LeakAgent coordinates four software agents to plan pressure-sensor placement and detect leaks in water-distribution networks. Across five benchmark networks, the study reports mean accuracy of about 96% on each of two simulated tasks: detecting whether a leak exists and identifying the affected district-metered area (DMA), a zone monitored as one unit. The result was measured at a leak magnitude equivalent to 20% of total demand, the largest tested level.
The researchers generated 24-hour pressure time series in simulation. LeakAgent first uses a hydraulic model to estimate pressure responses to leaks at candidate locations. Those estimates inform network zoning and sensor placement. A dual-branch graph model then combines the simulated pressure series with a representation of network topology to predict leak presence and the affected zone. A multimodal large language model coordinates the workflow from an operator's request; it does not make the leak classification itself.
The tests covered four synthetic networks and one confidential model of a real utility network. All leak cases were simulated, and the paper reports no trial on a utility's live sensor system. Zone classification narrows a leak to a district-metered area, not the exact pipe. The researchers archived the code and four synthetic benchmark networks on Zenodo.
Sources
- LeakAgent: a multi-agent system for leak detection using multi-modal large language model coordination in water distribution network, Nature Communications AI & Computing, published October 6, 2026.
- LeakAgent code and benchmark-network archive, Zenodo.
Photo: Sergey Mikhel, via Wikimedia Commons, CC BY-SA 3.0. A pipe-repair trench in Omsk, Russia; contextual image, not part of the study.
