AI assisted steel cleanness evaluation: Predicting the morphology of La-traced non-metallic inclusions using backscattered-electron images

Abstract

This study shows how computer vision, a field of artificial intelligence (AI), can be used to gain access to morphological information for a better process understanding and furthermore, how it can be implemented in the evaluation of active tracing experiments. Non-metallic inclusions (NMIs) play a crucial role in determining the chemical and physical properties of the final steel product. The active tracing technique allows for tracking the formation and modification mechanisms of detrimental NMIs throughout the steelmaking process by adding rare earth elements (REEs) to the melt. The interaction of REEs with NMIs results in the formation of traced multiphase particles and, therefore, lead to challenges for accurate evaluation by automated scanning electron microscopy with energy-dispersive spectroscopy (SEM/EDS). However, obtaining information about the REE distribution in traced NMIs without extensive, time-consuming human effort is not feasible using the existing method. A two-step classification process with AI was established, which included training and testing of two random forest classifiers (RFC). The first RFC achieved for the binary NMI/non-NMI classification an accuracy of 94.8 %. Four different morphology-based classes were considered in the second classification, whereas traced homogeneous NMIs and NMIs with a high La content got classified with a precision exceeding 90 %. However, caution is required in interpreting results for traced heterogeneous NMIs, as they have a higher prediction error of 37.2 %.

Keywords

• Active tracing
• Rare earth elements
• Automated SEM/EDS analysis
• Image classification
• Non-metallic inclusions
• Steel cleanness