Scholarlink

<p>Local site effects represent a major source of ground-motion variability; however, the direct prediction of horizontal-to-vertical (H/V) spectral ratios remains relatively underexplored in seismic hazard modeling. This study introduces a data-driven framework for estimating period-dependent PSA-based H/V ratios using a multivariate ensemble of machine learning models. A comprehensive dataset comprising 38,504 strong-motion records from the Türkiye Strong-Motion Database (1976–2023) was utilized to capture the influence of seismic source (moment magnitude, Mw), propagation path (hypocentral distance, Rₕᵧₚ, and focal depth), and site conditions (Vs₃₀ and station elevation) across 946 stations. Three regression algorithms Gaussian Process Regression (GPR), Gradient Boosted Regression Trees (GBRT), and Random Forest (RF) were combined within an ensemble framework, with optimal period-dependent weights determined using a Genetic Algorithm (GA). The optimized ensemble achieved a mean root mean square error (RMSE) of 0.68 over the spectral period range of 0.01–4 s, corresponding to a 5.6% reduction in prediction error relative to the best-performing standalone model (RF) and an improvement of approximately 20% compared to the GPR baseline. In addition to its quantitative performance, the model demonstrates consistency with established seismological principles, effectively capturing low-frequency resonance behavior in soft-soil conditions and flatter spectral responses in stiff-rock environments. These findings highlight the potential of optimized machine learning ensembles as a reliable and scalable alternative for site-effect characterization, particularly in regions where detailed subsurface information is limited.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

H/V spectral ratio estimation using ensemble machine learning models: a case study from Türkiye

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Ahmet Özdil,
  • Hasan Erbay,
  • Mohammadreza Najaftomaraei

摘要

Local site effects represent a major source of ground-motion variability; however, the direct prediction of horizontal-to-vertical (H/V) spectral ratios remains relatively underexplored in seismic hazard modeling. This study introduces a data-driven framework for estimating period-dependent PSA-based H/V ratios using a multivariate ensemble of machine learning models. A comprehensive dataset comprising 38,504 strong-motion records from the Türkiye Strong-Motion Database (1976–2023) was utilized to capture the influence of seismic source (moment magnitude, Mw), propagation path (hypocentral distance, Rₕᵧₚ, and focal depth), and site conditions (Vs₃₀ and station elevation) across 946 stations. Three regression algorithms Gaussian Process Regression (GPR), Gradient Boosted Regression Trees (GBRT), and Random Forest (RF) were combined within an ensemble framework, with optimal period-dependent weights determined using a Genetic Algorithm (GA). The optimized ensemble achieved a mean root mean square error (RMSE) of 0.68 over the spectral period range of 0.01–4 s, corresponding to a 5.6% reduction in prediction error relative to the best-performing standalone model (RF) and an improvement of approximately 20% compared to the GPR baseline. In addition to its quantitative performance, the model demonstrates consistency with established seismological principles, effectively capturing low-frequency resonance behavior in soft-soil conditions and flatter spectral responses in stiff-rock environments. These findings highlight the potential of optimized machine learning ensembles as a reliable and scalable alternative for site-effect characterization, particularly in regions where detailed subsurface information is limited.