Abstract Machine learning-driven automated replication micrographs analysis makes possible rapid and unbiased damage assessment of in-service steel components. Although micrographs captured by scanning electron microscopy (SEM) have been analyzed at depth using machine learning, there is no literature available on the technique being attempted on optical replication micrographs. This paper presents a machine-learning approach to segment and quantify carbide precipitates in thermally exposed HP40-Nb stainless-steel microstructures from batches of low-resolution optical images obtained by replication metallography. A dataset of nine micrographs was used to develop a random forest classification model to segment precipitates within the matrix (intragranular) and at grain boundaries (intergranular). The micrographs were preprocessed using background subtraction, denoising, and sharpening to improve quality. The method achieves high segmentation accuracy (91% intergranular, 97% intragranular) compared to human expert classification. Furthermore, segmented micrographs were quantified to obtain carbide size, shape, and density distribution. The correlations in the quantified data aligned with expected carbide evolution mechanisms. Results from this study are promising but necessitate validation of the method on a larger dataset representative of evolution of thermal degradation in steel, given that characterization of the evolution of microstructure components, such as precipitates, applies to broad applications across diverse alloy systems, particularly in extreme service.

Toward automated microstructure characterization of stainless steels through machine learning-based analysis of replication micrographs