Robust prediction of particle attachment efficiency for nanoparticle and microplastic environmental mobility assessment.

Particle attachment efficiency (α) controls whether suspended nanoparticles and microplastics remain mobile or aggregate, making reliable estimates important for environmental transport, retention, and treatment assessments. Earlier curated-data studies demonstrated machine-learning prediction, but did not jointly test stability across data partitions, physical plausibility, physicochemical interpretation, and transfer to unseen experimental groups. We benchmarked Support Vector Regression, Random Forest, LightGBM, CatBoost, baseline XGBoost, and Improved Harris Hawks Optimization (IHHO)-tuned XGBoost using 2565 published mono-particle and binary-particle aggregation experiments. A leakage-controlled workflow kept the external test partition untouched across ten random-seed repetitions and combined bootstrap confidence intervals, paired comparisons, out-of-range checks, grouped validation, and seed-averaged SHAP interpretation. Baseline XGBoost achieved the strongest mean held-out accuracy, with R=0.744±0.121 for mono-particle and R=0.555±0.088 for binary-particle systems. IHHO-XGBoost did not improve mean accuracy (R=0.699±0.092 and 0.546 ± 0.056) but showed 41.4% and 59.7% lower descriptive cross-seed R variance, respectively; paired tests did not establish decisive superiority. Seed-aggregated SHAP results identified salt concentration and critical coagulation concentration as the leading mono-particle model drivers, and primary-particle zeta potential and salt concentration as the leading binary-particle drivers. Performance deteriorated when complete sources, particle groups, salts, or methods were withheld. The study therefore contributes a reproducible reliability benchmark and shows that current descriptors support environmental screening and interpolation more strongly than extrapolation to new particle systems.