Method for analyzing yield in semiconductor industry

Abstract

The invention provides a yield analysis method in the semiconductor industry. The yield analysis method is characterized by comprising the following steps: S1, acquiring a data set; s2, establishing an algorithm input data set through ETL processing; s3, algorithm model operation iteration is carried out, confidence and lifting degree indexes are calculated for all paths through an association rule algorithm Apriori, and path sorting indexes output by the algorithm are formed; and S4, outputting an algorithm model result, and obtaining a suspicious path by taking the improvement degree as a sorting index. By analyzing the reject ratio on the equipment path, problems in the equipment manufacturing process can be found out, targeted improvement measures are taken, and the quality and reliability of products are improved.

Description

Technical Field

[0001] This invention is applied to the field of semiconductor manufacturing process optimization, specifically relating to a method for yield analysis in the semiconductor industry. Background Technology

[0002] In the semiconductor industry, there are a wide variety of products and many processing steps, which takes a long time. Finally, the quality of the products can be tested and feedback can be obtained more than half a month after the start of production.

[0003] During product processing, the product's path can become more diverse due to factors such as production capacity. This makes data analysis on factors affecting yield, particularly the product's path, increasingly difficult when working to improve yield. Summary of the Invention

[0004] The present invention aims to overcome the aforementioned concerns, help products select the highest yield path combination among various walking paths, and provide a scientific basis for this.

[0005] This invention provides a method for yield analysis in the semiconductor industry, characterized by comprising the following steps:

[0006] S1. Filter target analysis objects based on the subsequent detection data;

[0007] S2. Use the target analysis object to retrieve its corresponding production history;

[0008] S3. Concatenate the two data segments and perform data analysis using the Prim algorithm (a machine learning algorithm).

[0009] The data analysis methods are as follows:

[0010] a) Using data with product labels as evaluation criteria

[0011] b) Define the root node;

[0012] c) Use a set S to store the generated tree nodes and a set Q to store the ungenerated tree nodes. Add the root node to S and the remaining nodes to Q.

[0013] d) During the search process, traverse each process-equipment change node, calculate the edge weight between the node and its neighboring nodes, and if the sum of the distances from the node to the root node plus the edge weight is less than the sum of the distances from the neighboring nodes to the root node, then update the distances of the neighboring nodes and the parent node.

[0014] e) Find the node with the smallest distance: Find the node with the smallest distance and the smallest parent node from set Q;

[0015] f) Add the node to S and remove it from Q.

[0016] g) Repeat the above operation until set S contains all tree nodes;

[0017] The final S obtained is a minimum spanning tree.

[0018] Furthermore, the present invention provides a method for yield analysis in the semiconductor industry, characterized in that:

[0019] The root node represents the data with the highest percentage of good results.

[0020] Furthermore, the present invention provides a method for yield analysis in the semiconductor industry, characterized in that:

[0021] The process, including testing data and production data, is as follows:

[0022] a) Label data indicating the final state of the product, stored in area A, with the data source being the final inspection data.

[0023] Among them, the field used to identify the product name in the test data is product information;

[0024] b) Use product information to extract data from the production and processing history. The corresponding data is the production and processing history data, stored in area B, and the data source is the product processing history data.

[0025] Furthermore, the present invention provides a method for yield analysis in the semiconductor industry, characterized in that:

[0026] Each product's production history represents a complete path combination.

[0027] The testing data labels for different products vary.

[0028] Furthermore, the present invention provides a method for yield analysis in the semiconductor industry, characterized in that:

[0029] The data is also differentiated based on the product's form, as detailed below:

[0030] a) The product status varies in different process areas. To ensure that the label data that can characterize the product status is as accurate as possible, the concept of integration is introduced here, which refers to how many minimum detection units are integrated on the processed product.

[0031] b) Divide the data into zones according to the integration level of different processing units and perform calculations separately.

[0032] Furthermore, the present invention provides a method for yield analysis in the semiconductor industry, characterized in that:

[0033] The detection data is processed as follows:

[0034] a) Classify products according to different labels.

[0035] b) Minimum detection unit count: Count separately for groups labeled as good and bad.

[0036] c) Calculate the percentage of good products for processing units with different levels of integration.

