LeakGFM: Towards Graph Foundation Models for Water Leakage Detection

Leakage in water distribution networks wastes treated water and the energy used for its treatment and pumping, resulting in avoidable carbon emissions. Existing machine learning approaches to leakage detection typically train separate models for individual networks, leaving transfer to unseen distribution topologies largely unaddressed, and demanding labelled leak data and resources that utilities may lack. To address these challenges, we present LeakGFM, a pre-trained graph model for cross-network transfer and a step toward graph foundation models for leakage detection, which learns transferable representations from network topology and hydraulic behaviour across multiple simulated water networks. LeakGFM detects leakage by comparing an evaluated window with a leak-free reference, and adapts to an unseen network with as few as a single labelled hydraulic scenario. When pre-trained on 26 source networks and evaluated on four held-out networks, LeakGFM consistently outperforms both the same architecture without pre-training and a learning-free hydraulic baseline, with the largest gains when labelled adaptation data are scarcest.