Disaster Copilot: Agentic AI for Collective Human–Machine Intelligence in Disaster Management

Escalating disaster risks increasingly overwhelm traditional response systems due to fragmented data, siloed AI tools, and the loss of institutional knowledge. This study introduces Disaster Copilot, a multi-agent AI framework that orchestrates specialized disaster analytics into a unified, human-centered decision-support system, enabling adaptive, real-time disaster management across all phases of response and recovery.

Project Highlights:

• Propose a multi-agent, agentic AI architecture that coordinates domain-specific sub-agents (early warning, risk analysis, situational awareness, impact assessment, and resource allocation) under a central task planner.
• Integrate multimodal and multilingual data streams (satellite imagery, sensors, UAVs, social media, text, audio) to generate a coherent, real-time operational picture.
• Embed retrieval-augmented generation (RAG) and persistent memory to preserve institutional knowledge and support evidence-grounded decision-making.
• Advance Disaster Digital Twins from passive visualization platforms to active, reasoning, and adaptive intelligence systems

Check out the paper here!

Characterize Decadal Spatiotemporal Trend of Power Outages

This study provides a nationwide, longitudinal assessment on U.S. power system vulnerability.

Project Highlights:

• Construct a three-dimensional framework to characterize power outages from its frequency, duration, and intensity and develop multiple metrics.
• Investigate a high-temporal resolution, large-scale dataset of ~179 million outage records across 3022 U.S. counties during 2014–2023.
• Results display unbalanced spatial distribution and escalated trend of power outages.
• Also, we reveal disproportionate and enlarging gap of power outage impacts on socially-vulnerable populations.

Check out the paper here!

Relevant info:

• We have a related project establishing the national power system vulnerability index (PSVI) at county level. Learn more!

Incorporate Environmental Considerations into Infrastructure Inequality Evaluation

Empirical evidence show that infrastructure may trigger negative impacts on the environment (e.g. air pollution). Yet those environmental concerns have not been considered in infrastructure provision. This study applies interpretable machine learning to construct an environmental-integrated infrastructure provision index.

Project Highlights:

• Apply XGBOOST + SHAP to capture associations between multiple infrastructures and environmental hazards (i.e. urban heat and air pollution).
• Implement the model in five metropolitan areas in the U.S.
• Uncover spatial and income inequality in environmentally integrated infrastructure provision.
• Inform integrated urban design strategies to promote infrastructure equity and environmental justice.

Check out the paper here!

Unravel Fundamental Properties of Power System Resilience Curves

Resilience curves depict the fluctuation of system performance before, during and after a disruption. Resilience triangle has been the primary paradigm of resilience curve for over two decades. However, the model only provides a theoretical one-size-fits-all framework, with little empirical studies to delineate infrastructure curve archetypes and their fundamental properties based on observational data. This study applies time-series clustering method on outage data during extreme weather events to address the gap.

Project Highlights:

• Construct over 200 resilience curves based on power outage data during three extreme weather events.
• Apply time-series clustering method to identify two archetypes of power system resilience curve (i.e.trapzoidal and triangular).
• Trapezoidal curves are determined by (1) duration of sustained performance loss (2) constant recovery rate.
• Triangular curves are determined by (1) recovery pivot point (2) critical functionality threshold (3)critical functionality recovery rate.

Check out the paper here!