EvacuAIDi: An AI-Driven, Causal-Informed Framework for Probabilistic and Disability-Inclusive Evacuation Guidance

A comprehensive AI-driven framework to revolutionize building evacuations through standards-compliant automation, disability-inclusive modeling, causal evaluation, and probabilistic risk assessment.

PhD Dissertation 2025

Amir Rafe • Utah State University

Conducted under the supervision of Dr. Patrick Singleton

Research Overview

This dissertation addresses critical gaps in evacuation science through four interconnected research contributions that form the comprehensive EvacuAIDi framework.

The Critical Research Gaps

🔧

Standards-Compliant Evacuation Simulation Inputs Generation

Most existing simulation workflows rely on manual extraction of parameters from fire safety codes (e.g., NFPA 101, SFPE), which is error-prone and lacks causal traceability. This creates inconsistencies across studies and weakens the credibility of simulation results.

Gap: There is no widely adopted, automated pipeline to extract, reason over, and validate evacuation input parameters in a way that guarantees both standards compliance and causal coherence.

👥

Disability-Inclusive Behavioral Fidelity

While many evacuation models include agents with reduced mobility or altered walking speeds, few go beyond kinematic simplifications. Key behavioral aspects, such as leadership dynamics, marshal guidance, and probabilistic compliance, are poorly captured for individuals with disabilities.

Gap: There is a critical need for empirically calibrated models that realistically simulate both the physical constraints and behavioral responses of heterogeneous groups, including disabled individuals.

📊

Causal and Probabilistic Evaluation of AI Guidance

Studies that consider AI guidance during evacuation frequently rely on deterministic assumptions and average-case analyses, failing to account for uncertainty in key behavioral factors such as occupant compliance, as well as contextual variables like occupant load.

Gap: A robust, scenario-aware framework for evaluating the causal and probabilistic impact of AI guidance on evacuation safety under uncertainty remains underdeveloped.

Central Research Questions

RQ1: Automated Simulation Parameters Extraction

How can AI be leveraged to automate the extraction of simulation parameters from complex regulatory documents to ensure that models are both standardscompliant and causally coherent?

RQ2: Disability-Inclusive Behavioral Modeling

How can agent-based models be enhanced to realistically simulate the evacuation dynamics of a heterogeneous crowd, including individuals with disabilities, by incorporating empirically-derived social and behavioral forces?

RQ3: Causal Impact of AI Guidance Compliance

What is the quantifiable causal effect of occupant compliance with a personalized, accessibility-aware AI guidance system on evacuation performance across a range of hazard scenarios?

RQ4: Probabilistic Risk Assessment Framework

How can a probabilistic framework be used to formally integrate uncertainty in human compliance and occupant load to produce a robust assessment of evacuation risk?

Research Architecture

Introduction & Problem Formulation

Established the critical need for AI-driven, disability-inclusive evacuation systems by identifying major gaps in current evacuation modeling approaches and defining the four primary research questions.

Problem Definition Research Questions Framework Overview
Dissertation Structure

Research Framework Structure

Literature Review: Who Leads, Who Follows?

Systematic review of 233 peer-reviewed studies on evacuation guidance and leader-follower dynamics, covering human-led, environmental, technological, and hybrid guidance systems to establish the theoretical foundation.

Systematic Review Leader-Follower Dynamics 233 Studies
Literature Review Summary

Literature Classification Framework

Causal AI for Evacuation Modeling

Developed an AI pipeline that automatically parses regulatory documents (NFPA 101, SFPE Handbook, BS-9999, etc.) using fine-tuned LLM, GraphRAG, and causal reasoning to generate simulation-ready parameters with 92%+ accuracy.

Fine-tuned Gemma 3 Neo4j GraphRAG Causal Inference
92%+
Parameter Extraction Accuracy
Training dynamics of the fine-tuned Gemma 3 (4B) model

Training dynamics of the fine-tuned Gemma 3 (4B) model

Performance of four LLMs across evacuation-related question categories

Performance of four LLMs across evacuation-related question categories

Disability-Inclusive Social Force Model (DiSFM-GS)

Enhanced agent-based modeling by incorporating empirically-calibrated parameters for individuals with disabilities, using RFID trajectory data from controlled evacuation drills in a 4-story university building.

Social Force Model Transfer Entropy RFID Data
Walking Speed Analysis

Walking Speed Analysis

Leader-following rates

Leader-following rates

Objective function convergence

Objective function convergence

Causal-Intervention Scenario Design (CISD)

Used counterfactual simulations and Pearl's do-calculus to isolate the causal impact of AI guidance compliance on evacuation performance, demonstrating a 22.5% reduction in evacuation time across diverse scenarios.

