The 2023 US 50-State National Seismic Hazard Model: Overview and implications

A recent joint study led by the US Geological Survey (USGS) and Federal Emergency Management Agency (FEMA) forecasted direct average economic losses of $14.7 billion per year from ground-shaking–related damage to buildings across the United States (Jaiswal et al., 2023). The USGS National Seismic Hazard Model (NSHM, Petersen et al., 2020) forms the underlying ground-shaking hazard layer applied in this risk assessment and is made by developing the nation’s fault and fold deformation databases, interpreting crustal and volcanic seismicity, and evaluating the probabilistic ground-shaking levels for various earthquake sizes, distances, tectonic regimes, and site conditions. The projected rates of earthquakes and economic risk in some urban areas pose a threat to citizens and infrastructure. The USGS has continued to develop the NSHM over the past five decades and revises these models regularly to reflect newly published earthquake science on earthquake hazards (e.g. Algermissen and Perkins, 1976; Frankel et al., 1996, 2002; Petersen et al., 1996, 2008, 2014, 2015a, 2020, 2021 for the conterminous United States (CONUS); Wesson et al., 2007 and Appendix D of Electronic_Supplement_1.pdf for Alaska; and Klein et al., 2001 and Petersen et al., 2022 for Hawaii). In addition, the NSHMs are calculated for US territories (Puerto Rico: Mueller et al., 2010; Guam and Mariana Islands: Mueller et al., 2012; American Samoa: Petersen et al., 2012) that also face significant risk from earthquakes; the NSHMs for these territories are planned to be revised following this 50-state update. In this article, we describe the basis of the 2023 NSHM that is available publicly on the GitLab website (source models (HI: Powers and Altekruse, 2022b; CONUS: Powers and Altekruse, 2022a; AK: Powers and Altekruse, 2023) and nshmp-haz computer code (Powers et al., 2022)), OpenSHA website (Field et al., 2003), USGS ScienceBase Catalog (HI: Rukstales et al., 2021; CONUS and AK: Petersen et al., 2023), and through supporting information (e.g. logic trees) from papers listed in the references and Electronic_Supplement_1.pdf. The 2023 NSHM represents a living model which for this version applies the best available science as of the end of calendar year 2023. Any minor adjustments needed for policy applications (e.g., applications of V_S30 dependent tapered adjustments to central and eastern U.S. ground motion models and Alaska recurrences that are more compatible with geologic observations) will be updated and tracked via the USGS ScienceBase data releases and nshmp-haz computer code version numbers.

The NSHM seismic hazard products provide information needed to save lives and property through application in life-safety–based building design codes; by assisting the public and industry in identifying, preparing for, and mitigating against economic and infrastructure losses; and by supporting functional recovery and resilient community development through effective disaster management and seismic safety planning. Earthquake resistant building provisions rely on the NSHMs to guide construction of new and existing buildings, highways, bridges, railroads, dams, pipelines, energy systems, and hazardous waste disposal. Besides application in building codes, these models have also influenced policy and industry safety and continuity plans including FEMA risk assessments and mitigation grants, governmental and military disaster management strategies, state-wide and local government seismic safety plans and legislation, critical infrastructure assessments, business and industry planning tools, and insurance risk applications. Given the importance of these applications, we routinely engage and foster evaluation and feedback from the end-users of the NSHMs. For example, as part of this update, we hosted a user-needs workshop and several other user-specific meetings with participation from leaders from industry, academia, government, and insurance to enfranchise a larger user base.

Earthquakes are quite common across the United States with most activity focused within the western United States (WUS), Alaska, and Hawaii but with occasional earthquakes scattered across the central and eastern United States (CEUS). The seismic activity rates and ground-motion levels are thought to differ between the CEUS and WUS because of varying tectonic processes and distance to the plate margin, earthquake source and attenuation characteristics, and underlying geologic conditions. For example, CEUS earthquakes typically have larger stress drops and can be felt at larger distances than comparable sized WUS earthquakes due to differences in crustal properties (USGS, 2018). Disparities in earthquake faulting mechanisms (strike-slip, normal, or reverse motions) can also result in regional variability in ground-shaking levels. More than half of the United States (27 states) have experienced potentially damaging earthquakes of moment magnitude (M)5+ since 1900 according to the USGS Advanced National Seismic System Comprehensive Earthquake Catalog (USGS, Earthquake Hazards Program (ComCat), 2022; depicted as shaded regions in Figure 1). An additional 10 CEUS states have historical records of pre-1900 activity of M5+ earthquakes (e.g. depicted with hatched lines in Figure 1). Only 13 states have not experienced earthquakes above M5 in ComCat (depicted with lighter shading). Therefore, 37 of the 50 states (74%) have some past record of significant earthquake activity and face the potential of future damaging shaking. Understanding the historical and prehistoric earthquakes is important because these data are used to forecast locations, rates, and ground shaking of future earthquakes.

Figure 1 highlights several large or significant earthquakes that have occurred over the past few hundred years across the United States (described below). The largest US earthquakes are produced from plate convergence in subduction processes along the Pacific Northwest Cascadia subduction zone (M∼9 1700 (offshore Washington, Oregon, California)) and Alaska-Aleutian Arc (M9.2 1964; M8.7 1965; M8.6 1946 and 1957; M8.3 1906; M8.2 1938 and 2021; and M8.0 1949, 1985, 1986, and 1995). Large damaging earthquakes are also produced along major plate boundary transform (strike-slip) faults spread mostly across California (M7.9 1857 (Fort Tejon); M6.8 1868 (Hayward); M7.4 1872 (Owen’s Valley); M7.9 1906 (San Francisco); M6.9 1989 (Loma Prieta); M7.3 1992 (Landers); M7.1 1999 (Hector Mine); M7.2 2010 (El Mayor Cucapah, northern Baja, Mexico); M7.1 2019 (Ridgecrest)); and Alaska (M7.9 2002 (Denali)); reverse and strike-slip faults in California (M7.5 1952 (Kern County); M6.7 1994 (Northridge)); normal and strike-slip Basin and Range earthquakes (M6.9 1915 (Pleasant Valley, Nevada); M7.3 1932 (Cedar Mountain, Nevada); M7.3 1954 (Fairview Peak, Nevada); M6.9 1954 (Dixie Valley, Nevada); M7.2 1959 (Hebgen Lake, Montana); M6.9 1983 (Borah Peak, Idaho)); and along ancient re-activated tectonic margins in the CEUS, including three events near New Madrid, Missouri (M∼7.3–8.0 1811–1812) and an earthquake near Charleston, South Carolina (M∼7 1886). Geologic evidence of large prehistoric earthquakes has also been identified at sites across the United States, including large fault offsets and prehistoric evidence of damaging earthquakes along the San Andreas Fault, Cascadia subduction zone, Wasatch fault, Walker Lane, and Central Nevada seismic zone, as well as from liquefaction features dispersed across the central and northeastern parts of the country. The Island of Hawai‘I has also experienced large, damaging volcanic and fault-based earthquakes (M7.9 1868; M7.7 1975; M6.9 2018). The US territories have experienced large and damaging earthquakes in Puerto Rico (M7.7 1943; M6.4 2020), near Guam (M7.8 1993), and a tsunamigenic earthquake affecting American Samoa (M8.1 2009).

Along with the large earthquakes on plate boundaries, many smaller- to moderately large-sized earthquakes (3.0 ≤ M < 7.0) have ruptured along known or previously unidentified faults across the United States and can cause damaging levels of ground shaking over more limited areas. The maps and scale bar charts in Figure 2 show the locations and numbers of earthquakes of M3+ through time for four regions (CEUS, WUS, Alaska, Hawaii). Moderately large-sized earthquakes typically occur in regions that historically have been very active (e.g. WUS, Alaska), but they can also occur in places where earthquakes have not been common during the recent past (e.g. CEUS, 2011 M5.8 Mineral, Virginia; 1755 M∼5.9 offshore of Cape Ann, Massachusetts; and WUS 2022 M5.7 Magna, Utah). Although CEUS moderate-size earthquake rates are typically lower than other places, this is not always the case and rates substantially increased over the past decade due to thousands of induced earthquakes triggered by industrial fluid injection of wastewater with a peak about 2015 (Figure 2). Such induced earthquakes can cause damaging ground motions (e.g. 2011 M5.3 Trinidad, Colorado; 2011 M5.7 Prague, Oklahoma; 2016 M5.8 Pawnee, Oklahoma earthquakes). The induced earthquakes are not included in these policy-based models because they are generally transient features of the seismicity, and mitigation actions can translate into rapid changes in earthquake rates. Instead, we supplement these NSHMs with short-term earthquake shaking forecasts that account for industrial injection and hydrofracturing based earthquakes (Petersen et al., 2015b, 2016a, 2016b, 2017, 2018). Short-term (1 year) model forecasts have not been made since 2018 but may be reconsidered if seismicity rates ramp up substantially again. Even though induced seismic activity has subsided in many areas over the past few years due to management efforts, seismicity remains higher than normal in Oklahoma, Kansas, and the Permian Basin of Texas (refer to Appendix B of Electronic_Supplement_1.pdf). WUS earthquakes continue with persistently high numbers that fluctuate in time due to aftershocks following large earthquakes. The Alaska seismicity reflects earthquakes on a major subduction interface along the Aleutian arc, deeper intraslab events in the subduction zone, shallow crustal fault events, and volcanic activity. Earthquakes in Alaska appear to increase with time (Figure 2), but this observation likely is an artifact caused by the increased number of deployed seismographs rather than being associated with any underlying physical changes. The numbers of earthquakes across Alaska are quite high compared to other regions. The Island of Hawai‘i has recently had high earthquake counts, mostly due to recent volcanic eruptions of Kilauea and Mauna Loa. Figure 2 also shows the 2023 NSHM averaged gridded smoothed seismicity models for each of the 50 states along with M4+ earthquakes in the WUS, Alaska, and Hawaii and M2.7+ in the CEUS that are used to produce the seismicity models.

The USGS NSHM Project applies the new science on the potential locations, sizes, and rates of future earthquakes as well as the probable strength of ground motions and integrates this information into seismic hazard curves, ground-motion maps, and other supporting information. These hazard assessments are constructed using well-accepted time-independent probabilistic seismic hazard analysis (PSHA, Baker et al., 2021; Cornell, 1968) that allow for consideration of a broad range of alternative earthquake models available in the science and engineering communities. Because aftershocks are not an independent Poisson process as assumed in the standard PSHA methodology (Cornell, 1968), the NSHM hazard values may not be appropriate for return periods much less than about 475 years in some applications (Field et al., 2023). This could be corrected with additional research on time-dependent spatially and temporally dependent earthquake cluster models (Field et al., 2021). Difficulties in understanding prehistoric earthquake data also make it challenging to forecast seismicity beyond about 10,000 years without extensive fault specific studies as implemented in siting of nuclear facilities or important dams. Therefore, our PSHA estimates are best defined for return periods between about 475 and 10,000 years. Earth processes are naturally complex, and earthquakes are very difficult to forecast over short times, so our analyses here focus on long-term forecasts—50 years with probabilities of exceedance ranging from 2% to 10%.

The 2023 NSHM is based on a new earthquake rupture forecast (ERF) that describes the probabilistic locations, sizes, and rates of potential earthquakes and a ground-motion model (GMM), which defines probabilistic levels of ground shaking conditioned on an earthquake rupture and site characteristics. The ERF are described in detail by Field et al. (2023) for CONUS, this paper for Alaska, and Petersen et al. (2022) for Hawaii. The GMMs are described by Moschetti et al. (2023a) for an overview of CONUS, Rezaeian et al. (2023) for Cascadia subduction, this paper for Alaska, and Petersen et al. (2022) for Hawaii. This seismic hazard assessment considers several new ERFs and GMMs that have been developed since the previous NSHM update:

1. Seismicity-based earthquake grids are founded on new earthquake catalogs compiled with alternative new declustering algorithms, induced-seismicity identifications for CEUS, and gridded (smoothed) seismicity assessments (Field et al., 2023; Petersen et al., 2022; this paper for Alaska). The 2023 model considers full-catalog rates (as recommended by the ERF panel) with spatially smoothed seismicity kernels from declustered catalogs, rather than only using the declustered catalogs as in previous NSHMs.

