Lineaments and earthquake ruptures on the East Japan megathrust

There is a continuing debate as to whether science will ever allow prediction as to when and where earthquakes initiate. To make progress, we need to better understand how load is distributed during the build-up of tectonic stress prior to catastrophic failure, and how the complex material properties of rock affect the outcome. It is also important that we determine how geological structures in the loaded region influence where and when yield commences at the start of an earthquake, in map view (Bassett et al., 2016; Ikeda et al., 2009) as well as in three dimensions (Lister et al., 2008). This is difficult, because the way stress accumulates during loading is influenced by the natural heterogeneity and variable mechanical properties of rock. Nevertheless, the tectonic architecture is critical, and we need to determine its influence.

Just as an engineer will study the structure around a dam that has failed, to assess whether there were specific weaknesses that predicated that outcome, a geoscientist can forensically examine the structure of a region that has hosted a large earthquake. We can thus seek to determine whether controlling structures exist that localized the onset of failure, and if such focusing structures might have been monitored to allow accelerating precursor motion to be detected in the weeks, months, years, or even decades prior to a catastrophic event. In this paper, we report the outcome of such a forensic analysis for the earthquake geology of the 2011 Tohoku-oki Great Earthquake on the East Japan megathrust.

There have been few attempts at seismotectonic analysis at map scale. Nakajima et al. (2009) provide an in-depth analysis of the Tokyo region, and provide important data on slab geometry. Otherwise, research on this topic has focused on the seismic potential of different domains (e.g., Triyoso and Shimazaki, 2012), with a multivariate statistical approach adopted by Kumamoto et al. (2016) allowing a revision focused on moment magnitude. Ghimire and Kasahara (2009) focus on stress analysis, linking regional stress axes to P-axis orientation during earthquake moment release, but this analysis did not account for fault kinematics. Kyriakopoulos et al. (2013) use finite element models to examine heterogeneities in the margin architecture. Sato et al. (2013) examine the seismotectonic architecture from the point of view of the location of high-slip patches. Zhao et al. (2011) consider structural heterogeneity of the megathrust and the 2011 earthquake mechanism. The approach adopted here is outlined in Lister et al. (2014), and in Lister and Forster (2018), who applied similar methods to seismotectonic analysis of the significance of fault movements during the 2004 Great Earthquake in the Andaman Sea.

We began with data from the Global Centroid-Moment-Tensor (CMT) Project (Ekström et al., 2012; Nettles et al., 2011). This catalogue includes the strike, dip, and rake of two possible but conjugate fault planes for many earthquakes. These conjugate solutions are characterized by a simple geometric relation: the slip line on one fault plane is orthogonal to the fault plane of the other. Because the sense of relative displacement for one conjugate fault plane solution is opposite to the sense of relative displacement for the other, the type classification of the fault geometry in all but a small minority of cases is the same for both solutions, even though the possible (conjugate) fault planes dip oppositely at ninety degrees to each other. This geometry is made evident in a beachball representation of the centroid moment tensor, with the shaded area defining the orientations of ray paths for compressional moment pulses emanating from the rupture. This geometry also makes it possible to automatically classify the type of rupture associated with an individual earthquake, here using the eQuakes program to characterize earthquake type.

The eQuakes program allows ready analysis of the geometry of earthquakes (Lister et al., 2014; Lister and Forster, 2018) and thus allows recognition of patterns in the type of faulting taking place during an earthquake sequence. Figure 1A shows epicenters for reverse faults and thrusts (red), normal faults (light blue), and strike-slip faults (purple). The classic beachball representation is also possible, but statistical analysis of the significance of orientation clusters can best be accomplished by grouping earthquakes and plotting slip line directions and fault plane normals in a scatter diagram, using the lower hemisphere of a stereographic projection.

To allow such statistical analysis, we constructed stereograms for slip lines and fault plane poles in the different spatial clusters that can be recognized (Figs. 1B–1H). We chose to do this for three months prior, and nine months after the 2011 Tohoku-oki Great Earthquake (Fig. 1A), to include foreshocks, and to limit earthquakes to those occurring immediately after the main rupture. In this way, we limit the seismotectonic analysis to consider only the geodynamic state of the lithosphere immediately prior to the main rupture, and during the early to mid-stages of the aftershock sequence. Initially, epicenters for all earthquakes with moment magnitude M_w > 5 were plotted, since these are more accurately located, and centroid moment tensor data is available.