Counterfactual Analysis A* Pathfinding Pearl's do-calculus
22.5%
Evacuation Time Reduction
Posterior Evolution

Posterior Evolution

Compliance Impact Analysis

Compliance Impact Analysis

Hierarchical Bayesian Risk Framework

Developed probabilistic risk assessment integrating uncertainty in compliance and occupant load using Gaussian Process surrogates and Global Sensitivity Analysis, identifying occupant load as the primary risk driver (57% of variance).

Bayesian MCMC Gaussian Process Sobol Sensitivity
57%
Risk Variance from Occupant Load
MCMC Trace Plots

MCMC Trace Plots

Parameter Sensitivity Ranking

Parameter Sensitivity Ranking

Conclusions & Future Directions

Synthesized research contributions, quantified stakeholder implications, and outlined future research directions including integration with CFD models, multimodal guidance systems, and dynamic trust modeling.

Research Synthesis Stakeholder Implications Future Research

Key Research Findings

Empirical results demonstrating the effectiveness and impact of the EvacuAIDi framework across multiple dimensions.

🎯 Causal Impact of AI Guidance

22.5%

Reduction in mean evacuation time

Increasing AI compliance from 10% to 70% causes a statistically significant 22.5% reduction in evacuation time, with diminishing returns beyond 50% compliance.

♿ Equity Enhancement

15.9%

Stronger benefit for vulnerable populations

AI guidance provides 15.9% greater benefits for individuals with disabilities, actively helping to close the safety gap and promote equity.

📊 Primary Risk Drivers

57%
Occupant Load
29%
AI Compliance

Global Sensitivity Analysis identifies these two factors as primary drivers of evacuation risk variance.

🤖 LLM Performance

92%+

Parameter extraction accuracy

The RAG pipeline achieves over 92% accuracy in extracting simulation parameters from regulatory documents.

Research Implications & Future Directions

Transforming evacuation science through actionable insights for stakeholders and comprehensive recommendations for future research.

Implications for Key Stakeholders

👷‍♂️

For Fire Safety Engineers

  • Reproducible, causally-defensible simulation inputs
  • Inclusive performance-based design tools
  • Strengthened reliability of safety cases
🏢

For Building Managers

  • Clear justification for AI guidance investment
  • Integration with real-time sensor networks
  • Enhanced vertical evacuation infrastructure
💻

For Software Developers

  • Blueprint for next-generation simulation tools
  • Validated inclusive agent profiles
  • Robust probabilistic assessment features

Technology Integration Recommendations

🤖 Occupant Evacuation Elevators (OEEs)

Integrate AI guidance with NFPA 101-compliant OEEs for seamless vertical evacuation of individuals with disabilities.

  • • Real-time capacity and status monitoring
  • • Dynamic routing based on occupant mobility profiles
  • • Automated dispatch coordination

📡 Real-Time Sensing Integration

Deploy multi-modal sensor networks to provide AI systems with live situational awareness.

  • • LiDAR for precise occupant counting and density
  • • Thermal cameras for smoke-obscured visibility
  • • Wi-Fi/Bluetooth for crowd movement patterns

🏠 Smart Areas of Refuge

Transform static Areas of Refuge into intelligent, AI-managed safe zones with dynamic capacity management.

  • • Real-time occupancy monitoring
  • • Automated first responder notification
  • • Priority-based occupant routing

🎯 Assisted Evacuation Coordination

Enable AI systems to coordinate trained personnel for optimal assisted evacuation strategies.

  • • Automated triage of assistance requests
  • • Optimized staff dispatch routing
  • • Real-time progress tracking

Future Research Directions

🔥 Integration with High-Fidelity Hazard Models

Couple DiSFM-GS with validated CFD models like Fire Dynamics Simulator (FDS) for dynamic fire-occupant interaction studies.

📊 Expanded Empirical Data Collection

Conduct large-scale evacuation drills with diverse populations across multiple building types to improve model generalizability.

🎧 Multimodal AI Guidance Systems

Explore auditory, visual, and haptic guidance modalities using VR environments to optimize communication effectiveness.

🤝 Dynamic Trust and Social Influence Modeling

Develop agent-based trust models incorporating AI performance history and social network effects on compliance.

Ask the AI Researcher ✨

Have questions about the EvacuAIDi research? Chat with an AI assistant trained on the complete dissertation content to explore findings, methodology, and implications.

I'm an AI research assistant specialized in the EvacuAIDi dissertation. I can answer detailed questions about the methodology, findings, and implications of this AI-driven evacuation framework.

Please wait until you see "✅ Ready for Q&A!" before asking questions - I'm currently indexing the dissertation content for better responses.

Quick questions to get started:

×