2. Fault-based deformation models include new geologic (Hatem et al., 2022a, 2022b) and joint geologic and geodetic inversions (Pollitz et al., 2022; this paper for Alaska).

3. New magnitude-scaling equations provide estimates of sizes of earthquakes for shallow crustal, deep crustal, subduction, and stable continental environments (Shaw, 2023).

4. A new inversion-based methodology for inclusion of multi-fault ruptures, including a wider range of epistemic uncertainties, was applied across the WUS region (Milner and Field, 2023). New traditional (classic) rate models were developed for the CEUS (Shumway et al., 2023), Alaska (this paper), and Hawaii (Petersen et al., 2022).

5. New semi-empirical Next Generation Attenuation (NGA)-Subduction GMMs are implemented for pseudo-spectral accelerations (SAs) and peak parameters (Bozorgnia et al., 2022; Rezaeian et al., 2023). New region adjusted coefficients were implemented for Alaska.

6. New CEUS adjusted revisions to the stable continental GMMs (Moschetti et al., 2023a) are also implemented.

7. New three-dimensional (3D) ground-motion simulations for the Seattle and Los Angeles regions are considered in assessing ground motions from earthquake ruptures that have not been instrumentally recorded (Frankel et al., 2018; Moschetti et al., 2023a; Wirth et al., 2018a).

8. Basin-depth amplification models based on NGA-West2 are constructed and applied for the San Francisco Bay Area, Los Angeles, Great Valley of California, and Portland/Tualatin regions (Ahdi et al., 2023; Moschetti et al., 2023a).

9. New amplification models are constructed and implemented for the Atlantic and Gulf Coastal Plains that account for sediment thickness (Boyd et al., 2023).

10. New CEUS-WUS boundaries are developed and applied for earthquake catalogs, fault models, and GMMs.

The component models and overall hazard model updates were reviewed by several expert panels, together consisting of more than 50 scientists and engineers.¹ These expert panels deemed the new components to be reasonable for inclusion in the 2023 NSHM, and furthermore that they account for a more complete range of uncertainty than previous seismic hazard models (ERF review—Jordan et al., 2023, GMM review—Electronic Supplement of Moschetti et al., 2023a, NSHM Steering Committee review—Electronic_Supplement_1.pdf, Appendix H). The NSHM Steering Committee (NSHM-SC), which oversaw the review process, was requested to provide feedback on (1) whether the model was adequately reviewed, (2) whether the USGS response appropriately addressed the review comments and recommendations, and (3) whether the model is suitable for release and to serve as the basis for hazard mitigation. It was the consensus of the NSHM-SC that the level of technical review for the conterminous US, Alaska, and State of Hawaii were adequate, that the USGS responded acceptably to most panel comments and recommendations, and that the models are suitable for use in building code and similar applications at return periods of 475 years and longer. The NSHM-SC and the Alaska Review Panel reviewed the ERF and GMMs for Alaska and found them to be suitable for use in policy applications, subject to additional documentation and verification (Electronic_Supplement_1.pdf, Appendix H).

An example 2023 seismic hazard map developed as part of this study is shown in Figure 3 that is based on Modified Mercalli Intensity (MMI, a macroseismic measure of earthquake effects and damage that describes the strength of earthquake shaking inferred from intensity observations). The probabilistic seismic hazard and risk is quite variable across the United States. Earthquake hazard is high along West Coast population centers due to potential of earthquakes associated with the Cascadia subduction zone and San Andreas Fault System, across the densely populated Wasatch Front and up through the Teton Range/Yellowstone region, within the Walker Lane and central Nevada seismic belt, near the New Madrid seismic zone (NMSZ) and Charleston seismic zone, Aleutian Island chain of southern Alaska, and across the southern Island of Hawai‘i. Moderately high hazard and risk is found across many of the WUS states, and areas of pre-existing faulting across the CEUS, central Alaska, and central islands of the Hawaiian chain. Lower hazard and numbers of earthquakes are observed across northern and central portions of the Midwest and southern regions of the country as well as in the northern portion of Alaska and farthest northwest islands of Hawaii. Nevertheless 26 states have moderate to higher probabilities of shaking - 25 percent or reater chance of experiencing damaging shaking (MMI VI+) in 100 years (Figure 3).

One of the greatest challenges in constructing this 2023 NSHM has been to evaluate and then implement only the best available or applicable science into the policy model. Best available science is defined as including well-vetted and published hazard input component models with adequately documented assumptions, is consistent with open and timely science principles, encompasses a scientifically reasonable range of earthquake ruptures and ground-motion effects that improve the basis of the 2023 NSHM, assesses variability and uncertainties in such information, and ensures that the end-product is reasonable for and consistent with the intended purposes (e.g. Title 40 CFR § 702.33). The ERF review panel suggests the following elements need to be included as part of the best available/applicable science: relevance, inclusiveness, objectivity, transparency and openness, timeliness, and peer review; in their report, objectivity is replaced with verification and validation (Jordan et al., 2023). Best science evaluations require each of these science-based criteria and include well-vetted and published hazard information and models that are accepted through a comprehensive review process and encompass a reasonable scientific range of earthquake characteristics and effects for updating the 2023 NSHM. This version of the NSHM is meant for seismic policy applications, and we attempt to adhere to the definitions given above for best applicable science. In terms of peer review, openness, transparency, inclusiveness, and objectivity, the NSHM input models included in this update are evaluated with a thorough (over 50 reviewers), open (dozens of public workshops and a public comment period), and transparent review process (tools, data, and documentation available on NSHM Project website). In addition, the model incorporates comprehensive and defensible representations of epistemic uncertainty (lack of knowledge that is reducible through additional studies) and aleatory variability (intrinsic randomness in the earthquake process). Through scientific studies, we can sometimes convert apparent variability into epistemic uncertainty (e.g. consideration of non-ergodic GMMs). We plan to follow this policy-based study with research and development models that explore additional input parameters and models that were not considered in previous NSHMs so that end-users can evaluate impacts and explore implementation issues (refer to section “Discussion”).

To confirm the models are based on the current best science, that the results are reasonable, and the products are useful, the resulting NSHM outputs were tested using historical intensity data, evaluated based on published inclusion criteria (Rezaeian et al., 2015), and appraised extensively by the science and engineering communities through more than 25 public workshops and the publication process. The draft model was available in written form and in a public workshop for comment by any interested person during an open-review period in April and May 2023. The model was examined and evaluated by hundreds of earthquake science and hazard experts who participated in regional and topical workshops and members of several review committees (see Note 1) who analyzed details of component models. The NSHM research strategies have benefited extensively from direct involvement of the NSHM-SC, consisting of nine renowned seismic hazard experts—who review inputs and outputs of models, recommend model improvements, promote comprehensive uncertainty assessments and testing of the model, and advocate for improved science directions.¹ Reviews by the ERF review panel (Jordan et al., 2023), GMM review panel (Moschetti et al., 2023a Electronic Supplement), Alaska review panel (Appendix G of Electronic_Supplement_1.pdf), Tiger Team (also known as the GMM advisory panel; Appendix F of Electronic_Supplement_1.pdf), Hawaii review panel (Petersen et al., 2022), and NSHM-SC (Appendix H of Electronic_Supplement_1.pdf) are available for the public’s consideration.

In terms of timeliness for consideration of input components for the 2023 NSHM, any ERF and GMM models published before 31 December 2020, deadline were given full consideration for inclusion in the model. Information that became available after this date may or may not have been considered for model development. NSHM Project scientists, in consultation with the NSHM-SC, exercised their discretion and judgment regarding utilization of recent results, and do so when justified in terms of the result’s scientific credibility and significance to hazard. Concerns and objections were expressed by some reviewers dealing with the following issues: inclusion of input models that came in late, lack of adequate documentation during the review process, lack of adequate time for review, and desire for more discussion of review comments that were not accepted. These objections and the possibility that the model may be premature in some regions were considered by the review panels and NSHM project members. In those cases, we followed the consensus (but not unanimity) of the review panels and project members in making these decisions of whether to include models or not. USGS personnel had to make difficult decisions about which review comments were feasible for implementation and did not conflict with other guiding principles of the model development (e.g. weighting and epistemic uncertainty considerations).

This overview paper is written for a variety of audiences, and the scope is broader than in many papers. The introduction and overview of methodology sections were written for the informed scientists/engineers who may be interested in why we calculate seismic hazard, where damaging earthquakes occur, the types of data that are used to assess hazard, the intensity measures of hazard, and a description of how we can use these hazard maps and mitigate the effects of future events. The ERF and GMM sections are technical and written for scientists and engineers who want to understand the breadth of issues with input models—pointing to other companion papers for details, how models are derived and evaluated, and the basis for weighting models. The results section is written for the users of 2023 NSHM with descriptions of where, why, and how much the hazard has changed compared to the previous NSHM versions. The discussion section describes hazard issues that the project members, review committees, and other scientists debated including the evaluation of best available science for each input model type as well as the associated epistemic uncertainty and aleatory variability, descriptions of the input component model sensitivities and uncertainty assessments, and description of policy as well as research and development models that can be considered for various engineering purposes. The references are extensive and may be useful for anyone needing more detailed information than that provided in this overview article.

The highest hazard areas in the 2023 NSHM are concentrated in places where large damaging earthquakes have or are thought to have occurred in historic and prehistoric times as discerned through earthquake records, historical intensity recordings, and geologic and geodetic fault studies. Nevertheless, damaging earthquakes can also occur in places where they have not been observed in the recent past. For assessing the probabilistic hazard, we develop an ERF (Field et al., 2023) that specifies the probabilistic locations, sizes, and rates of future earthquakes, and a GMM (Moschetti et al., 2023a; Rezaeian et al., 2023) that specifies the probabilistic ground-motion distributions resulting from earthquakes of various sizes, distances, alternative rupture mechanisms, and site conditions. Both inputs control the shaking levels for pseudo-SAs for multiple oscillator periods, peak acceleration and velocity, and alternative site classes at hundreds of thousands of sites across the 50 states. In developing the forecast, we try to (1) ensure that input data are properly processed and robust for the purpose of policy hazard applications, especially in urban areas; (2) work with internal and external USGS partners to build probabilistic ERF and GMM input component models that consider available data, physics-based constraints, empirical-based regressions, computer simulations, and to ensure that models are consistent with known physics principles in extrapolating models into regions with poor data availability; (3) incorporate new input data, models, and methods through interaction with our NSHM-SC and an extensive review process; (4) ensure that NSHM captures the breadth of epistemic uncertainty and by assigning justifiable weights based on data, model residual analysis, and expert judgment; and (5) develop robust, reasonable, and useful seismic hazard models describing the expected frequency of exceeding a set of ground-motion thresholds.

