Wafer level adaptive test method based on dual predictor collaborative decision

Abstract

The application discloses a wafer-level adaptive testing method based on double-predictor collaborative decision-making, which comprises the following steps: 1, data preprocessing; 2, screening test items by using an RFECV algorithm; 3, establishing a quality predictor based on test items by using a random forest algorithm; 4, dividing wafer grains into different grades according to BNR, and establishing a quality predictor based on spatial information; 5, clustering wafer grains according to two kinds of prediction results by using a DBSCAN clustering algorithm, dividing wafer grains of different categories into quality grades, and determining the quality of grains of each category successively according to the distribution results of wafer grains of different grade categories. The application can reduce the testing cost of wafers, control testing escape and yield loss at a very low level, and ensure good testing quality.

Description

Technical Field

[0001] This invention belongs to the field of integrated circuit testing, specifically a wafer-level adaptive testing method based on dual predictor collaborative decision-making. Background Technology

[0002] As the electronics industry's expectations for integrated circuit testing quality increase, efficient testing has become crucial to meeting stringent market demands. Currently, due to Moore's Law, the number of transistors integrated into chips is increasing, thus raising the cost of chip testing. Cost constraints limit the time available to test each chip. Therefore, reducing integrated circuit testing costs while maintaining test quality has always been a worthwhile research topic.

[0003] To achieve low-cost, high-quality testing, adaptive testing technology has been proposed. It predicts product quality by mining test data, using partial test information to predict whether a chip will pass the full test. Standard testing loads all test items sequentially according to the test set; if all pass, the product is considered good; otherwise, it is considered defective. Adaptive testing, however, can adjust test content, test order, and test thresholds based on predicted results, thereby saving testing time and improving test quality. However, a trade-off needs to be explored between removing test items from the production test set and predictive accuracy. The goal is to reduce production costs without affecting quality inspection accuracy, as omitting actual test items often leads to a decrease in test quality. Ideally, adaptive testing should shorten testing time without impacting test quality. However, relaxing testing in adaptive testing can cause test quality problems; incorrect predictions can lead to greater revenue losses. Therefore, the trade-off between predictive accuracy and testing costs becomes a crucial consideration. Test quality is often measured by yield loss and test escape; predicting a good product as a failure results in yield loss, while predicting a defective product as a pass leads to test escape.

[0004] Adaptive testing primarily utilizes correlation information in test data to predict product quality and reduce testing costs. Common correlations in wafer testing include inter-test item correlation and spatial correlation. Inter-test item correlation arises because highly correlated tests are likely to yield consistent results, leading to significant correlation in the test data. In contrast, systematic changes in space are very similar between adjacent dies. Therefore, spatial correlation involves performing standard tests on a subset of dies on the wafer and using the clustering of quality among neighboring dies to predict the quality of the remaining dies. Quality prediction methods based on inter-test item correlation can achieve high levels of performance, but there is room for further improvement in terms of cost savings. While spatial correlation-based quality prediction significantly reduces testing costs, it comes at the cost of poor prediction performance in areas of abrupt quality changes, increasing the risk of test escape and yield loss.

[0005] Numerous inventions have focused on one particular type of correlation. Currently, combining wafer-level spatial correlation with test item correlation yields predictors with higher performance than those modeled using only one type of correlation information. It is noteworthy that most current research cannot accurately classify grain quality when multiple predictions conflict. The different predictions provided by multiple predictors also complicate the boundary between good and defective products, making it difficult to represent simply. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention proposes a wafer-level adaptive testing method based on dual predictor collaborative decision-making, aiming to reduce wafer testing costs while keeping test escape and yield loss at extremely low levels, thus ensuring good test quality.

