Abstract Accurate prediction of material properties is essential in steel manufacturing, as it enables optimized production processes. The scarcity of microscopy images in steel microstructure analysis has led to the adoption of pre-trained deep learning models for property prediction on smaller datasets. This study investigates prediction capabilities of a fine-tuned VisionTransformer for five mechanical properties on the microstructures of 47 steel specimen over six magnifications. While the model achieves competitive predictions for upper yield strength, yield strength, ultimate tensile strength and elongation at fracture, while the Charpy V-notch impact energy is predicted less accurate. Lower magnifications tend to improve the prediction of all properties. Our analysis of different popular data splitting strategies indicates that improper splitting of micrographs can lead to data leakage, resulting in an overestimation of model performance by up to threefold. Our findings underscore the need for reliable evaluation procedures and highlight the critical challenges of applying deep learning to steel microstructure analysis.
Wentzien et al. (Sat,) studied this question.