Out of bounds: The use of model ensembles to explore model structural uncertainty under extreme events

• Model ensemble disagreement for diagnosing model structure uncertainty. • Ill-posed boundary conditions limit the model explanatory power. • Climate change contributes to amplifying model structure uncertainty. The response of an urban drainage model to extreme events cannot be investigated independently from the abstraction process adopted in model development. The work explores structural model uncertainty, and particularly the impact of boundary conditions, using an Italian catchment that has been the subject of several modelling studies over the years. While previous research has focused on parameter optimization, limited attention has been devoted to model structure and boundary conditions. However, during extreme events, not only model parameters but also boundary conditions significantly influence system behavior. Model structural uncertainty is investigated through the development of four model configurations organized into two model structure classes: the first class reflects the original system structure, while the second incorporates downstream boundary condition based on a re-analysis of the system. Within each class, two models were developed and calibrated by including or excluding extreme events. Each configuration is represented by an ensemble of models, which were then subjected to synthetic extreme storms. Disagreement among their responses is used to diagnose structural uncertainty. Moreover, the study analyzed the risks associated with using a model calibrated on extreme events without properly accounting for system structure, underlining the potential consequences of such an approach. Given the expected increase in frequency and intensity of extreme events under future climate conditions, the work emphasizes the central importance of contextual knowledge and boundary conditions in model development. A thorough understanding of system structure and context is a necessary prerequisite for effective calibration of model parameters. Parameter calibration cannot be separated from knowledge of model structure: rather than being a purely automatic process, it must be guided by the modeler’s conscious choices.