| Title |
Development of a Generative Adversarial Network-Based Image Generator Trained on Grain Growth Simulation Data |
| Authors |
김도윤(Do-Yoon Kim) ; 최윤수(Yunsu Choi) ; 최지민(Jimin Choi) ; 권용우(Yongwoo Kwon) |
| DOI |
https://doi.org/10.3365/KJMM.2026.64.10.913 |
| ISSN |
1738-8228(ISSN), 2288-8241(eISSN) |
| Keywords |
Generative adversarial network; Grain growth; Microstructure generation; Phase-field model; Physics-informed learning |
| Abstract |
The phase-field method (PFM) provides highly accurate microstructure simulations but suffers from substantial computational costs that increase rapidly with system size, dimensionality, and physical complexity. To overcome this limitation, generative artificial intelligence can be utilized as an ultra-fast surrogate model for microstructure generation. In this study, we constructed a Wasserstein generative adversarial network with gradient penalty (WGAN-GP) trained on two-dimensional (2D) PFM grain growth data. The trained generator bypasses conventional sequential time integration and directly produces microstructures at desired timesteps, achieving a computational speedup of several orders of magnitude compared with the original PFM. To improve physical consistency, a spatial residual term derived from the Allen-Cahn equation was incorporated into the generator loss as a physics-inspired regularization term. During the early stage of training, the resulting physics-informed WGAN-GP (PI-WGAN) exhibited approximately 50% fewer dangling grain boundaries than the baseline WGAN-GP, indicating more rapid establishment of continuous and physically consistent grain boundary networks. PI-WGAN also suppressed anomalous disconnected boundaries throughout training, while the difference between the two models gradually decreased as training progressed. Furthermore, PI-WGAN reproduced the mean and median grain counts of the PFM data with mean relative errors of only 0.8?1.3%. The incorporation of physics-inspired regularization required only negligible additional computational overhead during training. These results demonstrate that a simple physics-based constraint can improve the structural consistency of generated microstructures, particularly during the early stages of learning. This 2D study therefore provides a proof of concept for the use of physics-informed generative models as computationally efficient microstructure generators and establishes a basis for their future extension to large-scale, three-dimensional (3D), and more complex phase-field problems. |