Modeling frameworks for resilience in socio-technical systems

Abstract

In the face of growing disaster frequency and severity, we aim to strengthen resilience across socio-technical systems. We develop several modeling frameworks to understand how communities prepare for and recover from disruptions that span interdependent infrastructures and communication channels. Although communities differ in structure and vulnerability, disasters uniformly strain both the physical networks and the social systems that support them. This dissertation examines resilience across two fronts: the operational resilience of communities within interdependent infrastructure systems and the communicative resilience of information flow within and between them. In the first part, we use a Markov Decision Process (MDP) framework to model failure-and-repair dynamics in binary hetero-functional graphs (HFGs) that couple power, water, and community infrastructures. We show that communities experiencing power outages respond more sensitively to prioritization strategies based on vulnerability or criticality, whereas those affected by water shortages are primarily influenced by service demand. To overcome the limits of binary HFGs, we introduce the Hetero-functional Agent-Based Infrastructure Toolkit (HABiT). HABiT enables fine-grained simulation of heterogeneous infrastructures and their interdependencies. We apply HABiT to a synthetic model of three communities with varying infrastructure access and social vulnerability. The simulation reproduces normal operations, disruptions, and recovery influenced by scarce resources, vehicle routing, and mobile repair crews. By incorporating stochasticity, we uncover variations in cascading failures and recovery patterns that deterministic models can miss. HABiT thus enables rapid evaluation of disruption scenarios and guides resource allocation and recovery planning under uncertainty. Recognizing that technical criteria alone may diverge from community priorities, we integrate community preferences into decision-making. We employ Large Language Models (LLMs) as proxy survey tools, generating simulated personas with diverse disaster experiences to obtain infrastructure-repair preferences. We aggregate these responses through a learning-to-rank algorithm, producing a total repair order that balances technical feasibility with social priorities. In the second part, we shift from infrastructure to information, examining how communication systems support or hinder resilience. Using survey data from three Midwest counties, we develop stochastic diffusion models that trace how information spreads through communities during both normal and disaster conditions. We integrate neighbors, online social networks, local news, cable news, and local government, mapping trust-based and interaction-based ties in each community. We find that trust in local government, frequent interaction with cable news, and social media strongly shape diffusion, with diffusion speed being independent of community size. We then extend this analysis to examine manipulation of information in online social networks. Using Twitter data from the 2016 United States elections, we analyze interactions between Russian troll tweets and public replies. We cluster historical user-interaction sequences and apply statistical tests to reveal diverse, unpredictable engagement behaviors. We find that trolls lacked a consistent strategy for provoking responses, although content with political figures generally attracted more attention.