Making waves: Model transfer as a key pathway to improving the generalization of machine learning in wastewater treatment engineering

Although machine learning (ML) achieves high predictive accuracy in wastewater treatment plant (WWTP) under stable conditions, its limited transferability in engineering practice often leads to repeated modeling and wasted resources. This Perspective reframes generalization as the central objective of ML in wastewater treatment engineering, rather than pointwise predictive accuracy. The recurring failure of ML model generalization in wastewater treatment systems should not be viewed as a temporary algorithmic limitation, but rather as a structural consequence of intrinsic system heterogeneity. We systematically organize existing model transfer approaches into a unified spectrum, encompassing parameter transfer, data transfer, feature transfer, and meta-learning, and clarify how prior knowledge is reused at different levels. In parallel, we synthesize the dominant forms of system shift encountered in WWTPs, including changes in influent characteristics, operating conditions, process configuration, plant scale, and sensing infrastructure. By explicitly linking characteristic shift scenarios with suitable transfer strategies, we propose a shift-aware engineering pathway that embeds model transfer across the entire ML lifecycle. Finally, we emphasize that model transfer should be guided with engineering judgement to clearly define its applicability boundaries, avoid negative transfer and overgeneralization, and be embedded as a core engineering principle that accounts for system heterogeneity, safety considerations, and institutional constraints, thereby enabling lifelong learning in wastewater treatment systems.