The Hindu Kush slab break-off as revealed by deep structure and crustal deformation

Break-off of part of the down-going plate during continental collision occurs due to tensile stresses built-up between the deep and shallow slab, for which buoyancy is increased because of continental-crust subduction. Break-off governs the subsequent orogenic evolution but real-time observations are rare as it happens over geologically short times. Here we present a finite-frequency tomography, based on jointly inverted local and remote earthquakes, for the Hindu Kush in Afghanistan, where slab break-off is ongoing. We interpret our results as crustal subduction on top of a northwards-subducting Indian lithospheric slab, whose penetration depth increases along-strike while thinning and steepening. This implies that break-off is propagating laterally and that the highest lithospheric stretching rates occur during the final pinching-off. In the Hindu Kush crust, earthquakes and geodetic data show a transition from focused to distributed deformation, which we relate to a variable degree of crust-mantle coupling presumably associated with break-off at depth.


Content
• Supplementary Table 1 Detailed explanation of input data subsets.

Supplementary Note 1. Joint inversion comparison to previous tomography studies
The comparison to either local or teleseismic tomography that had been calculated based on pre-2017 datasets is shown in Supplementary Fig. 2 In general, the velocity anomalies resolved in this study appear smoother compared to those resolved in other studies. This effect may arise as the grid geometry and smoothing and damping parameters in the joint tomography were tuned based on the combined local and teleseismic data.
However, compared to previous studies, only the joint inversion approach is capable to recover both, crustal and mantle anomalies. , no noise is added to the synthetic data and hypocentres are not disturbed. This is because the checkerboard test serves mainly as a proxy for ray coverage but does not resample realistic velocity anomalies. Results show a decrease of amplitude intensity at mantle depths. This is due to the relatively small nature of the implemented anomalies (80 km horizontal extent), which are resolved by teleseismic rays only at these depths. Synthetic tests with more realistic anomaly configurations (Figs. 3 and 4) showed a better amplitude recovery. White outline, blue/pink/purple circles and triangles represent resolution limit, local earthquakes and seismic stations as in Figure 2. Political boundaries are plotted in grey. The contours of the input synthetic model are highlighted in blue (positive anomaly) and red (negative anomaly).
Supplementary Fig. 4 Comparison of crustal event catalogues. a) Comparison of our crustal event catalogue to ref. 6 , a USGS compilation of published historic earthquake sources, and to ref. 7 . The ref. 7 catalogue was derived from a temporary seismic network located mainly in Tajikistan. The USGS compilation includes earthquakes of magnitude 5.5 and larger during 1900-63, and events with smaller magnitude since 1964. Dependent on quality, the USGS compilation groups events in three quality classes. For the best quality class, depth and horizontal location uncertainties are independent and less than 10 km. The worst quality contains events, which hypocentres have uncertain or unknown accuracy. The intermediate quality class includes all other events. Comparison of the herein presented event catalogue to the USGS compilation with longer observation duration suggests that the crustal seismicity pattern depends little on the observation period. Events highlighted in red are earthquakes derived in this study with location uncertainties larger than 10 km. b) Magnitude histograms for the three different event catalogues compared in a). The herein presented event catalogue contains more earthquakes at smaller magnitude compared to the USGS compilation, which highlights the advantage of a local seismic network in detecting such events. c) Location uncertainty histograms of the event catalogue derived herein, separated in vertical and horizontal errors. . All other plotted features as in Fig. 5b. b) Results of fault plane data inverted for stress tensors in the two subregions indicated in Supplementary Fig. 5a, plotted as beach-ball representations with highlighted compression-axes (P) and extension-axes (T). Small colour-coded circles show the spread of P-axes, T-axes and Null-axes (N) derived from bootstrap tests. The spread of these solutions is a measure for the robustness of the solution.
Supplementary Fig. 8 Trade-off curves between data variance and model variance dependent on different smoothing and damping values used in the inversion. Data variance is calculated from the travel time residuals after the final inversion step, model variance is the variance of the final velocity model in domains with hit-quality larger than 0.35. Each circle represents results of an individual model inversion with the colour-coding representing the smoothing value used. Note that local earthquake input used for the derivation of the L-curves varies slightly from the models shown in Figure 2, but output models are almost identical. The final set of parameters (highlighted as yellow star) was chosen as the best compromise between model roughness and data variance reduction. Insets show P-wave residuals prior to (grey) and after the final iteration (red) of the models shown in Figs. 2 and 3a. a) Model based on local data only. b) Combined teleseismic and local (joint) inversion. c) Model based on teleseismic data only. Data first published here. An event catalogue was automatically derived following the location chain of ref. 8 , then selection criteria were applied to obtain a high quality subset: -at least 2 S picks and 5 P Picks with quality class better than 3* -no rms greater than 1.5 s using the 1D velocity model of ref. 2 -no P-picks with quality class 3* As for the TIPAGE and TIPTIMON catalogues, additional declustering between ~160 km and 220 km depth was implemented.
After applying these selection criteria, the entire dataset was visually re-inspected in SeisComp3 and further picks were added. This step was introduced to increase the number of picks at stations in Afghanistan. * see ref. 8 for a description of pick quality classes.