Machine learning approaches to ground motion models for duration and velocity-based intensity measures in Türkiye with magnitude-dependent performance assessment

Impact of ground motion model selection on the benefit from monitoring data

In seismic hazard analyses, it is common practice to consider only earthquakes with magnitudes greater than 5 while assuming that smaller events cannot damage engineered structures. However, this assumption may not hold for induced seismicity, where earthquakes often occur at very shallow depths with rupture distances of only a few kilometers. The steep attenuation of ground motions within the first 10 km in small-magnitude events can yield median short-period accelerations that exceed those predicted by conventional ground-motion models. The combination of higher ground-motion amplitudes and elevated earthquake rates challenges the assumption that earthquakes smaller than magnitude 5 cannot cause damage to earth structures. This study evaluates the damage potential of small-to-moderate (magnitudes 3–5) induced earthquakes occurring at shallow depths on the seismic performances of earth canal dykes. The earthquake scenarios and ground motions selected for the assessments represent features expected from induced seismicity associated with wastewater injection. For dykes with a yield acceleration of 0.1 g , the results indicate that significant deformations (>10 cm) may occur during earthquakes with magnitudes between 3.0 and 4.0 at a rupture distance of about 3 km. The methodology presented herein also provides a framework for assessing the minimum threshold magnitudes of other engineered systems in regions affected by induced seismicity.

ABSTRACT Recently, generative models have become a computationally efficient alternative to physics-based numerical simulations of ground motions. Neural networks can learn from existing ground-motion data to generate unobserved ground-motion data at new source and site locations. A key challenge with generative models is ensuring that predicted ground motions remain within a physically realistic range. For this purpose, we developed an empirical, nonergodic ground-motion model (GMM) for small-magnitude earthquakes in the San Francisco Bay area based on about 5000 recordings per component for Mw ≤ 4 earthquakes. The nonergodic GMM predicts spatially varying median source, site, and path effects for both the Fourier amplitude spectrum (FAS) and the Fourier phase derivative (a proxy for duration), as well as the corresponding epistemic uncertainty for each term. For FAS, our model shows above-average source and site effects in the western part of the region and below-average effects in the eastern part, with regional effects exhibiting larger spatial correlation lengths with increasing frequency. For duration, the source term is negligible for small-magnitude earthquakes, and the site term leads to site-specific variations up to 5 s. Path effects for FAS and duration depend on the source–site pair and are extrapolated spatially using recent methods for path-effect modeling. The aleatory variability of the within-site within-path residuals is similar to the variability found in previous studies for other regions. The nonergodic model provides two key contributions: first, median adjustment terms that are transferable to larger magnitude earthquakes, further reducing aleatory variability in probabilistic seismic hazard analysis; second, region-specific criteria for validating machine learning-based ground-motion generators to evaluate whether synthetic ground motions exhibit physically realistic source, site, and path effects.

This study selected a simplified water intake tower model, simplifying the physical structure into a cantilever model, and MATLAB software (R2010b) was used to develop a rapid seismic response analysis program for the structure. Thirty near-fault pulse and non-pulse ground motions were selected as the input ground motions for this analysis. Peak ground velocity (PGV) was used as the intensity parameter for the ground motions. The acceleration, cross-sectional rotation, and lateral curvature of the simplified water intake tower model were calculated for ground motions modulated with different PGA amplitudes. The acceleration, maximum shear force, and cross-sectional rotation of the simplified water intake tower model were also calculated for ground motions modulated with improved effective peak acceleration (IEPA) and improved effective peak velocity (IEPV). The study showed that the seismic response of the simplified water intake tower model for near-fault ground motions modulated with different intensities of PGV amplitude modulation was closer to the unmodulated response order. PGV as an intensity parameter did not affect the acceleration response amplification factor of the water intake tower and hoist chamber. The AC coefficient indicated that PGV was less suitable for pulse-type earthquake amplitude modulation than PGA. Compared with PGA amplitude modulation, IEPA amplitude modulation is more suitable for pulse-type seismic motion, while IEPV amplitude modulation has less impact on pulse-type seismic motion.

After a major earthquake, seismometers are often deployed temporarily near the epicenter to monitor aftershock events. These aftershock records are valuable resources for nonergodic site term development, which improves ground motion prediction accuracy. The term ergodic site term refers to temporally and spatially averaged site conditions, while the nonergodic site term accounts for location‐specific variations. A ground motion model (GMM) predicts earthquake intensity measures (IM) at a site. The nonergodic approach reduces prediction uncertainty of the GMM, leading to more accurate IM estimates. Unlike the ergodic GMM method, the nonergodic approach could capture local variability using specific site terms. This study refines the IM map of the M 5.4 Pohang earthquake by incorporating nonergodic site terms derived from aftershock records. The detailed procedure includes: 1) development of nonergodic site terms including residual analysis and interpolation techniques; 2) derivation of site amplification modification factors by coupling a GMM with regional site amplification models; 3) validation using statistical metrics such as RMSE and MAE using two validation stations. As a result, the peak ground acceleration (PGA) shows 30% and 16% reductions in prediction error compared to the PGAs without nonergodic site terms at the two validation sites. The use of aftershock records for the assessment of the nonergodic site terms has a benefit in that the ray paths of earthquake waves to a site are shared among events so that the uncertainty of the path term is minimized. These findings highlight the importance of deploying temporary stations at sites of interest after a main event, which can refine the seismic intensity of the mainshock.

