Introduction

Deep earthquakes, within subducting slabs at depths greater than 300 km, occur at conditions of high temperatures and pressures, which should prohibit brittle failure1,2. Their mechanism remains enigmatic due to the limited experimental observations, unresolved thermomechanical structural properties, and unclear source rupture processes. Three main hypotheses have been proposed in the literature to explain the occurrence of deep earthquakes: thermal shear instability3,4, dehydration embrittlement5,6, and transformational faulting7,8,9,10. While these three mechanisms are still up for debate, recent studies argue that the “dehydration embrittlement” model, based on fluid overpressure, is not a suitable mechanism for deep intraslab earthquakes, especially concerning the sinking lithospheric mantle11,12,13,14. There has been an increasing number of seismic observations of metastable olivine wedges (MOW)15,16,17,18 and laboratory experiments9,19,20 that support the triggering of deep earthquakes via transformational faulting from α-olivine to β-spinel (wadsleyite) in the rim of the MOW. The geometry of the MOW is controlled by the thermal parameter of the subducting slab21,22,23. At the same depth, colder slabs such as Tonga are qualitatively inferred to have thicker MOW than warmer slabs, such as Japan and South America24. Hydrous defects should also be considered as water activates metamorphic transformations and thus could prevent olivine metastability25,26,27. Consequently, besides temperature, the structure of the MOW is also influenced by the degree of slab hydration, which can account for the intermediate-depth seismicity distribution and is controlled by the structural inheritance and fault distribution of the oceanic plate prior to subduction12,28. At intermediate-depths (30–300 km), the earthquake nucleation can originate from a dehydration-driven stress transfer (DDST), involving grain size reduction instead of dehydration-induced fluid overpressure11,27,29,30. However, it is unclear if slab dehydration and DDST can still be a viable explanation for deep-focus earthquakes.

The variability of the b value—slope of the Gutenberg-Richter distribution—provides important insight into the nature of seismic ruptures24,31,32,33. Especially for deep earthquakes, the b-value variations with slab thermal state and earthquake magnitude have been used to infer deep earthquake mechanisms: the nucleation can be due to transformational faulting within a metastable olivine wedge (MOW) and the rupture can propagate outside the MOW due to thermal runaway for larger earthquakes24,33. Although it is speculated that deep earthquakes should nucleate within the rim of the MOW15,16, there has been no seismic constraints on the dimensions of the rim, largely due to limited detection of small earthquakes. This prevents full understanding of both nucleation and dynamic propagation of deep ruptures.

Here we analyzed one of the most complete catalogs, the Japan Meteorological Agency (JMA) catalog, to estimate the b values of four clusters of deep earthquakes in the northwestern Pacific subduction zones, which host abundant deep earthquakes (Fig. 1; Fig. 2). We performed the earthquake clustering based on an unsupervised machine learning approach, i.e., the K-means algorithm, using the location of hypocenters. The b-value analysis of the deep seismicity in the Honshu and Izu segments of the Pacific slab reveals a kink at a magnitude of ~3.7–3.8 for earthquakes deeper than 300 km. The b value abruptly changes from 1.4–1.7 to 0.6–0.7 at the kink. This indicates a fractal dimension reduction for earthquakes above \({M}_{w}\) 3.7–3.8 in a distinct rupture domain within the MOW. The elevated b values greater than 1.5 are likely related to a highly hydrated rim of the MOW within the Pacific slab of the Honshu and Izu trenches. In comparison, no b-value kink is detected at the Kuril and Bonin trenches. Our unique observation of the b-value kink at \({M}_{w}\) 3.7–3.8 supports the existence of an unstable rim of the MOW with a characteristic rim thickness of ~1 km (between ≈ 1 and ≈ 4 km), in which many low-magnitude earthquakes are triggered by transformational faulting. Additionally, earthquakes with intermediate magnitudes (>3.7–3.8) can rupture inward from the unstable rim into the (meta)stable interior of the MOW, with a characteristic thickness of 10–20 km. We propose that the anomalously high b values for small earthquakes (\({M}_{w}\) < 3.7–3.8) highlight enhanced transformational faulting catalyzed by deep hydrous defects. The coincidence between enhanced seismicity within the MOW unstable rim and a denser distribution of pre-subduction hydrated faults would lead to the observed frequency-magnitude distribution anomaly at lower magnitudes.

