We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Machine learning prediction of organic contaminant partitioning to microplastics: A volume-normalized cross-polymer framework
Summary
Scientists built an AI tool that predicts which chemical pollutants are most likely to stick to microplastic particles, based on the plastic's properties and the chemical's structure — without needing to run a new lab test every time. This matters because microplastics can act like tiny sponges that soak up toxic chemicals from water and carry them into our food and bodies, so being able to quickly flag the highest-risk chemical-plastic combinations helps regulators prioritize which contaminants deserve the most attention.
Microplastics (MPs) can act as vectors for hydrophobic organic contaminants, altering their environmental fate and exposure pathways in aquatic systems. However, current understanding of contaminant-microplastic partitioning is still largely derived from fragmented batch sorption experiments, where heterogeneity in polymer types, contaminant structures, and test conditions limits cross-system comparison and predictive generalization. Here, we develop a machine learning framework to predict microplastic - water partition coefficients across diverse polymer materials and environmental conditions. The model integrates contaminant molecular descriptors (logK and Abraham parameters), microplastic properties (polymer type, density, particle size, surface area, and aging status), and environmental variables (pH, salinity, and temperature). To enable physically consistent comparison among polymers, a volume-normalized partition coefficient (logK) was introduced to correct for density-related bias. Under leave-one-polymer-out validation, the XGBoost model achieved R² values of 0.60-0.78 for the major hydrophobic polymers, indicating reasonable predictive performance under a stringent cross-polymer extrapolation setting, and showed promising generalization to contaminants not included in the training dataset (R² = 0.766). Model interpretation revealed that contaminant molecular properties dominated predictions (≈64 %), while microplastic characteristics and environmental conditions showed secondary nonlinear effects, including threshold-like responses and pH-dependent sign reversal. Application of the framework to 66 regulatory-relevant contaminants across three environmental scenarios identified 37 compounds (56 %) with consistent classification outcomes, including 25 high-priority and 12 low-priority substances. An interactive web-based platform (https://mp-insight.streamlit.app/) was developed to support rapid screening of emerging contaminants. Overall, this framework reduces reliance on repetitive experiments and provides a practical tool for data-limited assessments and evidence-based environmental management.