Physically Plausible Scenario Generation via \ Geo-Enhanced LLMs for Disaster Training

Scenario-based training is central to disaster preparedness, yet creating realistic scenarios is labor-intensive and limited by the weak spatial reasoning of large language models (LLMs). This study introduces a Geo-enhanced prompting framework that integrates elevation, slope, hazard maps, and breach distance into GPT-4.1 prompts to generate physically consistent, time-stepped disaster scenarios. The method was applied to flood response in Hirakata City, Japan, using publicly available geospatial datasets. A total of 740 scenarios were generated across ten locations and 37 timesteps per site, corresponding to realistic training intervals. Evaluation by a municipal crisis-management professional showed that all Geo-enhanced scenarios were physically plausible, while a geography-agnostic baseline produced numerous implausible outputs, including inconsistent inundation sequences. By embedding geospatial constraints directly into prompts, the approach improved temporal coherence and spatial ordering and removed the need for corrections due to physical contradictions. Because the framework relies only on open data and lightweight preprocessing, it is adaptable to other hazards and regions without the cost of model retraining. These findings demonstrate that geospatially grounded LLM prompting can mitigate structural weaknesses in spatial reasoning, offering a scalable pathway to more realistic and operationally useful training materials for disaster preparedness.