Assistant Professor University of Delaware
A Joint Synthetic Housing-Household Inventory
Co-Authors: Xiao Qian (University of Delaware) and Rachel Davidson (University of Delaware)
Abstract: Understanding how people interact with the built environment requires linking households to the specific housing units they occupy. Although high-resolution building and infrastructure datasets are widely available, our understanding of who lives within these structures remains limited. Existing methods face two major challenges: (i) the lack of synthetic populations that realistically capture sociodemographic diversity and spatial patterns, and (ii) the absence of approaches that learn and apply latent relationships between household characteristics and housing features when integrating population and building inventories. This gap preserves privacy but limits household-level assessment of social impacts, especially under policy interventions and environmental hazards.
We present an integrated framework for constructing a realistic joint housing–household inventory that addresses these challenges. First, we establish a systematic mapping between American Community Survey (ACS) population variables and National Structural Inventory (NSI) building attributes, enabling deep contrastive learning to capture transferable statistical relationships between households and housing units. These learned relationships quantify housing–household compatibility beyond rule-based or random assignment. Second, we introduce a scalable synthetic population generation strategy that avoids distributional mismatch between microdata and aggregated census statistics. We generate an oversampled pool of synthetic households and apply a hierarchical, multi-objective mixed-integer linear programming (MILP) optimization to select a subset that jointly satisfies building capacity, demographic consistency, and housing–household compatibility. The resulting joint inventory improves representation of household-level living arrangements and enables more accurate evaluation of social impacts from policies, climate-driven hazards, and physical disruptions such as hurricanes.
PhD Student Department of Civil and Environmental Engineering, University of California, Berkeley
Unveiling Occupant Response to Earthquakes: An Immersive Virtual Reality Study
Co-Authors: Yvonne Merino and Luis Ceferino
Abstract: Earthquake emergencies pose severe risks to occupant safety in indoor environments. Understanding human behavior during shaking is critical for enhancing seismic safety; however, obtaining empirical data remains challenging. Virtual Reality (VR) offers a high-fidelity, cost-effective tool to reconstruct indoor seismic scenarios, enabling the study of human decision-making under realistic post-earthquake conditions. This study designs an experimental framework that leverages virtual reality (VR) to capture residents' behavioral responses during earthquake events. Specifically, we developed a high-fidelity virtual environment replicating a typical two-bedroom, one-bathroom single-family residential apartment. To simulate realistic seismic impacts, we constructed indoor damage scenarios featuring non-structural component failures, the distribution of which was validated against earthquake records. Subsequently, we conducted immersive experiments using a between-subjects design to capture occupant actions throughout the shaking process. Participants were randomly assigned to experimental conditions, including the presence or absence of earthquake early warning, group versus individual settings, and different initial locations within the apartment. Behavioral data were captured through integrated eye-tracking and body-tracking capabilities in the VR headset. The results provide a complete picture of human behavior during an indoor earthquake emergency. The findings are expected to deepen the understanding of complex human behavioral dynamics during earthquakes, providing empirical insights to inform more effective emergency response strategies and, furthermore, facilitate the development of more accurate casualty estimation models.
PhD Student University of Kansas
Assessing Public Tornado Shelter Access Using an Agent-Based Model of Household Sheltering Behavior
Co-Author: Elaina Sutley (University of Kansas)
Abstract: Tornadoes pose a severe threat to U.S. communities. Approximately 990 tornado events occur annually, causing an average of 99 fatalities each year, disproportionately endangering low-income households and mobile home residents. Reducing loss of life and addressing inequities requires understanding resident sheltering behavior during warnings. This study presents a generalized Agent-Based Modeling (ABM) framework to simulate household sheltering behavior to assess sheltering access and inform mitigative actions.
Synthetic populations at the block-group level are generated from U.S. Census data and allocated to residential locations using dasymetric mapping. Stochastic tornado scenarios are sampled from historical distributions of path intensity, width, and warning lead time.
Over 2,000 responses from surveys administered to households in Kansas inform heterogeneous agent decision rules through multinomial and ordinal logistic regression models. Key behavioral factors—including trust in warnings, access to basements or safe rooms, travel time, income, and residence in a mobile home—shape household’s choice among four actions: doing nothing, sheltering in place, seeking public shelters, or leaving the area. Household agents’ route to shelters is simulated over real road networks using a shortest-path algorithm to assess timely arrival.
Repeated Monte Carlo simulations quantify spatio-temporal “protective gaps” and highlight vulnerable groups. What-if experiments explore alternative siting options for new shelters and educational programming impacts on sheltering behaviors. The results identify strategic shelter placements that reduce loss of life. By coupling empirical human behavior with the dynamic evolution of hazard and infrastructure constraints, this framework advances computational natural hazards engineering toward enhanced life safety and casualty reduction.
