Robust forecasting of packaging chemical migration into water and food

Chemical migration from food contact materials (FCMs) into food and water poses significant safety concerns. Accurate prediction of this migration is essential for risk assessment and regulatory compliance, yet experimental testing is time-consuming and costly. This research employs a varied array of machine learning (ML) techniques to predict packaging chemical migration, expressed as logKmW. A dataset of 1,847 experimental logKpf values covering 232 materials across 19 packaging compounds was used. Key input variables included material type, temperature (275–313 K), ethanol equivalency (0–100%), and logKow at 298 K. Preprocessing involved z-score normalization, one-hot encoding, and Monte Carlo Outlier Detection (MCOD). Fifteen ML models were tested, including XGBoost, LightGBM, Random Forest, SVR, ANN, and CNN. Correlation analysis showed that logKeq @ 298K (r = 0.63) and silicone rubber (r = 0.59) positively influenced migration, while EtOH-eq (r = –0.68) and temperature (r = –0.26) had negative effects. Among the models, XGBoost performed best with R2 = 0.9957, MSE = 0.0067, and MRD% = 17.29 on the test set. LightGBM and Random Forest also yielded high accuracy. Visualization and SHAP analysis confirmed the dominance of physicochemical variables in predicting migration behavior. The results demonstrate that advanced ML models, especially ensemble tree-based methods, can effectively predict chemical migration into food and water. This work provides a scalable and reliable framework for modeling migration and identifying key influencing variables.