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

Area of Science:

• Environmental microbiology
• Wastewater treatment technologies
• Virology

Background:

• Predicting viral particles in wastewater treatment plants (WWTPs) is difficult due to process variations.
• Existing methods struggle with unseen data across diverse wastewater matrices (WMs).

Purpose of the Study:

• To develop an accurate and robust model for predicting viral particles in aerobic membrane bioreactor (AeMBR)-based WWTPs.
• To evaluate the model's performance in handling effluent processing drifts and generalizing to unseen WMs.

Main Methods:

• Proposed data augmentation using Markov chain Monte Carlo (MCM), Markov chain and multivariate Gaussian (MMCM), Gaussian mixture (GMM), and Copula (CM) models.
• Developed a Dual-Attention Long Short-Term Memory Network (DA-LSTM) framework integrating generative models for viral particle prediction.
• Tested the DA-LSTM model on predicting pepper mild mottle virus, total viruses, and adenovirus in Saudi Arabian WWTPs.

Main Results:

• The DA-LSTM model demonstrated significant adaptability and robust performance across unseen WMs and effluent drifts.
• Achieved high zero-shot generalization performance with MMCM, yielding R² values of 0.91 and 0.97 for sand and MBR matrices, and 0.97 for chlorinated effluent.
• Confirmed effectiveness in predicting viral particles across municipal WWTPs in different regions, enhancing regional zero-shot generalization.

Conclusions:

• The DA-LSTM framework, combined with generative models like MMCM, offers a powerful solution for viral particle prediction in WWTPs.
• The model effectively handles process drifts and generalizes to new wastewater matrices, crucial for public health monitoring.
• This approach enhances the ability to predict viral loads and assess removal efficiencies in diverse wastewater treatment scenarios.