Call my agents – multi-agent LLM-based calibration of urban drainage models: performance and reliability across model scales

Graphical abstract in three panels. Left panel, titled “Urban Drainage Model Calibration?”: a map of an urban drainage model with seven subcatchments (S1–S7) draining to a pond with orifice and weir outlets, plus a small pre-development site. Below it, two icons show the current approaches. A pickaxe stands for manual calibration, which is labor-intensive. A code window stands for automated calibration, which requires programming knowledge. Middle panel, titled “Solution: Multi-Agent LLM Framework”: a hierarchy diagram with a Supervisor Agent on top, connected to three subordinate agents: Model Agent, Calibration Agent and Analysis Agent. Right panel, top, “Calibration results”: a 12-hour outflow hydrograph comparing observed data with the best-fit simulation (NSE = 0.97) and a parameter uncertainty band (NSE > 0.50). A callout notes that performance is comparable to automatic calibration (NSE ≈ 0.97–0.98) but depends strongly on model size. Right panel, bottom, “Comparison of different LLMs”: large LLMs are highly accurate but computationally costly, while small LLMs are fast but unstable. Urban drainage model (UDM) calibration remains a technically demanding and time-consuming multi-step process. While automatic calibration approaches have improved efficiency, they typically require advanced programming expertise and workflow customization. This study presents and evaluates a locally deployable multi-agent large language model (LLM) framework for agent-assisted calibration of urban drainage models. The proposed system integrates a supervisor and three specialized agents to autonomously execute model preparation, calibration, and evaluation tasks through iterative workflows. Successful detention pond calibrations produced consistently high Nash–Sutcliffe efficiency and low event scores, while workflow completion and LLM-selected calibration settings varied substantially between models and repetitions. In addition, six calibration runs of the Bellinge dataset improved the respective objective relative to the published model. Among the analysed LLMs, a calibration performance comparable to established automatic calibration routines could be reached. However, workflow reliability differed among LLMs under the tested setup. Higher computational costs were observed for the smaller detention pond benchmark compared with an automated reference routine, while this overhead became less relevant for the larger Bellinge model due to longer simulation runtimes. The results highlight different outcomes in workflow reliability and computational efficiency, and demonstrate the potential of conversational agent-based workflows for supporting UDM calibration routines.