Machine learning for complex microstructures - latent features as an alternative to physical features in steel structure-property relationships

Kurt Lejaeghere, Joshua Stuckner, Koenraad Theuwissen, Steven M. Arnold, Lode Duprez,
Machine learning for complex microstructures - latent features as an alternative to physical features in steel structure-property relationships,
Materials Characterization,
Volume 229, Part B,
2025,
115612,
ISSN 1044-5803,
https://doi.org/10.1016/j.matchar.2025.115612.
(https://www.sciencedirect.com/science/article/pii/S1044580325009015)
Abstract: As machine learning makes headway in materials science, the challenge of including microstructural information in process-structure-property models has gained increased relevance. It requires translating visual information into a feature set that is as compact, efficient, and relevant as possible. Physically inspired features are a popular option and recover much of the available domain knowledge. However, they rely on labour-intensive (semi-)manual measurements and/or the design of bespoke computer vision algorithms. In this work, we investigate whether the use of latent features obtained via pre-trained encoder-type neural networks can match the predictive performance and insight of engineered features derived from traditional methods. Both the choice of the encoder architecture and the pre-training dataset are considered, and advantages and disadvantages are weighed against traditional physical feature extraction. We show that pre-trained latent features work very well for classifying complex steel microstructures and predicting Brinell hardness up to 90 % accuracy and 0.96 R2 respectively. In many cases, excellent performance is obtained across various hyperparameter settings, reaching accuracies on par with or even exceeding the performance of physical features. The combination of ease of use and high performance is a powerful incentive for the adoption of image-based machine learning techniques in the materials industry.
Keywords: Microstructure; Martensite; Structure-property relationship; Artificial neural networks; Machine learning