To understand the spatial and size distributions of past earthquakes, we develop an earthquake catalog that lists 66,272 events M2.7–7.9 from 1568 to 2022 for CONUS, 36,878 events M3.5–9.2 from 1882 to 2020 for Alaska, and 32,782 events M2.5–7.9 from 1790 to 2019 for Hawaii (Figure 2). We remove suspected induced, mining, and other manmade earthquakes from the CEUS catalogs that occur within time periods of known earthquake inducing industrial activity and within spatially defined zones where known injection or other earthquake stimulating activities occurred (Field et al., 2023). The forecasted earthquake rates are assumed to follow a Gutenberg–Richter (GR; Gutenberg and Richter, 1944) exponential magnitude-frequency distribution (MFD) that is observed globally with b-values (slope of the MFD) that are considered based on regional earthquake observations. The distribution is truncated at a maximum magnitude (Mmax) based on sizes of historic earthquakes, physics-based considerations, or global analogs. Typically, this Mmax is set to be close to M8 for background earthquakes, which is higher than in early NSHMs and more consistent with global earthquakes in various tectonic settings. Larger earthquakes can also occur on known faults included in the model. The minimum magnitude where we begin integration of the ground-motion contributions is still set at the standard M5 even though we recognize that M4+ earthquakes can cause damaging shaking at nearby locations (we may consider revisions in the future). The gridded seismicity model accounts for moderate- or large-sized earthquakes along faults and in places where fault sources are unrecognized. We apply several declustering methods, some of which are applied for the first time in the 2023 NSHM. This model assumes that future large earthquakes will occur where past seismicity is concentrated. We develop a grid of earthquake rates over a regular interval of points with 6 km spacing. Smoothing grids are shown in Figure 2 and are based on smoothing the declustered earthquake catalogs and then scaling the seismicity rates at each grid point to be consistent with the full rate of earthquakes M ≥ 5.0 for WUS and M ≥ 4.7 for CEUS. This model is more effective for forecasting future earthquakes in places of high seismic areas where past seismic activity is concentrated but is less effective where seismicity is low. We also apply broad zones to account for a base level of seismicity across a wide area where seismicity is low but cannot be ruled out.

We consider geologic fault studies to assess the fault structure and seismicity rates for faults with geologically recent (late Quaternary or Holocene) activity. The ERF accounts for potential locations, rates, and sizes of future earthquakes on mapped faults (Field et al., 2023). Figure 1 shows locations of faults included in the model (Hatem et al., 2022a, 2022b). The new WUS fault database contains about 1000 fault sections and slip rates based on both geologic and geodetic interpretations, an increase of about 350 faults compared to the previous NSHM for CONUS. This fault assessment was coordinated with state geological surveys that have parallel and intersecting responsibilities for earthquake safety. Deformation models were constructed using geologic and geodetic deformation data and joint geologic and geodetic inversion models (Hatem et al., 2022a, 2022b; Petersen et al., 2013; Pollitz et al., 2022). We assess earthquake sizes by analyzing magnitude-scaling equations that consider a collection of data showing moment magnitudes as a function of empirical rupture areas and physics-based energy constraints (Shaw, 2023). The resulting fault geometry and slip rates, earthquake magnitudes, and several other filters, constraints, and modeling components are used to develop a broad range of earthquake rate models, which include multi-segment or multi-fault ruptures that are similar to complex ruptures observed globally. These models span from the classic model that is similar to prescriptive models that account for historical and prehistoric ruptures and were applied in most previous NSHMs, to the inversion-based fault-system solutions that apply generalized simulated annealing inversions initially developed for California (Field et al., 2009, 2013) and extended to ruptures across the WUS that account for a broad range of multi-fault ruptures (Milner et al., 2022). We account for alternative ERFs to populate logic trees that are applied for capturing and incorporating uncertainty into the probabilistic NSHM ground motions.

Semi-empirical GMMs account for probabilistic distributions of shaking levels using ground-motion recordings from past earthquakes, 3D numerical ground-motion simulations of large earthquakes, and amplification models that are conditioned on the characteristics of the rock/soils beneath the site (e.g. V_S30 is the time-averaged velocity in the upper 30 m, Z_1.0 is the depth to the 1.0 km/s shear-wave velocity, and Z_2.5 is the depth to the 2.5 km/s shear-wave velocity horizon). The 2023 NSHM uses semi-empirical GMMs including new NGA-Subduction models, 3D simulations for Seattle and Los Angeles, sedimentary thickness-based amplification models for the CEUS (Gulf and Atlantic Coastal Plains) and WUS (Seattle, Portland/Tualatin, San Francisco/California Great Valley, Los Angeles), and a revised CEUS-WUS attenuation boundary. Details of these models, implementations, and sensitivity studies are found in Rezaeian et al. (2023) for subduction zone GMMs, and in Moschetti et al. (2023a) for the remaining GMMs in the WUS and CEUS. We examined the distribution of implemented GMMs to ensure that the models are encompassing a reasonable level of epistemic uncertainty (higher in the CEUS and Hawaii where data are sparser than in the WUS and Alaska) and we also add additional epistemic uncertainty to account for earthquakes that may occur outside of the empirical databases applied in developing the GMMs. Some reviewers felt that the additional epistemic uncertainty and its application in the calculations need further refinement. These requests came late in the process and there are also applications in San Francisco and Los Angeles that show the importance of including these “unknown unknown” contributions (refer to Appendix B of Electronic_Supplement_1.pdf). Moschetti et al. (2023a) provide details on sediment-thickness- and seismic-velocity-based amplification models and 3D numerical simulations of large ruptures. Details of GMMs for Alaska and Hawaii NSHMs are provided in Appendix D of Electronic_Supplement_1.pdf and Petersen et al. (2022), respectively. Selected GMMs for all regions and their assigned weights are summarized in Table SA-4 of Electronic_Supplement_1.pdf and are based on the revised GMM selection criteria published in Rezaeian et al. (2021).

The weight assignments, as in previous cycles of the NSHMs, are based on a consensus building process that involves consideration of expert judgment, published selection criteria (e.g. Petersen et al., 2014; Rezaeian et al., 2015, 2021, 2023), data residual analyses (e.g. McNamara et al., 2020), and a thorough review of between-model dispersion as well as important characteristics and limitations. This process requires extensive analysis for both ERF and GMM components in developing scientific-based weighting schemes. In the past, we have applied a three-point distribution with weights 0.185, 0.63, and 0.185 representing the 95% confidence levels for ERF and GMM models, and sometimes, these were applied as one instead of two standard deviations in previous NSHMs. However, these branches do not represent commonly used percentiles of a normal distribution, and we currently apply these weights to μ − 1σ, μ, and μ + 1σ instead (Keefer and Bodily, 1983). We generally assign equal weights to models if we do not identify advantages to a specific model or significantly better fits with data. When input component models have advantages over other considered models, we typically assign ratio weights (e.g. 1/3–2/3, 1/4–3/4) or percentage-based weights (e.g. 20%–80%) reflecting preferences of the project members, NSHM-SC, review panels, and the informed community on the best science. It is important to recognize that outliers can significantly modify the mean of a distribution, especially toward the high end (Field et al., 2023). We try to include the best input component models that are justified by science-based principles; however, we sometimes down-weight newer or less studied models. In addition, we include models that are valuable to illuminate and broaden the range of epistemic uncertainty because the principal objective is to capture the true range of epistemic uncertainty (e.g. deformation models, Coastal Plain amplification models). Additional epistemic uncertainty has been added to the GMMs in the WUS because in some cases the range of uncertainty is artificially low for certain periods and the models are developed together in a consensus fashion using similar data sets and commonly agreed on methods. We sometimes classify alternative models by class or individual modeler and then assign weights to different suites of models based on expert judgment as in CEUS GMMs. We apply backbone models or more complicated models such as the Sammons mapping applied in NGA-East (Goulet et al., 2021a) for which a mean/median model is applied with a range of epistemic uncertainty models that is deemed reasonable by the science community (e.g. Atkinson et al., 2014). We retire older models as they are updated, proposed for exclusion by modelers, recognized as having deficiencies, or considered to be encompassed in the range of models already considered in the updated model. Occasionally, we phase out older models over multiple update cycles instead of eliminating them in one cycle when they are still viable but utilize less information compared to the newer equations. Below we provide an overview of the science input models described in detail for ERF (Field et al., 2023) and the GMM (Moschetti et al., 2023a; Rezaeian et al., 2023).

It is difficult to separate epistemic uncertainties from aleatory variability (especially for GMMs) but is necessary for development of hazard fractiles (Marzocchi and Jordan, 2018). For this analysis, we assess the center, the body, and the range of ground motions included in the NSHM to define a representative suite of models that capture the epistemic uncertainty. We spent a great deal of effort evaluating residuals from input data and models using various performance metrics to assess the best science data, methods, and weighting for models (e.g. McNamara et al., 2020). Inclusion of many additional proponent models in the hazard assessment does not necessarily capture a reasonable range of epistemic uncertainty, so we have also put an emphasis on assessing the scientifically defensible range of epistemic uncertainty that is consistent with available data, tectonic regime considerations, and other specific conditions (Atkinson et al., 2014). Where needed we have added additional epistemic uncertainty to account for a scientifically reasonable range of uncertainty that is dependent on the amount and distribution of data (empirical and synthetic), span of available models, physical understanding of the earthquake process, and assessment of variability in epistemic uncertainty that considers global analogs. For portfolio applications, one would need to consider the spatial correlations in the epistemic uncertainties (standard deviation) to avoid overestimating the risk.

For this update, we evaluated epistemic uncertainties in the ground-motion hazard for 2% and 10% probabilities of exceedance in a 50-year hazard level using a logic-tree approach, which sweeps through alternative weighted ERF and GMM epistemic inputs (weighted branches) to discern fractiles of potential ground motions at a handful of representative sites (Anchorage, Alaska; Honolulu, Hawaii; Los Angeles, California; New York City, New York; Salt Lake City, Utah; San Francisco, California). The formation of a logic tree ideally spans the mutually exclusive and completely exhaustive set of possibilities or logic-tree branches. However, sometimes this assessment of available models does not satisfy the wide uncertainties allowed by science-based principles and global observations (Atkinson et al., 2014). We evaluated full uncertainties at six sites and partial ERF uncertainties at dozens of sites but would have preferred to make these uncertainty assessments at all grid sites across the country. During our assessment, we realized that more effort is required to separate the aleatory and epistemic components, improve the computing facilities, and understand the correlations as described in the sections on uncertainties and discussion below.

Details of the input models and forecasts for CONUS, Alaska, and Hawaii are described in Field et al. (2023), this paper for Alaska, and Petersen et al. (2022), respectively. In this section, we discuss model issues, sensitivity, weighting schemes, and hazard results. The input ERF component models include the following: the earthquake catalog, declustering models, spatial smoothing applications, fault and deformation models (geologic and geodetic), magnitude-scaling equations, Mmax assessments, and earthquake rate models for all 50 states.

An ERF, also referred to as a seismic source characterization in some studies, gives a complete list of potentially damaging fault ruptures in a region and over a specified timespan (or suites of synthetic earthquake catalogs, especially for fully time-dependent models). This section summarizes the time-independent ERFs developed for the 2023 NSHM, restricting attention here to changes that are new and innovative and consequential with respect to implied hazard. The ERFs presented here represent updates to those utilized in the 2018 NSHM for CONUS (Petersen et al., 2020), which largely utilized unmodified ERF components from the 2014 NSHM for CONUS (Petersen et al., 2014, 2015a).

Several broader goals with respect to ERF developments are described by Field et al. (2023) including (1) a more comprehensive representation of epistemic uncertainties; (2) increased uniformity in model assumptions and methodologies across regions (involving de-regionalization of model component developments); (3) simplification wherever it can be achieved without degrading model usefulness; (4) more operationalization of model component creation; (5) model extensibility with respect to adding time dependencies (including spatiotemporal clustering, swarms, and induced seismicity); (6) utilizing more physics to make up for sparse data at larger magnitudes; and (7) providing more complete documentation to minimize the need to comb through previous publications when learning about or reproducing models. All these goals are aimed at improving the rate and efficiency with which future model improvements can be rolled out nationwide.