Assistant Professor Clemson University
Generative Agents to Model Household Adaptation in a Historically Flood-Prone City
Co-Author: Liming Lu (Clemson University)
Abstract: Coastal cities face mounting risks from more frequent and severe flood events as climate change drives sea‐level rise and intensifies storm activity. Traditional, regionally aggregated loss models provide valuable insights into broad exposure and vulnerability patterns, but they often overlook a critical dimension of flood risk: the capacity of individual households to dynamically adjust their exposure through adaptation actions. In this study, we introduce a novel, generative AI driven framework that couples a high‐resolution, spatially explicit flood model with a household‐level adaptation decision component, thereby capturing the feedback between evolving hazard conditions and autonomous adaptation behaviors. We apply this framework to the city of Charleston, South Carolina, a coastal municipality with a long history of flood exposure and a diverse mix of architectural typologies. Results indicate that, under a business‐as‐usual policy regime, rising flood damages could more than double by mid‐century. However, when enabling autonomous household adaptation, the expected losses decline by up to 30% relative to the no‐adaptation baseline. Spatial analysis further reveals targeted interventions such as zoning reforms or enhanced grant programs could yield outsized benefits. By integrating generative AI with process‐based flood modeling and explicit behavioral representations, our approach advances the state‐of‐the‐art in risk assessment and adaptation planning. It provides probabilistic, fine‐grained estimates of future flood risk that account for both hazard evolution and adaptive responses, allowing policymakers to identify where and when to deploy limited resources for maximum impact.
Assistant Professor Johns Hopkins University
Unraveling and Understanding Household-level Adaptation Decision Diffusion after Hurricanes across U.S.
Abstract: Coastal hazards like hurricanes flooding often cause widespread physical damage and trigger large-scale household repair and relocation decisions that reshape neighborhoods for years. Previous research suggests that post-disaster community recovery is inherently collective, i.e., a household's decision to repair or sell depends not only on its own exposure and resources, but also the responses of community neighbors and closed social ties. However, investigating what drives collective recovery -- and how strongly -- at scale remains challenging because the underlying influence networks are rarely observed and existing measures rely largely on small-scale survey data.
Here, we derive a new dataset of household-level house adaptation decision diffusion networks covering more than 3,000 census block groups across three hurricanes (Harvey, Ida, and Ian). To achieve this, we introduce a novel neural graph diffusion model that jointly reconstructs latent community influence networks and driving mechanisms from sparse sequences of property repair and sales data, achieving upto 98\% event prediction accuracy. Our analysis identifies two distinct decision diffusion regimes for house repair vs sale: repairs exhibit short-lived, coordination-like spillovers, whereas sales propagate more persistently and strengthen over time, consistent with self-reinforcing expectations about neighborhood viability. Meanwhile, cross-process effects are strongly asymmetric: neighbors' sales exert larger and longer-lasting impacts on individual household's repair decisions than collective repairs on sale decisions. We further investigate how neighborhood physical conditions and socio-demographic context shape the influence network structure, and we use counterfactual simulations to quantify downstream economic outcome. These findings provide an early-warning signal for communities at risk of cascading exit and suggest where timely support, by reducing early repair bottlenecks or stabilizing emerging sale clusters, may prevent tipping toward long-run decline.
Professor Florida Internationl University
Adaptation Investment Under Climate Uncertainty: A Dynamic Decision Model for Coastal Communities
Co-Author: Jamal Julien
Abstract: Adaptation investment for mitigating coastal flood damages may not only focus on cost effectiveness, but also the timing of irreversible infrastructure decisions under climate uncertainty. In this study we develop a dynamic decision model in which local governments choose whether to delay action, invest in protective infrastructure, or acquire additional information while facing uncertain sea-level rise and storm surge trajectories. Adaptation is modeled as a sequential decision problem with learning, finite planning horizons, and sunk investment costs, reflecting key constraints in real-world coastal planning. The framework is motivated by U.S. coastal communities making long-lived adaptation choices under federal programs such as FEMA’s Hazard Mitigation Assistance and related resilience investments, where the option value of waiting and the timing of subsidies play a central role. Using a belief-updating framework based on observed flood outcomes over time, the model characterizes how alternative investment designs interact with local decision-making. The analysis identifies conditions under which delaying investment is optimal and when early action becomes preferable as uncertainty resolves. Rather than assuming uniform effects, the framework highlights multiple policy channels—including cost effectiveness, accelerated learning, and enhanced flexibility—through which adaptation investment may influence outcomes and guide decision-making. By formalizing these mechanisms, the study generates testable predictions about how adaptation policies can reduce long-term coastal flood damage under deep uncertainty, offering guidance for future empirical evaluation and policy design.
Professor University of Missouri - Kansas City
Categorical Theoretic Decision-Making under Deep Uncertainties for Civil Infrastructure Systems
Abstract: Resilience-informed decision-making for civil structures and infrastructure across their lifecycles confronts profound climatic and socioeconomic uncertainties prior to the adoption of climate-scenario-dependent intervention measures. Recent advances in digital twin frameworks promise predictive and adaptive capabilities; however, excessive confidence and over-promising have been placed on these paradigms without adequately addressing how uncertainty, causality, and intervention propagate across heterogeneous spatial and temporal scales. Traditional uncertainty quantification approaches—whether probabilistic or imprecise-probabilistic—remain fundamentally limited in their ability to harmonize uncertainty across interacting entities, abstraction levels, and decision layers.
This work proposes a categorical-theoretic framework for infrastructure decision-making under deep uncertainty. By leveraging categorical structure, via functors, factorization, and diagrammatic reasoning, the framework provides a principled mechanism for mapping between system representations while explicitly tracking which causal claims survive abstraction. Using a scour-critical bridge system as a testbed, we reveal limitations of state-of-the-art workflows (e.g., those employing quoFEM and surrogate modeling) when climate-scenario-dependent interventions (e.g., scour countermeasures and structural retrofitting) are considered. We argue that categorical compositionality forms a pivotal new decision-making layer within modern digital twins, enabling robust and adaptive intervention planning while mediating between stakeholder-driven socio-economic objectives and the technical–physical dynamics of civil infrastructure systems.