Machine learning predictions and benchmarking of non-linear mechanical behavior of polymer composites

Abstract

The rise of environmental concerns and the worldwide transition to a circular economy is partly fueling the discovery of new materials with unique functionalities that satisfy industry requirements. Polymer composites are among the most popular materials in various metal-replacement applications, such as auto and aerospace light-weighting, electronics, optics and energy storage devices. Nonetheless, their vast compositional design freedom and sensitivity to ambient conditions require extensive physical experiments. To address this issue, the aim of this research is to develop machine learning (ML) algorithms which are able to characterize entire stress-strain curves of polymer composites based on their composition, processing and environmental conditions. Three distinct feature variables including temperature, filler content and strain were utilized to predict the output stress amounts. The results indicated that the artificial neural network (ANN) models, with a new train-test splitting strategy, accurately fit the training data, as evidenced by the root mean squared error (RMSE) values below 3 MPa for PET composites and below 1 MPa for PC composites. Furthermore, the developed ANN models revealed a satisfactory performance on the testing data, with RMSE of less than 1.5 MPa for PC composites and approximately 6 MPa for PET composites. These findings pointed out the effectiveness of ANN models in predicting complete stress-strain curves of polymer composites, especially when a sufficient amount of data is available. The outcomes obtained will pave the way for automated design and characterization of advanced multifunctional composites while minimizing extensive physical testing, thereby advancing the vision set out by Industry 4.0.