Postdoctoral Scholar Stanford University
Simulating Regional Recovery After the 2025 Palisades Fire: Implications for Recovery Management
Co-Author: Jack Baker (Stanford University)
Abstract: Regional recovery modeling is an emerging field at the intersection of engineering and social science, driven by efforts to enhance disaster resilience of the built environment. We demonstrate how regional recovery modeling can support recovery management following the 2025 Palisades Fire, with the City of Malibu as a case study. We utilize recovery data from the 2018 Woolsey Fire in Malibu, interviews with local experts, and recovery progress during the first year following the event to inform model development. Building recovery is simulated as a sequence of recovery steps including damage assessment, clean up, permitting, design, financing and construction. The duration of these recovery steps is conditioned on regional technical, administrative, and financial capacities. Using the resulting regional recovery model, we generate recovery timeline forecasts under alternative recovery management scenarios, identify prospective recovery bottlenecks to enable proactive decision-making, and estimate property tax losses associated with housing damage. Model outputs are compared against observed recovery progress to date, and forward-looking forecasts are provided to enable future validation as recovery unfolds.
PhD Student Texas A&M University
A Decision Framework for Equitable Use of Federal Funds for Voluntary Buyout Programs
Co-Authors: Zihao "Scott" Li (Texas A&M University), Chenchen Kuai (Texas A&M University), and Stephanie Paal (Texas A&M University)
Abstract: Established in 2018, the Structural Extreme Events Reconnaissance (StEER) Network coordinates the mobilization of research teams to gather perishable data following natural disasters. However, a significant bottleneck exists: while field deployment decisions may require action within hours, supporting Preliminary Virtual Reconnaissance Reports (PVRRs) average over 30 days to publish. A misconception is that this latency is driven by report writing speed. While recent Large Language Models (e.g., Zhou & Mosalam, 2025) successfully automate summaries, they function largely as news aggregators, lacking the deep technical tracing required for sound engineering decisions.
This study identifies the primary bottleneck as Information Retrieval Latency—the load placed on VAST teams sifting through high-volume, unstructured noise to locate technical evidence. To target this, we propose the Dynamic Assessment & Reconnaissance Targeting (D.A.R.T.) framework. Unlike approaches attempting to replace the human expert, D.A.R.T. functions as human-centered guidance to optimize the search process.
Utilizing a Retrieval-Augmented Generation (RAG) architecture anchored in over 60 expert-vetted StEER PVRRs, the workflow proceeds in three phases:
- Contextual Strategy Injection: Vector embeddings of historical analogs generate a dynamic “Technical Checklist,” transforming the team from passive searchers into active verifiers.
- Human-in-the-Loop Verification: Teams utilize targeted directives to validate evidence prioritized for deployment implications.
- The Circuit Breaker: An automated severity monitor triggers an immediate deployment recommendation if verified damage exceeds historical thresholds.
This shift from automated writer to automated decision support reduces the time-to-decision from weeks to a matter of hours, ensuring scarce engineering resources are deployed with precision.
PhD Student Stanford University
Simulating Firm-level Direct and Contingent Business Interruption after Disasters
Co-Authors: Nikola Blagojevic (Stanford University) and Jack Baker (Stanford University)
Abstract: Past disaster events reveal that businesses experience interruptions through various mechanisms. In addition to direct losses caused by damage to firm-owned facilities, contingent business interruption arises from disruptions to supply chains, lifeline infrastructure, workforce availability, and the customer base. The lack of quantification for the resulting losses remains a critical barrier to risk-informed insurance design and effective policies supporting business recovery and survival. We develop regional recovery models using the Pyrecodes framework to quantify the firm-level business interruption and recovery after disasters, explicitly accounting for interdependencies among built and social systems. The models simulate the dynamic recovery of housing, business buildings, and infrastructure at the community scale. The simulations are then used to evaluate evolving impacts on individual business operations through facility functionality, employee availability, customer access, and supply constraints. We demonstrate the model through a case study of business interruption in Alameda Island, California, under a Hayward earthquake scenario. The simulation quantifies firms’ lost revenue attributable to both direct and contingent business interruption. It also identifies the dominant drivers of business interruption across different recovery phases, which can inform targeted mitigations and business continuity plans to reduce impacts. While demonstrated for an earthquake scenario, the framework is extensible to other hazard contexts.
PhD Student Oregon State University
Building-scale observation and modeling of recovery from hurricane-induced damage and loss
Co-Authors: Mehrshad Amini (University of Rhode Island) and Dan Cox (Oregon State University)
Abstract: There is a need to develop longitudinal building scale studies for coastal systems to evaluate damage losses and recovery of infrastructure. The longitudinal dataset developed in this study provides a unique opportunity to examine recovery processes at a building scale and across multiple dimensions of community functionality. The novelty of this work lies in the development of this dataset through street-level video collection for all accessible streets in Fort Myers Beach, Florida capturing recovery following Hurricane Ian (2022). Data were collected quarterly, leading to an unprecedented twelve instances of full street-level imagery, covering of all phases of recovery of Fort Myers Beach from March 6, 2023 to December 6, 2025. For the physical systems (infrastructure), the trajectory is first characterized by the initial level of damage then the restoration of structural components and utility systems. For the social systems, the trajectory is quantified by re-occupancy and re-establishing social networks through access to essential services. For the economic systems, the trajectory focuses on the re-establishment of businesses, real estate values, taxes, and other publicly accessible metrics. Altogether, at the NHERI Computational Symposium we will present the methodology and applications of the recovery dataset to: (1) quantify building scale building recovery, leveraging virtual damage assessment, ongoing field observations of recovery, construction and occupancy permits, and sales and rental data; (2) evaluate the relationship of recovery trajectories between residential and commercial buildings; and (3) compare and validate existing recovery and restoration models.
