An efficient training methodology of a deep learning model for accurate stereological quantification of the ferrite–pearlite microstructure in steel

This study introduces an automated deep learning-based methodology for precise and comprehensive characterisation of ferrite-pearlite microstructures. By leveraging a UNET-based Convolutional Neural Network (CNN) architecture, which has been named by as PearlAI model, we segment individual pearlite colonies based on differences in lamellar orientation. A synthetic dataset, composed of pearlite micrographs with systematically varied geometric orientations and corresponding ground truth labels, was generated to train the model. Once trained, PearlAI accurately isolates pearlite colonies in real micrographs, enabling automated quantification of random, apparent and true interlamellar spacing and also colony size using established stereological methods. The apparent spacing, which is the perpendicular distance between alternate lamellae, is generally more difficult to estimate because of the fact that lamellae are neither perfectly parallel nor perfectly straight. For this an automated methodology was developed to obtain the best estimate of the apparent spacing. Further, an automated technique was developed to assess extent of isotropy or anisotropy of the microstructure vis-à-vis the lamellae orientation. The proposed framework not only enhances the reliability but also the accuracy of microstructural quantification, which in turn will assist in obtaining more meaningful correlation with properties and performance of engineering materials.