Tree-based ensemble machine learning model for nitrate reduction by zero-valent iron

초록

The use of zero-valent iron (ZVI) for nitrate reduction has long been a topic of interest owing to its excellent reactivity compared to other inorganic electron donors. However, there is currently a lack of comprehensive and sophisticated models that can effectively integrate and analyze the dynamic data generated by this type of experiment. Such modeling is necessary to facilitate information mining and improve our understanding of the underlying mechanisms. Taking advantage of available published experimental data, a gradient-boosting decision tree and extreme gradient boosting were applied to 288 data points collected from peer-reviewed publications and compared with conventional non-tree-based models, i.e., artificial neural networks (ANN) and support vector regression (SVR). First, a predictive analysis was performed for the rate constant and the percentages of ammonium and nitrogen conversion using seven predictors related to ZVI properties, reactant concentration, and experimental parameters. Both tree-based models accurately predicted the target variables with comparable performance (R2 = 0.88–0.97) and outperformed ANN and SVR (R2 = 0.60–0.89). In addition, SHapley Additive exPlanations analysis revealed that ZVI properties (44 % average contribution) and experimental condition (68.5 % average contribution) were the key variables affecting the nitrate reduction rate constant and conversion product, respectively. Aerobic conditions were favorable for the conversion of nitrate to ammonium, with anoxic conditions being beneficial for the selectivity of nitrogen gas conversion. This study demonstrates a promising approach using machine learning with complex cross-system data for predicting the performance of water treatment and for mining valuable insights into the process by model interpretation. This method presents a new approach to revolutionize data analysis and modeling in water remediation, which can simplify the experimental operational burden. © 2023 Elsevier Ltd

키워드