Deep learning aided evaluation for electromechanical properties of complexly structured polymer nanocomposites

Abstract

In the present work, two types of deep neural networks (DNNs) were employed to establish the structure–property relationship of polymer nanocomposites. The trained DNNs based on multiscale analysis results can not only overcome the limitations of the conventional clustering density-based model and multivariate regression models but also exhibit superior performance in evaluating the electromechanical properties of polypropylene matrix composites, wherein spherical SiC nanoparticles were randomly distributed and dispersed. A simple graph convolution network showed better capability than a complex artificial neural network, despite fewer features considered; this implies that the graph convolution network is more appropriate and user-friendly for evaluating the effect of nanoparticle distribution and agglomeration. In addition, the trained graph convolution network can effectively provide mechanical and electrical properties corresponding to large representative volume element (RVE) without a loss of accuracy. The present study demonstrates that deep learning techniques can be put to practical use for the design of next-generation polymer nanocomposite materials.