Fig. 1: Tectonic map and deep-focus seismicity distribution for the period 2002–2016 from the JMA catalog clustered by the K-means method.
figure 1

Earthquakes in four clusters (clusters 0–3, i.e., result of the K-means algorithm) are colored corresponding to the Bonin (red), Izu (cyan), Honshu (purple), and Kuril (yellow) regions. Each cluster is correlated with different slab hydration state controlled by oceanic plate features and fabrics orientation and distribution. Major plate boundaries are outlined by the double thin black lines and microplate boundaries by double dashed lines. The key features of the oceanic floor prior to subduction are annotated accordingly with the legends for faults and fabrics. KFZ Kashima Fault Zone, NFZ Nosappu Fault Zone.

Full size image
Fig. 2: Depth distribution of the earthquakes of the JMA catalog for the 300–700 km depth range.
figure 2

Mw > 2.4 (yellow); Mw > 3.4 (red); Mw > 4.4 (black).

Full size image

Results

The b values calculated using three different methods are listed in Table 1. It has been shown that 2000 events can suffice to calculate the b value to within 0.05 at 98% confidence and 1000 events can still provide a robust b value to within 0.1 at 98% confidence34. Another study also concluded that 1000 events yield small variation in the b value (roughly 0.1)35. There are 2108, 5338, 1571, and 1643 events in the four defined regions (Table S7). Thus, we argue each cluster has enough events for a robust b-value analysis, as supported by the error analysis (Table 1). For the Kuril and Bonin clusters, the GR distributions show a linear trend with calculated b values of around 1.0 (Fig. 3a, d). However, clear kinks (breaks in the slopes of the GR distribution) are observed at a \({M}_{w}\) of 3.7–3.8, for the Honshu and Izu clusters (Fig. 3b, c), with b values decreased by half from 1.4 and 1.7 (\({M}_{w}\) < 3.7–3.8) to 0.6 and 0.7 (\({M}_{w}\) > 3.7–3.8), respectively. This drastic b-values variation across the kink is much larger than the standard deviations (<0.03).

Table 1  b values and associated standard deviations using three different measurement methods.
Full size table
Fig. 3: Gutenberg-Richter distribution for earthquakes occurring between 2002 and 2016 deeper than 300 km in the four clusters.
figure 3

a Bonin (cluster 0), b Izu (cluster 1), c Honshu (cluster 2), and d Kuril (cluster 3). In b and c, very different slopes are observed below and above \({M}_{w}\) 3.7–3.8. The two straight lines b = 1.7 and b = 0.7 are the best fitting lines for 3.0 < \({M}_{w}\) < 3.8 (red) and 3.8 < \({M}_{w}\) < 5.5 (black) for cluster 1 (Izu), while b = 1.4 and b = 0.6 for 3.0 < \({M}_{w}\) < 3.7 (red) and 3.7 < \({M}_{w}\) < 5.8 (black) for cluster 2 (Honshu).

Full size image

To further validate the robustness of the above analysis, we resampled the earthquakes in each cluster using the Monte Carlo approach to conduct bootstrapping24 (Supplementary Note 1; Fig. S1). The Monte Carlo simulations of the b values are plotted on histograms (Fig. S1). Although the b values from the Monte Carlo simulations show some difference from those calculated via the maximum likelihood estimation in the Gutenberg-Richter (GR) plot (b1 in Table 1), the bimodal distribution of the b values in cluster 1 (Izu) and cluster 2 (Honshu) is obvious. Moreover, we confirmed that the kinks for the Izu and Honshu regions were not affected by the occurrence of specific large events affecting the regional stress field (e.g., Tohoku-Oki earthquake) (Supplementary Note 2; Table S1). These results further corroborate the robustness of the b-value analysis in each cluster and thus, the kinks observed in clusters 1 and 2 are significant and reflect structural characteristics.

