Estimation of Local Strain Fields in Two-Phase Elastic Composite Materials Using UNet-Based Deep Learning

The knowledge of the distribution of local micromechanical fields is crucial in the design of composite materials. Traditionally full-field methods (such as finite element methods) and fast Fourier transformation-based methods are used to obtain the local fields. However, full-field simulations are computationally expensive and time-consuming. Recently, there has been a push toward using the big-data-driven machine learning approaches to estimate the local fields and establish the structure–property linkages. In this work, we use one of the deep learning-based algorithms known as the UNet to predict the local strain fields in a two-phase composite material subjected to uniaxial tensile load. The model is trained and tested on 1200 two-phase microstructures comprising two-volume fraction categories and six different morphological classes. An R score of 94% is achieved on the test dataset. A detailed statistical analysis is performed to understand the role of the volume fraction and the ratio of elastic moduli of the phases in the deep learning model’s trainability. The insights drawn in this work are then discussed in the context of generating artificial datasets and training a robust predictive deep learning model for localization.