A Structured Evidence Synthesis and Interpretable Machine-Learning Framework for Predicting Nanoplastic Environmental Fate in Water and Soil

Nanoplastics behave as colloids: their environmental fate depends less on polymer identity than on the chemistry and structure of the medium through which they move. The evidence for this is scattered across studies that vary one factor at a time and rarely connect aquatic and terrestrial compartments. We address that fragmentation with an evidence-to-prediction platform built in three steps. A structured evidence synthesis first harmonised 346 study-condition cases from 36 DOI-tracked primary studies spanning aggregation, deposition, porous-media transport, and biologically mediated redistribution. Six classifiers were then benchmarked under cross-validation grouped by study, so that no publication contributed cases to both training and validation folds; the best model reached a cross-validated receiver operating characteristic area under the curve (ROC AUC) of 0.950 (95% confidence interval 0.934–0.965), an F1 score of 0.889, and a Brier score of 0.089, performing equally well in aquatic (n = 235) and soil (n = 111) subsets. The model was finally made deployable through probability calibration, SHapley Additive exPlanations (SHAP) attribution, conformal prediction sets, and retrieval of the evidence cases underlying each prediction. The central result is that the learned control hierarchy reproduces classical colloid theory from data alone: electrolyte concentration and cation valence dominate, followed by natural organic matter (NOM) regime, weathering state, eco-corona composition, and biological activity. This convergence supports a single reactivity–accessibility reading of nanoplastic fate, in which chemical attachment propensity and physical pathway availability jointly decide whether a particle travels or is retained, in water and in soil alike.