Electronic Theses and Dissertations Archive

Date

2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Civil Engineering

Committee Chair

Shahram Pezeshk

Committee Member

Arash Zandieh

Committee Member

Charles Camp

Committee Member

Roger Meier

Abstract

Seismic hazard analysis depends on reliable prediction of ground motion intensity and accurate estimation of earthquake magnitude. Data-driven approaches have gained attention as potential complements to traditional regression-based and waveform-based methods. This dissertation investigates the role of advanced machine learning frameworks in improving both ground-motion modeling and moment-magnitude estimation. The study first examines the current state of machine learning applications in ground motion prediction and identifies key methodological challenges, including model validation, residual behavior, and physical consistency. Building on this context, new machine learning-based GMMs are developed for small-to-moderate induced earthquakes in Central and Eastern North America. The models are trained on nearly 31,000 ground-motion records with moment magnitudes between 3.0 and 5.8 and hypocentral distances less than 200 km. Input variables include MW, Rhypo, and VS30, and the models predict peak ground acceleration (PGA) and multiple 5%-damped pseudo-spectral accelerations. Several nonlinear algorithms are evaluated, including artificial neural networks, kernel ridge regression, random forest regression, and gradient boosting regression. An ensemble framework is constructed to combine these models, resulting in improved prediction accuracy and smoother spectral behavior. Model performance is evaluated using error metrics, residual trends with magnitude and distance, and testing on independent events. The research further develops a machine learning framework for estimating moment magnitude using the NGA-West2 database. The input parameters include PGA, PSA at 21 periods, Rhypo, Ztor, fault mechanism, and VS30. Multiple algorithms are trained and combined using a stacking ensemble approach. After validating the model using records with known M_W, the framework is applied to events for which M_W is not originally reported. The results show that ground motion parameters contain sufficient information to support magnitude estimation and can be used to verify magnitudes and update earthquake catalogs. Throughout the dissertation, emphasis is placed on ensuring that machine learning models remain consistent with established seismic hazard principles. The results demonstrate that when properly validated, ensemble and nonlinear learning techniques can improve predictive performance while preserving physically meaningful trends. This work contributes to the integration of data-driven modeling within practical seismic hazard analysis and provides a structured framework for future applications.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest/Clarivate.

Notes

Open Access

Download

Share

COinS