Transfer learning approach with Voronoi-based synthetic data for high-accuracy segmentation of equiaxed grain boundaries

Accurate segmentation of microstructural images plays a critical role in quantitative materials analysis. However, it remains a challenging task due to the scarcity of annotated data and the inherent diversity of microstructures. While transfer learning is a promising solution to data scarcity, conventional approaches using models pre-trained on general-purpose datasets like ImageNet are often suboptimal due to their low similarity to microstructural images. To overcome this limitation, we propose a novel transfer learning approach utilizing synthetically generated Voronoi diagrams as a source dataset for equiaxed grain microstructure analysis. Unlike images produced by simulations or deep learning models, Voronoi diagrams can be created quickly with simple and intuitive parameters, while still resembling key structural features of equiaxed grain microstructures. This simplicity allows for efficient exploration of how generation conditions affect downstream model performance. Our results demonstrate that adding Gaussian noise to the Voronoi diagrams improves model performance, achieving a higher Intersection over Union (IoU) score than models pre-trained on ImageNet. Furthermore, the model learns microstructure-specific features such as triple junctions, enabling accurate grain boundary extraction even in low-contrast regions. This approach offers a simple and practical solution for pre-training task-specific models in materials science. • Synthetic Voronoi data mitigate the scarcity of annotated microstructure datasets. • Adding Gaussian noise improved grain boundary segmentation accuracy. • Voronoi-based pretraining outperforms generic ImageNet pretraining in IoU score. • The model learns microstructure-specific features, such as triple junctions. • This approach provides a practical solution for data-driven materials analysis.