UC Berkeley

Earthquake risk assessment of railway infrastructure remains a challenging engineering problem, particularly at the regional scale, where models must balance physical fidelity with computational scalability. Current regional simulation approaches are efficient but rely on simplified fragility-based representations that cannot fully capture asset-specific response, component-level damage, or train derailment mechanisms. More detailed train--structure interaction and nonlinear finite element models provide a richer description of seismic behavior, but their modeling effort, calibration needs, and computational cost limit their use for large infrastructure inventories and multiple earthquake scenarios. At the same time, the benefits of higher-fidelity simulations for regional applications remain unclear in the literature, limiting their development. This dissertation addresses this gap by developing a physics-informed framework for regional seismic risk assessment of elevated railway systems. The framework combines nonlinear train--structure interaction analysis, surrogate-based structural modeling, regional response-history simulation, and risk-informed earthquake scenario selection to enable more detailed yet computationally tractable assessment of the infrastructure's seismic performance.The research is motivated by the need for improved regional-level damage predictions to support estimates of earthquake-induced losses, enhance emergency response planning, and inform retrofit decisions. The study focuses on elevated railway infrastructure, where post-earthquake performance may be governed by both structural damage and train derailment, and where several components of the performance evaluation process may benefit from higher-fidelity simulation. Using the Bay Area Rapid Transit (BART) system as a case study, this dissertation develops tools and methods for physically meaningful, yet computationally tractable, simulation of large railway inventories subjected to spatially variable earthquake excitation.The first part of the dissertation develops an analytical two-dimensional train--structure interaction model that combines a nonlinear structural model, a multibody model of the train, and a wheel--rail contact formulation. This model is used to investigate derailment mechanisms under earthquake excitation and evaluate the influence of common modeling assumptions on the system's behavior. The results show that derailment risk is primarily governed by the motion transmitted to the train through the structure rather than by the input ground motion alone, and that train derailments are indeed influenced by nonlinear behavior in the structure. These developments provide the basis for creating structure-specific derailment fragility relationships that can be readily used in regional-scale analyses.The second part addresses the challenge of generating nonlinear models for large inventories of elevated reinforced concrete structures. A reduced-order model for cantilever-like bridge bents is combined with a novel Gaussian Process regression workflow to directly predict plastic hinge parameters from a small set of physically interpretable, non-dimensional descriptors. This surrogate modeling strategy enables rapid generation of nonlinear reduced-order models while preserving key behavioral features needed for nonlinear dynamic analyses. Training and validation of the model are performed using experimental data, and comparison with detailed finite-element models highlights the proposed approach's capacity to accurately reproduce the peak structural response for scalable nonlinear earthquake simulation.The third part of the dissertation integrates these developments into a regional simulation framework for elevated railway infrastructure. An inventory-based methodology is developed to transform as-built structural information into analysis-ready models, and a workflow is proposed for performing regional nonlinear response-history analyses under spatially correlated ground motions for earthquake scenarios. A risk-informed scenario-selection methodology is also presented, supporting large-scale applications of response-history analyses. The goal is to identify a reduced set of rupture scenarios that capture the dominant contributors to the system risk. Together, these components enable regional physics-based simulation while maintaining computational tractability. Comparisons with existing lower-fidelity models highlight the value of the proposed framework: by explicitly simulating structural response, the methodology provides more informative response and damage distributions, allows evaluating which assets are driving risk, identifies differences in the impact from different scenarios, and reveals where simplified fragility-based approaches may bias regional loss estimates or where they may be sufficient, depending on the loss metric of interest.Taken as a whole, this dissertation presents an end-to-end framework for regional seismic risk simulation of elevated railway systems. Through these developments, it advances regional railway earthquake assessment beyond conventional fragility-only approaches, enabling the explicit representation of both structural damage and derailment-related consequences within a physically informed simulation environment.