Hybrid machine learning and regression framework for automated phase classification and quantification in SEM images of commercial steels
Document Type
Article
Publication Title
Discover Materials
Abstract
This study presents an integrated framework combining supervised classification and composition-driven regression modeling for automated phase identification and quantification in steel microstructures. SEM micrographs of three commercially used steels EN3, EN353, and 20MnCr5 were acquired at magnifications of 5000×, 10,000×, and 20,000×. Images were segmented using the SLIC algorithm into 64 × 64 patches, from which six Gray Level Co-occurrence Matrix (GLCM) features were extracted: contrast, correlation, energy, homogeneity, dissimilarity, and angular second moment (ASM). The proposed framework provides a preliminary demonstration of interpretable classification and composition-linked regression modeling for phase prediction in steels, with future work required to validate its generalizability across broader steel systems. Using these features, a Random Forest classifier achieved 70% classification accuracy and a macro F1-score of 0.61 in identifying four phases: ferrite, pearlite, distorted pearlite, and bainite. Patch-wise predictions (972 in total) were aggregated to evaluate steel-specific phase trends. Distorted pearlite was predominant in EN3 and EN353, while bainite appeared mainly in 20MnCr5. A regression model was developed to predict global phase percentages from alloying elements (C, Mn, Cr, Ni) and magnification level, achieving strong agreement with machine learning predictions (R² = 0.88 for pearlite and 0.83 for distorted pearlite), moderate agreement for bainite (R² = 0.69), and weak agreement for ferrite (R² = 0.07). This hybrid framework exhibits potential for microstructural classification of texture-based classification and composition-informed modeling in capturing microstructural complexity. The approach lays groundwork for scalable microstructure analysis for steel evaluation and supports data-driven microstructure design and analysis.
DOI
10.1007/s43939-025-00323-6
Publication Date
8-1-2025
Recommended Citation
Hiremath, Pavan; Ambiger, Krishnamurthy D.; Suresh, Shilpa; and Ghadai, Ranjan Kumar, "Hybrid machine learning and regression framework for automated phase classification and quantification in SEM images of commercial steels" (2025). Open Access archive. 12819.
https://impressions.manipal.edu/open-access-archive/12819
