PhD Final Defense – Wendi Zhao

Data-Driven Modeling of Ground Response in Pressurized Mechanized Tunneling: From Empirical Evaluation to Domain-Specific Deep Learning Models

Advisor: Professor Youssef M. A. Hashash

Abstract

Urban rail expansion increasingly relies on Earth Pressure Balance Tunnel Boring Machines (EPB-TBMs) to construct tunnels beneath densely developed urban environments, where controlling tunneling-induced ground movement is essential for protecting buildings, utilities, transportation corridors, and other critical infrastructure. Although modern EPB-TBMs generate extensive operational and monitoring data, these records are often fragmented across projects and are rarely integrated to improve engineering understanding or support future tunneling applications.

This dissertation built a multi-project database by integrating operational, geological, and monitoring records from three Los Angeles Metro twin-tunnel EPB-TBM projects: the K-Line Tunnel, Regional Connector, and Purple Line Extension Section 1. The resulting database provides a reusable foundation for data-driven tunneling research and engineering applications. Using this database, the dissertation first investigates the relationship between EPB-TBM support pressure and tunneling-induced surface settlement under modern construction practices. Field observations from the three projects establish a consistent relationship between machine support pressure and ground response, providing new empirical insight into the mechanisms governing settlement and the influence of operational control during excavation.

Building on this unified dataset, the dissertation introduces the LA Basin Foundation Model, a domain-specific deep learning model for mechanized tunneling. The model integrates EPB-TBM operational histories, geological information, geometric characteristics, and field observations within a unified multi-task learning framework to simultaneously estimate surface settlement and excavation-face ground conditions. Explainable artificial intelligence is used to identify the operational parameters that most strongly influence model predictions, enabling the development of a simplified model that maintains predictive performance while improving interpretability and deployment efficiency. The model is further evaluated for its ability to transfer knowledge across tunnel drives and projects, demonstrating the potential to reuse historical tunneling data to support future construction and design-stage planning.

Overall, this research demonstrates how harmonized multi-project field data, empirical engineering analysis, and domain-specific artificial intelligence can advance the understanding of tunneling-induced ground behavior and support AI-enabled monitoring, engineering decision-making, and future mechanized tunneling projects.