AI AOI Software System Solves Rare Defect Detection in Semiconductor Advanced Packaging

As the core driving force of technological development, the semiconductor industry attaches great importance to the quality inspection of advanced packaging processes. However, the problem of rare defect detection with few samples has long troubled enterprises. The emergence of WeLinkirt's AI AOI software system has brought a turning point to this dilemma.

The semiconductor industry is the cornerstone of modern technology, and its advanced packaging process plays a decisive role in the performance and reliability of chips. With the continuous development of semiconductor technology, the integration of chips is getting higher and higher, and the packaging process is becoming more and more complex, which poses higher requirements for the appearance defect detection of packaged chips. In the advanced packaging production line of a leading semiconductor manufacturer, after the chip packaging process is completed, a comprehensive appearance defect detection of the packaged chips is required. The detection objects include rare defects such as scratches, cracks, and foreign objects on the surface of the chip package. Due to the extremely low probability of these defects occurring, the number of samples is very limited, which brings great challenges to the detection work.

Pain Points: Why Is It Difficult?

Current industrial vision defect detection methods have serious deficiencies in end - to - end integration operations, leading to various quantification dilemmas. In terms of the false negative rate, traditional methods have a relatively high false negative rate, and rare defects are difficult to be accurately identified. This is because the number of rare defect samples is limited, and the detection model is difficult to learn the characteristics of these defects. As a result, these defects are easily ignored in actual detection, which affects product quality. Statistics show that the false negative rate of traditional detection methods may be as high as over 10%, which means that a large number of defective chips may flow into the market, bringing huge losses to enterprises.

Frequent false alarms are also a major pain point of traditional detection methods. Due to the limitations of detection algorithms, environmental interference and other factors can easily lead to false alarms. For example, changes in lighting and tiny stains on the chip surface may be misjudged as defects, increasing the workload and cost of manual re-inspection. The false alarm rate of traditional methods may be as high as over 30%, which requires enterprises to invest a large amount of manpower and time in re-inspection, reducing production efficiency.

Traditional detection methods have a long model change time and cannot quickly adapt to the detection requirements of different types of chips. When changing the chip model on the production line, traditional methods need to be reprogrammed and debugged complexly, which may take several hours or even days, seriously affecting the production efficiency of the production line. In addition, in terms of data compliance, traditional detection methods also have certain risks and are difficult to meet the requirements of enterprises for efficient and accurate detection.

Technical Principle

WeLinkirt's AI AOI software system uses a feature recognition algorithm based on a visual foundation model, which can deeply analyze the appearance features of chips. By learning the features of good products, the system can establish accurate defect judgment criteria. Its core APDT positive-sample/few-sample learning mechanism is particularly crucial. Only 1-20 good products are needed to quickly learn the normal feature pattern of the chip. Different from traditional methods that require a large number of defect samples for training, this few-sample learning mechanism has obvious advantages in rare defect scenarios.

When facing rare defects, the system can keenly capture the differences from the normal pattern based on the learned normal feature pattern, thus accurately identifying defects. The semantic false-alarm filtering function analyzes the semantics of defects and eliminates false alarms caused by environmental interference and other factors. This function is based on natural language processing technology, which understands and judges defect information and effectively reduces the false alarm rate. The visual foundation model can automatically extract multi-dimensional features of the chip appearance, which is not affected by complex backgrounds and lighting conditions, enhancing the robustness of the system.

Typical Application Scenarios

• Scratch detection: Although the scratches on the surface of the chip package are small, they may affect the performance and reliability of the chip. WeLinkirt's AI AOI software system extracts the features of scratches through the visual foundation model and compares them with the features of the normal chip surface to accurately identify scratches. The difficulty lies in that the scratches may be very subtle and easy to be ignored, and the features of different types of scratches vary greatly.
• Crack detection: Cracks are relatively serious defects in chip packaging and may cause problems such as short-circuits in the chip. The system identifies the existence of cracks by analyzing the texture and structure of the chip surface. The difficulty lies in that the shapes and directions of cracks are complex and diverse, and they may be hidden in certain parts of the chip and are not easy to be found.
• Foreign object detection: Foreign objects on the surface of the chip package may affect the heat dissipation and electrical performance of the chip. The system identifies the position and size of foreign objects by analyzing the image of the chip surface. The difficulty lies in that there are a wide variety of foreign objects, with different colors and shapes, and they may be similar to the texture of the chip surface, increasing the difficulty of detection.
• Hidden solder joint detection: The quality of hidden solder joints is directly related to the electrical connection performance of the chip. The system combines DaoAI 2D / 3D AI AOI equipment to perform micron-level morphology detection on hidden solder joints. The difficulty lies in that hidden solder joints are located inside the chip and cannot be directly observed, and special imaging technologies and algorithms are required for detection.
• Coplanarity detection: Coplanarity refers to the flatness of the chip package surface, which has an important impact on the installation and performance of the chip. The system measures the height difference of the chip package surface through 3D imaging technology to determine whether the coplanarity meets the requirements. The difficulty lies in that the measurement accuracy of coplanarity requires high-precision imaging equipment and algorithm support.

Implementation Case

A large semiconductor manufacturing enterprise has multiple advanced packaging production lines and produces a large number of chip products every day. The enterprise faces the problem of rare defect detection in the chip packaging process. The false negative rate and false alarm rate of traditional detection methods are relatively high, and the model change time is long, which seriously affects production efficiency and product quality. After introducing WeLinkirt's AI AOI software system, after a period of online debugging and optimization, the system gradually runs stably.

WeLinkirt's AI AOI software system has brought new breakthroughs in rare defect detection for semiconductor advanced packaging, significantly improving detection efficiency and accuracy.

Before the implementation, the false negative rate of chip defects in the enterprise was about 12%, the false alarm rate was as high as 35%, and the model change time required 3-4 hours. After the implementation, the chip defect detection rate reached 98.5%, effectively reducing the risk of false negatives; the false alarm rate was reduced by -75%, greatly reducing the workload of manual re-inspection; the model change time was shortened from several hours to 5 minutes, significantly improving the production efficiency of the production line.

WeLinkirt's Solution and Product

WeLinkirt takes the AI AOI software system as the core. The system has the feature recognition ability of a visual foundation model and can realize 0-code automatic programming for a good product in 5 minutes, greatly shortening the model change time. Through APDT positive-sample/few-sample learning, only a small number of good products are needed to complete model training. The semantic false-alarm filtering function further improves the detection accuracy. The system supports 100% local private deployment of SDK/API/Docker to ensure that data does not leave the factory, meeting the data compliance requirements of enterprises. At the same time, it can be combined with DaoAI 2D / 3D AI AOI equipment to perform more comprehensive detection on chips, such as micron-level morphology detection of hidden solder joints and coplanarity.

Quantitative results: After using the AI AOI software system, the chip defect detection rate reaches 98.5%, effectively reducing the risk of false negatives; the false alarm rate is reduced by -75%, greatly reducing the workload of manual re-inspection; the model change time is shortened from several hours to 5 minutes, significantly improving the production efficiency of the production line.

This article was generated by AI. Customer cases are simulated scenarios based on real product capabilities and figures are illustrative; see product pages for official benchmarks.