A systematically designed open-access dataset for cross-microscope machine learning benchmarking in complex steel microstructure classification

Abstract

Abstract The establishment of robust machine learning workflows for microstructure analysis requires datasets that systematically capture the variance introduced by the imaging pipeline. While existing studies in this field predominantly rely on well-curated, homogeneous datasets optimized for high classification performance, these conditions rarely reflect the variance encountered in real-world settings. This work presents a publicly available, FAIR-compliant dataset of 5327 annotated micrograph patches, showing bainitic and martensitic microstructural constituents in quenched and quenched-and-tempered low-carbon steels, acquired on three different optical microscopes across two magnification levels (20x and 50x) and under deliberately varied acquisition conditions. Identical regions of interest were imaged on all platforms and aligned by image registration, enabling optimal comparability by pixel-wise alignment across instruments and magnifications. This correlative structure makes the dataset uniquely suited for systematic investigations of cross-instrument and cross-magnification domain generalization. Using this dataset, the influence of microscopy modality-induced domain shifts on the classification performance of deep convolutional neural networks was analyzed. A general model trained on the full dataset achieved near-perfect classification (F1 score of 99.91%), demonstrating that modern deep learning is capable of handling the full range of imaging-induced variance when sufficient data coverage is provided. Cross-domain experiments revealed that generalization is governed primarily by the physical fidelity of the input representation and the coverage of imaging-induced variance in the training data, rather than by model architecture or nominal magnification alone. The direction in which resolution gaps are bridged between training and validation domains proved more decisive than information density per se. Among single-instrument configurations, broad variation of acquisition conditions, such as aperture, exposure, and color temperature, contributed more effectively to cross-domain robustness than optical diversity between objective lenses alone. These results suggest that targeted variation of acquisition conditions, rather than purely morphological diversity, is a particularly effective strategy for improving model generalization. The dataset is released to the community as a benchmark resource for domain shift analysis, generative modeling, and domain-aware learning in microstructure-based machine learning.