Deep Learning Based Visual Recognition for Inline Defects in Production of Semiconductors
Semiconductor manufacturing is a complex process that involves different kind of mechanical, chemical and photo lithography treatments such as deposition, etching and impurity implantation. Quality inspection is vital during such process, since defects and anomalies introduced during the production...
Gespeichert in:
Veröffentlicht in: | IEEE journal of emerging and selected topics in industrial electronics (Print) 2024-01, Vol.5 (1), p.1-9 |
---|---|
Hauptverfasser: | , , , |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext bestellen |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | Semiconductor manufacturing is a complex process that involves different kind of mechanical, chemical and photo lithography treatments such as deposition, etching and impurity implantation. Quality inspection is vital during such process, since defects and anomalies introduced during the production process can lead to a variety of failures in later stages. The task of manually or automatically detecting and classifying defects is challenging due to the large variation in design, material, layout, colors and visual pattern. Existing automated optical inspection methods do not offer flexibility and scalability towards large numbers of defects or new designs. Therefore, visual inspection based on deep learning is considered as a long-term alternative for industry level quality assessment. In this paper, automated visual defect identification in patterned wafers during the production process is proposed. By employing transfer learning, we trained and evaluated a number of convolutions neural network architectures including VGG, ResNet and DenseNet, over a range of parameters for a binary and multiclass classification and determined optimal architecture. The experimental evaluation is carried out on real industrial data of patterned wafers comprising of 9 different defect types and from 14 production steps. The results show that the network based on DenseNet201 achieves a highest performance up to 98.17% on binary classification and 96.04% on multiclass classification. |
---|---|
ISSN: | 2687-9735 2687-9743 |
DOI: | 10.1109/JESTIE.2023.3326092 |