Microstructural images reconstruction of heterogeneous materials using generative adversarial networks

The macroscopic properties of heterogeneous materials are governed by their microstructures, yet acquiring high-quality structure–property data remains costly and time-consuming. This study develops a continuous conditional generative adversarial network (CcGAN) for the controlled reconstruction of diverse heterogeneous microstructures. Conditional accuracy is evaluated using parity plots and the R 2 coefficient of determination, while Two-Point Correlation Function quantify spatial similarity. Physical-response consistency is further evaluated through FEM-based homogenization by comparing effective thermal-conductivity distributions using the Wasserstein distance. Both Voigt–Reuss bounds and Hashin–Shtrikman bounds are used as the admissibility limits. The framework is demonstrated on fibrous, particulate, and experimental SEM microstructures of powder-metallurgy nickel-based superalloys. The results show that the proposed method reproduces the principal geometric and physical characteristics of the reference data, offering an efficient approach for microstructure generation and data-driven materials design.