NSF-BSF AI-Augmented Decision Support for Wildfire Evacuation and Infrastructure Planning

This U.S.-Israel Binational Science Foundation Collaboration (NSF-BSF) project develops new ways to evaluate and use artificial intelligence (AI) to improve decision-making in wildfire emergencies. Wildfires are increasing in frequency and severity, threatening lives and infrastructure, especially in the wildland–urban interface (WUI), where communities meet flammable landscapes. While current approaches focus on physical protection, infrastructure performance in emergencies depends critically on how people respond to risk, information, and evacuation orders. This project combines behavioral experiments with large language model (LLM)–based AI agents to simulate decision-making under stress. Rather than assuming AI outputs are reliable, the project develops a structured framework to test whether AI-generated decisions reflect real-world behavior. Engineers and researchers from Northwestern University and Ben-Gurion University will compare responses across decentralized (US) and centralized (Israel) emergency systems to understand how communication strategies and institutional structures influence evacuation and response. The project outcomes will provide validated AI tools to support safer evacuation planning and infrastructure decision-making under extreme conditions. This collaborative research project will develop and test a Hierarchical Decision Fidelity Framework (HDFF) to evaluate whether AI-generated decisions reproduce observed behavioral patterns across multiple levels of increasing rigor. Specifically, the project will: (1) design and implement cross-national experimental scenarios that systematically vary risk information, communication strategies, and constraints to generate controlled behavioral data; (2) evaluate AI-generated decisions against empirical data across aggregate outcomes, distributions, subgroup behavior, rare high-impact events, and underlying decision processes using statistical tests including equivalence testing, distributional divergence measures, and hierarchical modeling; (3) recover sparse and underrepresented behaviors, including delayed evacuation and noncompliance, using Bayesian Gaussian Process methods to construct empirically grounded benchmarks for tail events; and (4) develop an adaptive calibration framework that detects and corrects model misalignment by aligning AI outputs with observed behavioral relationships and response patterns. By integrating experimental design, statistical validation, and AI-based simulation, this project establishes a principled approach for using AI to model human decisions in infrastructure systems, with applications to evacuation planning, risk communication, and emergency response. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.