Compared to the kink at \({M}_{w}\) 6.5, previously reported from the global CMT catalog24, the kinks observed in this study are at much lower magnitudes (i.e., \({M}_{w}\) 3.7–3.8) for earthquakes deeper than 300 km. Because of the limited number of large deep earthquakes in each cluster from the JMA catalog covered by a short time-period between 2002 and 2016, the kink at \({M}_{w}\) 6.5 for earthquakes deeper than 500 km, is not observed. To confirm the possible existence of a kink at larger magnitude for earthquakes deeper than 300 km, we revisit the GR distribution of the CMT catalog with three more years of data compared to the previous study by Zhan24. The deep earthquakes in “warm” subduction zones combined (Japan-Kuril (JK), Izu-Bonin-Mariana (IBM), South America (SA), and Philippine (PH)) show a clear kink at \({M}_{w}\) 6.7 (Fig. S5a). This kink occurs at a slightly larger \({M}_{w}\) likely due to a thicker MOW at 300 km than at 500 km. The calculated b values are almost doubled, from 0.6 below to 1.1 above \({M}_{w}\) 6.7, consistent with previous observations24. The b value of 1.1 for the Tonga (cold) subduction without obvious kink also confirms previous results. The IRIS catalog with mixed magnitudes is also analyzed. The GR distribution shows a similar kink around \({M}_{w}\) 6.9, where the b value approximately doubles (Fig. S5b) for “warm” subduction zones, but remains constant (i.e., b value of 1.1) for the deep earthquake distribution in the colder Tonga subducting slab.

Discussion

Constant b values have been viewed as an indication of self-similarity for earthquakes at all sizes. However, changes in b values (kinks) have been reported for both shallow and deep earthquakes24,36,37. The departure from a constant b value in a GR distribution has been attributed to either the saturation of the fault width37 or a change of rupture mechanism at large magnitudes from transformational faulting to thermal shear instability24,33. Our observation of anomalously high b values (e.g., 1.4 and 1.7) below \({M}_{w}\) 3.7–3.8 in Honshu and Izu likely requires either high pore-pressure—unlikely at such depths—or strong heterogeneity that facilitates the occurrence of low-magnitude earthquakes31,38. Here the anomalously high b values and the existence of the kink at \({M}_{w}\) 3.7–3.8 happen to be correlated to the highly hydrated lithospheric segments prior to subduction (Figs. 1, 2). For the regions of Honshu and Izu, the oblique subduction of spreading faults and oceanic fracture zones (e.g., Kashima Fault Zone) relative to the trench leads to the formation of numerous bending faults at the outer rise parallel to the trench. These bending faults enhance water percolation and deep serpentinization12,39,40,41 resulting in highly hydrated slabs. In contrast, the slabs in Kuril and Bonin are relatively dry due to different reasons. In the Kuril region (Cluster 3), additional bending faults are not favored due to the pre-existing faults (i.e., spreading faults and Nozappu Fracture Zone) which are either already parallel or normal to the Kuril trench, which limits the opening of new bending faults and causes limited slab hydration. For the Bonin region, the dry slab is due to the warmer state of the Pacific lithosphere associated to the Cretaceous-Paleogene magmatism within the Ogasawara Plateau and its related ridges as results of the regional thermal anomaly (Fig. 1)42,43. We argue that the thermal anomaly associated with hotspot volcanism necessarily implies a warmer temperature profile in the lithosphere, and thus a shallower brittle-ductile transition as the bending faults penetrate only down to the brittle-ductile transition39. This implies that the Pacific Plate in Bonin is much drier prior to subduction compared to the three other lithospheric segments analyzed in this study. The dry MOW can be considered as relatively homogeneous, while its H-bearing transforming rim has strong intrinsic heterogeneity due to scattered transformation nuclei.

