Prediction of oxygen barrier properties of polymer-based nanocomposite membranes by using machine learning: comparison of random forest and support vector machine models

In this study, two machine learning models (ML), namely Random Forest (RF) and Support Vector Machine (SVM), have been trained to predict the oxygen relative permeability of polymer-based nanocomposite membranes. The dataset of more than two hundred samples covers a large range of nanocomposites matrices going from amorphous to semi-crystalline polymers and from rubbery to glassy polymers. In the same way, a wide variety of fillers with distinct shapes (spherical, elongated, lamellar), size and chemical nature have been considered. Six parameters related solely to base materials (filler type including possible filler orientation for lamellar fillers and aspect ratio, filler amount, filler polarity, matrix oxygen permeability and matrix surface energy) have been selected as input parameters. Additionally, the most prevalent techniques for the synthesis of nanocomposites (e.g.: melt-blending, solvent casting and in situ polymerization), have also been considered. The machine learning models were trained on 80% of the randomly selected samples from the dataset and the remaining 20% of samples were employed to assess the model accuracy by comparing the predicted outputs with the experimental results. In comparison with RF regression, SVM regression leads to relative permeability predicted values less consistent with experimental values for both training and testing. Despite the optimization of SVM machine learning approach, by including as a first step a classification between high and low relative permeabilities, the RF model remains the most appropriate to predict the barrier gain of nanocomposites with a mean absolute error of 11% and a mean coefficient of determination of 0.75.