The urban water networks that keep our cities running are increasingly turning to machine learning (ML) to predict leaks, manage pressure, and optimise treatment processes. But according to a new critical review published in *Water Research X*, many of these ML models are built on shaky ground—raising concerns about transparency, reproducibility, and real-world reliability.
Lead author Martin Oberascher, from the University of Innsbruck’s Unit of Environmental Engineering, and his team examined over 100 recent studies applying supervised ML to urban water infrastructure. Their findings reveal a troubling pattern: while many papers boast impressive performance metrics, few document the essential steps that ensure those models can be trusted beyond the lab.
“Performance numbers alone don’t tell the full story,” Oberascher notes. “Without details on feature scaling, hyperparameter tuning, or uncertainty quantification, we can’t truly assess whether a model will work in a different city—or even the same one a year later.”
The review highlights that fewer than one in four studies reported techniques for explainability—critical for operators who need to understand and justify automated decisions. And with no standardised way to document workflows, comparing models across research becomes a guessing game.
For the energy sector, which increasingly overlaps with water utilities through shared infrastructure and sustainability goals, this lack of rigour could translate into missed opportunities—or worse, costly missteps. Imagine a water utility deploying an ML model to reduce pumping energy, only to find it fails under real-world conditions because key methodological steps were overlooked.
To address these gaps, the authors propose a methodological checklist to guide future ML development in urban water networks. By standardising reporting, they argue, researchers and practitioners can build models that are not just powerful, but dependable.
The implications are clear: as ML becomes more embedded in critical infrastructure, transparency isn’t just academic—it’s a commercial necessity. And for a field racing to deliver smarter, greener systems, that’s a challenge worth taking seriously.