Advancing Analysis of Non-Metallic Inclusion Datasets

Non-metallic inclusions play a major role in defining the properties of steel and their examination has been greatly aided by automated scanning electron microscopes (SEM) coupled with energy dispersive x-ray spectroscopy (EDS). The output of SEM/EDS analysis generates a great deal of data that has not been utilized. The current work exploits this data using several data analytics tools for inclusion analysis tasks. Three inclusion datasets were compiled from different sources for the analyses carried out in this study.

The first part of this study addresses the problem of physical inclusion clusters, or agglomerates. An automated inclusion analysis tool was developed that can identify inclusion clusters from the spatial positioning data generated from SEM analysis. This was enabled using the machine learning clustering algorithm DBSCAN, which was developed for spatial clustering applications. The analysis was conducted on samples known by manual analysis to be clustered and non-clustered to evaluate the applicability of this technique. In addition, several cross sections were analyzed from each sample, to obtain a 3D representation of the inclusion distribution. The 2D and 3D results were consistent.

The technique proved to be a useful tool that can automatically detect inclusion clusters. It was then employed to study the evolution and types of clusters formed during the secondary metallurgy process, on samples from one the available datasets. Clustering of globular alumina inclusions was observed early in the ladle Further in the ladle, alumina clusters were detected as elongated non-globular inclusions. Spinel inclusions displayed a similar trend in cluster morphology, but to a lower extent. There was a significant number of clusters associated with the agglomeration of small micro inclusions around one large macro inclusion. The other aspect of the study focused on investigating the relationship between backscattered electron (BSE) yield and x-ray signals from SEM/EDS analysis. The complex interactions of an electron beam on an inclusion embedded in a steel matrix were simulated using PENEPMA, a Monte Carlo program for electron probe microanalysis simulations. Simulations were compared to inclusions filtered from the available datasets. Alumina (Al2O3), spinel (MgAl2O4), and CaS, C12A7 (12CaO.7Al2O3), and MnOSiO2 inclusions were simulated to examine the effect of inclusion size, embedded depth, and beam position. Comparisons between measured and simulated data were in agreement with inclusions in dataset B, while simulations suggested that inclusions in datasets A and C had on averaged higher BSE yield. This difference was attributed to the assumption of inclusion sphericity with the simulations. The average circularity of inclusions in dataset B were 1.08, 1.03, and 0.99, indicating spherical morphologies. Whereas the average circularity was 0.61 or less for inclusions in datasets A and C, implying large deviations from the assumption of sphericity.

More detailed analysis on the effect of accelerating voltage was carried out. This was achieved by analyzing the same scan area on sample rich with calcium aluminates at 10, 15, and 20kV accelerating voltages. BSE images were sharper at higher accelerating voltages. This in turn lead to more precise measurement of inclusion size at higher voltages, as a result the average inclusion size decreased at higher voltages. The Ca to Al ratio at 15kV was observed to be higher than at 10 or 20kV, for small inclusions ( < 4μm).

Furthermore, prediction models were generated using supervised machine learning to relate BSE image greyscale values (GSVs) to inclusion compositions. Initially, a univariate model was sued to predict this value from the inclusion’s chemical compositions and used as a quality control tool for inclusion SEM analysis. Other models were employed to predict inclusion types from BSE images, which was the fundamental aim of this part of the work. Several models were assessed and compared. The recursive model, which employs four sub-model, achieved promising results for the classification of 6 classes, with a testing accuracy of 74%. Preliminary investigations were also conducted to predict inclusion compositions from their BSE images. The results for the multivariate models were similar, the RMSEs were 18%, 13%, 8%, and 7%, for Al, Ca, Mn, and S, respectively. The models provided an estimate of the compositional distribution, which was more confined than the measured distribution.