I develop independent logic trees for aleatory variability for crustal and subduction‐zone (interface and intraslab) earthquakes for seismic hazards analyses in Puerto Rico and the U.S. Virgin Islands (PRVI) from existing suites of ground‐motion models (GMMs) and from ground‐motion datasets, including a regional PRVI dataset. The aleatory variability models are parameterized using a partially nonergodic partitioning of standard deviation that consists of independently developed between‐event (), site‐to‐site (), and event‐corrected single‐station () standard deviation components. The effects of nonlinear site response on aleatory variability are incorporated through additional terms that modify the standard deviation components. Because one goal of this work is to develop independent logic trees for aleatory variability that synthesize the aleatory variability models from GMMs, I make use of the functional forms of the input GMMs. The PRVI dataset contains a limited number of stations with high‐quality site metadata and does not contain records from earthquakes with magnitudes greater than 6.1, so I choose not to develop the aleatory variability models from the regional dataset alone. Instead, the standard deviation components from regional ground‐motion data are evaluated against the components derived from GMMs and from available global datasets, and regionalized standard deviation components are incorporated where there is evidence that regional effects exhibit substantial differences. The resulting logic trees for aleatory variability consist of models of and that are consistent with semiempirical GMMs for active crustal and subduction‐zone regimes, and two alternative models of , including one model that exhibits site‐to‐site variability informed by PRVI data, with values that exceed global models. The aleatory variability models may be considered in future hazards assessments in PRVI to simplify the hazard calculations, to incorporate regional ground‐motion variability effects, and to enable direct logic‐tree weighs of aleatory variability.

It is essential to improve the accuracy of ground‐motion predictions based on long‐term seismic observations because this allows for more accurate ground‐motion models (GMMs). Therefore, this study used a large number of ground‐motion records, focusing on existing seismic observation stations, to propose a GMM with improved accuracy by introducing spatially varying coefficients and a calibration method utilizing large datasets. The study also uniquely employed empirical site terms to eliminate errors due to site characteristic modeling, further enhancing the prediction accuracy at specific stations. Subsequently, a nonergodic GMM for predicting the maximum seismic wave acceleration was constructed using a practical study that incorporated a large dataset of ground‐motion records from a high‐density strong‐motion observation network in Japan. The introduction of spatially varying coefficients allowed the spatial variability of the propagation path and source characteristics to be considered, which effectively reduced ground‐motion prediction errors. The resulting GMM enables more accurate probabilistic analyses of seismic hazards.

This article presents a study aimed at deriving adjustment factors tailored for the median of the NGA‐East ground‐motion models (GMMs), specifically targeting regions within the Gulf Coast and the Atlantic Coastal Plain. The adjustment factors are formulated based on sediment thickness and R rup within these regions, known as the Coastal Plains. This research utilizes sediment thickness contour maps provided in another study.Residuals are computed by subtracting the natural logarithms of the observed data from those predicted by the median of the NGA‐East GMM and using a comprehensive dataset. This dataset combines the NGA‐East dataset, the Chapman and Guo dataset, and the newly compiled and verified dataset by USGS. Residual analyses are conducted through mixed‐effects regression to segregate total residuals into between‐event and within‐event components. We regress the within‐event residuals using a functional form incorporating sediment thickness and rupture distance to derive correction or adjustment factors for stations located within the Coastal Plains. Results suggest that, for stations within the Coastal Plains, the residual trend has largely been mitigated across most periods concerning site and path terms using the proposed adjustment factors. The results of this study hold applicability in seismic hazard and risk assessments for sites located within the Coastal Plains.

An indispensable component of any site‐specific probabilistic seismic hazard analysis (PSHA) is incorporating local site effects through site response analyses, rather than relying on generic amplification factors in ground‐motion models (GMMs). Traditionally, the seismic hazard is calculated at a buried rock horizon and then convolved with amplification factors for the overlying layers, which requires the definition and characterization of the buried rock profile (without which neither the model for rock motions nor the site response analyses can be developed). An alternative approach is to retain the host reference rock profile in the GMM for the hazard calculations and then perform the host‐to‐target site response adjustment for the full profiles in a single step. This approach is illustrated by its application to a site‐specific Senior Seismic Hazard Analysis Committee Level 3 PSHA for Idaho National Laboratory. The site adjustment factors are calculated using a logic‐tree approach that captures the full epistemic uncertainty in these factors. This is not necessarily achieved when the logic tree only captures uncertainty in the near‐surface shear‐wave velocity profile. The application illustrates the use of borehole and noninvasive shear‐wave velocity measurements, in conjunction with local recordings of weak earthquake motions, to constrain the logic‐tree branches and their associated weights.

