Foundational Models and Intelligent Agents for Urban Drainage Systems: The Road Ahead

Two emerging paradigms are redefining AI for physical systems: foundational models and intelligent agents. In the context of urban drainage systems, both offer the opportunity to move beyond task-specific tools toward more general, scalable, and operationally meaningful AI systems. This contribution discusses the state of the art, presents recent advances from our research, and outlines a realistic pathway toward AI-based urban drainage intelligence.

The first paradigm focuses on foundational, physics-aware models that aim to replace computationally expensive numerical simulators while fully exploiting real-world measurements. Such models have the potential to significantly strengthen digital twins by enabling fast, differentiable, and transferable representations of system dynamics. We present recent developments in graph neural network (GNN)–based surrogate models for urban drainage simulation, including autoregressive architectures capable of emulating both the dynamics of 1D numerical models and 2D shallow-water equation solvers. These results demonstrate that accurate and scalable learning-based simulators are now feasible for complex hydraulic processes.

Looking ahead, a key research challenge lies in integrating drainage network models with surface flow representations, enabling unified 1D/2D modelling of the coupled behaviour of sewers, floodplains, and urban catchments during extreme events. Another critical opportunity lies in exploiting the differentiable nature of AI models, which opens the door to assimilating real-world observations directly into model parameters. This offers a principled alternative to traditional calibration workflows, while also enabling continuous adaptation as new data become available. At the same time, the scarcity of high-quality real-world flood observations implies that pretraining on large-scale simulations will likely remain essential to developing robust and transferable models.

The second paradigm concerns AI agents for complex engineering tasks, where systems are designed not only to make predictions but also to reason, plan, and take actions within operational workflows. In the context of urban drainage, such agents could ultimately support activities ranging from model building and calibration to decision support, monitoring, and infrastructure management. As a concrete example of this broader direction, we present our ongoing work on automatic sewer defect detection, where we evaluate the limitations of current general-purpose vision–language models for infrastructure inspection. Our results indicate that meaningful progress will require domain-specific multimodal models, tailored to sewer imagery and engineering semantics. These models can naturally evolve toward vision–language–action systems, enabling compact, efficient agents suitable for deployment on robotic platforms and edge devices, with appropriate safety constraints.