synthetic travel times reproduced the gap-free structure without generating any spurious low-velocity anomaly. We then inverted the residuals between the observed travel times and those predicted by the gap-free model, using the gap-free model as the starting model; this inversion recovered a pronounced low-velocity anomaly at the original slab-gap location, demonstrating that the observed data cannot be explained without it.

3-D waveform simulations

We performed 3-D teleseismic waveform simulations to validate the tomographic inferences and assess the sharpness and necessity of the MTZ gap (Figs. 5 and 7). Simulations were carried out with SPECFEM3D_GLOBE⁵⁴, incorporating surface topography and anelastic attenuation from PREM⁵⁵. Event moment tensors and source-time functions were taken from the Global CMT catalog⁵⁶. Within the computational volume (24.5–45° N, 117.5–144° E, 0–800 km depth), we embedded the P‑ and S‑wave perturbations from our inversion into a 1‑D background model derived from regional waveform tomography⁶ (Supplementary Fig. 25). Outside this domain, the Earth structure followed PREM. Synthetic seismograms were band‑pass filtered to 0.02–0.1 Hz, consistent with similar previous cases, and computed at a virtual receiver grid with 0.2° spacing to map spatial amplitude and waveform variations. For reference, we also computed synthetics in the 1‑D PREM model. To isolate the role of MTZ structure, we designed a control experiment in which 3‑D perturbations were restricted to the upper mantle above 400 km (i.e., no MTZ anomalies), yielding the homogeneous, low‑contrast amplitude fields shown in Fig. 4d, g, j.

MTZ thickness compilation from previous receiver-function studies

To obtain a dMTZ model spanning our entire study area, we compiled published 410- and 660-km discontinuity depths from previous receiver-function studies in Northeast Asia (Supplementary Table 2). These studies report broadly consistent first-order patterns (Supplementary Fig. 26), but individual datasets cover only partial subregions and show minor differences in overlapping areas. We therefore formed a composite dMTZ model by averaging values across spatially overlapping regions. Because the contributing models were common-conversion-point stacked using 1-D velocity structures, we applied a 3-D velocity correction to each model with our 3-D velocity model to ensure self-consistency between the velocity and dMTZ datasets. Specifically, the correction was applied to the 660-km discontinuity relative to the 410-km discontinuity, which isolates MTZ thickness and homogenizes the dominant residual effect of lateral velocity heterogeneity across the contributing datasets³⁸. Within the region where the characteristic Type 1 structure is observed, 1° bins (matching the resolution of our tomography) are populated by more than 100 receiver-function measurements each, providing a reliable basis for the composite model (Supplementary Fig. 27).

Estimation of the MTZ temperature

We estimated MTZ temperature anomalies and water content by jointly inverting our seismic tomography (Vp, Vs) with previously published dMTZ measurements. This approach leveraged the complementary sensitivities of Vp, Vs and dMTZ to temperature and composition, allowing conservative bounds on the thermal state. Vp and Vs perturbations from this study were sampled every 20 km between 415–635 km depth. To obtain absolute velocities, we combined our lateral perturbations with region‑representative 1-D profiles⁶ (Supplementary Fig. 25).

For a pyrolitic composition, we computed equilibrium mineralogy and elastic properties with

Perple_X⁵⁷, then applied corrections for high‑temperature anelasticity^20,58 and hydration^15,59 (details in Supplementary Figs. 28 and 29 and Supplementary Text 4). Predicted dMTZ was obtained from the modeled depths of the 410 km and 660 km discontinuities.

For the 410‑km discontinuity (olivine transitions to wadsleyite), we used a Clapeyron slope of +2.9 MPa K⁻¹ (0.107 km K⁻¹), near the lower bound of experimental estimates^15,60 ( + 2.5 to +4.0 MPa K⁻¹). For the 660‑km discontinuity (ringwoodite transitions to bridgmanite + ferropericlase), we adopted −1.5 MPa K⁻¹⁶¹. Both choices reduce dMTZ’s thermal sensitivity, thereby yielding upper‑bound temperature anomalies from the thickness variations^62,63. Hydration was parameterized to raise the 410‑km boundary at low temperatures²⁷ and to deepen the 660‑km boundary by ~6 km per 1 wt% H₂O⁶¹, so that any contribution of water is explicitly accounted for and prevents over‑attributing dMTZ to heat. Taken together with our amplitude‑scaled velocities, the resulting ΔT values should be read as conservative upper limits.