The depth of the hypocenters analyzed was limited to <30 km, to allow a focus on crustal architecture. This cut-off depth was chosen because it encompasses the brittle-to-ductile transition in most crustal rocks, and thus includes almost all crustal seismicity in Japan. This choice of depth cut-off allows inclusion of all seismic activity associated with operation of the East Japan megathrust except that which occurs in the deep crust, and thus a focus on seismic activity in the down-going slab at its shallowest levels. Seaward of the megathrust, this choice of depth allows the geometry of normal fault earthquakes to be highlighted in the depth range that includes their most energetic seismogenic behavior. The Japanese lithosphere is ∼70 km thick (Kanamori and Press, 1970).

In almost all cases, the inherent geometrical ambiguity present in focal plane solutions could be resolved by bringing geology into the equation (Lister et al., 2008, 2014; Lister and Forster, 2018). Earthquakes that result from normal faults (Jackson, 1987) typically have a slip line that plunges at ∼60°, whereas classification of the fault plane normal and associated slip line for thrust faults can assume that the fault plane dips gently. In addition, for specific spatial groups of earthquakes, there are well-defined maxima on the scatter plot of earthquake slip lines and fault plane poles, so the faults causing these earthquakes are similarly oriented. This allowed ready classification of fault planes and slip line directions in orientation groups involving thrusts (Figs. 1B, 1C, and 1G) or normal faults (Figs. 1D, 1E, and 1F). It was also possible to resolve ambiguity for strike-slip fault solutions, by computing hypocenters and centroids for individual earthquakes using the Hypocenter-Centroid (H-C) method (Zahradník et al., 2005, 2008; Hejrani et al., 2010). This H-C method allowed classification of fault plane pole and slip line orientation clusters based on the fault plane solution that allows the alignment of the earthquake hypocenter and centroid in the same plane. The classification of fault geometry is shown on individual stereoplots.

Typically, a high degree of kinematic coordination is observed in movements associated with particular spatial clusters of earthquakes (Lister et al., 2014; Lister and Forster, 2018). This patterning is responsible for the orientation clusters on lower hemisphere stereographic projections when slip lines and/or fault plane poles are plotted for individual spatial clusters (Fig. 1). Such an analysis makes it obvious that there was a distinct and marked segmentation of tectonic modes above the megathrust. There are three spatial clusters of thrust earthquakes offshore East Japan (red circles in Fig. 1A). These each have a well-defined orientation group evident in the corresponding scatter plot (Figs. 1B, 1C, and 1G), and we thus infer similarly oriented principal compressive stress axes. Assuming Coulomb-Mohr failure, these three segments were subject to horizontal, margin-orthogonal shortening, and thus acted as compressional buttresses ramming the subducting Pacific Plate.

The compressional buttresses (Fig. 1) are named in turn: (i) the Sanriku buttress, with thrust faults that dip ∼20° toward the WNW, and slip directions trending ∼110° (shown in the scatter plot, Fig. 1B); (ii) the Tohoku-oki buttress, again with thrust faults that dip ∼20° toward the WNW, and slip directions that trend ∼110° (shown in the scatter plot, Fig. 1C); and (iii) the Joban buttress, with thrust faults that dip ∼25° toward the WNW, and slip directions trending ∼115° (again shown in a scatter plot, Fig. 1G).

There are prominent volcanic lineaments in East Japan (Fig. 2), and it is of interest that these approximate the seismotectonic bounds of the compressional buttresses. Most seismic activity in the Sanriku buttress is bounded to the north by the Aomori Volcanic Lineament and to the south by the Sanriku Volcanic Lineament. Similarly, the Tohoku-oki buttress is bounded to the north by the Chokai Volcanic Lineament, while to the south it is bounded by a seismotectonic lineament ∼15 km to the north of the Miyagi Volcanic Lineament. The Joban buttress is bounded to the north by a seismotectonic lineament ∼20 km south of the Fukushima Volcanic Lineament, and to the south by the Haruna Volcanic Lineament.

