The 6 February 2023 Mw 7.8 Pazarcık and subsequent Mw 7.5 Elbistan earthquakes generated strong ground shaking that resulted in catastrophic human and economic loss across south‐central Türkiye and northwest Syria. The rapid characterization of the earthquakes, including their location, size, fault geometries, and slip kinematics, is critical to estimate the impact of significant seismic events. The U.S. Geological Survey National Earthquake Information Center (NEIC) provides real‐time monitoring of earthquakes globally, including rapid source characterization and impact estimates. Here, we describe the seismic characterization products generated and made available by the NEIC over the two weeks following the start of the earthquake sequence in southeast Türkiye, their evolution, and how they inform our understanding of regional seismotectonics and hazards. The kinematics of rupture for the two earthquakes was complex, involving multiple fault segments. Therefore, incorporating observations from rupture mapping was critical for characterizing these events. Dense local datasets facilitated robust source characterization and impact assessment once these observations were obtained and converted to NEIC product input formats. We discuss how we may improve the timeliness of NEIC products for rapid assessment of future seismic hazards, particularly in the case of complex ruptures.

On 6 February 2023, at 01:17:34 UTC, a moment magnitude (Mw) 7.8 earthquake struck the Pazarcık area in the Kahramanmaraş province of south‐central Türkiye, near the northern border with Syria. Roughly nine hours later, at 10:24:48 UTC, an Mw 7.5 aftershock struck 88 km to the northeast, in the Elbistan district of Kahramanmaraş. The earthquakes occurred in a complex region tectonically controlled by the triple junction between the Anatolian, Arabian, and African plates and characterized by predominantly strike‐slip faulting (Fig. 1). The U.S. Geological Survey (USGS) National Earthquake Information Center’s (NEIC) estimate placed the Mw 7.8 hypocenter ∼28 km south of the East Anatolian fault (EAF), with a strike‐slip mechanism consistent with the orientation and left‐lateral sense of motion of the East Anatolian fault zone (EAFZ). The EAFZ is the major plate boundary that accommodates the westward extrusion of Türkiye toward the Aegean Sea. Geodetically derived slip rates along the EAFZ are estimated to be 10±1 mm/yr (Reilinger et al., 2006). The fault zone has caused destructive earthquakes throughout recorded history (e.g., Ambraseys, 1989), including an Mw 6.7 event in 2020, about 200 km northeast of the 2023 Mw 7.8, and at least three early historic Mw 7+ earthquakes along the southeast EAFZ (Taymaz et al., 2021). To the south of the EAFZ, the left‐lateral Dead Sea fault (DSF) accommodates northward motion of the Arabian Peninsula relative to the African and Eurasian plates. Location estimates for the Mw 7.5 earthquake are consistent with the Çardak fault (CF)—a left‐lateral fault that meets the EAF to the east of both the earthquakes near the village of Mestan, Adıyaman (Taymaz et al., 1991; Fig. 1).

The NEIC operates 24 hr/day to rapidly locate and characterize seismicity worldwide (Benz, 2017) and disseminates this event information (“products”) to emergency responders, national and international agencies, news media, and the public. Products are published via the product distribution layer (Guy et al., 2015) and appear on the publicly available USGS earthquake event pages (U.S. Geological Survey [USGS] Earthquake Hazards Program, 2017; see Data and Resources). The USGS earthquake event pages are often the first point of reference for researchers seeking earthquake source characteristic information, and these products are routinely cited in research papers about earthquakes of interest. Following major earthquakes, traffic to the event pages spikes; in the 24 hr after the start of the Kahramanmaraş sequence, there were 1,035,364 visits to the event pages for the Mw 7.8 Pazarcık and Mw 7.5 Elbistan earthquakes. The NEIC team also disseminates details of the near‐real‐time analysis to the public through television, radio, and print media requests. The NEIC’s rapid scientific analysis is utilized by several federal offices and agencies including (but not limited to) The White House, the U.S. Agency for International Development’s Bureau for Humanitarian Assistance, the National Oceanic and Atmospheric Administration, the National Aeronautics and Space Administration and the USGS’s partners in the National Earthquake Hazards Reduction Program; the National Science Foundation, the National Institute of Standards and Technology, and the Federal Emergency Management Agency. The NEIC impact assessments are used by U.S. government agencies and local humanitarian and emergency response agencies as a first‐order indication of the level of response likely required in the earthquake’s aftermath.