At each location, we searched for the pair (ΔT, H₂O) that best fits the observations by maximizing the likelihood L (Supplementary Fig. 28):

$$L\left(m\right)\propto \exp \left[-\frac{1}{2}\times \left\{\frac{{\left({V_S}_{obs}-{V_S}_{prd}\right)}^2}{{{\sigma }_{V_S}}^2}+\frac{{\left({V_P}_{obs}-{V_P}_{prd}\right)}^2}{{{\sigma }_{V_P}}^2}\right)\right)\left(\left(+\frac{{\left({dMTZ}_{obs}-{dMTZ}_{prd}\right)}^2}{{{\sigma }_{dMTZ}}^2}\right\}\right]$$ 1

where obs and prd denote the observed and predicted values, respectively. The data uncertainties (σ) were set to 1% for Vp, 1% for Vs⁶⁴, and 5 km for dMTZ^7,65. We performed a grid search over broad ranges of ΔT from −300 °C to +300 °C (0.5 °C step) and from 0.0 to 3.0 wt.% (0.2 wt.% step). The reference state (ΔT = 0) used a potential temperature of 1350 °C, consistent with a regional dMTZ ≈ 250 km. To remain conservative, we also scaled velocity amplitudes to account for tomographic underestimation (see “Methods”) before inversion. Spatial uncertainties were estimated from the standard deviation around the best fit (Supplementary Fig. 28). A harzburgitic reference composition yields similar ΔT patterns (Supplementary Fig. 30), indicating robustness to plausible bulk‑chemistry variations.

The inversion is more sensitive to temperature than to water content: typical uncertainties are 10 °C–27 °C for ΔT versus 0.3–0.7 wt.% for H₂O, because temperature exerts a stronger, coherent influence on both velocities and phase‑boundary depths, whereas elastic properties of highly hydrous assemblages remain less well constrained at MTZ conditions¹⁴. Accordingly, we present both ΔT and H₂O maps (Fig. 5b and Supplementary Fig. 29c) but base interpretations on the thermal structure, which is more tightly bounded by the data and our conservative assumptions.

Amplitude scaling of tomographic anomalies

Regularization in travel-time tomography (smoothing and damping) tends to underestimate anomaly amplitudes, and synthetics from our model accordingly reproduce the observed back-azimuthal travel-time variations at reduced amplitude (Supplementary Fig. 34). To compensate, we introduced a multiplicative factor as a proxy for this regularization-induced underestimation of MTZ velocity perturbations. A single uniform factor per wave type was applied to dlnVp or dlnVs when computing MTZ temperatures, since the back-azimuthal residuals primarily sample depth-integrated MTZ structure. The factor was determined against the characteristic residual contrast between back-azimuths crossing the slab gap and those sampling the surrounding slab (~1.5 s for P, ~2.5 s for S; Supplementary Fig. 24), using data at depths > 400 km and back-azimuths of 150°–300° where sensitivity to MTZ heterogeneity is strongest. Scaled travel-time perturbations were forward-modeled in 3-D for trial values between 1.0 and 2.6, and minimizing the L2 misfit between synthetic and observed residuals yielded best-fitting factors of 1.61 (P) and 1.69 (S) (Supplementary Fig. 34g, h). Applying these factors preserves the spatial pattern and regional mean temperatures while modestly increasing local contrasts (Supplementary Fig. 35, Supplementary Table 4). Because the adjustment biases Type 1 temperatures upward, we treat the corrected values as conservative upper bounds on the slab-gap thermal state.

Temperature uncertainty quantification

Total uncertainties in the temperature calculation were evaluated through a computationally intensive Monte Carlo ensemble sampling based on Latin hypercube sampling. Approximately 10⁵ realizations were generated by varying parameter values (i.e., dlnVp and dlnVs, tomographic regularization, amplitude scaling factors, reference velocities, dMTZ, anelasticity, rock composition, Clapeyron slopes, and water content) within physically meaningful bounds derived from this and previous studies (Supplementary Table 3). Mantle temperature was computed for each realization using the temperature conversion procedure described above, yielding an ensemble of temperature structures. The carefully determined parameter perturbation ranges and details of the sampling procedure are described in Supplementary Text 6.