Between the three compressional buttresses are “channels” in which normal-fault earthquakes dominated, implying margin orthogonal extension (Fig. 1). These channels are of particular interest for several reasons, as will be outlined in the remainder of the paper. The channels are: (i) the Tohoku-oki channel (which was the most active during the 2011 earthquake); and (ii) the Sanriku channel. Here we focus on the Tohoku-oki channel, which encompasses two important volcanic lineaments: (i) the Miyagi Volcanic Lineament in its northern extent; and (ii) the Fukushima Volcanic Lineament in its southern extent. The coincidence between the bounding seismotectonic lineaments and the volcanic lineaments is not exact: there is an offset, as noted.

We can relate inferred segment boundaries on the megathrust to the pattern of movement implied by the foreshocks and earthquakes. The contoured scatter plots (Fig. 1) show that slip lines for foreshock thrust fault earthquakes are statistically parallel, and close to the ∼NW-SE trend of the margins of the seismotectonic segments (Fig. 1). Each slip line marks the direction of relative movement on a fault rupture, shown using the relative motion of the hanging wall (or upper-plate) of the faulted rock mass. The trends of the slip lines are concentrated with an azimuth of 292–294°, a direction within a few degrees of the 286–287° trending volcanic lineaments that can be seen on land (Fig. 2). Symmetry requires that in the reverse fault segments the principal stress axis, σ₁, is subhorizontal, trending ∼5° clockwise from the trend of the volcanic lineaments.

Prior to the 2011 Great Earthquake, a foreshock cluster of compressional thrusts occurred, limited to the narrow extent of the Tohoku-oki buttress (Fig. 1C). Compressional earthquakes also define the southern buttress to the extensional channel and mark another switch in tectonic mode, south of the Fukushima Lineament. Structures parallel to the bounding lineaments thus separate different types of earthquakes, and there are abrupt changes from margin-orthogonal compression to margin-orthogonal extension across them. It is important to understand how this might have been possible: e.g., why margin-orthogonal horizontal compression caused thrust fault earthquakes from ∼20 km north of the Miyagi Lineament, while normal fault extensional earthquakes took place immediately to the south of this boundary.

It is also important to understand why normal fault earthquakes were prevalent in the subducting slab adjacent to the extensional earthquake channel. Such outer-rise earthquakes did continue northward, and southward, thus also forming seaward of the compressional buttresses, but occurred in markedly lower numbers in those regions. They concentrated in great number immediately adjacent to and in the slab beneath the crust in the Tohoku-oki extensional channel.

It is particularly significant that the fault plane appears to consistently dip landward for most of these normal-fault earthquakes (Figs. 1D, 1E, and 1F), so this aspect was examined in more detail. Their landward-dip was confirmed by more detailed analysis, as detailed in a subsequent section. In comparison, the hinterland of the extensional channel was marked by earthquakes on normal faults with scattered orientations (Fig. 1H). All normal faults in the hinterland of the extensional channel occurred in the top 30 km of the crust, above the megathrust as it shallows from ∼40 km to its eventual outcrop in the trench. Thrust fault earthquakes are few in number and occur near the megathrust, at the base of the extensional channel. The thrust fault solutions are almost all associated with failure on fault planes that dip northwest, i.e., with fault planes that likely merge with and/or are parallel to and/or coincident with the main megathrust (Figs. 1B, 1C, and 1G).

There are some strike-slip earthquakes within the subducting slab. The H-C method applied to these earthquakes implies left-lateral strike-slip faulting in the surface levels of the subducting Pacific Slab, where it meets the Miyagi Lineament. The slip direction trends parallel to the Median Tectonic Line, supporting Bassett et al. (2016). Such earthquakes would occur in the subducting slab if there was movement prior to or during the aftershock sequence, implying the Japanese crust had indented the subducting Pacific Plate.

To further assess the role of the volcanic lineaments, we considered the coincidence of the seaward prolongations of their trends with earthquake epicenters (Fig. 2A). The lineament trajectories were defined using topographic images, selecting peaks and ridge lines (see File DR1 in the Data Repository Item¹), defining for each point a position vector in 3-D Cartesian coordinates. The array of normalized position vectors for each lineament was contracted into a 3 × 3 orientation matrix, and the eigenvalues and eigenvectors of this matrix were then determined. The eigenvector orthogonal to the plane of best fit was used to determine how close any individual hypocenter was to an individual lineament. In all cases, the distance between the epicenter and the closest lineament was considered. This value was tabulated (Fig. 3) with the histograms showing the variation in distance of epicenters from the closest lineament, with moment magnitude cut-off at different values.