Mw 7.8 Pazarcık earthquake

The NEIC published a W‐phase scalar magnitude (Mww) of 7.8 for the 01:17:34 UTC Pazarcık earthquake 11.4 min after origin time. The moment tensor and focal mechanism were published 21.7 min after origin time, indicating motion on a nearly vertical strike‐slip fault striking either southwest–northeast (left‐lateral) or northwest–southeast (right‐lateral) (Figs. 1, 2).

NEIC uses the Wavelet and simulated Annealing SliP inversion code for rapid finite‐fault models (FFMs; Ji et al., 2002; Koch et al., 2019; Goldberg et al., 2022). We inverted broadband teleseismic body and surface waves between 30° and 90° from the hypocenter location. We modeled slip on a simple, planar fault consistent with the left‐lateral orientation of the W‐phase centroid moment tensor (WCMT; strike: 228°; dip: 89°) and pinned to the NEIC hypocenter location. The hypocenter, and therefore the fault plane, was ∼30 km south of the mapped EAF. This simplistic preliminary model showed a compact slip patch extending ∼100 km along strike and ∼60 km along dip (Fig. 2). A typical strike‐slip earthquake of Mw 7.8 ruptures an area of ∼190 km along strike and ∼20 km along dip (Blaser et al., 2010). Historically, this region experiences seismicity above ∼25 km depth, indicating that the deeper slip of the NEIC’s initial FFM was tectonically unlikely (Ambraseys, 1989; Taymaz et al., 1991, 2021; Tan and Taymaz, 2006; Fielding et al., 2013).

The USGS ground‐motion estimation product, ShakeMap (Wald et al., 2022), and loss estimation product, Prompt Assessment of Global Earthquakes for Response (PAGER; Wald et al., 2010), were released at 15.7 and 21.2 min after earthquake origin time, respectively (Fig. 2). The rapid PAGER estimate (Fig. 3a), based on the limited source characterization products available (location, magnitude, and mechanism), population exposure estimates, and knowledge of local construction practices and building vulnerabilities, was orange for estimated fatalities (100–1000 estimated casualties) and red for economic impact ($1000–10,000 million U.S.). The initial summary alert, which is the higher of the fatality and economic loss estimates, was red, indicating an international response‐level disaster (Wald et al., 2010).

Noteworthy updates of the evolving shaking and loss estimates included (1) the initial FFM rupture dimensions (PAGER v.3, 2.7 hr), (2) extended rupture length (v.4, 4 hr), and (3) Department of Earthquake, Disaster and Emergency Management Authority (AFAD) strong‐motion records (see Data and Resources) and a detailed rupture trace from satellite imagery (v.6, 49 hr; Fig. 3b). Commensurate median loss estimates were (1) 2K fatalities and $6.7B, (2) 2.5K fatalities and $6.7B, and (3) 18K fatalities and $33B. The PAGER losses were red alert levels for both fatalities and economic losses beginning ∼4 hr after the earthquake origin (v.4 onward).

Mw 7.5 Elbistan earthquake

The Mww estimate of 7.5 for the 10:24:48 Elbistan earthquake was published 18.1 min after the origin time. The full moment tensor was unstable, possibly due to long‐period noise contamination from the earlier Mw 7.8, and as a result was not released until 42.4 min after origin time, once consistency between the NEIC Mwb (body‐wave moment tensor) and WCMT solutions was observed (Fig. 2). Similar to the Mw 7.8 event, the moment tensor indicated a nearly vertical strike‐slip fault, striking either east–west (left‐lateral) or north–south (right‐lateral). However, the moment tensor had a double‐couple percentage of only 34%, indicating source complexity. Rapid moment tensors from other seismic monitoring centers indicated a poorly constrained magnitude (estimates vary between 7.5 and 7.8) and fault dip, further indicating source complexity (see Data and Resources).

