Investigating Deep Learning-Based Stress Prediction in Particulate Polymer Composites Using Multiple Quality Measures

The essential qualities for materials in various engineering and manufacturing industries require minimizing manufacturing costs and ensuring optimal design to reinforce safety and reliability. Particulate polymer composites (PPC) find wider applicability due to their versatility and exceptional structural and mechanical properties. To reduce the computational runtime of finite element modelling (FEM) software typically used for analysis of material mechanical properties, deep learning approaches have gained significant popularity. However, the stress fields generated by these approaches are inefficiently matched with FEM results, particularly when based only on a few parameters. In this study, a pix2pix conditional Generative Adversarial Network (cGAN) is used for predicting the stresses as 2D images in PPC corresponding to the results of ABAQUS CAE 2020 FEM simulations. The results are evaluated corresponding to several image matching quality metrics including Structural Similarity Index Measure (SSIM), Peak Signal to Noise Ratio (PSNR), Correlation coefficient, and Mean Square Error (MSE). In effect, these metrices reveal the deep learning model capabilities to imitate features and subsequently determine the optimal configuration of the model. The experimental results showed strong similarity based on these metrices between the modelled results and the FEM simulation results.