Systematic recording of data in earthquake catalogues has yet to reach its first half century, so only the most seismically active regions have experienced enough large events to allow statistical analysis. We set an upper limit as M_w ≥ 6.9 to ensure a significant number of major events could be analyzed, to recognize statistical trends. We found that almost two-thirds of the epicenters from the Global CMT Project database for earthquakes with moment magnitude M_w ≥ 6.9, since 1976, began within 15 km of one of the volcanic lineaments (Fig. 3). These concentrations of epicenters remain visible at lower moment magnitude cut-off values, but the correlation is most marked for the largest magnitude earthquakes.

We must conclude that there are physical structures associated with the volcanic lineaments, and that these played a significant role in triggering, or localizing the onset of major earthquakes on the East Japan megathrust. The next question, therefore, concerns their physical expression and mechanical strength. One can imagine the existence of faults that remain steep and thus continue the trends of the volcanic lineaments to shallower levels, as well as faults that curve from deeper structures while remaining parallel to these trends. We can infer that these structures must be weak because earthquake rupture planes never parallel these trends.

Numerous previous studies show the value of combining the analysis of satellite geodesy with mechanical modeling (Ikuta et al., 2012; Kato et al., 2012; Ozawa et al., 2012; Kyriakopoulos et al., 2013; Mavrommatis et al., 2014, 2015; Hasegawa and Yoshida, 2015; Johnson et al., 2016; Loveless and Meade, 2016; Hashima and Sato, 2017; Yokota and Koketsu, 2015). Several of these works demonstrate patterns that suggest the existence of margin segments similar to those recognized by the seismotectonic results presented above. For example, Ozawa et al. (2012), Hashima and Sato (2017), and Perfettini and Avouac (2014) highlight interesting aspects in the context of the segmentation inferred in this paper: (i) maximum coseismic and post-seismic displacement focused in the offshore prolongation of the Tohoku-oki compressional buttress (Fig. 1); and (ii) vertical acceleration focused within the Tohoku-oki buttress and the northern edge of the Joban buttress. However, the most compelling comparison comes from Loveless and Meade (2016) in diagrams that plot velocity changes in a (Lagrangian) reference framework based on the pattern of movement in their Epoch I (from 1996 to 1999). Their figures 5b and 5c show the velocity changes in four-year intervals up to the time of the 2011 Great Earthquake. For eight years prior to the main rupture, the Sanriku buttress accelerated westward (so material points were driven with ever increasing velocity away from the trench), in a direction roughly parallel to the Sanriku volcanic lineament. For the segments southward up until the southern margin of the Tohoku-oki buttress, as defined by the seismotectonic segmentation shown in Figure 1, there is little change in velocity. Southward of the Tohoku-oki buttress, however, for the eight years prior to the main rupture, the margin segment we refer to here as the Tohoku-oki extensional aftershock channel accelerated toward the southeast, or east-southeast (i.e., against the WNW-directed movement defined by Epoch I).

These observations are consistent with the conclusion that the buttresses focused loading during the decades prior to the Great Earthquake. Further, since the foreshocks to the Tohoku-oki Great Earthquake took place exclusively within the Tohoku-oki buttress, the southern margin of the Tohoku-oki buttress can be argued to have focused the accumulation of deviatoric stress. The foreshocks involved displacement on gently dipping thrusts (Ozawa et al., 2012; Sato et al., 2013) and can be taken as the first sign of imminent yield. This was registered by a reversal in the sense-of-shear across the buttress lineament, i.e., from left-lateral strike-slip in the eight years preceding failure, to right-lateral strike-slip in the time of precursor movements, and during the main rupture (Kamiyama et al., 2016). Later aftershocks within the northern reaches of the Joban buttress were also gently dipping thrust displacements, which explains the pattern of vertical uplift focused in or at the margins of the compressional buttresses. In contrast, extensional normal faulting at the seaward margin of the Tohoku-oki extensional channel requires either subsidence, or minimal vertical uplift, as observed (Hashima and Sato, 2017).

