Deep learning-based inverse design framework for property targeted novel architectured interpenetrating phase composites

Abstract

Interpenetrating phase composites (IPCs) are advanced multi-phase composites where each phase forms an entirely interconnected network leading to enhanced performance. This paper introduces a novel inverse design approach, wherein mathematically-known triply periodic minimal surfaces (TPMS) representations are combined with a deep neural network-based approach for generating IPCs with specified properties. The proposed approach utilizes a weighted combination of Schwartz P, Diamond D, and Schoen’s F-RD TPMS architectures to generate novel IPCs. Additionally, we outline a novel deep learning-based computational approach that predicts combinatorial TPMS-based IPCs for targeted effective elastic properties. Specifically, a five-layer deep neural network (DNN) that enables inverse mapping between five different material properties and six geometrical parameters defining the TPMS-based IPCs is outlined. It is also empirically shown that DNN accurately predicts combinatorial TPMS-based IPCs at a fraction of the computational cost and hence, can play a vital role in multiscale design problems and find extensive usage in IPCs design problems. The effective elastic properties (Young’s modulus, Poisson’s ratio, bulk modulus, and shear modulus) of proposed generative TPMS-based IPCs are evaluated using the finite element method and are compared against a single phase TPMS-based IPCs, analytical models, and elastic bounds. Numerical results demonstrate that IPCs generated by the proposed method shows superior mechanical behavior.