For the highly hydrated slabs in the regions of Honshu and Izu, the kinks observed at lower magnitudes (\({M}_{w}\) 3.7–3.8), with b values reduced by half from small to large earthquakes (Fig. 3b, c), suggest that the size of the \({M}_{w}\)-3.7–3.8 events corresponds to the characteristic thickness of a seismogenic zone of less than 1 km44,45. This seismogenic zone at depths greater than 300 km is very likely the rim of the metastable olivine wedge (MOW), where olivine would transform to its high-pressure polymorph wadsleyite through transformational faulting, in a narrow unstable zone composed of a mixture of different olivine phases characterized by a small grain size1,19, which is key for seismic rupture nucleation and propagation29,46,47,48. While the MOW should not be expected in wet slabs due to the water-enhanced kinetics of the olivine-wadsleyite transition25,26,49, experimental evidence shows that the MOW can still coexist with hydrous phases (e.g., Phase A) in water-undersaturated conditions, which allows transformational faulting within dry MOW50. Considering water-undersaturated conditions at greater depth for highly hydrated slabs51, it is possible that hydrous defects in the rim will locally trigger fast transformation and generate stress transfers toward drier rock volumes inside the MOW. We want to emphasize again that the mechanism proposed here is different from the “dehydration embrittlement” model often mentioned in the literature. At depths greater than 300 km as in this study, we only consider the remaining water defects within the nominally anhydrous minerals (e.g., olivine) and/or at grain boundaries, i.e., long after the dehydration of serpentine minerals within the subducting lithospheric mantle.

We propose that the highly hydrated slabs in Honshu and Izu are likely to preserve substantial hydrous defects at greater depths spatially coinciding with the MOW rim closer to the slab top (Fig. 4a). These hydrous defects can be nominal without producing fluid overpressure27,30 but should catalyze the transformational faulting by either minor water precipitation at grain boundaries52 or via transformation-driven stress transfers11,27. When \({M}_{w} < \) 3.7–3.8, the intrinsic heterogeneity introduced by hydrous defects in the rim can account for the observed high b values (1.4 and 1.7); in contrast, in dry slabs only negligeable hydrous defects should spatially coexist with the MOW rim (Fig. 4b). This leads to limited low-magnitude seismicity nucleated in the MOW rim and normal b values close to 1.0 in the Bonin and Kuril slabs. When\(\,{M}_{w}\) > 3.7–3.8, transformation-driven stress transfers (TDST)11,27 can trigger rupture propagation beyond the MOW rim and into the metastable olivine wedge (fault lengths > 1 km Figs. 3, 4). We want to point out that small events can well occur inside of the MOW. Only the abnormally numerous small events observed in some slabs (like the Izu and Honshu regions) in addition to the normal trend would be located within the rim.

Fig. 4: Schematic illustration of the hypothesis of transformational faulting enhanced by hydrous defects in the unstable rim of a metastable olivine wedge (MOW).
figure 4

a For a highly hydrated slab mantle, hydrous defects would coincide with the MOW transforming rim and make it a highly reactive halo that enhances the nucleation of small earthquakes, which leads to high b values (e.g., b > 1.5) for the Honshu and Izu deep-focus seismicity. b For a relatively dry slab mantle, there is limited seismicity inside the transforming rim and the b values are normal (e.g., b ≈ 1) for the Bonin and Kuril slabs.

Full size image

The break in the GR distribution observed at \({M}_{w}\) 6.7 from the global CMT catalog (Fig. S5) for deep-focus earthquakes suggests a fault size of 10–20 km44,45. The increase of b values at \({M}_{w}\) 6.5 for earthquakes at depths > 500 km was previously attributed to a change in the fractal dimension of earthquake size distribution from 1 to 2 due to a thin MOW (with a width smaller than 3–5 km) and the change in controlling rupture mechanism24. We argue here that changes in b values at different sizes should be determined by changes of both the rupture domain heterogeneity and the physical mechanisms controlling the rupture, which operate at different stages from rupture nucleation to dynamic propagation and arrest53. For large earthquakes (\({M}_{w}\) > 6.7) at depths > 300 km, it is possible for the earthquake to rupture through the entire width of the MOW and potentially propagate outside the MOW (but still within the slab) into the neighboring wadsleyite through thermal runaway instability46,54 (Fig. 5d). Assuming circular ruptures, the fault size of a \({M}_{w}\,\)6.7 event would be around 10–20 km44,45. In such scenario, the change of controlling mechanism and the increase of heterogeneity (and thus fractal dimension outside the MOW) can explain the increase of b values to 1–1.5 for \({M}_{w}\) > 6.7 (Fig. S5). Furthermore, enhanced transformational faulting in the water-bearing MOW rim would induce a heterogeneity level high enough to increase the b values for small events (\({M}_{w} < \) 3.7–3.8). Depending on the slab hydration state, the density of hydrous defects will modulate the b value at low magnitudes from 1 (dry slab with limited hydrous defects) to 2 (initially highly hydrated slab with abundant remaining hydrous defects; Fig. 5b).

