A triplet attention-enhanced deep learning approach to predict full-field stress of unidirectional CFRP composites with microvoids

Abstract

Stress analysis is a critical step in composite material design. The high computational cost of multiscale finite element analysis necessitates the development of efficient surrogate modeling frameworks. To this end, this study develops a U-Net deep learning model enhanced with Triplet Attention mechanism for efficient prediction of full-field spatially nonlinear stress distribution in unidirectional CFRP composites with microvoids. Firstly, a Python based parametric modeling script is developed by integrating greedy and hard-core algorithms, enabling the construction of a microstructural database with randomly distributed voids and fibers. Then, the high-resolution stress field are computed by micromechanics-based finite element method. Subsequently, a U-Net architecture embedded with a Triplet Attention module is designed to improve the model's capability in extracting critical stress features. Finally, the proposed method is compared with several mainstream attention mechanisms including NonLocal, CBAM, and CrissCross using weighted MSE, MS-SSIM, R², and SNR. The comprehensive evaluation demonstrates significant performance improvements over the traditional U-Net framework: Weighted MSE is reduced by 74.7 %, MS-SSIM improved by 4.7 %, R² increased by 21.3 %, and SNR enhanced by 25.2 %. Further validation through varying dataset size and five-fold cross validation also confirms the model's reliability and robustness. This integrated approach simultaneously meets the technical requirements for computational efficiency, prediction accuracy, and engineering applicability in composite multiscale simulations, and thus provides a promising tool for applications in multiscale simulation, microstructural optimization, and uncertainty quantification.