Machine learning in wastewater: opportunities and challenges — “not everything is a nail!”
Peter A. Vanrolleghem, Mostafa Khalil, M. G. Serrão, Jeff Sparks, Jean-David Therrien
Abstract
This paper highlights the potential of machine learning (ML) for wastewater applications, with a focus on key applications and considerations. It underscores the need for simplicity in ML models to ensure their interpretability and trustworthiness, cautioning against the use of overly complex ‘black box’ models unless absolutely necessary, especially with limited data. Not all modelling problems should be considered nails for which the ML hammer is the best-available tool. We emphasise the critical role of thorough data collection, including metadata, given its scarcity in some areas. Future research is encouraged to develop benchmark hybrid models to bridge the educational gap for environmental engineers and to establish best practices for managing data and model metadata, thereby improving ML’s accessibility and utility in wastewater applications. • Machine learning has arrived in a wide diversity of wastewater applications. • Simplicity must be pursued to ensure ML-interpretability and trustworthiness. • Successful ML needs high quality data, meta-data and bridging the educational gap. • Hybrid (ML-mechanistic) models tap into both data scientist and engineer expertise.