Undergraduate Student University of Nebraska Lincoln
Measuring Functional Recovery After Tornadoes: Linking Housing, Occupancy, and Socioeconomic Indicators
Co-Authors: Pramodit Adhikari (University of Nebraska Lincoln) and Milad Roohi (University of Nebraska Lincoln)
Abstract: Community recovery following tornadoes is shaped by the extent of physical damage and underlying socioeconomic conditions that influence how a community rebuilds. This study develops and applies a multi-resolution framework to analyze how socioeconomic conditions affect functional recovery after tornadic events. Two complementary recovery metrics are introduced. Method A provides high-accuracy parcel-level recovery measurements by linking severely damaged or destroyed residential structures using NOAA Damage Assessment Toolkit (DAT) data to municipal building permit records. Recovery is measured as the cumulative proportion of damaged residences receiving reconstruction-related permits over time, giving recovery trajectories and benchmark indicators such as time to 80% recovery. Due to data availability constraints, Method A is limited to a few well-documented case study cities. Method B extends recovery assessment to a national scale by using lower resolution but widely available proxies, including housing unit counts, occupancy rates, and other infrastructure and market data from the U.S. Census and related datasets. Method A results are used to validate and contextualize Method B, allowing scalable comparison across a larger sample of tornado-impacted communities. Together, these methods support analysis of the relationship between socioeconomic vulnerability and recovery outcomes. The framework establishes a foundation for future development of machine-learning-based predictive models that integrate physical damage and socioeconomic indicators to estimate recovery trajectories and inform disaster planning and resource allocation.
PhD Student Univeristy at Buffalo, SUNY
Resilience-Oriented Deep Reinforcement Learning for Post-Hurricane Recovery of a Large-Scale Transportation Infrastructure System
Co-Author: Teng Wu (Univeristy at Buffalo - SUNY)
Abstract: The post-hurricane repair sequence of the damaged transportation infrastructure critically influences the efficiency of system functionality restoration. Deep reinforcement learning (DRL) learns a repair policy through iterative state–action–reward interactions, enabling sequential, multi-purpose decision-making and emerging as a promising approach for transportation network recovery. While resilience is widely used for recovery planning, its quantification through functionality metrics from traffic simulations is computationally intensive, limiting its use as a reward in DRL for large-scale transportation networks. To address this issue, a learning-based surrogate model for rapid assessment of traffic-functionality metrics is needed; however, most surrogates approximate the mapping of origin-destination demand to traffic flows with a fixed network configuration. Existing configuration-aware surrogates dilute the link features and inter-link propagation by aggregating simulation at the node level, yielding limited accuracy under repair-driven, continuously changing network configurations. This study proposes a resilience-oriented DRL for post-hurricane recovery of large-scale transportation networks, with a link-based graph convolutional neural network (Link-GCN) as the reward surrogate. Specifically, the Link-GCN learns a mapping from link states to traffic flow by performing graph convolution over links and aggregates their features over link–link adjacency to capture nonlinear propagation under changing configurations. Ground-truth flows are precomputed via user-equilibrium assignment under diverse configurations to train the surrogate. With this reward surrogate, a deep neural network approximates the action-value function and is trained via deep Q-learning to learn an optimal recovery policy. The proposed method is evaluated on a thousand-node New York City transportation network, demonstrating good computational efficiency and accuracy.
Assistant Professor Texas Tech University
Data Driven Method for Structural Roof Damage Evaluation
Co-Author: Chuan Yu, Yimin Lu
Abstract: Tornadoes pose a critical threat to residential wood-frame structures, with roof sheathing removal serving as a primary failure mode that compromises structural integrity. This study proposes a novel framework bridging post-damage data and computational mechanics to evaluate structural performance. The methodology follows a three-stage pipeline. First, an Automated Roof Damage Ratio Detection Method is deployed. Utilizing high-resolution aerial imagery and computer vision techniques, this phase automatically quantifies the percentage of sheathing loss for individual buildings, creating a granular dataset of observed damage states. Second, the study integrates numerical tornadic field simulation to reconstruct the wind environment associated with the observed damage. By spatially correlating the automated damage ratios with the simulated flow field, the research establishes a quantitative relationship between specific levels of roof damage and local wind vectors, accounting for both magnitude and directionality. Finally, a Probabilistic Assessment of Roof Sheathing Vulnerability is conducted. Leveraging the derived wind-damage correlations, empirical fragility curves are developed to characterize the probability of sheathing failure under varying tornadic loads. This data-driven approach validates theoretical failure models against real-world reconnaissance data, offering a robust tool for performance-based design and community resilience planning against non-synoptic wind events.