Hybrid Ensemble Machine Learning Models with SHAP Explainability for Robust Prediction of Suspended Particle Attachment Efficiency in Complex Environmental Systems

The accurate prediction of aggregation attachment efficiency (α) is critical for suspended nanoparticle and microplastic fate in environmental systems, yet existing models struggle with nonlinear interactions and limited interpretability.  This study evaluates two recently proposed hybrid ensemble machine learning frameworks, Improved Harris Hawks Optimized XGBoost (IHHO-XGBoost) and AdaBoost-ExtraTrees, for predicting α across mono- and binary-particle systems. Using a curated dataset spanning diverse particle types and environmental conditions, we demonstrate that IHHO-XGBoost outperforms six benchmark algorithms, achieving test R² values of 0.865 (mono particle) and 0.797 (binary particle).  SHAP analysis reveals distinct mechanistic drivers: salt concentration dominates mono particle aggregation, while zeta potential asymmetry controls binary systems. By adapting these advanced ensembles to colloidal stability prediction, this work provides a computational framework for improving the prediction of particle interactions in complex environmental matrices.