Utilizing machine learning for reactive material selection and width design in permeable reactive barrier (PRB)

Abstract

Permeable reactive barrier (PRB) is an important groundwater treatment technology. However, selecting the optimal reactive material and estimating the width remain critical and challenging problems in PRB design. Machine learning (ML) has advantages in predicting evolution and tracing contaminants in temporal and spatial distribution. In this study, ML was developed to design PRB, and its feasibility was validated through experiments and a case study. ML algorithm showed a good prediction about the Freundlich equilibrium parameter (R² 0.94 for K_F, R² 0.96 for n). After SHapley Additive exPlanation (SHAP) analysis, redefining the range of the significant impact factors (initial concentration and pH) can further improve the prediction accuracy (R² 0.99 for K_F, R² 0.99 for n). To mitigate model bias and ensure comprehensiveness, evaluation index with expert opinions was used to determine the optimal material from candidate materials. Meanwhile, the ML algorithm was also applied to predict the width of the mass transport zone in the adsorption column. This procedure showed excellent accuracy with R² and root-mean-square-error (RMSE) of 0.98 and 1.2, respectively. Compared with the traditional width design methodology, ML can enhance design efficiency and save experiment time. The novel approach is based on traditional design principles, and the limitations and challenges are highlighted. After further expanding the data set and optimizing the algorithm, the accuracy of ML can make up for the existing limitations and obtain wider applications.

Keywords: Design methodology; Evaluation index; Permeable reactive barrier; SHapley Additive exPlanation; Total petroleum hydrocarbon; XGBoost.