[0007] The present invention adopts the following technical solution to solve the technical problem:

[0008] The present invention provides a wafer-level adaptive testing method based on dual predictor collaborative decision-making, characterized by the following steps:

[0009] Step 1: Utilize the test item dataset For Samples taken from the test wafer of each grain After testing each grain, the resulting quality test result dataset is denoted as: On the test wafer The position information of each grain is denoted as ;in, Indicates the number of test items for the grain. The first grain Information on each test item, Indicates the first The location information of each grain Indicates the first The quality test results of each grain are used as a label, if , indicating the first If the quality test result of each grain is qualified, , indicating the first The quality test result for each grain was a fault.

[0010] The remaining grains extracted from the test wafer are used as the test grains, and the position information of the test grains is recorded as follows: ,in, For the first The positional information of each grain to be tested;

[0011] Step 2: Use the RFECV algorithm to process the test item dataset. The test items are filtered to obtain a subset of filtered test items. ,in, This represents the information of the m-th test item after grain screening, where M is the total number of test items after screening, and ;

[0012] Step 3: Select the filtered test item subset The test result dataset will be used as input data. As labels, they are fed together into a random forest model for training to obtain a test item quality predictor. ;

[0013] Using test item quality predictor The quality of the grain under test is predicted, and the quality prediction result dataset of the grain under test is obtained. ,in, Indicates the quality predictor of the test item. For the Predicted quality results for each grain to be tested;

[0014] Step 4: Utilize the local spatial information of the grain Establish a space quality predictor It is used to predict the quality of the grains to be tested.

[0015] Step 4.1: Based on the dataset of quality prediction results for the grains to be tested Trial Results Dataset The local spatial information of the k-th grain on the test wafer is calculated using equation (1). This allows us to obtain the local spatial information of all grains on the test wafer. :

[0016] (1)

[0017] In equation (1), v represents the number of grains surrounding any given grain; Indicates the first The mass of the q-th grain surrounding a given grain, if the q-th grain is the grain to be measured, then let... , Indicates the quality predictor of the test item. For the The quality prediction results of the q-th grain; if the q-th grain To sample grains, let , Indicates the first Quality test results for each grain;

[0018] Based on the local spatial information of the grain The grains to be tested are classified; let These are upper and lower bound grading indicators, when When, it indicates the first The grade of each grain is easy to pass; when When, it indicates the first The grade of the first grain is faulty; otherwise, it indicates that the grade of the second grain is faulty. The quality level of each grain is ambiguous;

[0019] Step 4.2: Use the RFECV algorithm to filter the different grades of grains in the test item subset B to obtain the test item subset with the easiest pass grade. A subset of test items for high failure rate and a subset of test items for quality fuzziness levels ;

[0020] Step 4.3: Subset the test items for the three levels Together with BNR as input data, test the result dataset. As labels, they are fed into the random forest model for training to obtain the spatial quality predictor. ;

[0021] Using space quality predictors Quality prediction is performed on different grades of the grains to be tested, thereby obtaining the quality prediction results. ,in, Indicated by space quality predictor For the Predicted results for each grain to be tested;

[0022] Step 5: Utilize and Predicted total mass of the test grains Collaborative decision-making is employed to obtain the classification results of the grains to be tested;

[0023] Step 5.1: Use the DBSCAN clustering algorithm to process the total quality prediction result R of the test grain to obtain the cluster to which the quality prediction result of the test grain belongs. ,in, Indicates the first One cluster, Indicates the number of clusters;

[0024] Determine if there are any test grains that do not belong to any cluster. If so, the prediction result of the corresponding test grain is an outlier, and the corresponding test grain is added to the outlier grain set. ,in, Indicates the first grain in the test grain Outlier grains whose predicted grain values ​​are abnormal;

[0025] Step 5.2: Calculate the first equation in the Cartesian two-dimensional coordinate system using equation (2). Cluster center of gravity ):

[0026] (2)

[0027] In equation (2), The first Cluster The x and y coordinates of the u-th grain in the middle, Indicates the first Cluster The number of grains in the middle;

