Machine-learning informs mitigation strategies for nitrous oxide emissions from wastewater operations

This study focused on the development of machine-learning- (ML-) based strategies for mitigating nitrous oxide (N2O) emissions from various wastewater treatment systems in the United States measured using a benchmark USEPA-endorsed protocol. Results revealed that in general, poor process performance correlated with higher N2O emissions. Among the models evaluated, the Extreme Gradient Boosting (XGBoost) model demonstrated strong predictive performance for aerobic (R2=0.97), and anoxic (R2=0.87) zone process data. Specifically, the unsupervised principal component analysis (PCA), along with the trained ML models, demonstrated that local variables including zone-specific dissolved oxygen (DO), ammonia (NH3), and nitrite (NO2-) concentrations and global variables including effluent nitrite and nitrate as key contributors towards N2O emissions from both aerobic and anoxic zones of the process bioreactors. The operating regions associated with lower predicted N2O emissions included operations of aerobic and anoxic zones at DO < 4 mg O2 L-1 and < 1 mg O2L-1 respectively, coupled with appropriate solids retention times (SRTs) that maximize process performance. The scenario analysis using the trained ML model predicted 7.8-fold (95% CI: 2.2 to 52.8-fold) and 8.8-fold (95% CI: 3.83 to 21.24-fold) higher N2O flux at low DO/high NH3 and high DO/high NO2- conditions in aerobic and anoxic zones, respectively. This model predicted responses were consistent with the established nitrifier denitrification and N2O reductase inhibition pathways. Accordingly, our results underscore the utility of ML models interpreted in conjunction with bioprocess fundamentals for predicting and mitigating N2O emissions, while concomitantly improving wastewater treatment operations.