Machine Learning-Assisted Optimization of Mixed Carbon Source Compositions for High-Performance Denitrification

Appropriate mixed carbon sources have great potential
to enhance
denitrification efficiency and reduce operational costs in municipal
wastewater treatment plants (WWTPs). However, traditional methods
struggle to efficiently select the optimal mixture due to the variety
of compositions. Herein, we developed a machine learning-assisted
high-throughput method enabling WWTPs to rapidly identify and optimize
mixed carbon sources. Taking a local WWTP as an example, a mixed carbon
source denitrification data set was established via a high-throughput
method and employed to train a machine learning model. The composition
of carbon sources and the types of inoculated sludge served as input
variables. The XGBoost algorithm was employed to predict the total
nitrogen removal rate and microbial growth, thereby aiding in the
assessment of the denitrification potential. The predicted carbon
sources exhibited an enhanced denitrification potential over single
carbon sources in both kinetic experiments and long-term reactor operations.
Model feature analysis shows that the cumulative effect and interaction
among individual carbon sources in a mixture significantly enhance
the overall denitrification potential. Metagenomic analysis reveals
that the mixed carbon sources increased the diversity and complexity
of denitrifying bacterial ecological networks in WWTPs. This work
offers an efficient method for WWTPs to optimize mixed carbon source
compositions and provides new insights into the mechanism behind enhanced
denitrification under a supply of multiple carbon sources.