Efficient machine learning-based modeling for regional reliability analysis of infrastructure systems

Liu, Tong

Permalink

https://hdl.handle.net/2142/129448 Copy

Description

Title

Efficient machine learning-based modeling for regional reliability analysis of infrastructure systems

Author(s)

Liu, Tong

Issue Date

2025-04-27

Director of Research (if dissertation) or Advisor (if thesis)

Meidani, Hadi

Doctoral Committee Chair(s)

Meidani, Hadi

Committee Member(s)

• Cha, Eun Jeong
• Kontou, Eleftheria
• Talebpour, Alireza

Department of Study

Civil & Environmental Eng

Discipline

Civil Engineering

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Infrastructure Systems
• Graph Neural networks
• Reliability Analysis
• Transportation System
• Natural Hazards

Abstract

As infrastructure systems grow increasingly interconnected and vulnerable to disruptions from natural hazards and human-induced events, there is a critical need for scalable, adaptive, and physics-consistent modeling techniques. This dissertation investigates the reliability and resilience of structure and infrastructure systems. This work leverages deep learning methodologies, particularly physics-informed neural networks (PINNs) and graph neural networks (GNNs), to enhance infrastructure system analysis, addressing challenges related to system identification, reliability assessment, traffic assignment, dynamic flow forecasting, and infrastructure asset management. By integrating physics-informed constraints, heterogeneous graph representations, and data-driven optimization strategies, this dissertation provides novel solutions that improve computational efficiency, predictive accuracy, and decision-making in complex infrastructure networks. The primary contributions of this dissertation are threefold. First, physics-driven constraints are incorporated into data-driven models to address the limitations of purely data-driven approaches, embedding governing physical laws into the learning process to improve interpretability and robustness. Second, GNN-based frameworks are developed for multi-level regional infrastructure reliability analysis under probabilistic hazard scenarios, demonstrating high scalability and generalization across both in-distribution and out-of-distribution conditions. Third, generalized graph-based models are introduced and validated on diverse transportation applications to evaluate system behavior under disruption and inform equitable infrastructure planning. The research contributions span multiple domains, beginning with PIDynNet, a physics-informed neural network designed for nonlinear structural system identification. The dissertation then presents a rapid seismic reliability assessment framework for highway bridge networks, demonstrating the capability of GNN-based models to predict connectivity loss under probabilistic seismic scenarios. To improve post-disaster network analysis, a GNN-based shortest distance estimation framework is introduced, enabling scalable and accurate evaluation of roadway network performance under disruption. Furthermore, a curriculum-enhanced graph reinforcement learning model is designed for fast route recommendation in stochastic time-dependent networks, incorporating a graph-based actor-critic architecture and curriculum learning to enhance scalability and generalization. Additionally, a heterogeneous GNN model is introduced for static, dynamic traffic assignment, and multi-class traffic assignment, incorporating virtual links to improve demand propagation and flow estimation. Finally, a GA-GNN optimization framework is introduced to enhance transportation equity in seismic retrofit planning, balancing network resilience with social equity considerations. The dissertation concludes by summarizing key findings and outlining future research directions. Expanding GNN applications to multi-hazard risk assessment can improve infrastructure resilience across various disaster scenarios. Enhancing the generalization capability of GNN models will further improve adaptability to diverse infrastructure topologies and hazard conditions. Additionally, integrating real-time sensor data into GNN-driven frameworks will enable dynamic infrastructure monitoring, proactive risk assessment, and real-time decision-making. By addressing these challenges, future research can extend the impact of deep learning methodologies in infrastructure resilience, fostering the development of smarter, more adaptive, and disaster-resilient infrastructure systems.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129448

Copyright and License Information

Copyright 2025 Tong Liu

Owning Collections