Interpretable artificial intelligence for advanced oxidation systems: Principle, operations and performance

Abstract

Advanced oxidation processes have been widely studied and employed due to their potent mineralization capacity for pollutants. However, the intricate reaction mechanisms of these processes pose limitations for fitting and predicting performance. In this study, we comprehensively derived and assessed neural networks for three advanced oxidation methods: catalytic oxidation, catalytic wet air oxidation, and electrochemical oxidation. Our analysis encompassed multilayer perceptron principles, forward and back propagation process, strategies for handling overfitting, and performance evaluation matrices. Additionally, we utilized Bayesian optimization to probe the impact of network architecture on outcomes. Two conventional methods, multiple linear regression and response surface methodology, are employed for comparison. Our results demonstrate that neural networks exhibit more robust performance in fitting and predicting advanced oxidation processes as indicated by statistical indicators. Importantly, we tackle the "black box" issue of neural networks by incorporating the shapley additive explanations interpretable model of game theory to elucidate the impact of advanced oxidation features on outcomes. The superior performance of explainable artificial intelligence techniques implies their vast potential for broad applications in environmental science and technology.