The decadal signal showing landward acceleration of the Sanriku buttress (Loveless and Meade, 2016; Mavrommatis et al., 2015) can be interpreted as an effect caused by strain-hardening on the megathrust. Preliminary modeling we conducted shows that the effect can also result from increasing the down-dip extent of the locked area of the megathrust. Conversely, the seaward acceleration of the Tohoku-oki extensional channel to the south documented by Loveless and Meade (2016) can also result from decrease in the down-dip extent of the locked area of the megathrust. The switch in geodynamic behavior requires local increases in the deviatoric stress intensity, perhaps explaining why foreshocks and the main shock initiated close to the southern margin of the Tohoku-oki buttress. The megathrust may have been locking in the north, while unlocking was taking place in the south, defining a precursor to eventual catastrophic failure.

Strike-slip fault earthquakes are not observed because the accommodation structures are weak, and thus do not store strain energy. Such weakness is also a mechanical requirement to allow the abrupt transitions in tectonic mode that occur across them. Prior to the main shock, and during the aftershock sequence, the axis of principal compression north and south of the extensional channel was horizontal, trending orthogonal to the margin. Within the extensional channel, during the aftershock sequence, this principal compression axis is vertical. The abruptness of this transition requires the intervening zone of weakness be correspondingly narrow.

To better define the pattern of faulting in (and adjacent to) the Tohoku-oki extensional channel, the locations of the centroids and associated centroid moment tensors (CMTs) were recalculated (see Table DR2, ^{footnote 1}). This was done in order to improve the resolution on horizontal location and depth of the centroids reported in the Global CMT Project catalogue. We analyzed the broadband recordings by the F-net network, operated by National Research Institute for Earth Science and Disaster Resilience of Japan, selecting six high-quality stations with reasonable azimuthal coverage, noting that we had access only to on-shore stations. The CMT was computed for 84 events (Fig. 3), following methods as in Kikuchi and Kanamori (1986) and Zahradník et al. (2005).

A least-squares method was used to solve Eq. 2. One of the advantages of using this convention is that different constraints on the CMT components, such as deviatoric or pure double-couple inversion can simply be performed by choosing different subsets of CMTs in Eq. 1. The deviatoric CMT inversion assumed events that were pure double-couples, using the first five components as a basis. For each event, a spatiotemporal grid search over a selected number of nodes was performed, near the hypocenter, to find the CMT. The centroid time was searched by shifting the seismograms around the hypocenter time. A grid node location was then defined, and the time shift, which in combination provided the highest correlation between the observed and synthetic data as the optimum centroid location and time. The quality of all waveforms was checked manually and those with poor quality were removed prior to inversion.

The recalculation was made possible by a more detailed regional crustal structure model (Zhao et al., 1990) in comparison with the reference Earth model, PREM (Dziewoński and Anderson, 1981). This made it possible to predict ground motion down to 20 s, whereas the lowest period used in the Global CMT Project is 40 s. We could therefore constrain the depth for the crustal events confined within the top 25 km with more precision (as opposed to the Global CMT Project algorithm, in which the depth search stops at depths ≤12 km). Short period (<50 s) surface waves are sensitive to crustal structure and the short period body waves are sensitive to small-scale lateral heterogeneities in the mantle and crust. Using a low-pass filter to exclude periods shorter than 50 s, the Global CMT Project method avoids most effects of structural heterogeneity. This comes at a cost: at long periods, the ground motion due to dip-slip components trends toward zero for a shallow source. Therefore, the GCMT catalogue does not allow the depth parameter to be shallower than 12 km, and many of the earthquakes used in this study from the Global CMT Project database (Ekström et al., 2012; Nettles et al., 2011) had their centroid fixed at a depth of 12 km.

Lateral heterogeneity, even up to that present at the continental scale, can still affect waveforms filtered between 200 and 40 s, and thus influence the CMT solutions. Hejrani et al. (2017) have recently recalculated CMT solutions for 10 years of seismicity in Papua New Guinea and Solomon Islands using a 3-D crust and mantle model of Australia and its surroundings. The newly obtained continental-scale CMT solutions, in comparison with those in the standard Global CMT Project catalogue, have a significantly higher double-couple percentage and there is a considerably improved agreement in the inferred orientation of fault planes with the well-documented structural features that localized the earthquakes. This suggests that on local and regional scales (source-receiver distance <1000 km), a 1-D velocity model can be used to model data at 50–20 s. If the crust is not too heterogeneous, a range of 50–10 s would also work.

