Socio-technical regional disaster recovery modelling using generative LLM-based agents and physics-based models

Nikola Blagojević, Jack Baker

Understanding and simulating the interaction between people and infrastructure is critical for planning and managing recovery after disasters. Human decision-making in the post-disaster context is traditionally simulated using agent-based models, where agents represent households using pre-defined hardcoded rule sets. Rule sets can be difficult to calibrate due to data scarcity and become intractable as the number of parameters that affect households increases. The aim of this paper is to replace rule sets with large language models (LLMs), transforming classical agents into generative agents. LLMs learn from past rule sets, scientific literature, news articles, and general human behavior to simulate how people behave after disasters, allowing for simulation of complex decision-making. Generative, LLM-based agents are integrated with a regional recovery simulator of technical systems, capturing how households make decisions as the state of technical infrastructure changes during recovery. A case study considering post-earthquake recovery of buildings and the water supply system on the Alameda Island, CA, illustrates how generative agents can be used to forecast household behavior in a post-disaster context.