Automatic Defect Classification after Wafer AOI: From Manual Rejudgment to AI-ADC

In the complex process of wafer manufacturing, defect detection and classification are crucial steps to ensure product quality and production efficiency. Traditional automated optical inspection (AOI) technology can detect suspected defects, but the subsequent work of authenticity judgment and classification faces many challenges. WeLinkirt's AI-ADC automatic defect classification module provides an innovative solution to these problems.

Industry Background and User Scenarios: In the semiconductor wafer manufacturing industry, as the chip process continues to shrink, the requirement for the detection accuracy of wafer surface defects is getting higher and higher. Automated optical inspection (AOI) equipment, as an important tool for the initial inspection of wafer defects, is widely used in the front-end processes. Quality inspectors in wafer fabs need to further re-judge and classify suspected defects based on the results output by AOI equipment. However, traditional AOI equipment can only mark suspected defect points and cannot accurately distinguish between real defects and false alarms, which makes a large amount of subsequent work rely on manual labor. In the context of large-scale production, quality inspectors face huge work pressure. How to improve the accuracy and efficiency of defect classification has become an urgent problem to be solved for wafer fabs.

In - depth Analysis of Pain Points: Why is it Difficult?

From a quantitative perspective, traditional AOI equipment has many problems. Firstly, the over-kill rate is high. Long - term operation data of a wafer fab shows that its AOI over-kill rate remains above 20%. This means that one in every five alarms is a false positive, and a large number of qualified wafers are returned for re-inspection. For example, in a one-month production cycle, the additional re-inspection man-hours caused by the over-kill problem reach hundreds of hours, seriously consuming manpower and time costs. Secondly, the ratio of manual re-judgment to machine is too high. The front-end AOI equipment in this wafer fab produces a large number of suspected defect points per wafer, and all these points need to be manually re-judged and classified by quality inspectors on microscopic images one by one. Even with three shifts, it is still difficult to cover the production peak, and the ratio of manual re-judgment to machine in the re-judgment link is seriously unbalanced. Finally, the production line rhythm is affected. The backlog of re-judgment directly affects the shipment node, resulting in a reduction of the overall line rhythm, and about 30% of the production capacity is wasted.

The root cause of these difficult - to - solve problems lies in the limitations of traditional AOI technology itself. Its imaging and algorithms can only perform suspected defect detection based on simple pre-set rules and cannot effectively distinguish between real defects and false alarms such as scratch reflections and particle residues. Moreover, with the expansion of production scale and the increase of product complexity, defect types have become more diverse. The manual annotation and classification method is difficult to keep up with the changing speed, resulting in more and more onerous re-judgment work and also affecting the data accuracy.

Technical Principle

WeLinkirt's AI-ADC automatic defect classification module uses advanced algorithms and imaging technologies. In terms of algorithms, for categories with unbalanced historical samples, APDT positive sample learning is used in conjunction with the DaoAI World model for generalization. APDT positive sample learning can focus on low-frequency defects and learn and train through a small amount of positive sample data, so as to quickly identify these low-frequency defects. The DaoAI World model has strong generalization ability and can adapt to different types of defects and complex production environments. In terms of imaging, the module can analyze the images of suspected points output by AOI with microscopic/micron-level accuracy to accurately distinguish between real and false defects.

Compared with traditional methods, AI-ADC has significant advantages. Traditional methods mainly rely on manual experience and simple rules for judgment, which are easily affected by subjective factors and have low efficiency. AI-ADC can quickly and accurately filter false alarms through automated secondary judgment and classification, improving the accuracy and efficiency of defect classification. At the same time, it can also feed back the classification results to the process end, helping process personnel better locate real abnormal sources and optimize the production process.

Typical Application Scenarios

• Particle Detection: Particle contamination is a common problem in the wafer manufacturing process. AI-ADC can perform a secondary judgment on the suspected particle points detected by AOI, and accurately distinguish real particles from false alarms by analyzing the size, shape, and position of particles. The difficulty lies in that particles may be very small, requiring high-precision imaging and algorithms for identification.
• Scratch Detection: Scratches may affect the performance and yield of wafers. AI-ADC can detect different types of scratches, such as straight-line scratches and curved scratches, and classify them. The difficulty lies in that the shapes and degrees of scratches vary, requiring the algorithm to have strong adaptability.
• Pattern Defect Detection: With the increasing complexity of chip design, the detection of pattern defects has become more and more difficult. AI-ADC can analyze the integrity and symmetry of patterns to detect and classify pattern defects. The difficulty lies in the diversity and complexity of patterns, requiring the model to have high generalization ability.
• Hole Detection: Holes on the wafer surface may cause problems such as short-circuits in circuits. AI-ADC can detect information such as the size and depth of holes and classify them. The difficulty lies in that the features of holes may not be obvious, requiring a delicate algorithm for identification.

Implementation Case

Before introducing WeLinkirt's AI-ADC solution, a medium-sized wafer fab faced serious problems in defect classification. The front-end AOI equipment in this factory produced a large number of suspected defect points per wafer, the over-kill rate remained above 20% for a long time, the ratio of manual re-judgment to machine in the re-judgment link was too high, and the production line rhythm was seriously affected. During the implementation process, the WeLinkirt team first conducted a detailed analysis of the factory's production data and defect types, and then connected the AI-ADC automatic defect classification module to the AOI re-judgment workstation. After a period of debugging and optimization, the module gradually adapted to the factory's production environment.

AI-ADC transforms defect classification from cumbersome manual operations to intelligent automation, bringing significant efficiency improvements and cost reductions to wafer fabs.

After the implementation, compared with before, all indicators have been significantly improved. The over-kill rate has dropped from above 20% to below 2%, greatly reducing the number of re-inspections of qualified wafers and saving a large amount of man-hours and costs. The ratio of manual re-judgment to machine in the re-judgment link has decreased by about 90%, greatly reducing the workload of quality inspectors and enabling them to devote more energy to the handling of abnormal situations. The overall line rhythm has increased by about 30%, accelerating the product production cycle and improving the enterprise's market competitiveness. At the same time, the defect classification data has become available again, providing reliable data support for process improvement and quality control.

WeLinkirt's Solution and Product

The core of WeLinkirt's AI-ADC solution is to connect the automatic defect classification module to the AOI re-judgment workstation. The images of suspected points output by AOI directly enter the AI-ADC module. The model performs a secondary authenticity judgment on them with high accuracy, filters false alarms, and automatically classifies them into categories such as particles, scratches, and pattern defects, and then writes the classification results back to the manufacturing execution system (MES). At the same time, APDT positive sample learning is used in conjunction with the DaoAI World model for generalization, reducing the dependence on a large amount of manual annotation and shortening the implementation cycle. This solution can not only improve the accuracy and efficiency of defect classification but also provide accurate defect information for the process end, helping enterprises optimize the production process and improve product yield.

Quantitative Results

By introducing WeLinkirt's AI-ADC solution, the wafer fab has achieved significant quantitative results. The over-kill rate has dropped from above 20% to below 2%, effectively reducing the false judgment and re-inspection of qualified wafers and improving production efficiency and resource utilization. The ratio of manual re-judgment to machine in the re-judgment link has decreased by about 90%, greatly reducing the workload of quality inspectors and enabling them to focus more on the handling of abnormal situations. The overall line rhythm has increased by about 30%, accelerating the product production cycle and improving the enterprise's market competitiveness. At the same time, the defect classification data has become available again, providing reliable data support for process improvement and quality control.

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.