Abstract Egypt has adopted long-term strategies for urbanizing new cities across the country for resolving the overwhelming problem of overcrowded cities. New Mallawi is an example of such new cities. Egypt is occasionally being shaken by big-sized earthquakes such as the destructive 1754 local Cairo event and the strong ( 1995) Gulf of Aqaba earthquake; therefore, the buildings need to be designed to withstand the shaking of future damaging events. The hazard model of this research is based on an up-to-date catalogue of earthquakes. To obtain realistic results, lateral changes in soil characteristics are considered through using the parameter of average shear-wave velocity in near-surface layers. In addition, different state-of-the-art formulations of declustering temporal and spatial windows are used to represent the epistemic uncertainty in earthquake recurrence functions. Moreover, a set of ground motion models is used for better consideration of the uncertainty of the empirical estimates, and uncertainty in maximum earthquake size is treated. The resulting earthquake shaking is rather low, and the observed variability is affected by lateral changes in soil properties. Various hazard levels are investigated for reaching a better earthquake-resistant design that accounts for damaging earthquakes of hypothetically long return periods. This is done in an attempt to examine low-probability hazards after what was observed in the 2023 Turkey earthquake. The obtained results can be used for urban planning and risk mitigation at the New Mallawi city.

Abstract eGSIM is an OpenQuake-based Python library and web application for analysis and interpretation of ground motions (observed and/or simulated) and ground-motion models (GMMs). Developed in the frameworks of the European Facilities for Earthquake Hazard and Risk and the European Plate Observing System (EPOS), the eGSIM webservice can be used in many applications, from data visualization through its graphical user interface in the user's browser, to programmatic access to its application programming interface in the user’s code. Its core functionalities are based on two types of GMMs comparisons: “Model-to-Model,” in which the expected ground motions and aleatory variabilities are compared for a range of user-configurable scenarios, and “Model-to-Data,” in which the GMMs can be compared against databases of ground motion observations. We showcase eGSIM’s main features and highlight many of the potential applications, before illustrating its use in an exploratory analysis comparing ground motions from observations and physics-based simulations. These examples demonstrate not only the ease with which eGSIM can perform common analysis of GMMs for hazard and risk modelling, but also its versatility for application to other customized exploratory analyses and for integration in state-of-the-art high-performance-computing workflows.

Development of ground motion models using supervised learning: application to North and Northeast India

ABSTRACT This study develops a comprehensive framework for assessing time and state‐dependent aftershock damage accumulation under an M9.0 megathrust interface earthquake in the Cascadia Subduction Zone (CSZ). The framework integrates aftershock probabilistic seismic hazard analysis (APSHA) and state‐dependent fragility analysis (SDFA) within a Markovian model to quantify unit‐time probability transitions to higher damage states under aftershocks. Key advances and novelties include: (1) a full Poisson‐process formulation to deal with high aftershock occurrence rates, (2) CSZ geometry and rupture‐consistent simulations for developing conditional aftershock distance distributions, and (3) the utilization of ordinal regression to deal with the nonlinearity observed in the state‐dependent seismic demand data for efficient and reliable SDFA. The framework is applied to assess the damage accumulation of existing steel moment‐resisting frame buildings located in Vancouver and Victoria, Canada. The multi‐step application consists of site‐specific APSHA, the state‐of‐the‐art subduction ground motion model, a realistic CSZ geometry model, and the selection and pairing of subduction zone mainshock and aftershock motions for nonlinear response history analyses. Results show that the transition to higher damage states under aftershocks increases sharply within the first few days after the mainshock. The damage accumulation is more significant when the building sustains severe mainshock damage. This study provides a detailed multi‐step procedure for aftershock damage accumulation assessment when facing forthcoming megathrust interface earthquakes in the CSZ.

ABSTRACT Northeast India (NEI), one of the world’s most seismically active regions, lacks a region-specific ground motion model (GMM) spanning the full range of magnitudes and distances. To address this gap, we develop an LSTM-based GMM using magnitude, distance, depth, site class, tectonic setting, and a region flag as predictors, and PGA, PGV, and PSA as targets. Two approaches are evaluated: a hybrid model combining NGA-SUB and NEI data, and a transfer-learning model fine-tuned on NEI records. The hybrid model shows low bias, reduced sigma, and strong predictive performance, and preliminary PSHA results support its suitability for NEI regional seismic applications.

