RP-cGAN for microscale stress-strain field mapping in FRP with interphase-aware learning

The damage initiation mechanism of fiber-reinforced composites under transverse loading is highly dependent on local stress concentrations in the fiber-matrix interfacial region. Although traditional finite element methods can accurately characterize multi-field distributions with interfacial interphases, their high computational costs significantly limit design optimization efficiency. Existing data-driven models face dual challenges in predicting such complex microstructures: interfacial effects cause notable declines in the prediction accuracy of local field variables, and the frequency-domain conjugate symmetry of strain fields under shear loading induces systematic distortions. This paper proposes a Residual-Pixelshuffle conditional Generative Adversarial Network (RP-cGAN), which establishes a direct mapping from microstructures to field distributions by deeply integrating residual connections with sub-pixel up-sampling techniques. The method effectively overcomes the vanishing gradient problem in traditional convolutional networks, eliminates checkerboard artifacts caused by transposed convolution (reducing FID metric by 65.8 %), and incorporates a loading direction labeling mechanism to address the conjugate symmetry challenge under shear conditions. Based on refined modeling of a T300/epoxy system (including a 118 nm interfacial interphase), a multi-condition dataset covering 20 %-60 % fiber volume fractions was constructed. Validation shows that RP-cGAN achieves high-fidelity stress/strain field reconstruction under both tensile and shear loading, with a 21.34 % improvement in comprehensive prediction accuracy across multiple metrics (average PSNR >33 dB, SSIM >0.98, C-MSE <0.0008), significantly outperforming the cGAN models. The model exhibits strong adaptability to complex configurations such as non-periodic fiber distributions and densely packed fiber regions. This work provides an efficient and reliable computational tool for predicting damage initiation in composites, offering significant engineering value for lightweight aerospace equipment design. • Novel model predicts complex stress/strain in composites 21 % more accurately than existing tools, especially near fiber-matrix interfaces. • Eliminates "checkerboard" artifacts in predictions and handles shear loading distortions via directional labeling. • Enables rapid design optimization for aerospace composites under tension/shear without costly simulations.