The initial PAGER estimate was available 28.8 min after origin time and was orange for both estimated fatalities and economic loss (100–1000 casualties, $100–1000 million U.S.; Fig. 2). Loss estimates for aftershocks are especially challenging, because they do not account for the cumulative effects of repeated shaking. Furthermore, the first earthquake may cause people to relocate outside, changing casualties relative to the model assumptions. Updates between 2 and 3 hr after origin time only slightly revised the PAGER estimates. The detailed imagery‐based rupture (∼4 days) updated PAGER estimates to red alerts for fatalities and economic losses (v.9, Fig. 2). More accurate early loss estimates may have been possible if those algorithms had considered the shaking throughout the sequence to that point, rather than for a single event in the sequence. This sequence‐specific rather than event‐specific product is in testing and is referred to as a “composite ShakeMap” (Wald et al., 2023).

The preliminary FFM assumed rupture on a single plane aligned with the left‐lateral nodal plane of the WCMT and inverted broadband teleseismic body and surface waves. Although a typical strike‐slip earthquake of this magnitude ruptures an area ∼120 km along strike by ∼18 km along dip (Blaser et al., 2010), our initial FFM indicated a very compact rupture only ∼50 km along strike and ∼30 km along dip (Fig. 2). Because of the low double‐couple percentage of this event, we did not expect the rupture to be well‐fit by a single‐plane model. Teleseismic data are not very sensitive to local complexities (Goldberg et al., 2022), making it critical to gather additional observations to better assess the rupture characteristics of this pair of events.

Earthquake information derived from optical and radar imagery have been increasingly integrated into NEIC response products in the recent years. Optical and radar satellite images were instrumental in defining the fault trace of this sequence, providing detailed rupture mapping that goes beyond a typical NEIC response. Commercial optical satellite imagery of a portion of the affected region became available within 24 hr of the Mw 7.8 event (Fig. 2). The first evidence of surface rupture was observed on WorldView‐3 and GeoEye‐1 (2023 Maxar) optical images acquired on 7 February near the town of İslahiye. From these images, ∼30 km of surface rupture on the EAF was observed and mapped within 24 hr of the start of the sequence. On the third day after the Mw 7.8 earthquake, the Advanced Land Observing Satellite‐2 (ALOS‐2) provided the first radar observations, which spanned the two events but did not image the full spatial extent of the ruptures. The ALOS‐2 swath provided evidence for ∼160 km of rupture on the EAF and ∼90 km of rupture on the CF (Fig. 2). European Space Agency Sentinel‐1 acquisitions acquired 4.1 and 4.6 days after the Mw 7.8 origin time imaged the full extent of both Mw 7+ ruptures, ≥340 km on the EAFZ and ∼175 km on the CF (Fig. 4). Detailed mapping of the surface rupture was updated and posted to a public website map (Reitman et al., 2023) as high‐resolution (<1 m) optical images were acquired in the days following the earthquakes. By 11 days after the Mw 7.8 origin time, the optical images had covered at least 280 km of EAFZ surface rupture, including two splay faults, and ∼150 km of surface rupture on the CF (Fig. 4c). Rupture endpoints of the Mw 7.5 earthquake were confirmed by scientists on the ground for which image coverage was lacking two weeks after the earthquakes, whereas endpoints of the Mw 7.8 rupture remained somewhat uncertain. High‐resolution optical coverage was not yet complete for either rupture one month after the earthquake.

We performed subpixel offset tracking on Sentinel‐1 synthetic aperture radar (SAR) pairs spanning the earthquakes (Fig. 4). We used image pairs from ascending path 14 (28 January 2023–9 February 2023), descending path 21 (29 January 2023–10 February 2023), and ascending path 116 (4 February 2023–28 February 2023) and processed the subpixel offset fields following the approach of Shea and Barnhart (2022). We used the observed displacement discontinuities to infer the approximate locations of fault rupture traces. These observations confirmed that the NEIC epicenter location of the Mw 7.8 Pazarcık earthquake is ∼30 km south of the main rupture trace, indicating that the mainshock began on a northeast‐striking fault, whose location is consistent with the Sakçagöz‐Narlı segment of the DSF (Emre et al., 2018), before it ruptured the EAF. The satellite observations also indicate that the ruptures of the two Mw 7+ earthquakes do not connect at the surface.

