Preliminary Findings on the Application of mHVSR for Data-Driven Site Response Prediction in California

Ergodic ground motion models (GMM) typically use time-averaged shear-wave velocity in the upper 30 m (VS30) as the site parameter but cannot capture site-specific resonance or frequency-dependent site effects. To address this, we explore using microtremor-based horizontal-to-vertical spectral ratios (mHVSR) to better reflect these variations and reduce uncertainty. This involves identifying predictive attributes of mHVSR curves beyond obvious resonant peaks, which are not easily detected visually. To aid in this, we apply three machine learning models—random forest, gradient boosting, and adaptive boosting—to a dataset of 700 California sites. These models identify patterns in mHVSR that relate to site response. Results show a significant reduction in site-to-site standard deviation (~0.1–0.2 natural log units) at low frequencies (f < 2 Hz). However, the physical basis for this improvement remains unclear. These findings represent early insights from ongoing research.