ERF development is a system-level problem that uses constraints from a broad range of disciplines (e.g. earthquake geology, tectonic geodesy, statistical seismology, and earthquake physics). As such, a modular design is critical to managing workflows and to enable different groups of scientists to focus within their respective areas of expertise. Field et al. (2023) summarize the various components presently defined and utilized, including (1) fault models, defining physical attributes of explicitly modeled faults; (2) deformation models, specifying fault slip rates as constrained by geology and assessing the influence of each fault and in some cases “off-fault” deformation rates; (3) earthquake rate models, specifying the long-term rate of each earthquake rupture; and (4) probability models, which specify conditional probabilities based on elastic rebound, spatiotemporal clustering, swarms, and induced events. Given the general time-independent nature of NSHMs, the main topic here is the earthquake rate models, which generally (with exceptions) get applied using Poisson probabilities. This 2023 US NSHM ERF was constructed at the same time as update efforts in New Zealand, and although the two models are quite distinct from one another in methodologies and goals, many of the concepts were discussed together (Gerstenberger et al., 2023).

Each earthquake rate model is composed of two types of earthquake sources, which are generally categorized as fault-based (i.e. explicitly modeled faults) versus “off-fault” or “gridded” seismicity, with measures often being taken to avoid double counting between these two. It is important to clarify that off-fault earthquakes as described above still occur on fault sources, but they are on uncharacterized faults that have not been detected or studied sufficiently for inclusion in this model. It is difficult to define potential ruptures in areas between known faults where future ruptures could occur as fresh ruptures or subsidiary fault ruptures. Fault-based sources are also distinguished by (1) “Classic” fault models, generally applied to individual faults and with prescribed attributes (e.g. the shape of the MFD); (2) fault-zone sources where a distinct fault surface is not identifiable (also applied with prescribed attributes); and (3) fault-system solutions, which are aimed at acknowledging and including many plausible multi-fault ruptures. The latter are generally inversion based, as described below for WUS, but were also derived from multi-cycle physics-based simulators (Milner et al., 2022). We start with the CONUS region because it embodies all types of components, and then follow with differences and exceptions reflected in the Hawaii and Alaska ERFs. Each section also discusses epistemic uncertainty representation, treatment of aftershocks, and the review process.

The most consequential updates for the CONUS ERF are the addition of faults, new slip-rate constraints, changes in gridded seismicity components, and the inclusion of multi-fault ruptures in the WUS outside California. The innovation in the fault and deformation models is to allow for multi-fault ruptures, which have been observed in several recent large earthquakes in California, Mexico, New Zealand, and Turkey. We also expanded the representation of epistemic uncertainties for most components, and we were able to conduct extensive branch sensitivity studies with respect to WUS hazard. Epistemic uncertainties for the Cascadia subduction zone were based on several workshops prior to the 2014 NSHM update and on an extensive logic tree that included many alternative input component models. Epistemic uncertainties for the CEUS fault sources are based on extensive logic trees developed by the US Nuclear Regulatory Commission (CEUS–SSCn) (2012) project. The model update relied on several component developments, including those represented in the publications listed in Table 1 of Field et al. (2023).

Fault model and geologic constraint updates for the WUS were compiled by Hatem et al. (2022a, 2022b), with about 350 new faults being added (mostly outside California), which were previously excluded due to lack of site-specific geologic slip-rate constraints. Hatem et al. (2022a, 2022b) also compiled geologic slip-rate estimates at points where site-specific studies are available, and applied default categorical estimates for the faults that lack such explicit constraints. Some of the fault ruptures had narrow faults when considering fault creep. The ERF review panel discussed this issue but felt that the narrow faults were reasonable for inclusion in the 2023 NSHM.

The five WUS deformation models utilized in the 2023 NSHM are described in a special issue of Seismological Research Letters (e.g. refer to the overview paper by Pollitz et al. (2022)). Again, the main inferences from these models are slip-rate estimates for each fault section in the WUS Fault Model. One model is based almost entirely on geologic constraints (Hatem et al., 2022a), whereas the four others also utilize Global Positioning System (GPS) or geodetic constraints that were updated by Zeng (2022b). Two of these models represent updates to those utilized in the 2018 NSHM (Zeng (2022a) and Shen and Bird (2022)), both of which generally honor the geologic constraints unless GPS data strongly indicate otherwise. The other two models are new: one from Pollitz (2022) and another from Evans (2022) that effectively permit wider excursions from geologic values than the older models by relaxing geologic constraints and allowing for more freedom to follow the geodetic data. Further enhancements include corrections for potential viscoelastic effects (the “ghost transients” described by Hearn (2022)), and much more thorough inferences and corrections for fault creep processes (Johnson et al., 2022). Because the models have considerable dispersion between one another, it is important to point out that consideration of all these models can lead to substantial effects on the uncertainty assessments and can be important considerations for certain classes of end-users. These inter-model deformation differences are also important in assessing off-fault slip rates and their associated uncertainties.

Given the complexity and importance of these models, a special review team (Deformation Review Panel for CONUS chaired by Kaj Johnson (see Note 1)) was convened to advise on the reliability of each modeling approach (and to suggest logic-tree branch weights). This analysis involved scoring each model based on 15 different metrics. The review panel recommended relatively low branch weights for the two new models, Pollitz (2022) and Evans (2022), due mostly to their having more anomalous slip rates, but also emphasized that these models are nevertheless viable and allow the model to capture a broad range of epistemic uncertainty. However, and as discussed in Field et al. (2023), initial hazard calculations revealed that some slip-rate outliers were having a disproportionate and concerning effect on mean hazard. We use the results of the Deformation Review Panel for CONUS as input into our deliberations on weighting of the deformation models and generally follow these recommendations but modify them based on further considerations by the ERF Review Panel and others: Geologic 26%, Zeng 32%, Shen and Bird 32%, Pollitz 8%, and Evans 2% shown in Table SA-1. As noted below, further deformation-model development is one of our highest priorities with respect to planned future work, especially with respect to better quantification of slip-rate uncertainties and covariance between faults. Another priority research topic is the off-fault deformation estimates provided by some of these models, which were deemed too immature for application (e.g. as a gridded seismicity constraint) at this time.

The fault model and geologic constraints for the CEUS were updated by Thompson Jobe et al. (2022a, 2022b), including the addition of five new faults (Central Virginia, Saline River, Joiner Ridge, Crowley’s Ridge (South), and Crowley’s Ridge (West)). They also made fault-geometry adjustments to four previously utilized faults (Axial, Bootheel, New Madrid West, and Reelfoot), and added explicit surfaces to five faults previously represented with zones (Commerce, Eastern Margin (North), Eastern Margin (South), Crittenden County, and Meeman-Shelby, the latter of which was previously called River Picks). They also summarize the slip rate or paleo event-rate constraints available for each fault, with the latter typically involving some number of paleoliquefaction events inferred over some period. No WUS-style deformation models were developed for CEUS faults, owing mostly to the lack of deformation rate constraints.

The biggest innovation with respect to earthquake rate models is the inclusion of multi-fault ruptures throughout the WUS, made possible by a new inversion-based fault-system solution developed by Milner and Field (2023). This builds of the so-called “grand inversion” approach used in the Uniform California Earthquake Rupture Forecast version 3 (UCERF3; Field et al., 2014; Page et al., 2014), but with important improvements with respect to computational efficiency, better control with respect to fitting various data constraints, an expanded set of diagnostics and model-evaluation plots, and better reproducibility (our intent is to enable anyone to conduct such calculations themselves). Perhaps most importantly, we added variable segmentation constraints, enabling us to define a much wider range of models with respect to multi-fault ruptures. Specifically, we have a “Classic” segmentation branch that is consistent with previous UCERF1 and UCERF2 (Field et al., 2009; Petersen et al., 2007) and other models outside California (where extensive multi-fault ruptures are prohibited), three branches that apply a range of jump-distance–based penalties, and an unsegmented branch that applies no penalty for jumps up to 15 km (UCERF3 had one branch that applied no penalty for jumps up to 5 km, Field et al., 2023). Note that the extreme branches are partially a proxy for the fault model being incorrect with respect to the actual distance between faults (i.e. the classic branch might be most appropriate where jump distances are much larger than implied by the fault model, or the unsegmented branch might be appropriate where an unknown connector fault exists). After considerable deliberations, the classic model and unsegmented branch models were each given lower weight (10%) compared to the other fault segmentation models (low 20%, medium 30%, and high 30%). These weights were based on recommendations from the ERF Review Panel and observations that the classic model has a magnitude–frequency curve bulge and assumes that faults cannot rupture together more than that observed in the geologic data. The classic model does not generally allow ruptures that extend beyond geologic observations and results in shorter ruptures, smaller magnitudes, and shorter recurrence times compared to the longer ruptures, larger magnitudes, and longer recurrence times associated with the three multi-fault rupture models that allow various ranges of rupture jumps. Therefore, the classic model typically has larger ground-motion hazard than the multi-fault rupture models.

Another improvement is the ability to map out a wider range of models via a supra-seismogenic b-value constraint (defined further below), which sets the slope of a GR density distribution (not cumulative distribution) at large magnitudes. In UCERF3 (Field et al., 2014), inversion solutions were constrained to stay as close as possible to the previous model (UCERF2, Field et al., 2009), which might have been justifiable from a policy perspective (do not change the model any more than needed to fit the data), but this limited the range of models considered. In other words, a wider range of models may fit the data equally well, which represents epistemic uncertainty that warrants being accounted for. The supra-seismogenic b-value constraint is our way of doing this, with five discrete branches between values of 0.0 and 1.0, which is effectively a way of adjusting the total rate of events on each fault section; in fact, we do not constrain the entire implied MFD on each fault section in the inversion, but rather we constrain this implied total rate. We do, however, sum the fault-section MFDs to also provide a total regional MFD constraint in the inversion, which likely is an improvement over how the total fault MFD target was defined in UCERF3 and provides a path to check the model against historical data.

This inversion approach starts with the fault model and applies a new plausibility filter (Milner et al., 2022) to define the ruptures that have a non-zero likelihood of occurrence (to reduce an effective infinite number of ruptures to a manageable and representative set, resulting in 580,000 different WUS ruptures here). Figure 9 of Field et al. (2023) shows the implied inter-connectivity among WUS faults (where multi-fault ruptures can occur in the model). The magnitude and average slip for each of these ruptures are computed from one of six scaling relations, the options of which are based on new recommendations from Shaw (2023). The inversion then solves for the rate of each rupture by satisfying the chosen deformation model slip rates, as well as the paleoseismic recurrence intervals updated by McPhillips (2022) for sites in California and those compiled by Valentini et al. (2020) for the Wasatch fault. Because slip rates may be inconsistent with nearby paleoseismic recurrence intervals, we added three paleoseismic data-fit branches that adjust how well these data are fit (relative to slip rates), providing a wider range of models than utilized in UCERF3. Fault creep is assumed to reduce rupture area, except for highly creeping faults where it is applied as a slip-rate reduction as well.

In summary, we have five deformation models, six scaling relationships, five target supra-seismogenic b-values, five segmentation models, and three paleoseismic data-fit options represented on the logic tree, which amounts to 2250 different inversion-based fault-system solutions for WUS (Figure 4). This new ERF consists of rates of 580,000 alternative but possible fault ruptures across the entire WUS. The ratio plots show each model compared to the average of all models within a similar logic-tree branch. Large differences in earthquake rates are found between the geologic and geodetic deformation models, fault segmentation models, and the declustering and gridded (smoothing) components of the seismicity-based rate models. As such, we think we now have a fully comprehensive representation of the viable range of models, at least with respect to the propensity of multi-fault ruptures (i.e. leaving aside slip-rate uncertainty questions). This new ERF consists of rates of 580,000 alternative but possible fault ruptures across the entire WUS. A much more detailed summary of inversion-based fault-system solutions is available in Field et al. (2023). In addition, we point the users to further information on how branch weights were decided, references to other related studies, full implementation details, quantification of data-fit improvements, and various sensitivity analyses (Field et al., 2023).

