Dynamic fire conditions change the network itself.
Road accessibility, travel time, and safe routing can shift rapidly as fire and weather evolve.
An interdisciplinary research initiative at the University of Southern California.
When Wildfire Moves, Intelligence Must Move Faster.
AI-Enabled Evacuation for Resilient Communities
The EVAC-AI initiative brings together researchers in AI, transportation, wildfire science, control, and human-centered decision making.
Problem
Wildfire, mobility, and human behavior are deeply coupled.
Approach
Agentic AI with humans firmly in the loop.
Focus
Vulnerable populations, mixed autonomy, practical deployment.
Invitation
We want researchers, agencies, data partners, and communities at the table.
Why Now
Evacuation is a systems problem, not a single-routing problem.
Wildfire evacuation unfolds at the intersection of fire spread, road accessibility, institutional coordination, information quality, and unequal household resources. EVAC-AI is built around that reality, with special attention to vulnerable populations and mixed-autonomy transportation.
Core Thesis
Households do not evacuate with equal capacity.
Zero-car families, older adults, people with disabilities, and multi-lingual communities need different support.
Human coordinators need decision support they can trust.
We are building AI that helps people compare options, not automation that bypasses operational judgment.
Project Overview
Four integrated thrusts, one connected system.
EVAC-AI connects fire-network coupling, transportation optimization, information design, and human-centered AI into a practical research program for wildfire evacuation.
Research Design
Four thrusts, one mission.
The EVAC-AI research program is deliberately end-to-end: understand evolving hazard, optimize movement under constraints, communicate effectively, and support human coordinators with useful AI interfaces.
Fire-Network Coupling
Translate wildfire and weather dynamics into time-varying road accessibility and network state.
- Fire spread potential and accessibility estimation
- Retrospective event reconstruction and scenario generation
- Inputs for downstream routing and coordination tools
Network Optimization
Model evacuation when private vehicles, public fleets, and autonomous mobility all compete for limited capacity.
- Mixed-autonomy operations
- Routing and control under behavioral uncertainty
- Coordination across wildfire and transportation systems
Information Design
Study how AI-assisted alerts and disclosures can be timely, accessible, and usable across diverse communities.
- Communication for vulnerable populations
- Messaging under uncertainty and changing conditions
- Designing for comprehension, trust, and action
Agentic AI For Coordinators
Build closed-loop human-AI systems that help responders interpret state, weigh tradeoffs, and adapt in the loop.
- Human-centered coordination interfaces
- Scenario comparison and action support
- Practical translation to real operational workflows
The EVAC-AI Team
Four fields. One evacuation system.
EVAC-AI brings fire science, transportation, information design, and AI systems into one USC-based team built for operational translation.
Principal Investigator
Ruolin Li
Sonny Astani Department of Civil and Environmental Engineering; Thomas Lord Department of Computer Science
Thrust 2 · Network OptimizationFocus: Game-theoretic transportation optimization and guaranteed rescue access in mixed-autonomy systems.
Co-Principal Investigator
Costas Synolakis
Sonny Astani Department of Civil and Environmental Engineering
Thrust 1 · Fire-Network CouplingFocus: Dynamic fire behavior and road-accessibility inputs for evacuation decision support.
Co-Principal Investigator
Ketan Savla
Sonny Astani Department of Civil and Environmental Engineering; Daniel J. Epstein Department of Industrial and Systems Engineering; Ming Hsieh Department of Electrical and Computer Engineering; Aerospace and Mechanical Engineering
Thrust 3 · Information DesignFocus: Targeted information strategies for coordinated, capacity-aware departures.
Co-Principal Investigator
Yue Zhao
Thomas Lord Department of Computer Science
Thrust 4 · Agentic AIFocus: Safety-critical, closed-loop AI for coordinators and affected communities.
Los Angeles Living Lab
Real places. Real constraints. Real translation pressure.
The project uses Los Angeles-centered case settings to connect historical wildfire events, transportation systems, community needs, and emergency operations. These are research focus areas that help us test ideas where complexity is unavoidable.
Topanga
Canyon evacuation dynamics, field-oriented validation, and practical resilience questions around constrained road access.
Eaton / Altadena
Incident reconstruction and community-centered resilience research informed by recent wildfire experience.
Palisades
Road access, alerting, and multimodal response challenges that expose how hazard and mobility evolve together.
Regional Scale
Southern California mobility, planning, and inter-organizational coordination at the scale where policy and infrastructure decisions matter.
Collaboration Lanes
There are multiple ways to build with us.
We want this initiative to become a magnet for people who care about operational relevance. If you bring data, context, methods, testing environments, or on-the-ground knowledge, there is likely a meaningful role for you.
Shape the operational reality of the work.
- Coordinator workflows and exercise design
- Feedback on trust, usability, and failure modes
- Translation pathways into practice
Help us connect theory to real networks.
- Regional data and demand assumptions
- Network operations and policy context
- Scenario evaluation at multiple scales
Keep the work centered on people who need it most.
- Accessibility and inclusion needs
- Messaging design and local trust dynamics
- Co-creation with vulnerable populations
Expand the scientific and technical frontier.
- AI agents, HCI, optimization, wildfire modeling
- Mapping, mobility, simulation, and evaluation tools
- Open datasets, prototypes, and benchmarks
Translation Strategy
Designed to move beyond papers.
EVAC-AI is not just about producing models. It is about building datasets, interfaces, exercises, and relationships that make the research legible and useful outside the lab.
01
Co-create with practitioners and communities
Ground research questions in operational pain points and lived experience from the start.
02
Build reusable data and science assets
Create retrospective studies, benchmarks, datasets, and evaluation scaffolds others can build on.
03
Test through scenarios and exercises
Use structured review, tabletop settings, and operational feedback to challenge prototypes before they travel.
04
Turn lessons into capacity
Translate results into tools, resources, and training materials that practitioners can actually use.
Work With EVAC-AI
We are assembling the network around the mission now.
If you want to help shape how wildfire evacuation intelligence is researched, tested, and translated, this is a good moment to join. We are especially interested in collaborators who care about both rigor and real-world usefulness.
Contact
Ruolin Li, Principal Investigator
University of Southern California
ruolinl@usc.edu