A Data-Driven Approach to Evaluate Site Amplification of Ground-Motion Models Using Vector Proxies Derived from Horizontal-to-Vertical Spectral Ratios

Abstract

ABSTRACT This study develops a data-driven framework to improve the prediction of site amplification in ground-motion models (GMM) using horizontal-to-vertical spectral ratios (HVSR) proxies derived from strong-motion data. Three machine learning algorithms (multiple regression, random forest, and support vector machine [SVM]) were implemented in an automated data-driven workflow that included feature selection (filter-based, wrapper-based, and embedded methods), hyperparameter tuning, and cross-validation modules. The site-to-site variability of resultant data-driven models was then compared to a baseline GMM. In addition, sensitivity analysis on framework choices was carried out. The results show that data-driven models with HVSR proxies provide lower site-to-site variability than conventional GMM. The best performing data-driven model (i.e., SVMs) showed an average of 24.1% (and up to 50.8%) lower site-to-site variability than the baseline GMM, where the difference was more significant at lower frequencies. Finally, all feature selection methods favor using a vector of the HVSR spectrum over single-valued HVSR proxies such as peak frequency and amplitude.