A recent deep earthquake doublet in light of long-term evolution of Nazca subduction

Earthquake faulting at ~600 km depth remains puzzling. Here we present a new kinematic interpretation of two Mw7.6 earthquakes of November 24, 2015. In contrast to teleseismic analysis of this doublet, we use regional seismic data providing robust two-point source models, further validated by regional back-projection and rupture-stop analysis. The doublet represents segmented rupture of a ∼30-year gap in a narrow, deep fault zone, fully consistent with the stress field derived from neighbouring 1976–2015 earthquakes. Seismic observations are interpreted using a geodynamic model of regional subduction, incorporating realistic rheology and major phase transitions, yielding a model slab that is nearly vertical in the deep-earthquake zone but stagnant below 660 km, consistent with tomographic imaging. Geodynamically modelled stresses match the seismically inferred stress field, where the steeply down-dip orientation of compressive stress axes at ∼600 km arises from combined viscous and buoyant forces resisting slab penetration into the lower mantle and deformation associated with slab buckling and stagnation. Observed fault-rupture geometry, demonstrated likelihood of seismic triggering, and high model temperatures in young subducted lithosphere, together favour nanometric crystallisation (and associated grain-boundary sliding) attending high-pressure dehydration as a likely seismogenic mechanism, unless a segment of much older lithosphere is present at depth.


Directivity of Event 1
Position of hypocentre, centroid and the last rupture stop indicated that Event 1 was predominantly behaving as unilateral rupture. This is further corroborated here by inspecting directivity. At most of the regional stations used, the P-wave group is dominated by the direct wave (as checked by calculating a point-source impulse response); hence the apparent source duration is directly estimated by subtracting hypocentre time from the time of the P-wave end. In Fig. S1, we plot the apparent duration as a function of the station azimuth. For Event 1 the last rupture stop was located at a distance of L = 58 km from hypocentre, with an azimuth of 150°.
Therefore, in Fig. S1, we compare the observed duration with the duration predicted for such a unilateral model, assuming a horizontally propagating rupture. The only least-square fitted parameter is L/Vr (where Vr is the rupture speed), linearly related with the apparent duration, from which we obtain rupture speed Vr=3.0. The corresponding source duration of 19 s is consistent with the two-point source models. Note that similar (south-east) rupture propagation has been suggested for another deep event of the investigated region, Mw 6.8 of June 20, 2003 1 . Figure S1. Directivity of Event 1. The azimuthal variation of the apparent source duration (crosses) is in agreement with a unilateral model (diamonds).

Back-projection of synthetics
Before back-projecting real data we make a test of station coverage, following prior work 2 (their section 7). A point-source event is supposed to have its epicentre at point (0,0) in a NS-EW grid.
Fictitious stations are situated at the same position as 18 real regional stations of this paper. Also the velocity model is the same. The assumed source depth is 620 km. Two experiments are made, in which we back-project synthetic data onto a horizontal grid situated at the true source depth, Fig. S2a,b,c and also at an erroneous depth (20 km shallower), Fig. S2d. A temporal error is supposed.
In Fig. S2a we assume non-realistically accurate data (error +/-0.5 s); this option makes the station strips quite narrow and enables us to see that the real station coverage is very good.
When increasing the temporal error to +/-3s (in Fig. S2b), the brightness pattern is smoothed and the point source is imaged as an ellipse, ~100 km long, elongated in the SSW-NNE direction (Fig. S2b); this is caused by a small imperfection of the coverage. Nevertheless, the image remains centred at the true epicentre. When the error is +/-1 s, and the back-projection is made at the correct depth (620 km), the source image is good (Fig. S2c), but when we perform back-projection at a grid situated at the depth of 600 km, the same temporal error yields an important distortion (Fig. S2d): the image splits into a major and a minor patch, neither of them having the correct position (0,0). When projecting real data at regional distances, the signal coherence across stations is low, hence a smoothing is necessary; this is equivalent to a larger temporal error in these tests. Therefore, we must expect blurred images of point sources. Moreover, the brightness maxima might be spatially biased due to incorrect depth. Naturally, finite sources bring even more

Back-projection of Event 2
The same technique as used for Event 1 was applied to Event 2 (Fig. S4). The spatial extent of the brightness maxima in Fig. S4e is comparable to Event 1, or even larger. The figures seem to indicate unilateral rupture propagation toward the NE. However, when comparing with the synthetic test in Fig. S2d, it appears possible that the obtained pattern is affected by the inclination of the fault plane. This is why we put a question mark on the NE patch in Fig. S4e; it could be an artefact, because centroid of Event 2 is situated 20 km NW of hypocentre, not toward the NE.
The hypothesis that the NE patch is an artefact has been supported also by an additional test in which we back-projected seismograms on the nodal planes, instead of the horizontal planes. In that case we observed instability of Event 2; indeed, the bright spot was artificially 'moving' updip with progressive time. No instability like that was observed in Event 1. It is likely that Event 2 involved more non-horizontal rupture evolution than Event 1, but such a non-horizontal evolution is hardly resolvable.    (Table 1 of the main text). The station codes appear to the right; for the station map see Fig. 1a of the main text. The NS and EW components of ATAH station were removed from inversion of Event 2 due to instrumental disturbance. Time 0 is the origin time of Event 1.