Zero-shot generalization for predicting viral concentrations and evaluating removal efficiencies across wastewater matrices

Predicting viral particles on new unseen data across wastewater matrices (WMs) in aerobic membrane bioreactor (AeMBR)-based wastewater treatment plants (WWTPs) remains an open challenge due to the process drifts involved in the treatment stages. Efficient data augmentation approaches based on Markov chain (MCM), Markov chain and multivariate Gaussian (MMCM), Gaussian mixture (GMM) and Copula (CM) were proposed to generate synthetic data from physicochemical parameters, virometry, and PCR-based method. Dual-attention long short-term memory network (DA-LSTM) with new generative models was proposed to predict viral particles and evaluate the removal efficiencies across AeMBRs, thereby handling effluent processing drifts. The DA-LSTM combines attention mechanisms to adaptively adjust the weights of the features and increase the long-term memory, enabling accuracy and robustness across unseen WMs. DA-LSTM framework was tested for predicting pepper mild mottle virus and enteric viral pathogens such as total virus and adenovirus in two regions of Saudi Arabia. The log removal values were evaluated through the estimated viral concentrations. The DA-LSTM model demonstrated significant adaptability to unseen data across different WMs, maintaining robust performance despite the effluent drifts. The results showed that DA-LSTM zero-shot generalization achieved remarkable viral particles prediction performance using MMCM with a mean average coefficient of determination R2 of 0.91, and 0.97 across the sand, and MBR wastewater matrices in region R1, respectively, and R2 of 0.97 across the chlorinated effluent treatment process in region R2. Tests on total viral prediction across municipal WWTPs located in two other regions in Saudi Arabia confirmed the DA-LSTM's effectiveness in predicting viral particle across WMs and its ability to enhance zero-shot generalization performance at the regional level.