Fig. 5: Mechanisms proposed to explain the b-value kinks observed around Mw 3.8 from the JMA catalog and Mw 6.7 from the CMT catalog.
figure 5

a Schematic representation of the unstable rim (with an example thickness of ≈ 1 km) around the MOW (with an example thickness of ≈ 20 km), with deep earthquakes of different sizes: \({M}_{w}\) < 3.8 within the transforming rim, 3.8 < \({M}_{w}\) < 6.7 in case of stress transfer rupturing through the MOW, and \({M}_{w}\) > 6.7 for ruptures propagating beyond the MOW toward the stable slab material already transformed into wadsleyite peridotites. b Range of b values corresponding to the different segments of the Gutenberg-Richter distribution defined by the observed two kinks at \({M}_{w}\) 3.8 and \({M}_{w}\) 6.7. The slope of each segment is controlled by the hydration state of the slab and the thermal state of the MOW. c Illustration of the link between rupture size and ruptured materials for deep-focus earthquake: within the MOW rim (red), inside the MOW (green), and outside the MOW (blue). d Corresponding mechanisms required in each material: primary transformational faulting in the unstable rim of the MOW (red), transformation-driven stress transfer inducing secondary transformational faulting rupturing into the MOW (green), and the additional rupture propagation outside the MOW via thermal runaway instability (blue), with increasing rupture width and earthquake magnitude.

Full size image

The development of ruptures of different sizes can all be related to three distinct rupture domains: the unstable rim of the MOW, the interior of the MOW, and the warmer wadsleyite domain around the MOW including slab material that already endured the olivine-wadsleyite transformation. These three distinct materials are the locations of three different rupture mechanisms, which can operate at three different stages for large ruptures (Fig. 5a, c, d). Considering a large event, the initial stage involves low-magnitude earthquakes nucleating within the MOW rim via transformational faulting when (meta)stable olivine becomes unstable and transforms into wadsleyite. As the fault grows larger, at the second stage, the rupture can propagate into the interior of the dry MOW via a transformation-driven stress transfer. If the rupture passes the center of the MOW (i.e., half width), the shearing instability likely continues via thermal runaway in the slab-strike direction within the MOW and slightly beyond the MOW further into the wadsleyite within the slab (Fig. 5c, d). The variations of the b values with all earthquake sizes can be explained by differences in fractal dimension between these three distinct rupture domains controlled by the slab hydrous state and the thicknesses of the MOW and its unstable rim, determined at first order by the slab thermal parameter. It should be noted that the b-value changes across different subduction zones for deep-focus earthquakes may indicate changes in physical rupture mechanisms. For example, for the cold Tonga subduction zone, the b value shows no significant change at larger magnitudes. This would suggest that the MOW in the Tonga slab is thick enough to accommodate large earthquakes with magnitude 7 or above with the same mechanism of transformation-driven stress transfer. However, due to the earthquake detection limit, the kink at smaller magnitudes still needs to be verified in the future with more complete earthquake catalogs for other subduction zones such as Tonga.

Furthermore, the observation of the b-value kink in the subducting slab can provide additional insights regarding features that characterized the oceanic lithosphere before subduction. For instance, we have enough data in the Bonin region in the 300–550 km depth range to report the absence of kink in the b value. As shown in the recent paper of Hirano et al.42, the hotspot volcanism that overprinted the Pacific lithosphere in the region of the Ogasawara Plateau at ≈ 55 Ma also affected the Minamitorishima area at ≈ 35 Ma. Considering the relatively steep slab geometry in the Bonin subduction zone55 and a subduction velocity of ≈ 80 mm.yr−1 56, and assuming a fixed location of the hotspot volcanism42, a depth of 550 km would correspond to a subduction period of ≈ 15 Ma. This would suggest that the activity of the Ogasawara hotspot magmatism started at least around 80 million years ago, consistent with the literature42.

