Machine Learning Modeling Based on Microbial Community for Prediction of Natural Attenuation in Groundwater

Natural attenuation is widely adopted as a remediation strategy, and the attenuation potential is crucial to evaluate whether remediation goals can be achieved within the specified time. In this work, long-term monitoring of indigenous microbial communities as well as benzene, toluene, ethylbenzene, and xylene (BTEX) and chlorinated aliphatic hydrocarbons (CAHs) in groundwater was conducted at a historic pesticide manufacturing site. A machine learning approach for natural attenuation prediction was developed with random forest classification (RFC) followed by either random forest regression (RFR) or artificial neural networks (ANNs), utilizing microbiological information and contaminant attenuation rates for model training and cross-validation. Results showed that the RFC could accurately predict the feasibility of natural attenuation for both BTEX and CAHs, and it could successfully identify the key genera. The RFR model was sufficient for the BTEX natural attenuation rate prediction but unreliable for CAHs. The ANN model showed better performance in the prediction of the attenuation rates for both BTEX and CAHs. Based on the assessments, a composite modeling method of RFC and ANN was proposed, which could reduce the mean absolute percentage errors. This study reveals that the combined machine learning approach under the synergistic use of field microbial data has promising potential for predicting natural attenuation.