Scholarlink

<p>The automated SEM/EDS analysis is one of the most important methods for assessing the micro-cleanness of steels. Non-metallic inclusions are segmented from backscattered electron images based on their gray value and subsequently typified according to their elemental composition. However, the classification based on EDS data alone is not always sufficient. Artifacts such as surface contaminations (preparation residues, dust particles) and surface defects (cracks, pores) cannot be reliably distinguished from common non-metallic inclusions with this approach. The manual evaluation of the backscattered electron images is not practical due to the large amount of data.</p><p>In the present work, a&#xa0;multimodal machine learning model is introduced that combines the inclusion morphology from backscattered electron images with quantitative EDS composition to automatically detect and remove artifacts that have been misclassified as inclusions. The model was implemented as an extension of the established evaluation tool and trained on a&#xa0;dataset of 3849 manually verified features. On an independent test set, the model removed 94% of the misclassifications (artifacts classified as inclusions) that had previously been overlooked, while incorrectly removing only less than 2% of the actual inclusions. In combination with the conventional inclusion classification, a&#xa0;more precise analytical method for the characterization of steel cleanness was developed.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine Learning-assisted Characterization of Artifacts and Non-metallic Inclusions within Automated SEM/EDS Analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月1日

查看全文
  • Robert Musi,
  • Kathrin Thiele,
  • Susanne Katharina Michelic

摘要

The automated SEM/EDS analysis is one of the most important methods for assessing the micro-cleanness of steels. Non-metallic inclusions are segmented from backscattered electron images based on their gray value and subsequently typified according to their elemental composition. However, the classification based on EDS data alone is not always sufficient. Artifacts such as surface contaminations (preparation residues, dust particles) and surface defects (cracks, pores) cannot be reliably distinguished from common non-metallic inclusions with this approach. The manual evaluation of the backscattered electron images is not practical due to the large amount of data.

In the present work, a multimodal machine learning model is introduced that combines the inclusion morphology from backscattered electron images with quantitative EDS composition to automatically detect and remove artifacts that have been misclassified as inclusions. The model was implemented as an extension of the established evaluation tool and trained on a dataset of 3849 manually verified features. On an independent test set, the model removed 94% of the misclassifications (artifacts classified as inclusions) that had previously been overlooked, while incorrectly removing only less than 2% of the actual inclusions. In combination with the conventional inclusion classification, a more precise analytical method for the characterization of steel cleanness was developed.