跳至内容

半导体智能制造:无监督学习

Diminish the effort of labelling data

Unsupervised segmentation in semiconductor imaging has become an essential tool in modern manufacturing, enabling accurate defect detection and quality control without the need for labeled training data. 

This article explores the principles, techniques, and recent advancements in unsupervised segmentation, showcasing its application in semiconductor manufacturing with illustrative examples and references to current research.

Designed by Freepik

Challenges in the Semiconductor Industry

The semiconductor industry struggles with identifying and managing a wide range of defects, including scratches, particles, cracks, thermal damages and contamination on wafers and devices. The high precision requirements of manufacturing processes make defect detection challenging, and traditional methods often fail to anticipate rare or unexpected anomalies which can have negative impacts in production process. 

This necessitates advanced detection techniques to ensure the reliability and functionality of electronic components.

Defect Analysis

Defect Type

Description

Impact

Particles

Particles are small foreign objects that can contaminate semiconductor wafers during manufacturing processes. These particles might come from handling, processing equipment, or environmental sources.

Even tiny particles can disrupt circuitry, cause electrical shorts, or affect the reliability of semiconductor devices.

Scratches

 

Scratches are visible marks or grooves on the surface of a semiconductor wafer or device. They can occur during handling or processing steps.

Scratches may compromise the structural integrity of the semiconductor, potentially leading to mechanical failures or electrical issues. 

Cracks

Cracks are structural defects characterized by fissures or fractures in the semiconductor material. These can result from thermal stresses, mechanical shocks, or fabrication errors.

Cracks weaken the semiconductor's mechanical strength and can propagate under stress, leading to device failure or reduced lifespan.

Etching Issues

Etching defects include irregularities or inconsistencies in the etching process used to pattern semiconductor surfaces. This can result in incomplete or improper removal of material.

 Etching defects can alter circuit patterns, affect device performance, or lead to manufacturing yield losses.

Contamination

Contamination involves the presence of unwanted substances, such as chemical residues or particulates, on semiconductor surfaces or within materials.

Contaminants can change semiconductor properties, degrade device performance, or cause reliability issues over time.

 

Alignment Errors

 

Alignment errors occur when layers or features in semiconductor fabrication are misaligned during photolithography or other patterning processes.

Misalignment can lead to non-functional circuit elements, reduced device yield, or improper device operation.

 

To further illustrate these defect types, the following images provide visual examples of the defects described in the table. These images help to visualize the characteristics of each defect and their potential impact on semiconductor wafers and devices. By examining these examples, one can better understand the nature of these defects and the importance of effective detection and mitigation strategies in semiconductor manufacturing.

Unsupervised Segmentation Techniques for Semiconductor Manufacturing

Having reviewed various defect types and their impacts, it is essential to explore how unsupervised segmentation techniques can aid in detecting these defects. Unsupervised segmentation involves dividing images or datasets into meaningful segments without relying on labeled training data, making it a valuable tool in semiconductor manufacturing. These techniques are instrumental in defect detection, process monitoring, and quality control within semiconductor fabrication. Following table provides the state-of-the-art approaches for unsupervised segmentation.

Standard Methods:

 

1.K-means Clustering

2.Autoencoders

3.WaferSegClassNet

4.Principal Component Analysis (PCA)

5.CFA: Coupled-hypersphere-based Feature Adaptation for Target-Oriented Anomaly Localization

6.U2Seg: Unsupervised Universal Image Segmentation

7.Spectral Clustering

8.FRE: A Fast Method For Anomaly Detection And Segmentation

9.SimSAM: Simple Siamese Representations Based Semantic Affinity Matrix

10.Fuzzy C-Means Clustering

Demonstration Case

To tackle the challenges of defect detection in semiconductor images, we employ a systematic approach that includes two main phases:

Segment-wise Feature Extraction: In the first phase, we focus on extracting features from individual segments of the semiconductor images. This process involves dividing the image into smaller, manageable segments and analyzing each segment to capture relevant characteristics, such as cracks, particles, and thermal damage patterns. These features are crucial for differentiating between various regions and identifying potential defects. By isolating and examining each segment separately, we enhance our ability to recognize subtle anomalies that may be indicative of defects.

Unsupervised Clustering and Labeling: The second phase involves applying unsupervised clustering techniques to the features extracted from each segment. We utilize methods such as K-means clustering, Gaussian Mixture Models (GMM), and Mean Shift clustering to group segments based on their feature similarities. This clustering process helps in categorizing the segments into different clusters or labels without requiring pre-labeled data. The outcome is a set of clusters that represent different defect types or regions of interest within the semiconductor images.

The above image demonstrates our segmentation approach, showcasing the segmentation results from both the feature extraction and clustering phases. It provides a visual representation of how segments are analyzed and categorized, highlighting the effectiveness of our method in identifying and labeling defects in semiconductor manufacturing.

 Conclusion

In the demonstration above, defect detection in semiconductor images is efficiently managed through segment-wise feature extraction and unsupervised clustering without human efforts. This approach enhances real-time defect detection and process monitoring while seamlessly integrating with existing systems. As the technique advances, it promises to elevate semiconductor manufacturing standards, driving innovation and efficiency.

纺织品瑕疵检测