Following the fault rupture mapping, we simplified the inferred fault traces to three planar fault segments for each of the two Mw 7+ events (Fig. 5) and divided these segments into 5 km × 5 km subfaults. We use spatiotemporal Laplacian smoothing of slip over neighboring fault segments (Koch et al., 2019; Goldberg et al., 2022). To infer fault dip, we inverted profiles of the downsampled surface displacements derived from the Sentinel path 21 pixel‐tracking results following the approach of Barnhart et al. (2014). We extracted azimuthal and range displacements from 5 km wide profiles spanning 10 km on either side of the fault trace, then inverted the displacements for the best‐fitting dip direction using the Neighbourhood algorithm—a modified grid‐search approach for nonlinear inversions that optimizes multiparameter searches while trying to avoid local minima (Sambridge, 1999; Barnhart et al., 2014).

In the week following the start of the sequence, we retrieved local strong‐motion accelerometer and Global Navigation Satellite Systems (GNSS) data to improve the FFM. We processed the high‐rate (1 Hz) and low‐rate (30 s) GNSS data with the PRIDE PPP‐AR (Positioning Racers to Image and Decipher the Earth Precise Point Positioning with Ambiguity Resolution) software (Geng et al., 2019). Low‐rate data were used to estimate static offsets and later combined with the GNSS static offset estimates of the Nevada Geodetic Laboratory (see Data and Resources). We also inverted the SAR range and azimuth displacement estimates, downsampling the subpixel offset fields to achieve a computationally tractable number of data points (Lohman and Simons, 2005) and cropping these data around each rupture trace. We assume that the displacements within these cropped regions are due only to the Mw 7+ earthquake on that fault trace. Slip inversions of both teleseismic and near‐field observations were published to the USGS earthquake event pages on day 10 after the start of the sequence.

Mw 7.8 Pazarcık earthquake

The final FFM update considers 45 teleseismic broadband P waves, 27 teleseismic broadband SH waves, 72 long‐period surface waves (Rayleigh and Love waves), 13 strong‐motion accelerometer waveforms, 18 high‐rate GNSS displacement waveforms, 53 static GNSS offset estimates, and SAR range and azimuth displacements cropped to surround the Mw 7.8 rupture trace.

Because of the complex fault segment orientations, we bypass the software’s default calculation of rupture onset times, which relies on the straight‐line distance from the hypocenter to each subfault. Instead, we specify the rupture initiation location and time on each segment by measuring the along‐fault distance to each subsequently ruptured segment, to control how rupture transitions from one planar fault segment to the next. For example, we specify that rupture begins on the EAF at the intersection location with the Sakçagöz‐Narlı segment of the DSF and then is allowed to rupture bilaterally along the EAF. We allow rake values between 339° and 19°—a range of ±20° from the WCMT estimated rake value of −1°, consistent with predominantly left‐lateral strike‐slip faulting.

The resulting slip model shows low‐amplitude slip (<2 m) along the Sakçagöz‐Narlı segment (segment 1; Fig. 5a), consistent with the small displacements observed by satellite imagery (Fig. 4a,b). After arriving at the EAF (segment 2), the rupture continues bilaterally, extending ∼125 km to the northeast and ∼35 km to the southwest before turning southward along segment 3, where it ruptures an additional ∼140 km (Fig. 5a). In total, the slipped region comprises about 300 km of along‐strike distance. Slip is largely constrained to <20 km depth, and maximum slip of ∼11 m is located along the EAF, 20 km northeast of the intersection with the Sakçagöz‐Narlı segment. The source time function shows a slow moment release in the first ∼12 s, consistent with the low‐amplitude slip on segment 1. Peak moment rate occurs at ∼24 s after origin time, when the maximum slip asperity is growing on the EAF (Fig. 5b). We estimate the total source duration to be ∼89 s, with later modeled moment release attributed to fitting of waveform noise.

The WCMT estimates utilize waveforms far enough away that the earthquake can be idealized as a point source. Therefore, the WCMT is insensitive to fault orientation complexity, as revealed by the fault rupture mapping. For consistency, we sum the contributions of each subfault in our three‐segment model to demonstrate that the total contribution combines to produce a moment tensor solution consistent with the NEIC WCMT (Fig. 5c).

Mw 7.5 Elbistan earthquake

The final FFM update considers 38 teleseismic broadband P waves, 28 teleseismic broadband SH waves, 48 long‐period surface waves, 24 strong‐motion accelerometer waveforms, 18 high‐rate GNSS displacement waveforms, 40 static GNSS offset estimates, and SAR range and azimuth displacements cropped in the area surrounding the Mw 7.5 rupture trace. We removed the observation at GNSS station EKZ1 (4.5 m horizontal offset), because our model was unable to fit this large observation without unreasonably high maximum slip (>20 m).

