Benchmarking geometric lamellar orientation: A large-scale synthetic dataset for quantification of ferrite-pearlite steels

Quantitative metallography of ferrite-pearlite steels is essential for establishing structure-property correlations, yet manual characterization is labour-intensive and prone to bias. This article presents a large-scale synthetic dataset designed to train and benchmark deep learning models for the automated segmentation of pearlite colonies and ferrite grains. The dataset was generated using a computational pipeline that superimposes experimentally obtained ferrite and pearlite morphological textures onto simulated polycrystalline templates generated via nucleation and growth phenomena. The primary parameter investigated was the geometric lamellar orientation of pearlite colonies, which was categorized into 10 distinct classes (20° angular bins and a background ferrite class) relative to the image frame. The resulting dataset comprises 10,499 synthetic micrographs (512 × 512 pixels) paired with pixel-perfect ground truth segmentation masks. This data provides a robust resource for developing computer vision algorithms capable of discerning pearlite colonies based on the geometric orientation of their lamellae, thereby facilitating high-throughput quantitative analysis in materials science.