The size of parts in the electronics manufacturing industry is small and the production cycle is short, which makes the abnormal diagnosis ability of equipment and products very important, and the strength of its identification ability has even become part of the competitiveness of enterprises. How to quickly and efficiently identify these abnormalities, and do a good job of early warning and analysis, is an indispensable part of improving the yield rate.
Today, we take a case in the electronics manufacturing industry as an example to see how JMP’s powerful neural network modeling and data visualization can help data analysts greatly reduce the identification and detection time of abnormal chips, and help companies improve production efficiency and quality.
Why is data analysis so important?
For high-tech companies, quality means manufacturing cost, delivery cycle and brand value. Therefore, almost every company regards product quality as the top priority. So how to improve the quality of enterprise products becomes a difficult task.
(Figure 2) is the typical production steps of common high-tech/category enterprises, in which:
1.wafer fabrication
Wafer manufacturing refers to the silicon wafer used in the production of semiconductor integrated circuits. Because of its circular shape, it is called a wafer; it can be processed into various circuit element structures on the silicon wafer and become an integrated circuit with specific electrical functions. product.
2.Package Test
Packaging is the process of processing the tested wafers according to the product type and functional requirements to obtain independent chips; testing is the process of detecting the quality of each independent chip through a series of inspection items.
Especially in the "test" link, the traditional method is to use manual identification, that is, to record the problematic wafer images one by one through personal judgment, and then do follow-up processing. However, in actual work, artificial recognition often faces the following three problems:
1. Missing graphics
When everyone is faced with hundreds of wafer inspection pictures, it is almost impossible to read each picture one by one. It is very likely that some important pictures are missed due to factors such as fatigue, and these pictures often contain important information.
2. takes too long
The entire manual identification process is very time-consuming, and time is not spent on solving core problems.
3. Standards are not uniform
For graphics that have never appeared before, the judgment standard will vary with individual standards, which will easily cause the loss of potential information.
Based on the above inevitable factors, JMP visual analysis comes in handy.
For example, we correspond the inspection results of each independent chip according to their positions on the wafer, and import the obtained inspection results into JMP, and the following wafer inspection diagram can be easily generated through the JMP graphics generator, so as to quickly Problematic products can be accurately identified, as shown in Figure 3.
Figure three
Among them, different colors represent the size of the final test detection value. Red is the largest, descending in turn, and blue is the smallest.
As can be seen from Figure 3 above, the detection value of the blue part at the lower right of the wafer is low. What does this mean? It means that during wafer manufacturing, there is an abnormality in this area, which needs to be investigated and resolved as soon as possible. This allows engineers to immediately start a quick follow-up on this problematic chip. Here we give an example to illustrate.
Case Study: Using JMP Quickly Identify Wafer Defects
case background
The engineer of an electronics company collected defect data of 1559 chips during the final test stage, and wanted to analyze whether the ratio of a certain defect showed a regularity with the change of position.
- If the defect rate of a certain position is '0', it means that there is no such defect in this position.
- If the defect rate at this location is high, it indicates that defective products are concentrated at this location.
Through the search for such defect patterns, another quality detection method or indicator can be established for the many previous processes, which can be used as a breakthrough in overall quality improvement.
Figure four
In the data table in Figure 4, the LocX and LocY columns are row and column coordinate values respectively, and the value type is nominal value. The last column is the defect rate, which is the rate at which a position has been defective over a period of time. In order to better judge the problem, we choose the neural network algorithm here. There are several reasons for this:
- The project requires algorithms that can automatically discover the features that need to be detected from the raw data.
- The algorithm of the neural network generally has multiple levels, and each level is transformed once, so that there are multiple opportunities to capture different features.
- Generally speaking, the more data used for training, the better its accuracy can be improved.
Next, let's explain how to analyze step by step in JMP.
Step1. data import JMP
First, the data is imported into JMP, and the collected positional defect rates are converted into data values and presented through the graph generator.
Open JMP -> Analyze -> Predictive modeling -> Neural, put the defect rate to be observed in "Y, Response", put Loc X and Loc Y in "X, Factor ", as shown in Figure 5 below:
Figure five
Enter the model configuration window and enter the following parameters. This example starts with the default neural network model and gradually adds more complex models with two hidden layers and additional activation functions. Finally, the following configuration is selected, as shown in Figure 6:
In the fitting option section, tell JMP how many times to repeat the model fitting algorithm. Ultimately JMP chooses the iteration with the best validation statistics as the final model. This paper finally uses 10.
