Expanding Design Spaces in Digital Composite Materials: A Multi-Input Deep Learning Approach Enhanced by Transfer Learning and Multi-kernel Network

Abstract

This study presents a novel approach to designing digital composite materials with desired mechanical properties by exploring a broad design space based on the spatial arrangements of binary constituents with a variety of materials' properties. Deep learning (DL) models that are trained on limited volume fraction (VF) ratios and limited materials' properties often struggle to accurately predict the mechanical responses of configurations that are not encompassed in the training data. To address this issue, an advanced multi-input deep learning approach is proposed, enhanced by transfer learning and a multi-kernel method. This approach can predict the stress field for both seen and unseen configurations in terms of the material properties and VF ratios, while accurately pinpointing stress concentrations at the interface. It can predict stress distribution from finite element method (FEM) accurately with significantly low computational cost, making it an efficient tool for the rapid design and optimization of composites. The incorporation of multiscale kernels in the model enables better capture of local and global features, resulting in more precise predictions. Transfer learning (TL) in the model proves to be an exceptional strategy for exploring new design spaces with minimal data augmentation. This research underscores the potential of DL models in advancing composite materials.