[0028] Step 5.3: Calculate the distances of the centroid positions of all clusters from the ideal points (0,0) and (1,1) respectively; and denote the cluster closest to (0,0) as the cluster of the fault standard class, denoted as... Its center of gravity is remembered as The cluster closest to (1,1) is designated as the qualified criterion cluster, denoted as... Its center of gravity is remembered as );

[0029] Clusters meeting the pass / fail criteria and clusters meeting the failure criteria are designated as standard clusters, while clusters not meeting the standard criteria are designated as edge clusters. ,in, Indicates the first An edge cluster, Indicates the number of edge clusters, and ;

[0030] Edge clusters The position of the center of gravity is denoted as ;in, Indicates the first The central position of an edge cluster;

[0031] Step 5.4: Calculate the first step using equation (3). Edge clusters center of gravity The centroid positions of the standard clusters respectively Distance:

[0032] (3)

[0033] In equation (3), It is the first Edge clusters Distance from the qualified standard category, It is the first Edge clusters Distance from the fault standard class;

[0034] Step 5.5: Use equation (4) to determine the first... Edge clusters Quality:

[0035] (4)

[0036] In equation (4), Indicates the first Edge clusters For qualified grain clusters, Indicates the first Edge clusters This is a cluster of faulty grains;

[0037] Step 5.6: For outliers, calculate the r-th outlier. center of gravity Distance from all cluster centroid locations ,in, express To the Cluster center of gravity The distance of ); and the r-th outlier grain The outlier is assigned to the cluster corresponding to the minimum distance, and its mass is used as the mass of the cluster corresponding to the minimum distance. The quality.

[0038] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the wafer-level adaptive method, and the processor is configured to execute the program stored in the memory.

[0039] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program is executed by a processor to perform the steps of the wafer-level adaptive method.

[0040] Compared with existing technologies, the beneficial results of this invention are reflected in:

[0041] 1. This invention proposes an adaptive testing method for wafer testing using a dual-predictor collaborative decision-making approach. The predictor can select the most effective test items for testing during wafer testing, thereby significantly reducing wafer testing costs, shortening testing time, and improving wafer testing efficiency.

[0042] 2. This invention utilizes the test items and spatial information in wafer testing to train test item predictors and spatial predictors to predict die quality. This not only makes fuller use of the correlation information in the test data, but also ensures high accuracy in the classification of the die to be tested, eliminates the adverse effects of quality mutation regions in the wafer on the predictor, and ensures better test quality.

[0043] 3. This invention addresses the decision-making conflict between predictors by clustering the results of the two predictors using a clustering algorithm. Then, it classifies the dies step by step by utilizing the centroid positions between clusters, thereby screening out dies with faults and determining the quality prediction result of each die. This solves the problem of not being able to ideally divide the irregular boundaries between good and bad chips, and keeps test escape and yield loss at extremely low levels.

[0044] 4. This invention utilizes test items and spatial information from the grain portion to train two quality predictors. Based on the prediction results, the grains are clustered, and quality levels are assigned to grains in different clusters, thereby successively determining the quality of each cluster of grains and outliers. While ensuring test quality, this invention improves prediction accuracy, significantly reduces test time and cost, and this cost-reducing solution has no hardware overhead and good versatility. Attached Figure Description

[0045] Figure 1 This is a flowchart of the adaptability testing process for the present invention;

[0046] Figure 2 This is a flowchart of the collaborative decision-making process of the present invention. Detailed Implementation

[0047] In this embodiment, a wafer-level adaptive testing method based on dual predictor collaborative decision-making is described, such as... Figure 1 As shown, the procedure is as follows:

[0048] Step 1: Utilize the test item dataset For Samples taken from the test wafer of each grain After testing each grain, the resulting quality test result dataset is denoted as: On the test wafer The position information of each grain is denoted as ;in, Indicates the number of test items for the grain. The first grain Information on each test item, Indicates the first The location information of each grain Indicates the first The quality test results of each grain are used as a label, if , indicating the first If the quality test result of each grain is qualified, , indicating the first The quality test result for each grain was a fault.

