Classification of non-metallic inclusions
Task definition
To determine the chemical composition of submicroscopic, non-metallic inclusions (NMI) in steel, analysis using a scanning electron microscope (SEM) with energy-dispersive X-ray spectroscopy (EDS) is widely used (resolution limit approx. 200-300 nm)
However, since information about the chemical composition can also be derived from the backscattered electron (BSE) detector image of the SEM or the particle shape, it was investigated to what extent machine learning (ML) can be used for purely image-based classification without relying on complex EDS measurements
Realization
- Data set: BSE images of the NMI, with associated ground truth is determined from the EDS spectrum
- Use of Deep Learning (DL): classification model based only on the BSE images
- In this application example, 5 classes are considered: Carbonitrides, oxides, sulphides, mixed oxide-sulphides as well as “non-inclusions” (e.g. pores, scratches, contaminations)
BSE recordings of the 5 classes
Results:
The DL model achieved an 84.9% accuracy in determining the chemical composition of the NMI using only the BSE image
Specifically, carbonitrides, sulphides and “non-inclusions” can reliably be classified
Fields of application
- Time saving by substituting EDS measurement with pure image evaluation
Cooperation partners
- Aktien-Gesellschaft der Dillinger Hüttenwerke, Dillingen/Saar
- Department of Materials Science and Engineering, Carnegie Mellon University, Pittsburgh
Contact for questions
Dr.-Ing. Dominik Britz
Deputy Head MECS Saarbrücken
Dr.-Ing. Tobias Fox
Chief Operating Officer
Have we sparked your interest in working together?
Please feel free to contact us! We look forward to talking to you and finding out together how we can help you with your project.