Unlike the Mw 7.8 event, each fault segment in this model is connected at its along‐strike edges, so the algorithm’s default method to calculate rupture onset times is sufficient. Because of the complexity of this event, both in the low double‐couple percentage of the WCMT and the inconsistency between moment tensor solutions from international earthquake monitoring centers, we allowed a wide range of rake values, ±75° from a purely left‐lateral rake value of 0°.

The resulting slip model shows a long, bilateral rupture that extends ∼55 km in each direction from the hypocenter and to a depth of ∼20 km (Fig. 5a). The maximum slip of ∼11 m is located just east of the hypocenter, at hypocentral depth. The source time function has a relatively simple triangular shape (Fig. 5b), consistent with most slip occurring on the hypocentral segment (segment 1). Peak moment rate occurs at ∼9 s after origin, and the source duration is ∼37 s, with later moment release attributed to fitting of waveform noise. The cumulative moment tensor calculated from the contribution of the three fault segments is consistent with the independently calculated NEIC WCMT (Fig. 5c).

Sequence seismotectonics

That the aftershock’s magnitude—with estimates from different monitoring centers ranging from Mw 7.5 to 7.8—was so close to the mainshock magnitude of Mw 7.8 seems a distinguishing characteristic of this earthquake sequence. Båth’s law states that the largest aftershock of a mainshock will be, on average, ∼1.2 magnitude units smaller than the mainshock (Båth, 1965). However, doublets and multiplets (triggered earthquakes within ±0.4 magnitude units) are sometimes observed in earthquake sequences. Statistical analysis of the NEIC earthquake catalog indicates an average doublet rate of 0.0694 doublets per mainshock, or a 7% chance that a mainshock will trigger an earthquake of similar magnitude (Felzer et al., 2004), indicating that this sequence behavior is not anomalous. This devastating sequence highlights the value of efforts to capture the likelihood of doublets or multiplets in time‐dependent seismic hazard analysis. Although the NEIC does not currently analyze specific interactions between earthquakes in its response timeframe, all FFMs are published with an accompanying file of the modeled slip distribution for input to Coulomb3—software that allows users to calculate the stress change the earthquake’s slip distribution causes to the surrounding faults (Toda et al., 2011; see Data and Resources). Future publications from users of these products may clarify the stress relationship between earthquakes in this sequence.

The rupture extent and location of both Mw>7 events may have implications for future seismic hazard in south‐central Türkiye. For example, there remains an apparent ∼12 km long gap on the EAF between the 2023 rupture and the 24 January 2020 Mw 6.7 Doğanyol‐Sivrice earthquake rupture (Bletery et al., 2020; Melgar et al., 2020; Taymaz et al., 2021). Similarly, the rupture mapping (Fig. 4a,b) and relocated aftershock distribution (Melgar et al., 2023) indicate a gap between the Mw 7.8 Pazarcık and Mw 7.5 Elbistan ruptures. Further efforts to characterize changes in the stress regime in these apparent gaps are critical for future hazard assessment, and deeper understanding of where and why large ruptures arrest.

Multisegment ruptures

The hypocenter of the Mw 7.8 maps onto a short fault south of the main EAF, likely the Sakçagöz‐Narlı segments of the DSF (Emre et al., 2018). In addition, the Mw 7.8 rupture linked together three segments of the EAF—the Erkenek, Pazarcık, and Amanos segments, as defined in the Turkish seismic hazard model source characterization (Demircioğlu et al., 2018). None of these fault segments, when treated independently, accommodates magnitudes of 7.8 in Türkiye’s seismic hazard model; the largest estimated magnitude of the four segments that ruptured was 7.46 along the Amanos segment (Emre et al., 2018). Given ample literature to indicate that strike‐slip earthquakes terminate at structural complexities along strike (e.g., Barka and Kadinsky‐Cade, 1988), delineating ruptures along geometric segments has been standard practice in many probabilistic seismic hazard analyses. However, the multisegment Mw 7.8 rupture did not observe these segment boundaries, highlighting the importance of incorporating multifault ruptures in probabilistic seismic hazard analyses (e.g., Field et al., 2014).