[0049] The remaining grains extracted from the test wafer are used as the test grains, and the position information of the test grains is recorded as follows: ,in, For the first The positional information of each grain to be tested;

[0050] During data preprocessing, since wafer testing follows the "stop on first failure" strategy, the test data usually contains a large number of missing values. For missing test data, the median of the test item information to which the data belongs is used to fill in the missing values.

[0051] Step 2: Use the RFECV algorithm to process the test item dataset. The test items are filtered to obtain a subset of filtered test items. ,in, This represents the information of the m-th test item after grain screening, where M is the total number of test items after screening, and The obtained test item subset B can reduce the test time of wafer testing and is also the input data for the subsequent test item predictor and spatial predictor.

[0052] Step 3: Select the filtered test item subset The test result dataset will be used as input data. As labels, they are fed together into a random forest model for training to obtain a test item quality predictor. The predictor is a functional mapping relationship between some test item information and grain quality. The quality of the grain to be tested can be determined by simply inputting the test item information into the predictor.

[0053] Using test item quality predictor The quality of the grain under test is predicted, and the quality prediction result dataset of the grain under test is obtained. ,in, Indicates the quality predictor of the test item. For the Predicted quality results for each grain to be tested;

[0054] Step 4: Local Spatial Information This represents the relevant test information of surrounding grains for any given grain, and is a concrete manifestation of spatial correlation in wafer testing. Based on the characteristic of clustered distribution of grain quality within a wafer, it utilizes the local spatial information of the grains. Establish a space quality predictor It is used to predict the quality of the grains to be tested.

[0055] Step 4.1: Based on the dataset of quality prediction results for the grains to be tested Trial Results Dataset The local spatial information of the k-th grain on the test wafer is calculated using equation (1). This allows us to obtain the local spatial information of all grains on the test wafer. :

[0056] (1)

[0057] In equation (1), v represents the number of grains surrounding any given grain; Indicates the first The first grain around the first The mass of each grain, if the q-th grain is the grain to be tested, then let , Indicates the quality predictor of the test item. For the The quality prediction results of the q-th grain; if the q-th grain To sample grains, let , Indicates the first Quality test results for each grain;

[0058] Based on the local spatial information of the grain The grains to be tested are classified; let These are upper and lower bound grading indicators, when When, it indicates the first The grade of each grain is easy to pass; when When, it indicates the first The grade of the first grain is faulty; otherwise, it indicates that the grade of the second grain is faulty. The quality level of each grain is ambiguous;

[0059] Step 4.2: Use the RFECV algorithm to filter the different grades of grains in the test item subset B to obtain the test item subset with the easiest pass grade. A subset of test items for high failure rate and a subset of test items for quality fuzziness levels The purpose here is to screen out a more effective subset of test items for different grades of grains, thereby improving the accuracy of grain quality prediction.

[0060] Step 4.3: Subset the test items for the three levels Together with BNR as input data, test the result dataset. As labels, they are fed into the random forest model for training to obtain the spatial quality predictor. ;

[0061] Using space quality predictors Quality prediction is performed on different grades of the grains to be tested, thereby obtaining the quality prediction results. ,in, Indicated by space quality predictor For the Predicted results for each grain to be tested;

[0062] Step 5: Utilize and Predicted total mass of the test grains To make collaborative decisions, such as Figure 2 As shown, the classification results of the grains to be tested are obtained.

[0063] Step 5.1: Use the DBSCAN clustering algorithm to process the total quality prediction result R of the test grain to obtain the cluster to which the quality prediction result of the test grain belongs. ,in, Indicates the first One cluster, Indicates the number of clusters;

[0064] Determine if there are any test grains that do not belong to any cluster. If so, the prediction result of the corresponding test grain is an outlier, and the corresponding test grain is added to the outlier grain set. ,in, Indicates the first grain in the test grain Outlier grains whose predicted grain values ​​are abnormal;

[0065] The goal here is to identify the grain clusters and outliers when the predictions of the dual predictors conflict, and to determine the final quality prediction results of these grains.