Conclusions

We have analyzed the Gutenberg-Richter distribution for deep earthquakes from one of the most complete regional earthquake catalogs: the JMA catalog of 15 years of data, in the northwestern Pacific subduction zones. An unsupervised machine learning approach, the K-means algorithm, divides all the deep-focus earthquakes (depths ≥ 300 km) into four clusters corresponding to subducting slabs with different hydration state prior to subduction. For the first time, we observe kinks of b values reducing from 1.4–1.7 down to 0.6–0.7 at \({M}_{w}\) 3.7–3.8 for the Honshu and Izu clusters, while normal b values (1.0) for the Bonin and Kuril clusters are observed at low and intermediate magnitudes 3.3 < \({M}_{w}\) < 6.0. The variation of b values at low magnitudes (\({M}_{w}\) < 3.7–3.8) indicates their dependence on the slab hydration state, with higher b values (1.4–1.7) correlated with highly hydrated slab coinciding with the expected location of the MOW rim. The hydrous defects favor the nucleation of small earthquakes via the mechanism of transformational faulting within the rim. Such mechanism operates for small events (\({M}_{w}\) < 3.7–3.8), with rupture lengths ≤ 1 km, which tells us the thickness of the MOW rim. Through transformation-driven stress transfer, the rupture can propagate into the MOW, which is more homogeneous with smaller fractal dimension, corresponding to b values between 0.5 and 1.0.

Combining with the b-value analysis from the latest CMT catalog, the kink at \({M}_{w}\) 6.7 suggests that the mechanism of thermal runaway instability operates for larger earthquakes rupturing through and propagating outside the MOW with increased heterogeneity in the new rupture domain. The change of controlling mechanism and rupture domain heterogeneity can explain the spatially varying b values due to the slab hydrous state and thermal state. The \({M}_{w}\,\)threshold of 3.7–3.8 provides a seismic constraint on the MOW rim thickness \(\sim\) 1 km, a narrow zone where hypocenters of low-magnitude earthquake should concentrate in, while the centroids of larger earthquakes below \({M}_{w}\) 6.7 should coincide with the geometry of the MOW with a thickness of 10–20 km. Furthermore, considering cold subduction zones such as Tonga and moderate hydration, it is possible that a more complete earthquake catalog reveals a similar kink and high b values at low moment magnitudes in case of the existence of a thin unstable rim.

Data and methods

The previous b-value analysis of deep earthquakes involves using the global CMT catalog with moment magnitude of completeness (\({M}_{C}\)) around 5.324. The high \({M}_{C}\) prevents the full understanding of earthquake nucleation at lower magnitudes. The depth distribution of earthquakes recorded in the JMA catalog is shown in Fig. 2. To study the distribution of low-magnitude earthquakes and its implications on deep rupture mechanism, we performed clustering and b-value analyses of deep-focus earthquakes using the JMA catalog, i.e., the most complete catalog of the Kuril, Japan, and Izu-Bonin subduction zones (Fig. 1) with a much lower \({M}_{C}\) around 3.0 (Fig. 3). There are a total of 10,660 earthquakes recorded at depths greater than 300 km in Northwestern Pacific subduction zones between 2002 and 2016. The reason to select the 300–700 km depth range for this study is detailed in the Supplementary Note 3 (Tables S2 to S6). To cluster these deep earthquakes more objectively, we employed the K-means clustering, an unsupervised machine learning method based on clustering analysis. The b-value analysis was then performed in each of the four clusters, which turned out to also correspond to different segments of the Pacific Plate with different slab hydration degrees (Fig. 1).

Clustering analysis

Cluster analysis or clustering is the process of finding similarities in the dataset and grouping similar data points together. It is one of the important tools for the pattern recognition57 and generally regarded as unsupervised learning58. Because the K-means algorithm is easy to implement and does not require high computational power, it is the most widely used algorithm for cluster analysis59,60,61. Cluster analyses of earthquakes provide insights into seismic hazard in different geological or tectonic settings62,63,64. In this study, we used the earthquake hypocenter locations as the features for the K-means clustering, with the intention of minimizing human interference. The optimal number of clusters is selected through silhouette analysis (Supplementary Note 4; Fig. S2). It evaluates how well the resulting clusters are separated, with higher silhouette scores (closer to 1) indicating better clustering. In this study, the optimal cluster number is four because it leads to not only a local maximum silhouette score (Fig. S2) but also clusters in excellent agreement with the existing segments of the subducting Northwestern Pacific Slab, i.e., Kuril, Honshu, Izu, and Bonin (Fig. 1). The robustness of the K-means clustering algorithm has been further corroborated by the spectral clustering and Gaussian mixture models (GMM) clustering algorithms. The clustering analysis results are similar for all three different algorithms and included in the SI (Supplementary Notes 5-7; Figs. S3 and S4. Tables S8 to S11).

