Development of an LLM-agent-based physics–AI hybrid Model for urban flood hazard prediction

 Recent increases in urban flooding necessities the development of real-time autonomous prediction models to support forecasting and decision-making. Urban flooding is caused by the combined effects of multiple factors, such as the expansion of impervious surfaces and limitations in drainage pipe capacity. Due to this complexity, predicting urban flooding only with rainfall–runoff is insufficient. Physics–data-based hybrid approaches have been proposed to improve prediction reliability by compensating individual’s limitations. However, hybrid models that require repetitive physics-based simulations face limitations in real-time applications due to high computational costs and long execution times. To overcome these limitations, this study applies an agent-based approach that integrates and coordinates the urban flood hazard estimation process within hybrid modeling frameworks. The proposed framework consists of three agents interconnected through a graph-based orchestration structure to form an iterative analytical workflow of execution, validation, and improvement. The first agent performs physics-based hydrological modeling to reproduce rainfall–runoff processes and the temporal response of urban drainage systems. It also automates model calibration, validation, and optimal model selection. The second agent spatially predicts flood susceptibility using machine learning models based on topography, land use, soil characteristics, drainage infrastructure, and historical flood data. The models applied in this process include random forest, extreme gradient boost, artificial neural networks, long short-term memory, and tabular data–oriented foundation models (TabPFN). The final agent integrates the results of the first two agents to conduct a hazard assessment that simultaneously reflects the probability of urban flooding and its potential intensity. The integrated flood hazard modeling framework enables automated, near-real-time prediction of urban flood hazards. It can serve as a foundational dataset for advancing future urban inundation forecasting and warning systems and decision-support frameworks.