Hybrid ML-driven simulation exploring the fate of environmentally relevant nanoplastics in complex aqueous matrices: Mechanistic insight into colloidal behavior via explainable AI.

Nanoplastics (NPs) are globally recognized as pervasive emerging contaminants with demonstrated toxicological risks, yet predicting their environmental behavior under realistic conditions remains a major scientific challenge. Addressing this gap, this study integrates controlled aggregation experiments with a hybrid machine-learning (ML) framework to enhance mechanistic and predictive understanding of NP fate across diverse aqueous matrices. Experimentally derived aggregation kinetics were used to train a stacked ML architecture, while Explainable AI (XAI) methods elucidated the underlying drivers of model predictions, providing essential post-hoc interpretability for an otherwise black-box system. Results based on Critical Coagulation Concentration and XAI analyses demonstrate that increasing pH enhances NP colloidal stability, whereas elevated temperature and particle concentration accelerate aggregation. Under environmentally relevant conditions, particularly those reflecting weathering-induced surface transformation and dissolved organic matter-mediated complexation, divalent cations with larger ionic radii exert a dominant destabilizing influence, rendering contributions from monovalent cations effectively negligible. Model simulations indicate that groundwater, soil porewater, and riverine systems generally preserve dispersed NP behavior, supporting potential subsurface mobility and fluvial transport. In contrast, the sharp saline transition within estuarine environments markedly promotes NP agglomeration, increasing the likelihood of deposition and retention within benthic sediments. Global-scale simulations based on secondary water chemistry datasets suggest that most freshwater systems, including groundwater, favor colloidally stable NP suspensions, whereas coastal and saline regimes promote destabilisation. Collectively, this integrated experimental-computational framework provides a robust and generalizable predictive tool that bridges laboratory observations with field complexity, advancing understanding of NP fate, transport, and ecological risk.