Performance evaluation of deep learning approaches for predicting mechanical fields in composites

This paper presents a rigorous and critical methodology for evaluating the performance of deep learning (DL) techniques in predicting mechanical responses within the microstructural representation of composites. In the past few years, deep learning has emerged as a powerful tool and an efficient surrogate for finite element analysis in computational mechanics. This research addresses questions regarding the suitability of common error metrics for evaluating the accuracy of DL techniques in predicting full-field mechanical responses of composites. Through comparative analysis, we evaluate the performance and identify the limitations of two DL frameworks in predicting the linear von Mises stress distribution within the microstructure of the selected fiber-reinforced composite. The first DL method is based on the residual network, while the second utilizes U-Net architecture. We use several evaluation metrics, including different types of error and statistical measures and examine their suitability. Additionally, this study also investigates the influence of the size of the training and validation dataset, ranging from 50 to 2000 samples, on the predictive performance of the employed ResNet and U-Net based approaches.