Abstract Ground-motion simulations of notable earthquakes in the central and eastern United States are limited and typically assume 1D Earth structure. In this study, we use a 3D seismic velocity model to better constrain the depth and focal mechanism of the 5 April 2024, moment magnitude 4.8 Tewksbury earthquake and investigate the spatial variability of earthquake ground motions and the effects of nearby sedimentary basins. We perform earthquake ground-motion simulations up to 0.5 Hz using the 3D spectral-element wave-propagation solver SPECFEM3D over a region 280 km wide by 260 km long by 77 km deep. Topography and subsurface geophysical structure are assigned using the U.S. Geological Survey (USGS) National Crustal Model with a minimum shear-wave velocity of 200 m/s. We use earthquake time series from 13 broadband seismic stations in the region that have a uniform azimuthal distribution and epicentral distances ranging from 76 to 131 km to compare with synthetics and explore the effects of 1D versus 3D seismic structure on focal mechanism and depth solutions. Ground-motion intensity metrics are also presented relative to the Next Generation Attenuation-East Project (NGA-East) ground-motion models (GMMs) currently used in seismic hazard assessments for the region. We find that the 3D model, which reveals a wide spatial variability of period-dependent ground motions, yields better predictions of earthquake ground motions relative to the 1D model and the NGA-East ergodic GMM, with a 76% reduction of residual variance in observed ground motions averaged over 3, 5, 7, and 10 s periods. Use of the 3D model to solve for a focal mechanism yields a shallower focal depth at 4 km and a shallower east-dipping focal plane relative to the USGS regional moment tensor and Global Centroid Moment Tensor. Our study demonstrates that use of 3D seismic velocity models can improve estimates of earthquake focal mechanisms, ground motions, and seismic hazard.

ABSTRACT In this study, we propose a ground-motion model (GMM) to estimate vertical ground-motion components in central and eastern North America (CENA) using the referenced empirical method (REM) introduced by Atkinson (2008). The GMM developed in this study is the first vertical GMM for CENA using REM. To account for epistemic uncertainty in choosing a reference model, we considered three alternative models for the host region: Stewart et al. (2016), Gülerce et al. (2017), and Bozorgnia and Campbell (2016). We began by computing the vertical response spectrum for recorded motions in the Next Generation Attenuation-East Project data set. Next, we calculated residuals between the vertical-component pseudo-spectral accelerations from 0.01 to 10.00 s and the average predictions of the three reference models. We then applied a mixed-effects regression technique to derive adjustment factors for refining these predictions and to estimate the standard deviation components. The proposed GMM is assessed through residual analyses and comparisons with recorded data as a final step. The proposed GMM does not exhibit any discernible residual trends with distance and magnitude and appropriately accounts for site effects.

ABSTRACT Simulated ground motions play a crucial role in seismic engineering applications. The present study introduces a novel conditional generative adversarial networks (CGAN) based model to generate the time‐frequency representation of the ground motions and to use the iterative power and amplitude correction (IPAC) algorithm, referred to as the IPAC‐CGAN method, to simulate the nonstationary non‐Gaussian ground motions for scenario events. The proposed CGAN‐based model is trained using time‐frequency representations derived from recorded ground motions originating from strike‐slip fault earthquakes, obtained via the S‐transform. It effectively captures the power distribution in the time‐frequency domain for specific scenario events. Unlike prior GAN‐ or CGAN‐based models in the literature, which assume fixed durations across scenarios and lack control over the marginal probability density function (PDF) of sampled records, our approach innovatively adapts ground motion duration to earthquake magnitude, rupture distance, and site conditions. The IPAC‐CGAN method enables the generation of ground motions with durations that vary according to earthquake magnitude, rupture distance, and site conditions, which were usually fixed for all scenarios in the model developed based on GAN or CGAN in the literature. The combination of the IPAC algorithm and CGAN ensures accurate replication of the nonstationary, non‐Gaussian characteristics observed in actual seismic records, while in the literature, there is no control on the marginal PDF of the sampled record. Validation of the proposed method involves comparing pseudospectral acceleration statistics between simulated and actual ground motions. Predicted values from ground motion models are also considered for comprehensive assessment. Additionally, ductility demand comparisons for single‐degree‐of‐freedom systems using the Bouc–Wen hysteretic model are conducted using simulated and recorded ground motions. In all cases, the results show that the statistics of the linear and nonlinear responses of the simulated ground motions agree well with those of the recorded seismic ground motions. This validation underscores the effectiveness of the proposed method in accurately simulating realistic seismic ground motions.

Unsupervised characterization of site conditions from detailed information for improved ground motion modeling

New conditional ground motion model for permanent displacement in near-fault zones