A Systematically Designed Open-Access Dataset for Cross-Microscope Machine Learning Benchmarking in Complex Steel Microstructure Classification

Machine learning (ML) for microstructure analysis has seen rapid progress, yet a fundamental challenge remains largely unaddressed: models trained on one imaging configuration frequently fail to generalize to others, a problem known as domain shift. This issue is particularly acute for complex steel microstructures — specifically bainite and martensite — whose visual appearance is highly sensitive to microscopy modality, magnification, and acquisition settings, and for which no standardized nomenclature or ground truth assignment procedure exists. This dataset was designed to systematically capture and enable the study of these imaging-induced variances. Its defining characteristic is the correlative, pixel-wise registration of identical regions of interest (ROI) imaged across three distinct light optical microscopes — an Olympus LEXT confocal laser scanning microscope, a Leica DM6000, and a Zeiss Axio Imager M2 — at multiple magnifications (20x, 50x, and 100x for the LSM) and under deliberately varied acquisition conditions, including aperture, exposure, and optical filter settings. Since registration to a common reference is performed prior to annotation, ground truth labels apply consistently across all imaging variants of each ROI, ensuring label coherence and eliminating the annotation bottleneck typically associated with multi-condition datasets. The dataset consists of 5,327 annotated image patches extracted from 20 steel samples representing a broad range of quenched and quenched-and-tempered microstructures, covering typical bainitic and martensititic constituents - simplified to two classes. Patches are provided at standardized sizes (256 px at 50x, 96 px at 20x), ready for direct use in standard convolutional neural network (CNN) architectures without further resizing. All micrographs are accompanied by harmonized metadata encoding microscope identity, magnification, pixel size, aperture, exposure level, and filter configuration, assigned via unique identifiers that allow full traceability of each patch to its imaging origin. In the accompanying publication, the dataset was used to systematically evaluate the influence of microscopy modality and magnification on CNN-based bainite/martensite classification, revealing a resolution-dependent asymmetry in cross-domain transfer, systematic class bias inversions under cross-scale validation, and a hierarchy of imaging variance types in their contribution to model robustness. Beyond this initial analysis, the dataset is well suited for cross-microscope and cross-magnification domain adaptation studies, generative image-to-image translation and super-resolution (enabled by the spatially aligned multi-condition image pairs), unsupervised representation and embedding analyses, the development and validation of domain coverage metrics, the study of metadata-informed or physics-aware ML approaches and much more. The dataset complies with FAIR data principles and is intended as a community benchmark for reproducible and transferable ML workflows in microstructure-based materials science.