Application of Data-Driven Approaches for Estimation of Site Response Utilizing mHVSR and Soil-Based Proxies in California

Francisco Javier G. Ornelas, Christopher A. de la Torre, Tristan E. Buckreis, Chukwuebuka C. Nweke, Scott J. Brandenberg, & Jonathan P. Stewart

Submitted September 7, 2025, SCEC Contribution #14767, 2025 SCEC Annual Meeting Poster #TBD

Ergodic models for predicting linear site amplification commonly rely on simple scalar proxies, such as the time-averaged shear-wave velocity in the upper 30 meters of the subsurface (VS30). These parameters are incorporated into ground motion models (GMMs) through regression against predefined functional forms to best fit observational data. While these functional forms may not always perfectly capture the data trends, they extrapolate in a physically reasonable manner beyond the data range, which can be important for practical applications. In this study, we investigate the potential of data-driven approaches, specifically machine learning (ML) techniques, to model site amplification without relying on explicit functional forms. Our aim is to explore the impact of incorporating a broader range of site characterization parameters—both scalar and vector-based—within a flexible, data-adaptive framework. We focus on microtremor-based horizontal-to-vertical spectral ratio (mHVSR) metrics, including the full spectral shape and peak-based features such as the predominant frequency (f0) and corresponding amplitude (a0). These are evaluated in conjunction with traditional proxies such as VS30 and depths to shear-wave velocity isosurfaces (z1.0 and z2.5). Using a random forest regression model, we assess predictive performance across various proxy combinations for 685 sites in California. Results indicate significant improvements in accuracy and reductions in epistemic uncertainty relative to the benchmark GMM BSSA14. While further analyses are required to fully understand the behavior of models of this type before practical application, this study provides valuable insights into how different parameters influence site response, which can guide the development of more robust site amplification models in the future.

Key Words
Site Response, HVSR, Machine Learning

Citation
Ornelas, F. G., de la Torre, C. A., Buckreis, T. E., Nweke, C. C., Brandenberg, S. J., & Stewart, J. P. (2025, 09). Application of Data-Driven Approaches for Estimation of Site Response Utilizing mHVSR and Soil-Based Proxies in California. Poster Presentation at 2025 SCEC Annual Meeting.

Related Projects & Working Groups
Ground Motions (GM)