Model generalization paradigms for predicting viral particles and evaluating removal efficiencies in anaerobic membrane bioreactor plants

The dynamic changes in influent and effluent streams and the shifts in effluent quality across filtration layers of membrane bioreactors (MBRs) are major challenges that hinder the generalization of machine learning (ML) models developed to predict bacterial and viral contaminants in unseen data. This paper proposes two model generalization paradigms based on lifelong and zero-shot generalization frameworks for predicting viral particles and assessing log removal values (LRVs) across two anaerobic MBRs (AnMBRs) based WWTPs located in different cities of Saudi Arabia using physicochemical parameters, virometry, and PCR-based data. The lifelong learning approach integrates a knowledge-based adaptation module with a shared dictionary and a local ML predictor for streaming and predicting viral particles with delayed output measurements. The zero-shot generalization approach is based on a dual-attention transformer model and adaptively prioritizes key input water features through temporal and input attention mechanisms for estimating viral pathogens. Both approaches ensured generalization and robustness guarantees across unseen AnMBR-based wastewater matrices (WMs) and WWTPs. We validated them by predicting adenovirus, coliphage, CrAssphage, pepper mild mottle virus, and total virus concentrations and estimating contaminant removal performances through the LRVs across various WMs and WWTPs.