Operator Learning for Predicting Mechanical Response of Hierarchical Composites with Applications of Inverse Design

Materials-by-design to develop high performance composite materials is often computational intractable due to the tremendous design space. Here, a deep operator network (DeepONet) is presented to bridge the gap between the material design space and mechanical behaviors. The mechanical response such as stress or strain can be predicted directly from material makeup efficiently, and a good accuracy is observed on unseen data even with a small amount of training data. Furthermore, the proposed approach can predict mechanical response of complex materials regardless of geometry, constitutive relations, and boundary conditions. Combined with optimization algorithms, the network offers an efficient tool to solve inverse design problems of composite materials.

Automated summary

In this paper, a deep operator network (DeepONet) is proposed to bridge the gap between the material design space and mechanical behaviors in the development of high-performance composite materials. The network efficiently predicts the mechanical response such as stress or strain directly from material makeup, and achieves good accuracy even with limited training data. Moreover, it can predict the mechanical response of complex materials regardless of geometry, constitutive relations, and boundary conditions. Combined with optimization algorithms, the network serves as an efficient tool for solving inverse design problems of composite materials.