Multi-university AI research may revolutionize wildfire evacuation

As wildfires grow wilder, the University of Florida and two other universities are developing large language models to make evacuations safer and more efficient. 

Armed with a nearly $1.2 million National Science Foundation grant, UF, Johns Hopkins University and the University of Utah are creating these AI-based models to simulate human behavior during evacuations – information that will help emergency managers shape more effective evacuation plans. 

“Strengthening wildfire resilience requires accurate modeling and a deep understanding of collective human behavior during evacuations,” said UF project lead Xilei Zhao, Ph.D., an associate professor with the Engineering School of Sustainable Infrastructure and Environment. “There is a critical need for simulation models that can realistically capture how civilians, incident commanders and public safety officials make protective decisions during wildfires.”

Existing simulation models face limitations, particularly with reliable predictions under various wildfire scenarios. New AI models can simulate how diverse groups of people behave and interact during the hurried scramble to seek safety.

Zhao’s team is developing a convergent AI framework for wildfire evacuation simulations powered by psychological theory-informed large language models. The project will produce simulation methods to promote teaching, training and learning, and support wildfire resilience by allowing public safety officials to use open-access tools. 

“This research seeks to be a transformative step toward improving the behavioral realism, prediction accuracy and decision-support capability of wildfire evacuation simulation models,” Zhao said.  

Zhao partnered with John Hopkins professor Susu Xu, Ph.D., and University of Utah professors Thomas Cova, Ph.D., and Frank Drews, Ph.D. The preliminary results of the study were recently presented at the 63rd Annual Meeting of the Association for Computational Linguistics. 

“In that paper, we started to train the model on the survey data we collected to see how we can accurately predict people's evacuation decisions with LLMs,” Zhao said.   

Research objectives include extending the Protective Action Decision Model for civilians and public safety officials, developing psychological theory-informed large language model agents for protective modeling and generating a realistic synthetic population as input for the simulation platform. The team also plans to develop learning-based simulations and predict human behavior under scenarios such as fire spread, warning and infrastructure damage. 

This research comes at a critical time, as the number of wildfires has significantly increased globally. About 43% of the 200 most damaging fires occurred in the last decade leading up to 2023, according to a recent study in Science. The intensity, size and volume of wildfires are threatening more urban areas.  

“If you go into the urban area, many people do not have cars, or they need additional mobility support,” Zhao said. “For example, the LA fires impacted nursing homes with a lot of elderly people, many of whom are immobile or lack the ability to drive. That's a big problem. This would be very relevant to them.” 

The large language models will provide important context for evacuation planning as well as real-time decision making. 

“We envision this tool being used during planning,” Zhao said, “so emergency managers can test different kinds of scenarios to determine how to draw the evacuation zones, where to issue the orders first and how to design the communications messaging.”