Stress Field Prediction in Particulate Polymer Composite Materials Using Paired Image-To-Image Translation Approach

Stress analysis is essential in understanding the mechanical behavior of designed composite structures. Conventionally, Finite Element Modeling (FEM) is employed for the computational analysis of newly fabricated materials. Ideally, for composite structures, the stress observations from the computational analysis using software such as ABAQUS must agree with the experimental yields of stresses. However, traditional computational techniques are known to have higher complexity and longer runtimes while further requiring intensive labor. Alternatively, in recent studies, to reduce the constraints of software such as ABAQUS, various machine learning approaches have been explored for examining the mechanical behavior of Particulate Polymer Composite (PPC). This study demonstrates the advantages and potential of using advanced deep learning architectures for the investigation of polymer stresses. In the proposed study, a paired image-to-image translation with Generative Adversarial Networks (GANs) is employed. The PPC stresses are developed using ABAQUS at six time steps and used for training the GAN. The results reveal that the implementation of the paired image-to-image translation with GANs aids an accurate and efficient investigation of stress fields in newly fabricated PPC materials. The performance of the approach is evaluated using the correlation between the reference data and the results. Further, the least mean square error (L2 norm) is investigated to analyze the performance of the network. The maximum correlation for the test results for epoxy, and polymer fibers are observed to be 0.9975, and 0.99975. The value of L2 Norm is observed to be less than 0.005 for stresses in x direction in both the materials used in the preparation of particulate polymer composite two-dimensional micromechanics model.