Hazard-Responsive Digital Twin for Climate-Driven Urban Resilience and Equity

Technical Summary: Hazard-Responsive Digital Twin for Climate-Driven Urban Resilience and Equity

Problem Statement
The paper addresses the growing complexity of climate-related disasters, which are increasingly characterized by compounding hazards (e.g., wildfires causing power outages that exacerbate heatwaves) rather than isolated events. Traditional risk assessments and existing Digital Twin (DT) technologies often fail in these scenarios due to their reliance on stable connectivity, complete datasets, and deterministic assumptions. Current DTs are typically fragmented, focusing on single assets or isolated hazards, and lack the capacity to integrate physical modeling with social equity considerations under conditions of data disruption and cascading infrastructure failures. There is a critical need for a framework that can operate adaptively under uncertainty, fuse heterogeneous data streams (including when sensors fail), and explicitly account for social vulnerability in resilience planning.

Methodology
The authors propose the Hazard-Responsive Digital Twin (H-RDT), a unifying framework structured as a three-layer cyber-physical ecosystem:

1. Physical Layer: Represents real-world entities (buildings, populations, infrastructure) instrumented with distributed sensors (IoT, UAVs, satellites) and supplemented by crowd-sourced data when formal networks degrade.
2. Digital-Intelligence Layer: The computational core utilizes a hybrid approach:
   • Physics-Informed Neural Networks (PINNs): These models embed governing physical laws (e.g., 2R2C thermal models, advection-diffusion equations for smoke, damage evolution laws) into machine learning to ensure predictions remain consistent with thermodynamic and structural principles, even with sparse data.
   • Reinforcement Learning (RL) Fusion: An adaptive module dynamically reweights data sources (IoT, UAV, Satellite) based on reliability and freshness, allowing the system to maintain situational awareness during sensor outages.
   • Graph Reinforcement Learning (GRL): This module models evolving interdependencies among community assets, identifying critical nodes and latent support pathways (e.g., schools acting as shelters) under stress.
3. Decision Layer: Visualizes probabilistic forecasts and integrates Equity-Adjusted Risk metrics. This formulation combines physical hazard exposure with socio-economic sensitivity and physical vulnerability to generate risk indices that highlight disparities in high-vulnerability clusters (e.g., low-income housing, clinics).

The framework also includes a Relief and Intervention Modeling module that simulates strategies (e.g., microgrids, cooling centers) to quantify their impact on population-weighted risk, tail risk (95th percentile), and overheating hours.

Case Study and Results
The H-RDT was evaluated in a synthetic urban district comprising 120 buildings with diverse archetypes (schools, clinics, housing) under a simulated 72-hour compounding wildfire–outage–heatwave scenario. Key findings include:

• PINN Performance: The PINN model accurately predicted indoor temperature trajectories (peaking at ~31–33°C during outages) with an overall RMSE of 1.93°C. It demonstrated robustness, maintaining stability and physical realism even when IoT sensors experienced dropout during power outages.
• Adaptive Data Fusion: The RL-based fusion module successfully shifted reliance from failing IoT sensors to UAV and satellite data streams. Fusion weights for IoT decreased from ~0.32 to ~0.22 during outages, while UAV and satellite weights increased, preserving spatiotemporal coverage.
• Network Resilience: The GRL module identified high-betweenness corridors and coverage-aware critical facilities, effectively reinforcing the network backbone for resource flow during stress.
• Equity and Interventions: Equity-adjusted risk analysis revealed that high-risk clusters (lower deciles of the population) bore a disproportionate burden. Five intervention strategies were tested:
  • Preemptive Cooling Centers (I1) and Microgrid expansion for Vulnerable Blocks (I4) proved most effective, reducing population-weighted risk by 11–13%, tail risk by 7–17%, and overheating hours by 7–9%.
  • Reactive Opening (I2) offered moderate benefits (~4% risk reduction), while targeted outreach and clinic-first microgrids provided smaller, focused improvements (1–3%).
  • Pareto-front analysis indicated that distributed microgrids offered the highest cost-efficiency, while preemptive cooling centers provided operational flexibility.

Key Contributions and Significance
The paper claims to advance the field of urban resilience by:

1. Unifying Framework: Establishing a transferable foundation that links physical hazard modeling, network-based interdependency discovery, and equity-aware reasoning within a single adaptive framework.
2. Robustness under Uncertainty: Demonstrating that hybrid physics-AI models can sustain predictive accuracy and physical consistency despite partial sensor loss and data incompleteness, a common feature of compound disaster scenarios.
3. Equity-Centered Decision Support: Moving beyond static asset monitoring to provide quantitative, equity-adjusted risk metrics that explicitly prioritize vulnerable populations and critical facilities in resilience planning.
4. Actionable Insights: Providing a methodology to evaluate and prioritize interventions (e.g., anticipatory activation of cooling centers vs. infrastructure hardening) based on their ability to reduce both average and tail-end risks.

The authors position the H-RDT not merely as a diagnostic tool but as a "cognitive collaborator" that advances digital urban resilience toward adaptive, learning-based, and equity-centered decision support for climate adaptation. The study acknowledges that while the current implementation uses a synthetic district and simplified thermal models, the framework establishes a scalable foundation for real-city implementation, pending future validation with municipal datasets and field pilots.