CEUS fault-based sources, described by Shumway et al. (2023), are based on the updated fault models and geologic constraints discussed above (Thompson Jobe et al., 2022a, 2022b). All are either represented with a fault zone (area source) or with an explicit fault surface. Figure 1 shows faults and zones for CEUS sources. All these models also adopt the Repeating Large-Magnitude Earthquake (RLME) hypothesis introduced by the US Nuclear Regulatory Commission CEUS–SSCn (2012), which assumes that each source produces only a narrow range of magnitudes (or a single-magnitude event in the case of past and present USGS NSHMs). For this 2023 NSHM, we use a flexible definition of RLME because some faults are only characterized by evidence from a single earthquake rupture rather than evidence of repeating ruptures. The magnitude of these characteristic events is inferred from a scaling relationship using either fault-surface area, the maximum length of a fault zone, or the spatial extent of paleoliquefaction deposits. The mean rate of the RLME source is either inferred by moment balancing if a slip-rate estimate is available or from an observation of N events having occurred in some timespan T (e.g. from paleoliquefaction deposits) using calculations of likelihood. Although there is no aleatory variability in source magnitude, considerable epistemic uncertainties are accounted for with respect to what the characteristic magnitude is, the recurrence rate implied by the geologic constraints, and sometimes the area assigned to fault-zone sources.

With respect to fault-zone sources (Figure 1), the following changes have occurred since the 2018 NSHM for CONUS: three have been added in the NMSZ (Joiner Ridge, Crowley’s Ridge (south), and Crowley’s Ridge (west)); two were added elsewhere in the CEUS (Saline River and Central Virginia); five were converted from zones to explicit fault surfaces (Commerce, Eastern Rift Margin (North), and Eastern Rift Margin (South), which includes three fault segments); and four were unchanged (Wabash Valley, Charleston, Charlevoix, and Marianna). We weighted these five new rupture models with 50% on the new ERF model and 50% on the gridded seismicity characterization based on discussions with the authors on limitations of the paleoliquefaction studies and uncertainty on whether these sources represent unique sources relative to the background gridded seismicity sources in the model (Field et al., 2023; Shumway et al., 2023).

There are 12 CEUS faults modeled with an explicit fault surface, none of which is new, none of which has had any changes in source magnitudes or rates, and only one of which is outside the NMSZ. The latter, the Meers fault in Oklahoma, is now known to extend farther to the west, which increased computed hazard in that area. The only other consequential modification is with respect to the Axial fault, which was extended to the southwest and produced a computed hazard increase in that area too (Field et al., 2023; Shumway et al., 2023). Although the other fault-based sources in the NMSZ did not undergo any other consequential changes since the 2018 NSHM, it is worth noting that they embody additional complexity with respect to some single and multi-fault ruptures (each of which still honors the RLME assumption) and there are branches that assume ruptures always come as doublets or triplets. A more extensive summary of these fault-based sources is given in Field et al. (2023), and full details, including logic-tree descriptions for each, are given by Petersen et al. (2008, 2014) and Shumway et al. (2023).

Field et al. (2023) also describe some improvements that could be made to these models. One is to relax the RLME assumption to acknowledge a wider range of possibilities (i.e. increase aleatory variability in source magnitudes). Another would be to streamline logic trees to reduce the number of logic-tree branches and avoid over-complication of the modeling. Finally, the interconnected fault system in the NMSZ seems like a prime candidate for an inversion-based fault-system solution, although lack of mapped surface faults with accompanying slip rates will lead to a much broader range of fault-based solutions. Furthermore, assumptions on the earthquake clustering and recurrence times between sequences of New Madrid events could be improved with more information on paleoliquefaction timing constraints and additional global analogs.

The Cascadia subduction zone model is essentially a fault-based source with a more sophisticated 3D representation of the subducting interface that includes three down-dip widths. Updates relative to the 2018 NSHM for CONUS, summarized in Field et al. (2023), are based on new paleoseismic constraints and discussions at a virtual workshop in February 2021. These led to the addition of an alternative segmentation model proposed by Goldfinger et al. (2017), which has ruptures extending farther north, and swapping out a 2000-year M ≥ 8 recurrence branch with two alternatives (800-year vs 2300-year recurrence) inferred from on-shore data by Nelson et al. (2021). We also added a cluster model for which, 10% of the time, a sequence of smaller ∼M8 ruptures sweep down the subduction zone over a short period of time (Electronic_Supplement_1.pdf, Table SA-2 and Figure SB-1). Finally, Goldfinger et al. (2017) and Nelson et al. (2021) include an optional time-dependent probability for full subduction zone ruptures assuming a Brownian Passage Time distribution. This model assumes a recurrence interval of 529 years and a coefficient of variation of 0.5, with the last event having occurred in 1700. The model applies the magnitude-scaling equations from the 2018 model as well as a new equation by Shaw (2023, M = logA + 4). This time-dependent assessment produces a 50-year probability of 12.5%, which is a factor of ∼1.4 greater than the time-independent probability of 9% (Petersen et al., 2002). There was a consideration of whether to include time-dependent probabilities for recurrence of the largest Cascadia rupture M9; however, in the end, the review panels felt that these additions should not be included here. Changes are caused by new turbidite results, new magnitude-scaling equations, and new subduction GMMs. The logic tree is displayed in Electronic_Supplement_1.pdf (Figure SB-1). The changes in computed hazard produced by these modifications are generally less than 10%. Hazard sensitivity calculations have also been performed of implied along-strike variation in slip for the three different down-dip width branches of the logic tree, revealing that the choice of the location of the up-dip limit of rupture has a large impact on implied slip rate, and adjusting this rupture bound may be considered in future updates.

Subduction intraslab earthquake sources are modeled as a gridded seismicity source in the Cascadia model. The 2018 NSHM, and prior NSHMs, used three different depth slice horizons to mimic the down going geometry of the subducting slab. In the 2023 NSHM, we have updated this approach using depths from the Slab2 Cascadia model (Hayes et al., 2018). This smoothly varying depth model is an improvement over the prior “stair-step” model and is consistent with the intraslab implementation for the Alaska–Aleutian subduction zone in the 2023 NSHM for Alaska. The intraslab rate model has not been updated for the 2023 NSHM. The Mmax model has also not been updated, and normally caps at M7.2, but some branches have Mmax as high as M8 (GR branch with Mmax = 7.95).

Gridded seismicity sources are meant to account for the fact that our fault models are incomplete with respect to representing all possible fault-rupture surfaces, especially at smaller magnitudes. These sources are represented in the NSHM using the following: (1) a polygon defining the region and a spatial discretization interval (0.1 degrees here) to define the grid cells; (2) a spatial probability distribution defining the relative rate of earthquake nucleation within each grid cell; (3) a total M ≥ 5 rate and b-value for the region; (4) an assumed maximum magnitude for the region or a set of subregions; (5) a probability distribution of focal mechanisms for each grid cell; and (6) rules for converting a nucleation point into a finite-rupture surface. Steps are also taken to ensure that gridded seismicity sources are not double counted with fault-based sources. Campbell and Gupta (2018) suggest an efficient method of modeling random orientation of virtual faults that avoids biases and could be considered in future NSHMs.

The models utilized here are based on observed seismicity, so the first step was to update the earthquake catalog through 2022 using the methodology of Mueller (2019), taking care to exclude induced-seismicity events. Due to catalog completeness heterogeneities and other differences, the following steps were conducted separately for the WUS, CEUS, and the deep seismicity near the Cascadia subduction zone. The total M ≥ 5 rate and b-value for each region were inferred by Field et al. (2023), including a best estimate and 95% confidence bounds (a rate and b-value pair for each). The spatial probability distribution of grid-cell rates was calculated by Field et al. (2023), which involved declustering the earthquake catalog to avoid bias from aftershock sequences, spatially smoothing the resultant declustered catalog, and normalizing by the total rate so that grid-cell values sum to 1.0. This process results in a spatial probability density function (PDF) map. Additional epistemic uncertainty was accounted for in the declustering process, because no methodology is universally accepted as the best way to decluster earthquakes from the catalog, with some models removing more large earthquakes and some removing fewer events. The algorithms of Reasenberg (1985) and Zaliapin and Ben-Zion (2020) were added to the traditional approach (Gardner and Knopoff, 1974), resulting in three weighted logic-tree branches. We reduced the weight on the Reasenberg model to half (20%) that of the other two declustering models (40% each) because the ERF review panel felt that the Reasenberg model did not adequately remove aftershocks for some events and it likely spreads seismicity too far in some areas (Field et al., 2023). The full and declustered catalogs, for both the CEUS and WUS, are available at the USGS ScienceBase Catalog (Petersen et al., 2023).

The same two spatial smoothing algorithms applied in the 2018 NSHM for CONUS were again utilized here, both of which implement a two-dimensional Gaussian kernel (Field et al., 2023). The first applies a “fixed” correlation length that, using the convention of Frankel (1995), is $2$ times the standard deviation of the Gaussian, which was set as 50 km in the WUS and between 50 and 70 km in the CEUS (depending on assumed minimum magnitude of completeness). The second “adaptive” approach sets the smoothing width as the distance to the Nth nearest event (N = 3 in the WUS and N = 4 in CEUS), providing greater spatial resolution in higher seismicity areas. These nearest neighbor numbers were reoptimized using likelihood tests for the updated catalogs but are similar to the 2014 NSHM values (Field et al., 2023; Moschetti, 2015; Petersen et al., 2014). The fixed smoothing approach can produce dubiously low seismicity rates in areas that have had few or no observed events, in which case rates are prevented from going below a specified floor rate; this is now only applied in CEUS. The adaptive weighting approach was given more weight (60%) compared to the fixed weighting model, which was given less weight (40%) in the 2023 NSHM compared to that in 2018. This change allows for the use of more nearest neighbor earthquakes than in 2018 NSHM for CONUS (based on comments from the ERF review panel, Jordan et al., 2023). A new boundary between CEUS and WUS catalogs was suggested at longitude −104°W, which also led to new differences in the catalogs and applicable processing (Field et al., 2023). This boundary was changed so that both faults and gridded seismicity sources modeled using the WUS fault-system solutions would have the same areal extent.

Three branches are utilized for gridded seismicity Mmax, with the options and weights for the WUS being adopted from UCERF3 and those for the CEUS being adopted from the 2018 NSHM for CONUS (Figure 2). Likewise, the spatial distribution of focal-mechanism likelihoods remains unchanged since the 2018 NSHM for CONUS, as is the procedure for assigning finite-rupture surfaces to gridded seismicity events. Finally, adjustments were made to avoid double counting seismicity on faults in the WUS.

As noted above, the removal of aftershocks is arguably necessary when inferring the long-term spatial distribution of seismicity rates. The question addressed here is whether aftershocks should be included in time-independent hazard assessments. Previous NSHMs applied declustered models to honor the Poisson hazard assumption, a decision that had only a minor impact on results because the declustering algorithm used at that time, Gardner and Knopoff (1974), does not generally remove events above M6.5. The problem is that more modern, best available declustering algorithms (e.g. based on the ETAS model of Ogata (1988, 1998) or the Zaliapin and Ben-Zion (2020) approach as applied here) tend to remove about half the events at all magnitudes, which would be highly inappropriate for hazard assessments. A significant body of literature asserts that, for 2% or 10% in 50-year hazard, keeping aftershocks and assuming a Poisson process are better than declustering with antiquated or biasing methodologies (Field et al., 2021; Marzocchi and Taroni, 2014; Michael and Llenos, 2022; Wang et al., 2022).

