Machine Learning to Predict Pathogen Removal in Potable Reuse MBRs Based on Microbial Surrogates

Applicability Potable reuse treatment requires advanced processes to significantly reduce pathogen concentrations. A novel approach includes membrane bioreactors (MBRs), reverse osmosis (RO), and ultra-violet advanced oxidation processes (UV/AOP). This is an improvement to the conventional approach, often referred to as the 'gold' standard for reuse consisting of activated sludge (CAS), ultrafiltration (UF), RO, UV/AOP. MBRs offer benefits like smaller footprint, reduced sludge, and more reliable effluent quality. Yet, MBR-based reuse raises concerns about pathogen passage due to potential membrane damage, necessitating vigilant pathogen reduction monitoring. Models for pathogen removal in MBR systems have several practical applications, such as an early alert system for damaged membranes or to adjust downstream disinfection. Utilizing online sensors or microbial surrogates is quicker and more cost-effective than direct monitoring of pathogens like Cryptosporidium and Giardia. Our goal was to develop a sufficiently accurate model using online sensors or lower cost surrogates to reduce the frequency of Cryptosporidium and Giardia sampling while ensuring public protection. While the focus of this study was Giardia in the context of potable reuse, the methods discussed in this presentation would also be applicable to E. coli or total coliforms for NPDES compliance. Demonstrated Results and Outcomes Methodology We developed machine learning (ML) models to predict post-MBR treatment pathogen levels, focusing on Cryptosporidium and Giardia. Data was collected from a demo-scale MBR treating secondary effluent, referred to as a 'tertiary MBR' due to its polishing biological treatment of secondary effluent. Some MBR fibers were intentionally cut during the pilot study to allow microbe passage. This helped assess the effectiveness of surrogates and techniques for monitoring and predicting Cryptosporidium and Giardia levels across various MBR effluent turbidity and operational setpoints. In the preliminary dataset, 76 Cryptosporidium and Giardia samples were analyzed, along with paired surrogates. We used regression ML to predict Cryptosporidium or Giardia concentration and classification ML to predict if they exceeded 4 or 4.5 log reduction values (LRVs) alert thresholds, respectively. Six microbial surrogates (aerobic endospheres, C. perfringens, total coliforms, E. coli, male-specific coliphage, and somatic coliphage), pressure decay rates, and turbidity served as input variables (Figure 1). The data was split 80:20 with 61 datapoints for model development and 15 for model evaluation as shown in Figure 2. Results In terms of whether LRVs would be below or over the alert thresholds, conventional multiple linear models achieved 80% testing set classification accuracy for Cryptosporidium and 87% for Giardia. Classification ML algorithms like ada and svmLinearWeights2 reached 87% accuracy for Cryptosporidium, while pcaNNet and loclda models achieved 93% accuracy for Giardia, all better than the linear models (Table 1). Regression supervised ML models, xgbTree and ridge quantitatively predicted the Giardia LRVs with testing set R2 of 94.4% and 91.9%. However, linear models performed similarly for Giardia in this context. Conclusions Depending on the data analysis task (classification or quantitative prediction) and target pathogen (Giardia or Cryptosporidium), ML performance was marginally above or comparable to linear models. This may have been due to collinearity among the features, or the limited sample size to support ML. Nonetheless, our findings so far are a proof-of-concept for the efficacy of ML in developing pathogen monitoring and alert systems for MBRs. As part of 2024 schedule for this project, we will incorporate additional microbial sample size from the partner utility. We also plan to scale this project by making the ML models available as an interactive webtool that facilitates real-time pathogen prediction. This study was part of a US Bureau of Reclamation funded project including an Independent Expert Panel facilitated by the National Water Research Institute. This interdisciplinary collaboration brought together insights from environmental engineers, microbiologists, and data scientists, enriching the project's depth and applicability. Our methodology could be extended to MBR-based treatment plants discharging to surface water with microbial or disinfection parameters in their NPDES compliance regulation, offering a valuable approach for addressing wastewater concerns in diverse contexts. Relevance to Audience Attendees of this presentation will leave with a comprehensive understanding of an innovative approach to potable water reuse treatment, the integration of ML models, cost-effective monitoring techniques using surrogates, and the potential ML applications in various advanced treatment settings. The insights shared can disseminate innovation in efficient wastewater treatment monitoring for public safety and regulatory compliance.