Machine learning for kinetic constant of organic pollutant removal in ultrasonic-activated oxidant processes: Scenario simulation for interpretation and validation

초록

Machine learning (ML) techniques have increasingly been adopted to model degradation kinetics in ultrasonic activated advanced oxidation processes (US-AOP). However, current ML-based studies commonly exhibit several methodological limitations: insufficient consideration of intrinsic oxidant characteristics, lack of systematic rationale in algorithm selection, and inadequate examination of model validation strategies. To address these critical gaps, the present study developed a robust ML modeling framework specifically tailored for ultrasonic activated oxidant (US-oxidant) systems. A comprehensive dataset containing 170 data points collected from 22 literature sources was established, incorporating 17 diverse input features covering ultrasonic parameters, oxidant descriptors, pollutant molecular characteristics, solution chemistry, and operational conditions. Three widely used gradient-boosting algorithms were systematically benchmarked across varying cross-validation (CV) strategies and training-to-test ratios. Results demonstrated that a gradient boosted decision trees (GBDT) model with optimized hyperparameters achieved superior predictive accuracy and excellent generalization capability. The optimal GBDT model was further interpreted using shapley additive explanations (SHAP) and validated via scenario simulations. In addition to conducting inverse design using the particle swarm optimization (PSO) algorithm, a new method for selecting the optimal system based on the electrical energy per order (E<inf>EO</inf>) approach was also proposed. By addressing previously neglected methodological aspects, this study provides a systematic and reproducible ML modeling protocol, serving as a robust reference framework to facilitate future ML-based investigations of US-oxidant degradation systems. © 2025 Elsevier Ltd.

키워드