[0037] The function and effects of this invention:

[0038] The advantages of this invention are: by utilizing machine learning algorithms to effectively partition and segment large amounts of data, it enables scientific, effective, accurate, and efficient analysis of scenario-based data. This overcomes the previous limitations where the analysis accuracy was too low due to excessive data volume, overly complex scenarios, and varying levels of personnel experience, making it unsuitable as a scientific basis for improving production yield. Attached Figure Description

[0039] Figure 1 For flowcharts;

[0040] Figure 2 This is a schematic diagram of a specific example. Detailed Implementation

[0041] This invention is capable of various modifications and embodiments, and therefore specific embodiments are illustrated and described in the accompanying drawings. However, this is not intended to limit the invention to specific implementations, but should be understood to include all modifications, equivalents, and even substitutions that fall within the spirit and scope of this invention.

[0042] 1. Large-scale data processing

[0043] Typically, the data is divided into two parts: one is the label data used to mark the final state of the product (good label / bad label), and the other is the data on the corresponding product's processing and production history. These two data sets are stored separately as A and B, and corresponding logical processing relationships are applied to each part.

[0044] □ Zoning: Inspection data / Production processing data. Inspection data is used to differentiate production data, mark the inspection results of production processing data, and find the optimal production path based on the inspection results.

[0045] a) Label data indicating the final state of the product, stored in area A, with the data source being the final inspection data.

[0046] b) The field used for product name identification in this test data is product information.

[0047] c) Extract data from the production and processing history of the product information. This data corresponds to the production and processing history data, which is stored in area B. The data source is the product processing history data (information is associated through product name).

[0048] □ Scenarios: Dimension 1: Path Combination

[0049] a) The production history corresponding to each product information is a complete path combination.

[0050] b) The test data labels differ for different product information. Since the good / bad labels are percentages, the test data labels for different product information will vary.

[0051] c) Prepare the data for this scenario to prepare for subsequent algorithm analysis.

[0052] □ Segmentation: Data is differentiated based on the product's form (processing unit)

[0053] a) The product status varies in different process areas. To ensure that the label data representing the product status is as accurate as possible, the concept of integration degree is introduced here, which refers to the number of minimum inspection units integrated on the processed product. For example, in different process areas, in area A, the calculation method for the inspection data label of product information M is as follows: integration degree 100, that is, product M has 100 minimum inspection units N, the number of minimum units used for inspection is 70, the number of good units is 50, the number of defective units is 20, and the data label is 50 / 70.

[0054] b) Divide the data into zones based on the integration level of different processing units and perform calculations separately. For example, in different process zones, zone A, the calculation method for the detection data label of product information M is as follows: integration level 100, that is, product M has 100 minimum detection units N, the minimum number of units used for detection is 70, the number of good units is 50, the number of defective units is 20, and the data label is 50 / 70.

[0055] 2. Different algorithms are used for different data.

[0056] □ Test data

[0057] a) Classify products according to different labels;

[0058] In some specific instances, the partitioning methods include: a one-size-fits-all approach; or using a normal distribution; or random sampling.

[0059] In some specific examples, the grades can be divided into: good products, generally good products, medium products, bad products, and scrapped products (this classification is based on different data quality levels and should be selected according to the actual situation).

[0060] b) Minimum detection unit count: Count separately for groups labeled as good and bad.

[0061] c) Calculate the percentage of good products (good / (good + bad)) for processing units with different levels of integration. In some experimental cases, this refers to the same process but different equipment, and the combination of different equipment units processed by different equipment. This is to distinguish the data labels under different paths.

[0062] Note: The number of "good" + "defective" items represents the number of tests conducted per processing unit, not the integration level.

[0063] □ Production Data

[0064] a) By using the downstream testing data (i.e., the testing data at the last stage of production and processing), the target analysis objects (e.g., time / location / product name / label name) are screened. The purpose is to screen out the label data, integrate it according to the product information, and calculate the label data (good / bad).

[0065] b) Using the target analysis object, retrieve its corresponding production history: station (process name) / equipment / equipment unit / equipment processing chamber / time / action name

[0066] c) Concatenate the two data sets (e.g., production data and tag data, linked by the product's unique identifier ID), and apply the Prim algorithm (a machine learning algorithm) for further data analysis.

