Predicting nitrification status in aerobic membrane bioreactors by interpretable machine learning models: Proof of concept for advancing greywater reuse through process monitoring with cross-scenario model validation

Membrane bioreactor (MBR) systems for on-site greywater treatment and reuse have been widely investigated, where robust nitrification monitoring is critical to prevent operational disruptions and maintain effluent standards. This research proposed an interpretable machine learning-based framework to predict nitrification efficacy (classified as "sufficient" or "insufficient") in MBR systems, involving model-based and post hoc interpretability. Three interpretable ML algorithms—logistic regression (LR), random forest (RF), and eXtreme Gradient Boosting (XGB)—were systematically evaluated, accounting for model bias-variance trade-offs during binary classification. Six input features were selected based on direct measurability, compatibility with commercial sensors, and operational relevance. Model optimization criteria focused on maximizing precision to prevent erroneous controls for insufficient nitrification, achieved by low false positive rate (FPR: misclassified insufficient nitrification) and high true positive rate (TPR: correctly identified sufficient nitrification). Results showed that both LR and XGB achieved identical precision scores >0.85 under standard conditions. Cross-scenario validation demonstrated RF's superior generalizability (precision=0.87) when predicting nitrification in biocarrier-amended MBRs using training data from non-biocarrier systems, reflecting the potential transferability of the model. The post hoc interpretability and stability analysis further identified prediction sensitivity to data limitations and distributional biases, providing critical insights for improving and even properly deploying models in practice. • Interpretable machine learning models can predict the nitrification status in MBRs. • Effluent water quality is feasible for nitrification process monitoring in MBRs. • Sensor-compatible inputs show potential for nitrification monitoring in practice. • Model enables cross-scenario nitrification prediction with and without biocarriers. • Post hoc interpretation shows aeration-centric with multi-N-load training strategies