An accelerated strategy to characterize mechanical properties of polymer composites using the ensemble learning approach

Abstract

Screening new materials and decoding their structure-property relationships is a time-consuming and costly task in the laboratory. This research proposes the use of machine learning (ML) algorithms with a new train-test splitting strategy to predict the complete stress-strain curves of polymer composites based on their compositional, environmental, and processing conditions using data extracted from an open-source database. Both the artificial neural network (ANN) and ensemble learning (EL) models demonstrated acceptable fit to the training datasets, yielding root mean squared errors (RMSEs) below 3.2 MPa. However, EL models demonstrated superior performance by reducing the RMSE of the testing data by an average of 61% for polyethylene terephthalate (PET) composites and 30% for polycarbonate (PC) composites, compared to the wide ANN models. Additionally, the developed EL models reduced the training time from a few hours (∼4 h on average) for wide ANNs to just a few seconds. The results of this study imply that relying solely on ANN models may not always provide the best solution for a given problem in materials science, and that EL models should be considered as a viable alternative, particularly in cases where data are limited. This study will pave the way for the automated design and characterization of advanced composites, while reducing the costly and laborious experiments in keeping with the vision of smart manufacturing and Industry 4.0.