Water Permeability Coefficient Prediction for Polymers Applying Explainable Machine Learning

The water permeability coefficient of polymers is an important physical property and a critical factor in the quality of food products in the packaging industry, which sometimes can be difficult to measure experimentally. Data-driven machine learning (ML) approaches are important alternatives to assess the water permeability coefficient in a high-throughput way. In this study, a data set of homopolymers and copolymers with reported water permeability coefficient (P) values was collected from various sources to develop an explainable machine learning (ML) model to analyze a structure–property relationship. Sequential feature selection and ML modeling were applied for the selection of molecular descriptors and model development to find a correlation with water permeability. The best explainable ML model that includes data for both homopolymers and copolymers shows a R2 equal to 0.84 and 0.81 for training and test sets, respectively. A comprehensive analysis of ML-QSPR models was done to understand key factors influencing the model’s performance. Moreover, factors like molecular volume, ionization potential, and the presence of specific structural motifs, such as oxygens within specific topological distances, appeared to be crucial factors affecting the water permeability of polymers. Finally, the best ML model was implemented in a user-friendly web application to facilitate access to the developed model, enabling researchers to predict water permeability coefficients for polymers and reducing the time and effort required for experiments.