The model presented in Zhao et al. (1990) shows that a simple crustal structure exists from the slab interface to the Japanese mainland. This is the region that contains the earthquakes and stations used in this study. Our waveform modeling shows that the Zhao et al. (1990) model is detailed enough to allow higher frequencies to be successfully modeled. We have therefore computed Green’s functions for a layered 0.1° spaced latitude-longitude grid, with a depth interval of 4 km. Best fit was determined based on waveform analysis (see below).

The relocated centroids (Fig. 4A) confirm the abrupt nature of the change in earthquake mode across the lineaments, and improved the definition of some orientation groups (Figs. 4B and 4C), lessening the scatter of slip lines and fault plane poles on the stereoplots. Significantly, this reveals well-defined orientation clusters, again with normal faults dipping ∼60° landward. Further, the depth relocation allows a cross section (Fig. 4D) that reveals that the extensional channel is bounded at depth by the subducting Pacific Plate. In addition, the depth relocation of the cluster of normal faults at the trenchward-end of the extensional channel (Fig. 4) makes it evident that these earthquakes initiated within the down-going Pacific slab, implying that these faults cut and offset the megathrust, thus enhancing the role of gravity in driving motion of the overlying crust (McKenzie and Jackson, 2012). The slab faults also dip consistently landward, with slip vectors having the same trend as thrusts in the adjacent compressional buttresses.

Our results open the question as to why lineament-bounded extensional channels mark segments of the East Japan megathrust with different geodynamic behavior to that of the adjacent compressional segments. This pattern implies that movement in the extensional channel was driven by seaward gravitational collapse of the Japanese crust, supporting the conclusions of previous authors (Cubas et al., 2013; McKenzie and Jackson, 2012; Scholz, 2014; Tsuji et al., 2013). Moreover, the implied switch in geodynamic mode represents abrupt transitions from mode-I to mode-II behavior on the megathrust, where: (i) mode-I involves the classic “push from behind” action (Scholz, 2014), and leads to crustal shortening; and (ii) mode-II requires movement on a weakened basal thrust, with motion associated with concomitant crustal extension in the overriding crust (Lister and Forster, 2018). In the latter circumstance, motion above the megathrust must be driven by gravity, at least during the aftershock sequence, and this requires the rupture to have offered negligible resistance.

Since the mode of deformation involves margin-orthogonal compression in the Tohoku-oki and Joban buttresses (i.e., mode-I megathrust behavior) to the north and south of the extensional channel, and changes to margin-orthogonal extension within it (i.e., to mode-II megathrust behavior), the relevant question is what weakened the basal megathrust between the two compressional buttresses. It is difficult to explain the margin segmentation without calling on the effects of a deep crustal structural control. The effects of deep-seated fluid activity could reduce effective stress on the overlying megathrust, and produce lubricating mineralogy as the megathrust slowly unlocked in the decade preceding catastrophic failure (Johnson et al., 2016; Mavrommatis et al., 2014, 2015; Yokota and Koketsu, 2015). The migration of fluids and magma upward along these deeper structures would explain why the volcanic lineaments developed in the first place (Tatsumi and Eggins, 1995). Fault splays from such deeper structures would also explain why the northern and southern boundaries of the extensional channel are offset ∼15–20 km from the Miyagi and Fukushima Volcanic Lineaments.

An additional aspect needs to be considered, however. The cluster of normal faults at the trenchward-end of the extensional channel (Fig. 1E) appears to have initiated within the down-going Pacific slab, implying that these faults cut and offset the megathrust (McKenzie and Jackson, 2012), thus enhancing the role of gravity in driving motion of the overlying crust. The slab faults dip consistently landward, with slip vectors having the same trend as thrusts in the adjacent compressional buttresses. Although outer-rise normal faults are usually interpreted as the consequence of bending the slab as it enters the subduction zone, the consistent landward-dip of these structures suggests the existence of a slowly slumping slab sheet (Fig. 5), with a strike dimension comparable to the width of the extensional channel.

