Inverse design of interpenetrating phase composites with targeted stiffness through deep learning

Abstract

Interpenetrating phase composites (IPCs) contain at least two interconnected phases. It is still challenging to establish the topology-property mapping and inversely design IPCs with targeted properties. A deep learning model is proposed to design new IPCs with various material symmetries. Hybrid triply periodic minimal surface (TPMS) scaffolds are mathematically designed as the reinforcement phase to construct the IPCs. The IPCs display orthotropic, tetragonal and cubic material symmetries, whose stiffness tensors are determined using a numerical homogenization method. A robust dataset is generated to establish the topology-property mapping. Then, a tandem dual-network model, including a forward sub-model and an inverse sub-model, is developed to predict the stiffness tensor and design the topologies. In addition, the hyperparameters are optimized to improve the accuracy of the model. The dual-network model provides excellent prediction and design capabilities. Moreover, the inversely designed IPCs can fulfill the requirements of the targeted stiffness both in and beyond the original dataset, which include orthotropic, tetragonal and cubic material properties. The current study provides a feasible approach to the forward prediction of elastic properties and inverse design of topologies through deep learning.