Tracing by means of the light rare earths (REs), particularly La and Ce, is a state‐of‐the‐art method used to track deoxidation products during the steelmaking process. Traced heterogeneous multiphase inclusions are analyzed using scanning electron microscopy with energy‐dispersive spectroscopy (SEM/EDS) to perform a 2D characterization. The sequential chemical extraction technique is implemented for a 3D investigation to determine traced particles’ actual sizes and shapes. The automated SEM/EDS measurement must be optimized since RE oxides appear brighter in the backscattered electron images due to their high atomic numbers. Therefore, two grayscales are implemented for the detection of RE‐containing multiphase inclusions. Within this technique, individual RE‐traced heterogeneous nonmetallic inclusions (NMIs) are counted as separate particles. Thus, the measured NMIs must be recombined, which is achieved using a self‐developed MATLAB tool. The extracted particles are also analyzed by automated and manual SEM/EDS measurements to determine the 3D morphologies and sizes of traced NMIs.

Nonmetallic inclusions have strong influence on final steel properties. An important characterization tool to make a comprehensive analysis of nonmetallic inclusions is the scanning electron microscope equipped with energy‐dispersive spectroscopy (SEM‐EDS). A major drawback which prevents its use for online‐steel assessment is the time taken for analysis. Machine learning methods have been previously introduced which circumvents the usage of the EDS for obtaining chemical information of the inclusion by classifying inclusion based on their back scatter electron images. This study introduces a method based on a simpler tabular data input consisting of morphological and mean gray values of inclusions. Naive Bayes and Support Vector Machine classifier models are built using the R statistical programming language. Two steel grades are considered for this study. The prediction results are shown to be satisfactory for both binary (maximum 89%) and 8‐inclusion class (maximum 61%) categorization. The input dataset is further improved by optimizing the image settings to distinguish the different types of nonmetallic inclusions. It is shown that this improvement results in a higher rate of correct predictions for both binary (maximum 98%) and 8‐class categorization (maximum 81%).