Figure six
Click the small red arrow to the left of the Model and select Save Formulas, mainly to store the transformed values for future use.
Figure seven
In this way, after row-column conversion, the following new data table is obtained, as shown in Figure 8:
Figure eight
If you use graphics to briefly illustrate the conversion process, it will be shown in Figure 9 below:
Figure 9
Step2. use JMP Graph Builder Analysis
Next, use JMP's unique "Graph Builder (Graph Builder)" for analysis.
Here we put LocX on the horizontal axis, LocY on the vertical axis, and Transform Value on the Color in the upper right corner, click the small red arrow below Graph Builder, and select Local Data Filter, as shown in Figure 10:
Figure ten
Select Label under Local Data Filter:
Figure Eleven
Then, select the small red arrow - "Animation" to get the following dynamic graphics:
This animation mainly shows how JMP scans the entire wafer one by one during the internal operation process, and finally finds the optimal solution. Where H1_1 refers to the first node of the first hidden layer, and so on.
This is our thinking and strategy at the model building level. So, specific to the application level of the actual project, how should it be deployed and analyzed? Let's move on.
JMP Landing application: project deployment and actual results
According to the model generated by JMP above, we can then deploy it at the application level.
At the deployment level, in fact, we only need to establish a ranking model, that is, a ranking model. This model does not need to ensure that every calculation result is very accurate, the most important thing is to be able to capture the most prominent graphics in multiple wafer maps. Due to the large amount of data and the server is a Linux system (relatively stable), the python program is used to extract data from the database, and the problematic wafer map is sorted according to RSquare according to the above neural network method.
Figure Fourteen
The reason for choosing RSquare sorting is to represent whether the pattern has been caught according to the significance of the model.
That is, the more significant the model, the greater the possibility of catching the pattern.
For how to calculate Generalized R-Square (as shown in the screenshot), please refer to the previous article " Increase speed and efficiency, neural network helps enterprises operate more efficiently "
For example, here are the following 7 types of defective products (from Defect A to Defect G), each row represents the X, Y coordinates on the wafer map, and the cumulative defect rate (in a certain week) of different types of defective products at the corresponding coordinate positions.
Figure 15
According to the algorithm introduced in the previous section and RSquare sorting as follows, we found that Defect C and Defect A have the highest median value of all types of defective products (if the defined threshold is >0.7, of course the threshold can be defined according to the case).
Figure sixteen
Then the Defect C and Defect A defective graphs generated by Graph Builder are really worthy of study, and Defect D, which has the lowest value, is really not obvious after the wafer map is drawn by Graph Builder.
Defect C (high RSquare value) - Right edge graph:
Figure seventeen
Defect A (high RSsquare value) - center upper right graph:
Figure 18
Defect D (low RSsquare value) - almost invisible graphics
Figure 19
The following is a schematic diagram of the system architecture:
Figure 20
It should be noted that, according to the sample data, the number of nodes in the first layer and the second layer needs to be updated for the model every month through JMP calculations (for example, for new patterns, whether the current configuration can be recognized, you need to consider adjusting number of nodes, etc.) in order to make the calculation of the model more efficient.
Through this method, after actual operation, it is proved that each data analyst in the enterprise reduces the pattern recognition time by an average of 60% per week, and the average detection time of abnormal patterns is shortened by 70% compared with before, making a contribution to the improvement of the overall level of quality control made a huge contribution.
Through the above actual combat cases, we can see that relying on manual labor for yield detection of products is not only time-consuming and labor-intensive, but also prone to errors and missed inspections, while important quality risks are often hidden in some places that cannot be identified manually .
This is like an iceberg. People often pay attention to the part above the water surface, but the part below the water surface is often invisible due to various reasons. Through advanced data analysis and technology, problems can be effectively solved, and more human energy can be put on solving problems, which fundamentally saves the cycle time of quality inspection.
It's better to sit back and do it, try JMP for real, and start your first analysis!
Further reading:
5 reasons to tell you why JMP software is more suitable for you?
Analysis of Abnormal Analysis of SPC
Yield rate of SPC regulated products——quantile method to set product control limits and regular review
Using JMP to purify outliers, taking semiconductor analysis of WAT data as an example
