Integrating interpretable machine learning with multi-algorithm and encoding strategies for electrooxidation process optimization in rifampin degradation

In this study, a novel and systematic prediction framework is proposed to optimize electrooxidation (EO) process parameters (anode material, applied current, reaction time, initial pH, type and concentration of supporting electrolyte, initial RIF concentration, and different real water matrices) during rifampin (RIF) degradation. This framework integrates machine learning (ML) algorithms (random forest, decision tree, gradient boosting, multilayer perceptron, support vector machine, and K-nearest neighbors) with various encoding strategies (one-hot, target, and label). Furthermore, hyperparameter optimization techniques (grid search, grid search with noise-augmented data, random search, and Bayesian optimization) were evaluated to enhance predictive performance and provide an in-depth examination of RIF degradation. The boron-doped diamond (BDD) anode achieved 100% RIF degradation under optimal conditions (i = 0.2 A, C RIF = 10 mg/L, SEc = 51 mM NaCl/L). In the performance evaluation of ML algorithms, the one-hot encoding for the DT algorithm yielded the best predictive accuracy, with the highest R 2 = 0.952 and Adjusted R 2 = 0.920 scores, alongside the lowest error metrics: MAE = 3.909, RMSE = 7.594, RSR = 0.217, AAD = 0.0624, and MAPE = 10.391. However, future importance analysis using Shapley Additive Explanations (SHAP), Gini impurity, and Mean Decrease in Accuracy (MDA) revealed that reaction time and anode materials are the most influential parameters for the EO process, whereas the electrolyte types and concentration demonstrated limited impact. Overall, the findings provide important insights into the EO process optimization and demonstrate the statistical robustness of the proposed ML-based framework.