The stereoplots (Figs. 1 and 3) show that the headwall master faults dip landward, with the overall geometry comparable to that of the arrays of tilt-blocks due to margin orthogonal horizontal extension in the Basin and Range Province of the continental United States (Davis et al., 1986; Lister and Davis, 1989). Down-dip movement on these structures may connect (via a basal detachment) to the zone of seismicity that extends to ∼150 km depth in the lowermost of the pair of seismic zones recognized in the down-going slab beneath East Japan. Data from CMTs can be interrogated to similarly allow inference as to the orientations of rupture fault planes and associated slip lines at these depths, allowing structural analysis in the double seismic zones. Interestingly the inferred orientations of the principal stress axes are consistent with a slumping slab sheet, particularly if Coulomb-Mohr failure is assumed at the “cold” top of the subducting slab, and ductile failure in its depths where the slab is dewatering (Peacock, 2001). Earthquakes in the double seismic zone have been shown to have gently dipping attitudes (e.g., Warren et al., 2007; Kita et al., 2010) with a geometry that is consistent with semi-brittle failure of a stretching slab-sheet with limited ductility.

Fault planes for intermediate depth earthquakes in the double seismic zone are never parallel to the locus of the double seismic zones (Fig. 6), implying that aseismic movement takes place both on the slab interface (where it couples to the weak asthenosphere) and at depth, where the lowermost of the doubled-planed seismic zones reflects material weakness. In this case, the behavior of the slumping slab-sheet is no different to what is observed at mesoscopic scale in most crustal-scale ductile shear zones. Figure 6 shows the geometry of the subhorizontal fault planes that can be inferred using CMT data, consistent with the analysis of Warren et al. (2007) and with the locus of high resolution relocated hypocenters (Kita et al., 2010). These data suggest that slip on both interfaces is aseismic, but that the slab sheet itself fails from time to time due to stretching as it slumps down the slab interface.

It should also be noted that the spatial coincidence between the extensional channel and the swarm of outer-rise normal fault earthquakes in the adjacent subducting Pacific slab suggests dynamic interactions between structures in the down-going slab with structures in the overriding Japanese lithosphere. The longevity of the volcanic lineaments implies that the dominant structural control might well have been exerted by mechanical properties of the overriding plate throughout the five million years during which volcanic activity has occurred in the observed lineaments (Tatsumi and Eggins, 1995). Mechanically, the implications of our observed juxtaposition of regions with distinctly opposing stress orientations is important, since it requires abrupt spatial switches from margin-orthogonal horizontal shortening, to margin-orthogonal horizontal extension. This increases the rate of deviatoric stress build-up on adjacent megathrust segments, which explains the repeated focusing of the onset of failure.

In any case, the extensional Tohoku-oki earthquake channel at surface levels is limited in its extent, so it cannot be explained by general horizontal extension as would be possible if there was subduction roll-back. A more local explanation is required. We suggest that movement on outer-rise normal faults created a gravitational potential well, and this drove seaward flow of the Japanese crust in the adjacent extensional channel (cf. McKenzie and Jackson, 2012). This does require reduced basal friction on the megathrust to allow this segment to act in this way. Perhaps the slab dewatering from the double seismic zone below lubricated the fault plane (through the development of soft mineralogy) or lenses of pressured pore fluid may have reduced the effective stress in the extensional segment of the megathrust, thus maintaining its weakened state.

To understand why the segment boundaries have focused the onset of the largest earthquakes in the past four decades in East Japan, we need to bring geology into the equation, at the scale involved in modeling subduction zone interactions (Bassett et al., 2016; Freed et al., 2017; Hashima and Sato, 2017; Hu et al., 2016). Geological architecture responds to loading just like any man-made construction. A failure that leads to catastrophic release of energy in an overloaded structure can be triggered by a stress concentration associated with movement on a simple weakness. Irrespective of the magnitude of the displacements involved, such structures are often expressed as geomorphic features (e.g., the volcanic lineaments). Less obvious seismotectonic lineaments coincide with the boundaries of the extensional earthquake channel, even though these delineate fundamental structural switches in behavior and act as controlling features that determine the earthquake geology. If there are faults associated with these bounding seismotectonic lineaments, they outcrop in valleys adjacent to the volcanic lineaments.

A key finding is that the sense of movement on the southern boundary of the Tohoku-oki buttress reversed prior to the catastrophic earthquake of 11 March 2011. Movement across the southern boundary of the Tohoku-oki buttress involved left-lateral strike-slip displacement in the decade prior to the 2011 Great Earthquake, but the sense of relative motion became right-lateral as the buttress accelerated seaward in the days prior to the main shock. Precise monitoring across the lineaments is thus essential, for this change in movement pattern within the geological architecture can be interpreted as a pre-yield phenomenon, implying that its early detection may have been of some use in seismic hazard assessment. Perhaps other lineaments (and associated compressional buttresses) deserve closer attention, e.g., in the Sanriku region (Nomura et al., 2017; Ye et al., 2012). The Sanriku Lineament marks the southern bound of a compressional buttress, with an extensional earthquake channel to its south. It is therefore directly analogous to the structure that localized onset of the 2011 Great Earthquake. Large tsunamis (Tanioka and Sataka, 1996) related to normal faults (Uchida et al., 2016a, 2016b) imply the possible existence of a second slab sheet slump seaward of the Sanriku extensional earthquake channel.

