Rare Defects in Advanced Packaging: APDT Positive Sample Learning with Few Samples

Advanced packaging technology, a crucial part of the semiconductor industry, faces a severe challenge of insufficient rare defect samples in the early stage of mass-production of new processes. Microchain DaoAI's DaoAI product provides an effective solution to this problem with its innovative APDT positive sample learning technology.

In the semiconductor industry, advanced packaging technology plays a crucial role in improving chip performance and achieving system integration. With the continuous development of chip manufacturing technology, new advanced packaging processes emerge in an endless stream. However, in the early stage of mass-production of new processes, various complex situations often occur. Due to the characteristics of new processes, defect types appear sporadically, and the number of single-type samples is extremely limited, often less than ten. Against this background, traditional negative-sample - driven supervised models face great difficulties in data acquisition.

Pain Points: Why Is It Difficult?

From a quantitative perspective, in terms of sample quantity, the number of single-type rare defect samples is often less than ten, while traditional supervised defect classification requires hundreds or thousands of negative samples for each type to make the model converge. In terms of classification effect, the classification rate of rare defects is extremely low, and traditional methods are difficult to achieve the ideal classification effect, resulting in a large number of rare defects being in the model blind area. In terms of risk, the escape rate of rare defects is relatively high. Once they escape to the client, the traceability cost is extremely high, which may be dozens or even hundreds of times the production cost.

The root cause of this dilemma is that the defect forms of new processes are constantly evolving, and some rare defects only appear several times in a whole batch of products. Traditional supervised classification methods rely on a large number of negative samples, while rare defects, due to their extremely low occurrence frequency, cannot collect enough samples in a short period. This leads to these rare defects being unable to be effectively recognized by the model for a long time, thus forming a model blind area. In the past, to deal with the problem of sample shortage, the only ways were to relax the criteria or add manual intervention. However, both methods have serious drawbacks. Relaxing the criteria will lead to a sharp increase in over-killing, that is, misjudging a large number of qualified products as defective products; adding manual intervention will increase labor costs and still cannot avoid the escape of rare defects.

Technical Principle

The APDT positive sample learning route adopted by DaoAI has a unique algorithm mechanism. The model mainly learns the feature distribution of qualified samples, comprehensively extracts and analyzes various features of qualified samples, and forms a standard feature model. In the actual detection process, any sample that deviates from this standard feature distribution is regarded as a suspicious sample. The advantage of this method is that there is no need to collect a large number of negative samples in advance for each rare defect. Compared with the traditional negative-sample - driven supervised model, the traditional model needs a large number of negative samples for training to distinguish between defective and non-defective samples, while APDT positive sample learning only needs to focus on the features of qualified samples, greatly reducing the dependence on the number of samples.

At the same time, DaoAI cooperates with the DaoAI World model to generalize the defect forms. The DaoAI World model can summarize the general features and change rules of defects through learning and analyzing a large number of known defect forms. Even if there are only a few samples of a certain type of defect, with the generalization ability of the DaoAI World model, it can be stably identified and classified with microscopic/micron-level accuracy. This generalization ability enables the model to adapt to the continuously evolving defect forms in new processes, effectively improving the detection ability of rare defects.

Typical Application Scenarios

• Flip - chip soldering process: In the flip-chip soldering process, rare defects such as false soldering and missing solder balls may occur. Traditional detection methods are difficult to accurately identify these defects due to insufficient samples. DaoAI uses APDT positive sample learning to first learn the feature distribution of qualified solder joints, and any solder joint that does not meet this feature distribution is regarded as a suspicious point. The difficulty lies in that the tiny differences in solder joints may affect the detection results, and accurate identification is required at the microscopic level.
• Wafer - level packaging process: In wafer-level packaging, rare defects such as bubbles and cracks in the packaging layer may occur. These defects appear with a relatively low frequency in a whole batch of products, and it is difficult for traditional methods to collect enough negative samples for training. DaoAI learns the features of qualified packaging layers and uses the DaoAI World model to generalize the possible forms of bubbles and cracks, realizing the effective detection of these rare defects. The difficulty lies in that the forms of bubbles and cracks are complex and diverse, and the model needs to have strong generalization ability.
• Chip stacking process: When stacking chips, rare defects such as stacking offset and chip damage may occur. Since the probability of these defects is relatively small, traditional supervised models are difficult to converge. DaoAI is based on the features of qualified stacked chips and identifies samples that deviate from these features. The difficulty lies in that the tiny offset of chip stacking is difficult to accurately judge at the micron-level scale, and high-precision detection technology is required.
• Wire bonding process: In the wire bonding process, rare defects such as wire breakage and poor bonding may occur. Traditional methods cannot effectively detect these defects due to insufficient samples. DaoAI learns the features of qualified wire bonding and classifies suspicious samples. The difficulty lies in that the tiny morphological changes of wires may lead to defects, and accurate detection is required at a fine scale.

Implementation Case

There is a medium-sized advanced packaging factory that faced the problem of rare defect detection during the introduction of a new process. Before adopting the DaoAI product, the classification accuracy of rare defects in this factory was extremely low, only about 30%, the escape rate of rare defects was as high as 10%, and the cycle of incorporating new defects into the model was as long as several weeks. During the process of launching the DaoAI product, the technical team first collected and analyzed qualified samples and established a feature model of qualified samples. Then, they used the DaoAI World model to generalize the possible defect forms. After a period of debugging and optimization, the DaoAI product was successfully launched.

The application of the DaoAI product has achieved a qualitative leap in rare defect detection for the advanced packaging factory.

Microchain DaoAI Solution and Product

Microchain DaoAI's DaoAI product provides a complete solution. Its core is the APDT positive sample learning route, which effectively solves the problem of rare defect sample shortage by learning the feature distribution of qualified samples. At the same time, the generalization ability of the DaoAI World model enables the model to adapt to the continuously changing defect forms in new processes. This product can achieve unified classification output of rare and common types and access the existing grading process, which is convenient for enterprises to manage and make decisions.

Quantitative Achievements

After the implementation of the DaoAI solution, significant quantitative achievements have been achieved. The classification accuracy of rare defects has exceeded 96%, a significant increase compared with the previous 30%; the escape rate of rare defects is about 0.2%, a significant reduction compared with the previous 10%; the cycle from the appearance of new defects to their incorporation into the model has been compressed from several weeks to days, greatly improving the enterprise's response speed to new defects. These achievements have effectively controlled the unknown defect risks during the mass-production ramping period, enabling the process team to focus on process improvement rather than collecting samples, providing strong support for the enterprise's production efficiency and product quality improvement.

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.