Machine learning prediction of mechanical and optical properties of uniaxially oriented polymer films

Abstract

Improving properties of polymers can bring about tremendous opportunities in developing new applications. However, the commonly used trial-and-error method cannot meet the current need for new materials. We demonstrate the utility of Machine Learning (ML) algorithms in creating structure-process-property models based on industrial data in polymer processing. In this study, ML algorithms were used to predict the optical and tensile strength of multi-layer co-extrusion polyethylene films as a function of material structures and process parameters. The input features to predict the mechanical and optical properties are the composition of five-layer polyethylene film, polyethylene molecular properties like the amount of long chain branching $\left(\mathrm{LCB}\right)$, and the extrusion process conditions. Different data featuring steps are conducted to improve the quality of the input data: (1) feature importance scoring using an ensemble algorithm (XGBoost); (2) application of autoencoder to reduce the dimensionality; (3) replacing the categorical inputs with molecular characteristic properties. We then use this data to build an Artificial Neural Network. Finally, the prediction capability of the resulting model was investigated. This project demonstrates a successful end-to-end execution of a material data science project; from understanding material science, data engineering, algorithm development, and the model evaluation.