We follow UCERF3 in terms of including the rates of all M5+ events in the time-independent ERFs presented here, a decision our review panel agreed with. UCERF3 included an optional Gardner–Knopoff filter for hazard calculations, which was applied in the 2018 NSHM for CONUS for consistency with other regions, but there is no longer any scientific rational or sensible procedure for perpetuating this practice, especially given the available additional declustering algorithms. The consequence of this modification may be minimal for building code applications that apply 2%, 5%, or 10% probability of exceedances (PEs) in 50 years; however, the impact may be more significant for probabilities of exceedance greater than 10% or for smaller probabilities where gridded seismicity controls the hazard (Field et al., 2023). Therefore, modifications for applications that require these higher exceedance probabilities (e.g. risk community) may be warranted.

Model validation is an important exercise for ensuring the model results are robust and acceptable to the wider science community and demonstrating that the model is consistent with all kinds of data. Model evaluations were performed to validate the ERF model against different data sources. Several model-evaluation metrics are described in Field et al. (2023), including model-implied MFDs and their comparison with observations for various regions, inversion misfits with respect to slip-rate and paleo recurrence interval constraints, implied segmentation and rates of multi-fault ruptures, implied hazard curves and hazard maps, an explanation of changes with respect to the 2018 NSHM for CONUS (summarized below), and sensitivities of the above with respect to the WUS logic-tree branches. Figure 23 of Field et al. (2023) exemplifies the latter with respect to maps of the PGA that has a 2% chance of exceedance in 50 years. However, we cannot create such plots for CEUS fault-based sources due to the logic-tree complexity and heterogeneity described above.

In addition to the Deformation Model Review Team for CONUS discussed above, we also convened a 19-member ERF review panel chaired by Tom Jordan (see Note 1). Their feedback led to model improvements, including assessments of the slip-rate outliers and additional constraints on the rate of very long ruptures (≥700 km) (Jordan et al., 2023). The WUS fault-system solutions were also carefully scrutinized by an ad hoc group of USGS geologists, which also led to a number of model adjustments as well as some future recommendations (Hatem et al., 2022a, 2022b). All these model evaluations allow us to conclude that this new model represents a substantial improvement in assessing epistemic uncertainties (Jordan et al., 2023).

The 2023 Alaska NSHM, previously updated in 1999 and 2007 (Wesson et al., 1999, 2007), is documented in this paper and Appendix D of Electronic_Supplement_1.pdf and updates all ERF and GMM components. The Alaska NSHM was subject to review by an independent panel of Alaska hazard experts (i.e., AK Review Panel chaired by Mike West)¹ and a Tiger Team of GMM developers (chaired by Norm Abrahamson) enlisted to evaluate the Alaska subduction GMMs¹. Reports from the Tiger Team and Alaska Review Panel can be found in Electronic_Supplement_1.pdf, Appendix F and G, respectively. The 2023 Alaska ERF describes earthquake rupture potential in the active-crust, subduction interface, and subduction intraslab tectonic settings. The 2023 crustal fault model includes over 100 faults, and we use estimates of slip rates derived from geodetic data for both crustal fault and subduction interface ERF components. The 2023 Alaska NSHM also includes a new subduction segmentation and recurrence model, new approaches to developing gridded seismicity, and the structure of the Alaska-Aleutian subduction interface is constrained to Slab2 (Hayes et al., 2018).

The crustal fault model (Bender et al., 2021; Haeussler et al., 2023) consists of 105 fault sections, increased from 9 in 2007 and was developed following the same methods applied in the CONUS region (Hatem et al., 2022a,b). Important additions include a connector fault through the Chugach and Wrangell-St. Elias Ranges, which permits slip transfer from the Queen Charlotte—Fairweather system to the Totschunda and Denali faults. The model also adds numerous reverse and strike-slip faults in the complex zone of deformation where localized strike-slip motion on the Queen Charlotte—Fairweather system gives way to compressional deformation at the east end of the Alaska-Aleutian arc. The crustal fault ERF assigns both geologic (Haeussler et al., 2023) and geodetic block model (Elliott and Freymueller, 2020) slip rates to crustal faults. We downweight the geodetic block model (1/3 weight) because it provides lower estimates of slip rate on the Denali and Queen Charlotte faults where the long-term geologic rates are better constrained. Faults that do not coincide with block boundaries are assigned geologic rates with full weight. For the fast-slipping strike-slip faults of the interconnected Denali–Totschunda–Fairweather–Queen-Charlotte system, the ERF accommodates multi-fault ruptures and along-strike variations in slip rate. The crustal ERF also includes several area source zones in regions where numerous active structures are mapped, but for which slip-rates are poorly constrained. Rupture magnitudes and areas in the crustal fault model are computed using a width-limited magnitude-area scaling relation by Shaw (2023) and we impose a regional Mmax of 7.9±0.34.

The Alaska ERF includes gridded (smoothed) seismicity models for all tectonic settings that are based on an earthquake catalog that includes events through 2020. For crustal and intraslab settings, the gridded seismicity model forecasts earthquake rates for M5 up to M7.9 events. For active-crust, the Mmax distribution is M7.5 (0.9 weight) and M7.9 with a tapered MFD (0.1 weight), consistent with the gridded seismicity model for active-crust in the 2018 CONUS NSHM (Petersen et al., 2020). For the subduction interface, the gridded seismicity model accounts for the numerous small M57 earthquakes not covered by the finite fault model used for M7+ earthquakes. Development of the gridded seismicity models included the novel application of a probabilistic earthquake classifier to segregate the catalog into crustal, subduction interface, and subduction intraslab events. Each catalog was then processed using the same three declustering, two smoothing, and rate modeling methods as used in the development of the CONUS ERF update (Field et al., 2023; Table SA-1) and applied with identical weights as the CONUS model. Point earthquake sources for the interface and instraslab grids are modeled at depths derived from Slab2 (Hayes et al., 2018). Their spatial extent is also limited to the extent of the Slab2 contours with the interface grid further limited to depths shallower than 50 km. Similarly, gridded seismicity model rates (a-and b-values in the Gutenberg-Richter MFD) for Alaska are based on the full-catalog rate and not the declustered catalog rate that was used in prior NSHMs.

For the Alaska-Aleutian subduction zone, large magnitude (M8+) megathrust earthquakes occur on updated structural and segmentation models that consider both geologic and geodetic recurrence rates (Briggs et al., 2023). We give the geologic branch 80% weight and the geodetic branch a lower 20% weight because it implies very high rates of M8 earthquakes which are plausible but are only rarely observed in the historic record. Future analysis could focus on the higher M8 rates implied by the geodetic model. We performed sensitivity studies showing the relative contributions of the ERF and GMM models and find that for most sites the GMM modifications are much more substantial in increasing the hazard compared to the ERF modifications (e.g., site in Anchorage shows that overall changes are based on a 2/3 contribution from GMM changes and 1/3 contribution from ERF changes, Appendix D of Electronic_Supplement_1.pdf). The ERF branches without the geodetic models result in higher ground motions than in the 2007 NSHM for Alaska and the geodetic branch increases this further. The geologic branch of the megathrust model considers two down-dip widths (narrow and wide) constrained to depth contours of the Slab2 model (Hayes et al., 2018). Magnitudes are computed using the scaling relations applied for the 2018 Cascadia model supplemented with a new alternative model by Shaw (2023; logA+4). For the geodetic branch, rupture areas are constrained to the modeled locked portions of each subduction interface section. On the geologic branch, we allow ruptures to consist of either 1, 2, or 3 contiguous rupture sections of the megathrust (with equal weights of 1/3; refer to Table SA-3 in Electronic_Supplement_1.pdf). Four of the multi-section models match the rupture extents of historic interface earthquakes in 1788, 1957, 1965, and 1964, the Great Alaska earthquake. The multi-section rate models are constrained to not exceed the single-section geologic recurrences with any residual rate assigned to single section ruptures.

As in other parts of the United States, the Hawaii seismicity source model was developed using gridded seismicity and fault-based deformation models. To update the gridded seismicity model, we applied a new earthquake catalog, declustering methods, spatial smoothing algorithms, and earthquake rate models for moderate to large earthquakes and caldera collapses to update the previous model developed in 2001 (Klein et al., 2001). The Hawaii model prescribes full weight on the classic model that is built from geologic observations and paleoseismic studies of prehistoric ruptures. The new Hawaii earthquake catalog was first separated into two regions (north and south) and five subregions based on alternative earthquake mechanisms and depths (shallow summit, shallow non-summit, deep summit, deep non-summit, and caldera collapse, with shallow and deep defined as less than or greater than 20 km, respectively). Rate models were computed for each subregion independently. The catalog was further subdivided into three different time periods (1840–1899, 1900–1959, 1960–2019) to account for spatial and time-dependent variability in the earthquake rates. We used the same decluster methods now considered in CONUS and Alaska but have excluded Gardner and Knopoff (1974) because it does not fit the clustered earthquake behavior observed across Hawaii. Full and declustered catalogs can be found at the USGS ScienceBase Catalog (Rukstales et al., 2021). Seismicity across the Hawaiian island chain varies spatially from the dense distribution of earthquakes on the Island of Hawai‘i to the much sparser distribution of earthquakes found across the islands located to the northwest. In 2001, the spatial distribution of this seismicity was characterized using fixed spatial smoothing combined with a ramp-type smoothing model that allowed for a linear spatial decay of earthquake rates toward the less seismically active northern islands (Klein et al., 2001; Petersen et al., 2022). For the 2021 NSHM, we only applied the adaptive smoothing model because it naturally results in decaying rates from the southern to the northern islands and does not require the complex ramp (fixed smoothing) modeling techniques of Klein et al. (2001). We also apply a caldera collapse shaking model to account for ground motions near the Kilauea caldera that caused damage to nearby structures during the 2018 Kilauea eruption (Michael and Llenos, 2022). We apply maximum magnitudes ranging from M7.5 to M8 that account for seismicity that was observed on the southern and western flanks of the islands as well as a new interpretation of a large earthquake with M 7.5 located near Lanai (Butler, 2020). The earthquake rates are based on a GR distribution of earthquakes with b-values derived from each declustered catalog (Petersen et al., 2022). As an alternative, the rates and b-values were also calculated using the full catalog combined with declustered catalog spatial distributions (e.g. Field et al., 2021; Marzocchi and Taroni, 2014).

The Hawaii fault-based model was updated with 20 years of additional seismicity data, new fault deformation GPS data, and a geologically recent (Quaternary) fault database (Klein et al., 2001; Petersen et al., 2021). The new fault model considers moderate to large earthquakes on low angle dipping décollement faults recognized or inferred on the southern and western portions of the Island of Hawai‘i at depths between 8 and 12 km based on seismic reflection studies and seismicity patterns. The ERF fault model allows for several alternative (epistemic) south and west flank sources derived from seismic moment of large historic earthquakes, slip rates across the southern and western flank regions, and historical inter-event earthquake recurrence information. Forecasted earthquakes on major low-dipping décollement sources are similar to the 1868 Ka’ū (M7.9, M7.0 aftershock) and the 1975 and 2018 Kalapana earthquakes (M7.7, M6.9) on the southern coast and the 1929 M_s6.5 near Hualālai and the 1951 M6.9 earthquake near Kona along the western coast of the Island of Hawai‘i. The geologic and geodetic deformation model is based on slip rates inferred across a network of 66 telemetered GPS stations on the Island of Hawai‘i and shows significantly higher potential for large earthquakes in the southern flank region compared to the western flank of other areas of the island. Maximum magnitudes were increased to M7.5–8.2 across the southern and western flanks of the Island of Hawai‘i to account for potential large earthquakes, similar to those observed historically on the south flank. The décollement zones were weighted with recurrence information described in Petersen et al. (2021). The models were reviewed by the Hawaii Earthquake and Tsunami Advisory Council for use in the 2023 NSHMs (Petersen et al., 2021).