Possible influences on interpreting the gap structure

Peak‑to‑peak variations within the MTZ in our model reach ~3% in Vp, ~4% in Vs, and ~1.2% in Vp/Vs. If these were attributed solely to temperature (using standard anelastic corrections appropriate for MTZ depths)⁶⁶, they would imply a domain‑wide ΔT on the order of 400–500 °C. This is unrealistically high in the eastern Eurasian margin. Several non‑thermal processes can reduce seismic speeds without requiring such extreme heating: (1) Grain‑size reduction expected in cold, deforming slabs²³ enhances attenuation and lowers velocities; if not accounted for in the modeling, a temperature‑only mapping will overestimate ΔT. (2) Major‑element compositional variability in plausible upper‑mantle ranges has only minor effects on Vp, Vs, and Vp/Vs (generally <1%)^67,68 and thus cannot explain the observed amplitude of the anomalies. (3) Explaining the anomalies by hydration alone would require order‑of‑magnitude lateral contrasts of ~2 wt% H₂O, which is unlikely within adjacent portions of a stagnant slab even allowing for localized hydration/dehydration during subduction⁶⁹. (4) Anisotropy is also an inadequate explanation. For subvertical teleseismic paths, commonly reported negative radial anisotropy in the MTZ would tend to increase Vp and decrease Vs, thereby raising Vp/Vs relative to an isotropic mantle. This is the opposite to what we observed in the Type 1 domain, where both Vp and Vs are reduced while Vp/Vs is elevated. Reported azimuthal anisotropy in the MTZ is typically weak (≈1%)⁷⁰, far too small to reproduce our amplitudes and unable to match the combined Vp, Vs, and Vp/Vs pattern. Furthermore, homogenizing the back‑azimuthal sampling of our travel‑time data produced no meaningful change relative to using all paths (see “Methods”). (5) Partial melting is not a dominant process in the MTZ. If it exists, however, it can reduce Vp and Vs by more than ~4 % and ~5.5 %, respectively, and significantly increase Vp/Vs at the top of the MTZ. This is inconsistent with conditions in the slab gap (Type 1).

In principle, a combination of several processes could also reproduce a similar range of observed values. However, such a multi-factor scenario is less likely because it requires a contrived convergence of conditions that are usually mutually incompatible. For instance, invoking significant partial melting or extreme hydration to achieve the velocity drops would demand anomalously high temperatures and water content, which conflicts with the known cold and relatively dry nature of stagnant slabs. Moreover, it is difficult to account for the confined spatial distribution of the Type 1 anomaly. We consider it more plausible that a single dominant factor is responsible for the observed anomaly, rather than an improbable fine-tuned combination of several smaller effects. Taken together, these considerations indicate that the mapped velocity perturbations are primarily thermal, consistent with the independent dMTZ constraints and our conservative thermometry.

Author contributions

J.-H.S.: conceptualization; methodology; software; validation; investigation; data curation; visualization; writing – original draft. S.K.: conceptualization; funding acquisition; supervision; writing – review & editing. B.T.: writing – review & editing. J.R.: supervision; funding acquisition; writing – review & editing.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Funding

J.-H.S. acknowledges support from a postdoctoral research fellowship of the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (RS-2025-02373070). S.K. and J.R. were supported by the Korea Meteorological Institute (Grant KMI2022‑01010 and Grant KMI2022‑00910, respectively).

Data availability

Seismic travel‑time residuals, waveforms, and the final 3-D velocity models and MTZ thickness data, along with the code used for T_p calculations and detailed dataset documentation, are available at Figshare⁷¹ (10.6084/m9.figshare.30370507). Waveform data from stations in South Korea are available via the Korea Meteorological Administration (KMA) portal (https://necis.kma.go.kr) and KIGAM. Data from southwest Japan are available via NIED (Hi‑net/F‑net; https://www.hinet.bosai.go.jp/). Seismic data from JMA and the Global Seismographic Network (GSN) are accessible through FDSN web services (https://service.iris.edu/fdsnws/).

Code availability

The tomography code FMTOMO (Fast Marching Tomography) and its adaptive stacking tools are available on Prof. Nicholas Rawlinson’s webpage (https://nickrawlinson.com/fmtomo). The waveform simulation code SPECFEM3D_GLOBE is open‑source and available from the Computational Infrastructure for Geodynamics (CIG) website (https://geodynamics.org). Software used for plate reconstructions is available on GPlates (EarthByte Group). The mineralogical modeling code Perple_X used to calculate seismic velocities from mantle compositions is available, along with documentation and installation instructions, at https://www.perplex.ethz.ch/. The modeling code for evaluating the effects of anelasticity on seismic velocities as a function of temperature is available at http://web.mit.edu/hufaul/www/Anelasticity.html.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-75246-8.