An automated optical inspection (AOI) system based on pre-set rules and conventional image processing techniques can usually inspect quickly and correctly. However, sometimes the system detects cosmetic abnormalities caused by process variations which do not affect die performance. They are labeled as defective simply because they fail the AOI system's detection criteria.

Routinely discarding dies with acceptable defects could lead to significant and unnecessary waste. More and more wafer foundries are looking for ways to reduce these overkills in order to avoid wastage and improve yields.

Foundry Case Study

Let's consider an example from a real use case. In the process of wafer manufacturing, cracks are usually considered as serious defects that may lead to poor reliability of the end product. However, it is possible that there are other visual variations that result in AOI overkills at the same position. Taking spillover as an example, spillover usually does not affect the performance of the die, and is considered an acceptable defect. However, cracks and spillover look very similar in color or pattern, which causes traditional AOI to frequently misjudge spillover as cracks due to a lack of abstraction capability, resulting in unexpectedly high overkill. After review, the proportion of real cracks may account for only 10% of the defects, while acceptable spillovers could be as high as 90%, resulting in an incredibly high number of overkills.

Combining AOI with AI to Avoid Die Misclassification

When conventional rule-based AOI is increasingly unable to distinguish actual defects from acceptable cosmetic abnormalities, more versatile artificial intelligence (AI) tools are needed to strengthen the AOI's capabilities.
The main difference between AI and AOI inspection is that AI does not require rules to be established. Through effective defect labeling and Deep Learning, the computer can learn interconnected information from a myriad of variable features, enabling the system to make accurate judgments.

In this case, 'cracks' and 'pattern spillover' are first labeled and learned separately. During the automated production inspection, AOI performs inspections in various regions of interest (ROI). Defects detected by AOI may include both cracks and pattern spillover. Then, by using AI inference to perform precise classification, spillover can be reclassified into the 'good products' category to avoid overkill and resultant wastage.

By combining AI and AOI, the Chroma 7945 successfully solves this long-standing problem faced by many wafer foundries. This solution achieves the following goals:

1. Significantly reduced overkill rate from 90% to 2%
2. One-stop solution for reducing AOI overkills, can reclassify overkills into pass dies in real-time
3. Reclassifying specific dies helps to minimize impact on AI processing time

The Chroma 7945 In-Process Wafer Inspection System implements AI deep learning to accurately classify AOI detection results. This effectively solves the problem of AOI being unable to incorporate abstract rules and prevents overkill caused by misjudgment of acceptable variation.

To learn more about our products and solutions, please visit our website and leave your contact information. We will be happy to be of service!