The 2023 NSHM applies new and modified GMMs (NGA-Subduction, CEUS NGA-East), 3D numerical ground-motion simulations (Seattle-M9 and Los Angeles region-CyberShake), sedimentary thickness-based amplification models for the CEUS (Gulf and Atlantic Coastal Plains) and WUS sedimentary basins (Seattle, Portland/Tualatin, San Francisco, California Great Valley, Los Angeles), and a revised CEUS-WUS attenuation boundary. Details of these models, implementations, and sensitivity studies are found in Rezaeian et al. (2023) for subduction zone GMMs and in Moschetti et al. (2023a) for the GMMs in WUS and CEUS. Past NSHMs applied semi-empirical GMMs that were based on abundant moderate-size earthquakes recorded in the WUS and were supplemented with larger and great earthquakes recorded globally to characterize ground shaking. Stochastic- and physics-based simulations and hybrid- and reference-empirical GMMs were applied for earthquakes in the CEUS. The 2023 NSHM includes 3D-simulated ground motions for the first time for earthquakes on the Cascadia subduction zone and for faults across the Los Angeles basin (Moschetti et al., 2023a). Moschetti et al. (2023a) also provide details on sediment-thickness- and seismic-velocity-based amplification models and 3D numerical simulations of large ruptures. Details of GMMs for Alaska and Hawaii NSHMs are provided in this paper and Appendix D of Electronic_Supplement_1.pdf and Petersen et al. (2022), respectively. Selected GMMs for all regions and their assigned weights are summarized in Tables SA-4 to SA-9 of Electronic_Supplement_1.pdf and are based on the revised GMM selection criteria published in Rezaeian et al. (2021). The weight assignments, as in previous cycles of NSHMs, are based on a consensus building process that involves consideration of expert judgments, data residual analyses, and a thorough review of each model’s characteristics such as magnitude and distance scaling and limitations such as applicability ranges of the model parameters. Example logic trees for GMMs with weights are shown in Figure 5.

We made several additions and modifications to GMMs applied in the 2023 NSHMs. For subduction zones, we have implemented new semi-empirical NGA-Subduction for Cascadia and Alaska subduction sources (Rezaeian et al., 2023). We supplement these NGA equations with two older GMMs for Cascadia that provide additional epistemic uncertainty at longer periods; these are similar and developed using computer simulated data for longer periods (T ≥ 1 s) but are downweighted because they are older and did not have the benefit of the most recent data. We implement adjustment factors to the CEUS GMMs to reduce bias with regional records; these adjustment factors are incorporated with reduced weight because they are new and less tested. New region-specific GMMs are implemented for Hawaii with one model (Wong et al., 2015) being downweighted because the residuals with data were much larger than for the other GMMs. Weights were presented at regional workshops and the CONUS GMMs were evaluated by the review panels.

Comparisons of the average weighted GMMs with some of the older GMMs are shown in Figure 6 for the median response spectral ground motions and standard deviations. Generally, the median ground motions for shallow crustal, subduction interface, and subduction intraslab earthquakes are similar to one another within about a factor of 2. However, the CEUS model medians and sigmas are considerably higher than for WUS GMMs, especially at short periods. In particular, Alaska interface and intraslab standard deviations are much higher than the earlier versions, which substantially affects hazard calculations.

To evaluate earthquake ground shaking for the hazard model, it is important to understand details of how ground motions are influenced by earthquake magnitude, distance, shallow soils (⁠$VS30$⁠), deep sedimentary basin deposits (Z_1.0 and Z_2.5) and other factors. Tens of thousands of strong ground-motion records from earthquakes have been recorded over the past century that provide insights into ground-shaking characteristics. The Pacific Earthquake Engineering Research (PEER) Center has developed ground-motion databases for earthquakes located within different tectonic regimes. For shallow crustal earthquakes, the NGA-West2 project (Bozorgnia et al., 2014) used 21,192 three-component records from 599 shallow crustal events M3–M7.9 (Ancheta et al., 2014). For stable continental earthquakes, the NGA-East project used 27,000 records from 82 earthquakes M2.5–M6.9 (Goulet et al., 2021a, 2021b) enhanced by simulations for larger magnitude earthquakes. For subduction zone earthquakes, the NGA-Subduction project (Bozorgnia et al., 2022) used 71,340 three-component recordings from 1883 events M4–M9 (Mazzoni et al., 2021). We do not apply all the available GMMs for this 2023 modeling but instead implement models that form a reasonable distribution of alternative input models that are broad enough so that we can capture epistemic uncertainty (i.e. center, body, and range). Figure 5 shows logic trees for WUS and CEUS GMMs that are described in detail in Moschetti et al. (2023a) and described more generally in the sections below.

For this 2023 NSHM, we modify the site response terms of some existing GMMs with data from numerical simulations and from seismic observations (Moschetti et al., 2023a). These GMMs and modifications to those equations apply realistic 3D basin-specific geometries and computer simulations using shear-wave velocity profiles, mostly from community velocity models, to guide GMM development. We consider region-specific measurements and proxy-based maps of time-averaged shear-wave velocities in the upper 30 m of the crust (V_S30) (e.g. Heath et al., 2020; Wald and Allen, 2007) and community velocity models that define the deeper sedimentary sequences quantified by Z_1.0 and Z_2.5—the depths to the 1.0 and 2.5 km/s shear-wave velocity horizons (Aagaard, 2023a, 2023c; Aagaard and Hirakawa, 2021; Ahdi et al., 2023; Lee et al., 2014; Magistrale et al., 2008; Shah and Boyd, 2018; Simpson and Louie, 2020; Stephenson et al., 2017)—to test basin-scaling models developed in the NGA-West2 GMMs to the Great Valley of California, to the Reno, Nevada, and Portland/Tualatin, Oregon basins, and to the Atlantic and Gulf Coastal Plains. In addition to the empirical GMMs, 3D numerical simulations are now used in Seattle (Frankel et al., 2009, 2018; Wirth et al., 2018a) and Los Angeles (CyberShake, Graves et al., 2011) to model basin effects. The effects of these new GMMs and sedimentary basin models are described in the sections below and in Moschetti et al. (2023a).

We commissioned a GMM review panel chaired by Jon Stewart to review the input models for the 2023 NSHMs. This review is presented in Appendix C in Electronic_Supplement_1.pdf of Moschetti et al. (2023a).

We apply the same GMMs in the WUS active crustal regions as considered in the 2018 NSHM using the NGA-West2 (Abrahamson et al., 2014, hereafter ASK14; Boore et al., 2014, hereafter BSSA14; Bozorgnia et al., 2014; Campbell and Bozorgnia, 2014, hereafter CB14; Chiou and Youngs, 2014, hereafter CY14). In addition to what is listed in Table SA-4 of Electronic_Supplement_1.pdf, we considered inclusion of the Graizer (2018) GMM for WUS. This GMM was not incorporated in this update cycle due to obstacles in implementation, but there are no known reasons why this GMM would not be included in future NSHMs—with additional implementation inputs. We retain the model of additional epistemic uncertainty developed for the 2014 NSHM to account for models that could vary from the NGA-West2 versions included in this update (Petersen et al., 2015a, 2015b; Rezaeian et al., 2015). Example logic trees for the WUS are shown in Figure 5.

We considered implementation of directivity models into the 2023 NSHM and active crustal GMMs, but decided to delay their use in NSHM until we have more complete vetting and agreement in the ground-motion modeling community concerning how to represent these earthquake effects. Withers et al. (2023) evaluated the effect of seismic directivity on hazard by implementing a median and aleatory adjustment to ground motion that accounts for the averaged effect of possible hypocenter locations along a fault plane. They focused on strike-slip faults with near-vertical dips, where directivity model predictions are generally consistent and implemented the directivity adjustments for the UCERF3 rupture forecast. Withers et al. (2023) found that azimuthal amplification can substantially increase hazard at the end of individual fault rupture traces but is substantially reduced when applied probabilistically, where there is interference from multiple, overlapping seismic sources. In order to further investigate directivity effects from complex fault geometries and better understand epistemic uncertainty, we support continued investigations into the directivity models and their effects on seismic hazard.

Compared to previous subduction GMMs, the NGA-Subduction project represents a substantial improvement in uniform data processing, multiple modeling platforms for alternative regions across the globe, and extensive review and evaluations. We updated the subduction GMMs using three of the new NGA-Subduction models (Abrahamson and Gülerce, 2022, hereafter AG20; Kuehn et al., 2020, hereafter KBCG20; Parker et al., 2022, hereafter PSBAH20). Various weighting schemes were considered that allow for combining regional (Cascadia and Alaska) and global versions of the new NGA-Subduction models and some of the older 2018 NSHM models. We considered a broad range of potential GMMs because the regional Cascadia models of AG20, KBCG20, and PSBAH20 suffer from very limited Cascadia strong-motion data. The final selected models include the three new NGA-Subduction GMMs while maintaining two of the three older models to better account for epistemic uncertainty at periods greater than 1 s (Atkinson and Macias, 2009, hereafter AM09; and Zhao et al., 2006, hereafter Zhao06). The AG20 GMM was considered as an update to the Abrahamson et al. (2016) which was applied in 2018. The NGA-Subduction models were supplemented with lower weighted older models (AM09 and Zhao06) to allow for additional epistemic uncertainty, especially for long-period ground motions. These older models are more consistent with the AG20 and M9 computer simulations than the new NGA-Subduction models.

A magnitude-scaling break point parameter is new in the NGA-Subduction GMMs (Campbell, 2020). This can result in substantial differences to ground-motion forecasts for large earthquakes where the rate of increase flattens beyond the break point. We did not incorporate variations of this parameter in this cycle, but sensitivity analyses are provided in Rezaeian et al. (2023). We could not adequately implement this improvement in the 2023 NSHM; however, this might be a topic for modifying epistemic uncertainty for large ground motions in future versions of NSHM.

We considered weighting the median GMMs and the lognormally distributed aleatory variabilities separately in logic-tree branches, but this was considered at a late stage in the update process and consensus on how to weight these models was not readily achieved. To represent epistemic uncertainty, we applied combinations of GMMs that were thought to be consistent with Cascadia and Alaska strong-motion data. For Cascadia subduction interface earthquakes, we apply the adjusted version of AG20 with weight 25% and the KBCG20 and PSBAH20 Cascadia versions with equal weight (25% each) along with the AM09 and Zhao06, which together receive 25% weight. For intraslab earthquakes, we apply AG20 Cascadia and Cascadia-adjusted equations, which receive a total weight of 25%—with most weight assigned to the adjusted version. The rest of the weight for intraslab events is given to KBCG20 and PSBAH20 Cascadia along with the older Zhao06 (equally weighted with 25%). An important improvement is the incorporation of epistemic uncertainties on the median ground motion, which were implemented as a three-point representation of a lognormal distribution for all three GMMs. Table SA-4 in Electronic_Supplement_1.pdf lists GMMs and weights.

The 2018 NSHM applied 17 final NGA-East models and 14 updated seed models with weights described by Petersen et al. (2020) and Rezaeian et al. (2021). For this 2023 NSHM, the CEUS ground-motion modeling implementations include two changes. One change involves a new correction to the nonlinear portion of the amplification model (Hashash et al., 2020) that is used in applying the NGA-East GMMs (V_S30 = 3000 m/s) to other V_S30 site conditions. The linear portion of the site amplification model as described by Stewart et al. (2020) has not changed. The second change involves a period-dependent modification to NGA-East GMMs to improve the fits of the residuals to updated compilations of CEUS strong-motion data (Moschetti et al., 2023a). Data misfits with the GMMs indicate an overprediction of ground motions for oscillator periods less than about 1 s and an underprediction at periods greater than 3 s (Moschetti et al., 2023a). These corrections apply to all V_S30 soil classifications. It is still unclear whether the GMM adjustment factors for the CEUS should be partitioned between NGA-East and site amplification models or whether the factors should be completely based on the NGA-East models alone. Adjusment factors were applied to reduce the bias of the model compared to regional observations (misfit). We convened a Tiger Team (also known as a GMM advisory team, see Note 1) to evaluate how we implement the NGA-East models. The recommendation of the Tiger Team (see Appendix F in Electronic_Supplement_1.pdf) was to taper the VS30 term between 1000 and 2000 m/s based on the residuals (Ramos-Sepulveda et al., 2023). In the current version of the model, we have not made this modification because the changes are relatively small. However, this could be considered for future updates. Figure 5 shows as example logic tree for CEUS GMMs. Table SA-4 in Electronic_Supplement_1.pdf provides a list of GMMs and weights.

