Rare Defect Few - Sample Detection Solution for Semiconductor Advanced Packaging

The detection of rare defects with few samples in the semiconductor advanced packaging process has always been a difficult problem in the industry. WeLinkirt (DaoAI) provides an innovative solution to this problem with advanced technologies and products.

Industry background and user scenario: The semiconductor industry, as the core driving force of modern technology, its advanced packaging technology is crucial for improving chip performance and reliability. In the advanced packaging process, accurate detection of the tiny structures after chip packaging is a key link to ensure product quality. An advanced packaging production line of a leading semiconductor manufacturer mainly produces high-performance chips. The detection objects in the packaging process include tiny structures such as solder joints and pins after chip packaging, and the size of these structures is usually in the micron level. At the same time, there are also some rare defects, such as tiny foreign objects and cracks. The probability of these rare defects is extremely low, but once they flow into the subsequent processes, they will have a serious impact on the performance and reliability of the entire chip.

Pain points: Why is it difficult?

Traditional detection methods face many difficulties when dealing with the situation of rare defects with few samples. From the dimension of sample quantity, due to the extremely small number of rare defect samples, the model training is insufficient. Taking this leading semiconductor manufacturer as an example, the miss-detection rate under the traditional method is as high as 5%. This means that a large number of defective products flow into the subsequent processes, which not only increases the production cost but also brings serious quality risks. Analyzing from the dimension of misjudgment, the false-alarm rate remains high, reaching 15%. This makes a large number of good products misjudged as defective products, requiring manual re-inspection, which greatly increases the labor cost and reduces the production efficiency. Considering from the dimension of production line flexibility, when changing the production line model, the traditional method needs to spend a lot of time on reprogramming and debugging, and the model-changing time is as long as 2 hours, which seriously affects the flexibility and efficiency of production.

The root cause of these problems lies in the limitations of traditional detection methods. Traditional methods often rely on a large amount of sample data for model training, and the scarcity of rare defect samples makes the model unable to fully learn the characteristic information of defects. In addition, traditional methods lack the ability to effectively acquire and analyze three-dimensional information when dealing with complex tiny structures, resulting in insufficient detection ability for hidden defects.

Technical principle

WeLinkirt (DaoAI) uses an advanced visual basic model and APDT positive-sample/few-sample learning algorithm to solve this problem. The visual basic model has a powerful feature-recognition ability, and it can deeply mine and learn image features. Through the analysis of a small number of samples, the model can learn the characteristic information of defects, thereby improving the recognition ability of different types of defects. Compared with traditional methods, traditional models need a large number of samples to achieve a good recognition effect, while the visual basic model can perform well with a small number of samples.

The APDT positive-sample/few-sample learning algorithm can quickly and accurately identify defect patterns through the learning of a small number of positive samples. This algorithm uses a small number of positive samples to quickly train the model to meet the detection needs of rare defects. In terms of imaging, WeLinkirt's self-developed 3D camera combined with three-dimensional morphology reconstruction technology can obtain the three-dimensional information of the detection object. Most traditional detection methods can only obtain two-dimensional information, and it is difficult to find the defects hidden inside the tiny structures. The self-developed 3D camera and three-dimensional morphology reconstruction technology can clearly present the morphological characteristics of tiny structures, which is helpful for finding hidden defects and micron-level morphological changes. This multi-dimensional information acquisition and analysis method enables the algorithm to more accurately judge the existence of defects, greatly improving the detection accuracy.

Typical application scenarios

• Solder joint detection: In chip packaging, the quality of solder joints directly affects the electrical performance of the chip. During detection, WeLinkirt uses the self-developed 3D camera to obtain the three-dimensional information of solder joints, and analyzes the shape, size and position of solder joints through the visual basic model and APDT algorithm. The difficulty lies in the small size of solder joints and the possible hidden defects such as false soldering, which are difficult to detect accurately by traditional methods.
• Pin detection: The coplanarity and integrity of pins are crucial for the connection between the chip and the circuit board. WeLinkirt detects the coplanarity and micron-level morphological changes of pins through three-dimensional morphology reconstruction technology, and uses the algorithm to judge whether there are defects such as bending and breakage of pins. The difficulty lies in the large number of pins, and small deformations may affect the normal use of the chip.
• Tiny foreign object detection: Tiny foreign objects may affect the electrical performance and heat dissipation performance of the chip. WeLinkirt uses the visual basic model to analyze the image and identify the characteristics of tiny foreign objects. The difficulty lies in the extremely small size of foreign objects, and they may be similar in color to the background, making it easy for traditional methods to miss detection.
• Crack detection: Cracks after chip packaging will reduce the reliability of the chip. WeLinkirt obtains the three-dimensional information of the chip surface through the 3D camera and uses the algorithm to analyze the characteristics of cracks. The difficulty lies in the fact that the cracks may be very subtle, and traditional two-dimensional detection methods are difficult to find.

Implementation case

A leading semiconductor manufacturer has a large-scale advanced packaging production line and produces a large number of high-performance chips every day. Before introducing WeLinkirt's detection solution, it faced problems such as high miss-detection rate, high false-alarm rate and long model-changing time. During the implementation process, first, a small number of good product samples are collected and analyzed, and the DaoAI AI AOI software system is used for rapid programming and model training. Then, the trained model is deployed to the DaoAI 2D / 3D AI AOI equipment for real-time detection of products on the production line. Before implementation, the detection rate was 95%, the miss-detection rate was 5%, the false-alarm rate was 15%, and the model-changing time was 2 hours. After implementation, the detection rate increased to 99.2%, the miss-detection rate decreased to <0.8%, the false-alarm rate decreased to 3%, a reduction of -80%, and the model-changing time was shortened from 2 hours to 5 minutes.

WeLinkirt's detection solution has brought significant improvements to the semiconductor advanced packaging production line and effectively solved the problem of rare defect few-sample detection.

WeLinkirt's solution and products

WeLinkirt provides the DaoAI AI AOI software system and DaoAI 2D / 3D AI AOI equipment. The DaoAI AI AOI software system has a powerful positive-sample/few-sample learning ability. Only 1-20 good product samples are needed, and 0-code automatic programming can be completed within 5 minutes. At the same time, the system also has a semantic false-alarm filtering function, which can effectively reduce the false-alarm rate. The DaoAI 2D / 3D AI AOI equipment uses the self-developed 3D camera and three-dimensional morphology reconstruction technology, which can detect hidden solder joints, coplanarity and micron-level morphological changes. In practical applications, the software system is first used to analyze a small number of good product samples and train the model, and then the trained model is deployed to the equipment to realize real-time detection of products on the production line.

Quantitative results: After adopting WeLinkirt's solution, the detection rate increased from the original 95% to 99.2%, and the miss-detection rate decreased to <0.8%, greatly reducing the risk of defective products flowing into the subsequent processes. The false-alarm rate decreased from 15% to 3%, a reduction of -80%, reducing the workload of manual re-inspection. The model-changing time of the production line was shortened from 2 hours to 5 minutes, improving the flexibility and efficiency of production.

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.