It is worth noting that the Bonin segment is characterized by a remarkable magmatic anomaly that overprints the oceanic lithosphere in the region during the Cretaceous and the Paleogene42,43. The Ogasawara Plateau corresponds to the area affected by this magmatic activity, which necessarily has a prolongation within the lithosphere subducting at the Bonin trench. As the magmatic activity mobilizes water, we argue that it necessarily left the lithospheric mantle in a drier state than without additional magmatic activity (portion without overprinting subducting at the Izu trench). The lower water content in such a drier lithosphere (Bonin) is expected to reduce the probability of nucleation of metamorphic transformations13,25,26,27; therefore, we argue that the separation of the Izu-Bonin Arc seismic events into two individual clusters, as done by the unsupervised machine learning in this study, is appropriate.

b-value analysis

To carry out the Gutenberg-Richter (GR) distribution analysis for each cluster, we first converted the JMA magnitude scale \({M}_{j}\) to the physics-based moment magnitude \({M}_{w}\) scale using the following empirical relationship65:

$${M}_{w}=a{M}_{j}^{2}+b{M}_{j}+c$$
(1)

where a = 0.053 ± 0.003, b = 0.33 ± 0.02, and c = 1.68 ± 0.03. The magnitude of completeness (\({M}_{c}\)) is approximately estimated from the non-cumulative (binned) distribution (Supplementary Note 8; Table S12). Since the \({M}_{c}\) determined from all approaches are similar, the completeness determination in this study is reliable. Due to the seismic network coverage variations, \({M}_{c}\) changes across the different clusters, e.g., \({M}_{c}\) 3.3 for the Bonin cluster with fewer stations in its vicinity, while around 3.0 for the rest of the clusters with more station coverage (Fig. 3).

We then calculated b values (\({b}_{0}\)) by means of the maximum likelihood estimation66. We have also estimated the more accurate b values (\({b}_{1}\)) following the modified formula67,68,69. We note that Marzocchi and Sandri stated that either the original estimation (\({b}_{0}\)) or the modified estimation (\({b}_{1}\)) will be influenced by the possible error on the magnitude35. Therefore, we also applied another b-value estimation (\({b}_{2}\)) from Tinti and Mulargia70, which is much less affected by the magnitude error35:

$${b}_{0}=\frac{1}{{{{{{\rm{ln}}}}}}\left(10\right)* (\mu -{M}_{c})}$$
(2)
$${b}_{1}=\frac{1}{{{{{{\rm{ln}}}}}}\left(10\right)* (\mu -({M}_{c}-\triangle M/2))}$$
(3)
$${b}_{2}=\frac{1}{{{{{{\rm{ln}}}}}}\left(10\right)* \triangle M}* {{{{{\rm{ln}}}}}}\left(1+\frac{\triangle M}{\mu -{M}_{c}}\right)$$
(4)

where \(\mu\) is the average of the magnitudes, \({M}_{c}\) is the magnitude of completeness, and \(\triangle M\) is the magnitude interval (0.05 in this study).

We calculated the corresponding standard deviations of all three b-value estimation methods:

$${\sigma }_{{b}_{0}}=2.3{{b}_{0}}^{2}\sqrt{\frac{\mathop{\sum }\limits_{i=1}^{N}{({M}_{i}-\mu )}^{2}}{N(N-1)}}$$
(5)
$${\sigma }_{{b}_{1}}=2.3{{b}_{1}}^{2}\sqrt{\frac{\mathop{\sum }\limits_{i=1}^{N}{({M}_{i}-\mu )}^{2}}{N(N-1)}}\,$$
(6)
$${\sigma }_{{b}_{2}}=\frac{\frac{\triangle M}{\mu -{M}_{c}}}{{{{{{\rm{ln}}}}}}\left(10\right)* \triangle M* \sqrt{N* (1+\frac{\triangle M}{\mu -{M}_{c}})}}$$
(7)

where \(N\) is the total number of earthquakes.