[0067] d) First, use data with product labels as the evaluation standard. The evaluation standard can be a percentage calculated based on the label, and then classify the grades. Considering that for the same product information, not every processing unit may have testing data, labels without testing data can be removed.

[0068] e) Define the root node (i.e., the node with the highest percentage of good products). In this case, the corresponding process-equipment-equipment unit-chamber is the initial solution O1.

[0069] f) Use a set S to store the generated tree nodes and a set Q to store the ungenerated tree nodes. Add the root node to S and the remaining nodes to Q.

[0070] g) During the search process, traverse each process-equipment change node, calculate the edge weight (i.e., the percentage of good quality) between the node and its neighboring nodes. If the sum of the distances from the node to the root node (which can be stored using a dist[] array or a map) plus the edge weight is less than the sum of the distances from the neighboring nodes to the root node, then update the distances of the neighboring nodes and the parent node.

[0071] h) Find the node with the smallest distance: Find the node with the smallest distance and the smallest parent node from the set Q (i.e., an important factor affecting the good rate).

[0072] i) Add the node to S and remove it from Q.

[0073] j) Repeat the above operation until the set S contains all the tree nodes.

[0074] The final S obtained by k) is a minimum spanning tree.

[0075] l) This weighted tree can also be transformed (by proportionally converting the label data and weight factors) into a process-equipment tree that affects the yield rate (from top to bottom), with the influence factors decreasing continuously.

[0076] Because high-end semiconductor manufacturing and panel industries require analysis of product yield and different stations in the entire process path, the current situation is that manual analysis of too many stations is difficult to achieve. By using the aforementioned machine learning or statistical methods, it is possible to analyze the overall path that the product takes during the manufacturing process, derive the optimal path, and identify problematic nodes in the path.

[0077] like Figure 2 The image shows the experimental results for a specific scenario. Based on conventional experience, performing five layers is sufficient to identify the problematic nodes.

[0078] While the foregoing has focused on embodiments, these are merely illustrative and do not limit the invention. Those skilled in the art will understand that various modifications and applications not illustrated above can be made without departing from the essential characteristics of these embodiments. For example, the constituent elements specifically shown in the embodiments can be implemented through modifications. Furthermore, various differences related to such modifications and applications should be interpreted as being included within the scope of the invention as defined in the appended claims.

Claims

1. A method for yield analysis in the semiconductor industry, characterized in that: It includes the following steps: S1. Obtain the dataset; S2. Establish the algorithm input dataset through ETL processing; S3. Algorithm model operation and iteration: The confidence and lift indices of all paths are calculated using the Apriori association rule algorithm to form the path ranking index output by the algorithm. S4. The algorithm model outputs results, using lift as the ranking metric to identify suspicious paths.

2. The method for yield analysis in the semiconductor industry as described in claim 1, characterized in that: In step S1, the dataset includes equipment history data and production data for each process on the equipment path.

3. The method for yield analysis in the semiconductor industry as described in claim 1, characterized in that: In step S2, the collected data is cleaned and processed to remove abnormal data and incomplete data with missing values.

4. The method for yield analysis in the semiconductor industry as described in claim 1, characterized in that: Step S3 includes the following specific data processing steps: S3-1. Based on the preset frequency of defective labels, obtain all devices {device, NG} that may cause defects, and form a 1-path set; S3-2. Traverse all bad paths that contain suspicious devices in the 1-path set, and add all suspicious devices in sequence according to the preset frequency of the bad tags to form the 2-path set; S3-3. Continue in this manner to form an N-path set, until there is no N+1-path set that meets the preset frequency of occurrence of bad tags; S3-4. Based on all paths from S3-1 to S3-3, obtain their confidence index; S3-5. Obtain the lift index based on the confidence index, and make it into a path ranking index.

5. The method for yield analysis in the semiconductor industry as described in claim 4, characterized in that: The confidence index is calculated as: Number of defective products passing through the current path / Total number of products passing through the current path.

6. The method for yield analysis in the semiconductor industry as described in claim 4, characterized in that: The lift index = confidence index / absolute defect rate; The absolute defect rate is calculated as: (Number of defective products / Total number of products)

Images