A key issue thus becomes how to determine which aspects of the geological architecture deserve more focus. A key consequence of the 2011 Great Earthquake is that it occasioned a major release of stored elastic energy, and it may take decades or centuries to once again build up deviatoric stress to critical levels on these segments (Nomura et al., 2017). So, while the southern boundary of the Tohoku-oki buttress localized the onset of mechanical failure for the main shock of the 2011 Great Earthquake, and the initiation of several past large earthquakes on the East Japan Megathrust, including the 2005 M_w ∼7.2 Miyagi Earthquake, this bounding lineament may not exhibit stress concentration for a considerable time into the future. This suggests that other lineaments now deserve more attention. The most recent large earthquake offshore East Japan initiated close to the prolongation of the Fukushima Lineament (Fig. 2), for example.

Routine detection of precursor movements in the geological architecture would advance the science of earthquake forecasting (Barbot et al., 2012; Hori et al., 2014; Kaneko et al., 2010; Stuart, 1988). The Japanese Median Tectonic Line is already recognized (Bassett et al., 2016; Ikeda et al., 2009) as a focusing structure, but this is for strike-slip earthquakes. The prolongations of the lineaments of East Japan may be more significant since these attest to the existence of natural weaknesses that repeatedly focus the onset of megathrust failure. Precise geodetic monitoring of relative displacement across the prolongations of such structures is thus warranted in order that continuous signals might inform policy and response.

There are numerous cautionary tales in respect to the “science” of earthquake forecasting. Many argue that it should not be attempted. Risk analysis based on precursor (?) movements in the geological architecture will require the light of decades in order to develop quantitative constraints. Nevertheless, our broader investigation implies that improved accounting for structural controls might well have allowed a few days warning in advance of the March 2011 earthquake, at least providing the opportunity of taking precautionary measures in respect to the operation of major infrastructure, e.g., the Fukushima nuclear reactor. To recognize focused relative displacements on structures within the geological architecture, we recommend increased attention to the geodetic network. There are solutions that would allow monitoring of the dominant lineaments for the types of shuffling movements recognized in this paper, thus providing continuous signals to inform policy and response in this seismically active region.

Structural weaknesses localize the onset of catastrophic failure in any loaded architecture. Our study has demonstrated this relation in detail for the East Japan Megathrust. The visible volcanic lineaments on land are parallel to seismotectonic lineaments. The associated fault structures must be weak, since they are not seismogenic. Mechanically, this allows them to separate segments of the megathrust above which the crust exhibits distinctly different geodynamic behavior, evidenced by the contrasting seismotectonic response of compressional buttresses and intervening extensional earthquake channels. For a decade prior to the 2011 Great Earthquake relative to the Tohoku-oki buttress, the extensional earthquake margin segment to its south accelerated seaward, potentially reflecting an unlocking segment of the megathrust. Three days before the Tohoku-oki Great Earthquake, the Tohoku-oki buttress began to accelerate seaward. The sense of relative movement across the bounding seismotectonic lineament reversed. This shuffling motion on an accommodating structure that focuses margin segmentation may be an intrinsic part of the seismic cycle, and the reversal of shear-sense across the southern boundary of the Tohoku-oki buttress was a precursor signal that reflected the onset of catastrophic failure.

The research reported here was supported by the Satellites, Seismometers and Mass Spectrometers initiative, a joint venture between the Structure Tectonics team and the Geodesy and Geodynamics team in the Earth Dynamics Group, and the Seismology and Mathematical Geophysics Group at the Research School of Earth Sciences. Additional support derived from Australian Research Council Discovery Project DP120103554 and Linkage Project LP130100134. EJR acknowledges Australian Laureate Fellowship FL120100050. We would like to thank Nick Rawlinson and one anonymous reviewer who considerably improved the manuscript through thorough, systematic, and careful review.