Considerations for future earthquake response

The availability of local seismic and geodetic data has proven critical for rapid earthquake source characterization and loss estimation. Although strong‐motion accelerometer data from the Kandilli Observatory Regional Seismic Network were available in near‐real time through the International Federation of Digital Seismic Networks, the AFAD strong‐motion data had to be accessed separately in the days following the start of the sequence (see Data and Resources). New data format conversion scripts were written on the fly to facilitate ingestion of the AFAD data into ShakeMap ground motion and PAGER impact assessments. Efforts to standardize these data formats and access methods would allow for more rapid data ingestion in the future.

It took 10 days to publish the slip models that included near‐field data. However, general knowledge of the tectonic setting and pinning to the previously mapped faults can facilitate more realistic slip models even in the absence of local observations. For example, the depth of the seismogenic zone inferred from historical earthquakes could have been used to constrain the allowable depth of the early FFMs. Such a constraint may have resulted in early models with a more realistic along‐strike dimension—a critical parameter for ShakeMap and PAGER updates.

Impact assessment for earthquake sequences that include large, damaging aftershocks is vital, yet remains challenging due to the complicating factors that follow a mainshock (e.g., population sheltering and migration, weakened structures, reduced building exposure due to prior collapses). Given the substantial potential contribution of aftershock shaking to overall losses, we have introduced a composite ShakeMap—depicting the maximum shaking at each location for the entire sequence—to improve loss estimates. Although the composite ShakeMap approach does not accommodate complexities of population migration, improved total loss estimates for the entire sequence are achieved by considering the maximum shaking during the sequence. Total deaths and economic losses estimated with a composite ShakeMap as the hazard input layer to PAGER were 30,000, and $51 billion U.S., respectively.

The rapid characterization of the southeast Türkiye earthquakes, including moment tensors, kinematic slip models, shaking and loss estimates, and detailed rupture mapping has improved our understanding of large, continental, strike‐slip earthquakes and their impact. Iterative integration of many disparate observational datasets facilitated updates to these products, contributing to robust source characterization and impact assessments within two weeks of the start of the February 2023 Türkiye sequence.

Strong‐motion accelerometer data were retrieved from Department of Earthquake, Disaster and Emergency Management Authority (AFAD) of Türkiye through the Turkish National Strong Motion Network (doi: 10.7914/SN/TK), and Kandilli Observatory and Earthquake Engineering Research Institute of Boğaziçi University (KOERI, Istanbul; doi: 10.7914/SN/KO). The CORS‐TR (TUSAGA‐Aktif‐Türkiye) Global Navigation Satellite Systems (GNSS) network is administrated by the General Directorate of Land Registry and Cadastre and General Directorate of Mapping. Preliminary GNSS coseismic offsets estimated by the Nevada Geodetic Laboratory are available for the Mw 7.8 and 7.5 events at and, respectively. The USGS event pages are available at The event pages for the Mw 7.8 and 7.5 events, specifically, are available at and, respectively. Rapid moment tensors from various monitoring centers can be found at Input data and model fits for the finite‐fault models (FFMs), as well as the synthetic aperture radar (SAR) subpixel offsets, are available at doi: 10.5066/P9R6DSVZ (Goldberg et al., 2023). The supplementary figures include the finite‐fault model results published by the National Earthquake Information Center throughout the sequence response period. All websites were last accessed in March 2023.

The authors acknowledge that there are no conflicts of interest recorded.

The authors appreciate the Department of Earthquake, Disaster and Emergency Management Authority (AFAD) seismological data services for providing an initial catalog of aftershocks and phase arrival data acquired from their bulletin resources. Tuncay Taymaz (T. T.) acknowledges the Istanbul Technical University Research Fund, Turkish Academy of Sciences (TÜBA) in the framework for Young Scientist Award Program (TÜBA‐GEBİP), and the Alexander von Humboldt Foundation Research Fellowship Award for financial support and for further providing computing facilities through the Humboldt‐Stiftung Follow‐Up Program. T. T. appreciates Beyza Taymaz’s phenomenal support during chaotic days and sleepless nights, trying to organize international scientific collaborations and to coordinate global media requests. The authors also acknowledge Ben Brooks, Brian Shiro, and two anonymous reviewers for their valuable comments.

Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government.

This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Supplementary data