Abstract
Accurate prediction of full-field nonlinear mechanical responses in heterogeneous materials remains a fundamental challenge in computational engineering, where conventional finite element analysis (FEA) is often too costly for large-scale design and inference tasks. This work introduces a physics-guided conditional Generative Adversarial Network framework to predict full-field nonlinear plastic strain distributions in aluminum-silicon carbide nanocomposites under shear loading. By embedding physical symmetry properties of the boundary value problem through symmetry-consistent data augmentation, the surrogate learns robust mappings from microstructural geometry to localized plasticity patterns, including sharp shear bands and fiber–matrix interactions. The framework is validated across microstructures of varying complexity, achieving a coefficient of determination (R2) of 0.99 for 20-fiber systems and maintaining robust scalability with an R2 of 0.96 for significantly more complex 57-fiber systems, along with low global mean squared error across domains. Once trained, the surrogate produces full-field predictions in milliseconds, enabling orders-of-magnitude speedups relative to traditional FEA. This study advances engineering informatics by demonstrating how generative AI, when coupled with physics-consistent training data and framework, can transform nonlinear field prediction into a computationally efficient and scalable inference task. The resulting methodology supports digital twin development, design optimization, predictive maintenance, and real-time decision support in advanced manufacturing and structural analysis.