Adaptive Flood-Scenario Generation via LLM-Guided Closed-Loop Simulation

Urban flooding is one of the most damaging consequences of urbanisation and climate change, yet assessing green-infrastructure interventions requires evaluating large numbers of computationally expensive coupled hydrological–hydraulic simulations. We propose an Large Language Model (LLM)-guided closed-loop framework that treats flood-scenario generation as a sequential decision-making problem. The LLM proposes physically constrained scenarios, evaluates them using a coupled Storm Water Management Model (SWMM) and Cellular Automata Fast Flood Evaluation (CA-ffé) flood model, and uses spatial flood-depth feedback to guide subsequent decisions. A reward combining damage-weighted impact, novelty, and exploration focuses the simulation budget on scenarios that are both informative and relevant to flood-damage reduction, while Annual Exceedance Probability (AEP)-based mitigation targets define stopping criteria. The framework is evaluated against Random Sampling and Latin Hypercube Sampling using scenario-space coverage, diversity, and efficiency in identifying high-impact interventions. This framework enables faster flood-risk and adaptation assessments, supporting timely, evidence-based infrastructure planning and the development of climate-resilient cities.