In the case of Alaska, older GMMs that were used in the 2007 Alaska NSHM do not meet our GMM inclusion criteria and are therefore not considered. For earthquakes in active crustal regions, we apply the four NGA-West2 GMMs and a model of epistemic uncertainty in a logic tree of equally weighted branches consistent with that used for active crustal earthquakes in the CONUS model over the last few update cycles. Average crustal GMM medians in the 2023 Alaska NSHM are lower and aleatory variability (sigma) is higher than in the 2007 model (Figure 6). For Alaska subduction earthquakes, we implemented and evaluated both the global and Alaska regionalized NGA-Subduction interface and intraslab equations (AG20, KBCG20, and PSBAH20) along with the epistemic uncertainty and aleatory variability models specified in each. Our analysis of the NGA-Subduction GMMs for interface raised some questions and we therefore enlisted the help of a Tiger Team consisting of NGA-Subduction and other GMM experts to advise and make recommendations on the application of these new GMMs in the Alaska NSHM. The Tiger Team noted that the Alaska ground motion database is limited to moderate magnitude events at large distances and that there are issues with the Alaska regionalized models. At the recommendation of the Tiger Team, we applied the three global NGA-Subduction GMMs for interface with 1/6 weight each and bias-corrected versions of each model, also with 1/6 weight each. The bias corrections were developed using records that post-date the NGA-Subduction database and which are located closer to the population centers of south-central Alaska than the events used to develop the GMMs. Median ground motions are generally lower for the NGA-Subduction models, except at shorter distances and periods where they are higher (Appendix D of Electronic_Supplement_1.pdf). Average aleatory variability of the NGA-Subduction GMMs is higher (Figure 6), with individual NGA-Subduction models having sigma as high as 0.9 (e.g., PSBAH20). At the recommendation of the Tiger Team, we modified the KBCG20 aleatory variability model to match the published CB14 model to account for nonlinear site effects. The Tiger Team recommended the use of the Alaska regionalized NGA-Subduction intraslab GMMs as published because they are constrained by more data. For this update, we did not include amplification of long period ground motions over deep sedimentary basins as in the CONUS NSHM. Appendix A of Electronic_Supplement_1.pdf provides full details and weights (Table SA-4) of the GMMs selected for the 2023 AK NSHM.

We also updated the Hawaii shallow and deep GMMs in developing the 2023 NSHM (Petersen et al., 2022). The new 2023 NSHM considers five shallow GMMs (Atkinson, 2010; A10, ASK14, BSSA14, CB14, CY14) and three deep GMMs (A10, Zhao06, Wong et al., 2015; W15) that were compared to the strong ground-motion records on Hawaii to identify the equations that best fit the data. We used a total residual approach that compares the PDF for both mean and standard deviations and a log-likelihood test that considers the best GMMs for use in the model (McNamara et al., 2020). The new equations for Hawaii are shown in Table SA-4 of Electronic_Supplement_1.pdf with weights described in Petersen et al. (2022). We weighted the models based on residual analyses (McNamara et al., 2020; Petersen et al., 2022). In addition, we applied additional epistemic uncertainty similar to the 2018 NSHM and applied the empirical corrections to the A10 model for estimating ground motions from caldera collapses (A10_Caldera, Petersen et al., 2020). Site effects are implemented by evaluating soil profiles for Hawaii and the available strong ground-motion database. Hawaii sedimentary basin models are not available for amplifying long-period ground motions. Moreover, the V_S30 models do not seem to perform as well in Hawaii as they do in the WUS.

The 2023 NSHM provides maps for several values of V_S30 and sedimentary basin depth. These factors are important for many of the largest western cities. Relevant basin-specific data and models to implement these factors are not universally available, but NSHM applies them where there is confidence that available data are sufficient to result in an improvement in hazard estimates. For the 2018 NSHM, we applied sedimentary depth-based amplification models for WUS basins (Seattle, San Francisco, Los Angeles, and Salt Lake City); well-defined depth-based amplification models were not available for the CEUS or most other regions for this analysis. The semi-empirical models are conditioned on depth to a particular velocity horizon to account for amplification effects related to deeper velocity structure. Basin-depth models modify the active crustal, stable continental, and subduction GMMs for a proxy depth parameter included in the model: depth to the 1.0 km/s (Z_1.0) or the 2.5 km/s shear-wave horizon (Z_2.5). For active crustal and subduction GMMs, these are typically defined by each of the modeling teams for their GMM. These basin-amplification factors are correlated with the V_S30 amplification factors and cannot be used independently. We only applied the long-period basin amplification and not the deamplifications in several basins (Wasatch near Salt Lake City, Seattle, Portland, Tualatin, Seattle, and the Great Valley of California) where velocity data and available models have not been applied for this purpose, strong-motion recordings are lacking, and amplification models need to be tested further. We allow for deamplification in Los Angeles and San Francisco, as described below using a logic-tree approach. We did not amplify or deamplify ground motions within the Reno basin for similar reasons. For Reno the improvement was not substantial when using these local depth-based models, uncertainty in published velocity models is high, and there is not much additional amplification due to basin-depth scaling compared to that captured by V_S30 scaling model.

For this 2023 NSHM, we apply basin-depth amplification models for CONUS where local basin-depth models are available, but not for Alaska or Hawaii where there are no accepted community velocity models available. We also use a model that can be applied to the NGA-East GMMs and seed models to account for the deep-sediment effects of the Atlantic and Gulf Coastal Plains. For updating the basin-amplification models in the WUS, we considered updated community seismic velocity models in the San Francisco Bay Area (Ahdi et al., 2023), new basin-depth models for the California Great Valley (Ahdi et al., 2023), Reno (which was not ultimately implemented), and Portland–Tualatin basins (Ahdi et al., 2023), new treatment of basin-amplification models at shallow-basin and out-of-basin sites, as well as the M9 simulations for the Puget Lowlands region (Frankel et al., 2018; Wirth et al., 2018b) and CyberShake simulations for the greater Los Angeles metropolitan region (Graves et al., 2011). The M9 and CyberShake simulations are not directly used to establish ground-motion levels. However, these data are used to modify the GMMs through implementation of amplification factors that are a function of sediment depth. This modification was also suggested by Nweke et al. (2022a) with further discussion presented in Moschetti et al. (2023a).

We apply the basin depths from seismic velocity models for Seattle, San Francisco, Great Valley, Los Angeles/Ventura, Portland/Tualatin, and Salt Lake City (Aagaard, 2023a, 2023c; Aagaard and Hirakawa, 2021; Lee et al., 2014; Magistrale et al., 2008; Simpson and Louie, 2020; Stephenson, 2007; Stephenson et al., 2017) to assess Z_1.0 and Z_2.5 depth parameters. The basin-amplification models are derived along with the NGA-West2 equations (ASK14; BSSA14; CB14; CY14). We evaluated the basin-depth models in Reno (Aagaard, 2023b; Simpson and Louie, 2020) and found that because the basins were relatively shallow, the equations did not predict substantial amplifications in those regions. Using the local basin-depth models did not provide additional predictive capability (Ahdi et al., 2023)—similar to our 2018 analysis for Salt Lake City (Petersen et al., 2020). Ground-motion amplifications have been observed in each of these regions but scanty strong-motion data from large earthquakes are available to assess basin amplifications. Additional work would be beneficial to evaluate geometric basin influences and amplification characteristics and evaluate differences between these shallow basins and the deeper basins in California where most of the strong-motion data are located that underpin the development of these amplification models. We weight each of the basin-depth models with 100% weight because we are improving the information using Z_1.0 and Z_2.5 depth information that is not applied in the pre-2018 NSHMs. We could not add additional epistemic uncertainty to the amplification models (e.g. depth uncertainties of Z_1.0 and Z_2.5) because the methodology is currently not available.

We give full weight to the Great Valley Z_1.0 and Z_2.5 models because no alternatives are available. We applied the Z_1.0 and Z_2.5 in zones shown in Figure 7. Basin-depth models for Reno, the Wasatch Front, and Portland/Tualatin are shallower than the basins in California, resulting in amplification at very few sites, relative to the basin-amplification effects implicit in V_S30-scaling models built into the NGA-West2 GMMs. The Great Valley of California is much deeper and results in amplifications for long periods, especially in the southwestern portion of the zone.

In preparing the 2018 NSHM, we debated whether to modify ground motions from the NGA-West2 models based on sedimentary thickness inside and outside of basins. This led to many discussions with our NSHM-SC and in open workshops (Petersen et al., 2020). In the end, we decided that we lacked sufficient information to warrant lowering predicted ground motions from GMMs at the basin-edge regions where past earthquakes caused extensive damage and ground-motion variability was large (e.g. Santa Monica, California; West Seattle, Washington). Thus, for the 2018 hazard assessment, we amplified ground motions only in the deepest portions of the sedimentary basins and reverted to average GMMs everywhere else with no basin-amplification terms being applied.

For the 2023 NSHM, we revisit this decision considering new region-specific analyses for the San Francisco Bay Area and the greater Los Angeles region presented in Moschetti et al. (2023a) and in Nweke et al. (2020, 2022b). We tested the new California models for their effectiveness in predicting amplifications throughout these urban regions and found that the NGA-West2 model with basin response terms Z_1.0 and Z_2.5 perform well in these two regions (Moschetti et al., 2023a). For this 2023 NSHM, we develop a logic tree that accounts for modified ground motions in areas outside of basins, at basin edges, and inside of basins (Moschetti et al., 2023a). We consider a logic-tree uncertainty with half weight given to the 2018 model that only amplifies ground motions within the deepest portion of the basins, and half weight given to empirical basin-depth amplification models (Z_1.0/Z_2.5) of NGA-West2 that deamplify long-period ground motions for relatively shallow sediments and amplify motions in deeper sediments (refer to Table SA-6 for the San Francisco GMM logic tree and Table SA-7 for the Los Angeles GMM logic tree in Electronic_Supplement_1.pdf). Deamplifications of ground motions have been observed in relatively small or shallow sedimentary structures (e.g. the valley province in Nweke et al. 2022b).

We apply these logic trees because our studies indicate that the NGA-West2 model is appropriate for use with the current database of strong motion in the San Francisco and Los Angeles regions, which lower ground motions by about 30% in both of these urban areas (Moschetti et al., 2023a). However, it is important to recognize that there are large epistemic uncertainties and aleatory variabilities that have not been explored that could influence these results in the future. For example, (1) this is the first time we are using the community velocity models for this purpose of amplifying ground motions in shallow soils and we need to ensure that the models are appropriate for this purpose; (2) we do not have strong-motion recordings from large San Andreas system earthquakes, so these effects have not been recorded or explained; and (3) there exist several unmodeled uncertainties such as at basin-edge sites due to converted waves and in recent work by Withers et al. (2023), which shows about 5% increases in the hazard at the 2% in 50-year level when applying the Watson-Lamprey (2018) directivity equations for the San Francisco Bay Area. For other regions we will need to carry out a similar analysis as we did in the urban areas of California to ensure that these basin models are consistent with the NGA-West2 amplification models. Therefore, for this 2023 update, we applied a logic-tree structures to account for the NGA-West2 models as well as unknowns associated with the unmodeled uncertainties.