[0066] Step 5.2: Calculate the first equation in the Cartesian two-dimensional coordinate system using equation (2). Cluster center of gravity ):

[0067] (2)

[0068] In equation (2), The first Cluster The x and y coordinates of the u-th grain in the middle, Indicates the first Cluster The number of grains in the middle;

[0069] Step 5.3: Ideally, the centroid of the qualified category is close to point (1,1), indicating that the probability of the product being predicted as qualified by both predictors is 100%; conversely, the faulty category will be close to the origin (0,0); calculate the distances of the centroid positions of all clusters from the ideal points (0,0) and (1,1) respectively; and take the cluster closest to (0,0) as the faulty standard category cluster, denoted as... Its center of gravity is remembered as The cluster closest to (1,1) is designated as the qualified criterion cluster, denoted as... Its center of gravity is remembered as );

[0070] Clusters meeting the pass / fail criteria and clusters meeting the failure criteria are designated as standard clusters, while clusters not meeting the standard criteria are designated as edge clusters. ,in, Indicates the first An edge cluster, Indicates the number of edge clusters, and ;

[0071] Edge clusters The position of the center of gravity is denoted as ;in, Indicates the first The central position of an edge cluster;

[0072] Step 5.4: Calculate the first step using equation (3). Edge clusters center of gravity The centroid positions of the standard clusters respectively Distance:

[0073] (3)

[0074] in, It is the first Edge clusters Distance from the qualified standard category, It is the first Edge clusters Distance from the fault standard class;

[0075] Step 5.5: Use equation (4) to determine the first... Edge clusters Quality:

[0076] (4)

[0077] In equation (4), Indicates the first Edge clusters For qualified grain clusters, Indicates the first Edge clusters This is a cluster of faulty grains;

[0078] Step 5.6: For outliers, calculate the r-th outlier. center of gravity Distance from all cluster centroid locations ,Right now ,in, express To the The distance between the centroids of each cluster; [record] for The minimum value represents the r-th outlier grain. To the Cluster If the distance is minimized, then the r-th outlier grain will be... The quality belongs to the first Cluster The quality, and thus based on Cluster The quality of the r-th outlier grain is obtained The quality.

[0079] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0080] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A dual-predictor based collaborative decision making method for wafer level adaptive testing, the method comprising: is performed by the following steps: Step 1: using the test item dataset on the test wafer After testing the extracted quality test result dataset is obtained and recorded as The position information of the crystals on the test wafer is recorded as ; wherein represents the number of test items of the crystal, represents the test item information of the crystal, represents the position information of the crystal, represents the quality test result of the crystal and serves as a label, if , it indicates that the quality test result of the crystal is qualified, if , it indicates that the quality test result of thecrystal is faulty. Extract the remaining dies on the test wafer as the to-be-tested dies, and record the position information of the to-be-tested dies as wherein, is the position information of the i-th to-be-tested die; the i-th to-be-tested die. Step 2: Test item dataset is screened by using RFECV algorithm Step 3: Test item subset is obtained after screening Wherein, represents the mth test item information of the screened crystal grain, M is the total number of test items after screening, and ; Step 3: Subsets of test items are filtered As input data, the test result dataset As labels, and together input into the random forest model for training, to obtain the test item quality predictor ; Utilizing test item quality predictor performing quality prediction on the test dies to obtain a quality prediction result dataset of the test dies wherein, indicates a quality prediction result of the i-th test die by the test item quality predictor the test dies ​ Step 4: Utilizing local spatial information of die , establishing a spatial quality predictor , for quality prediction of a die under test; Step 4.1: Calculate the local spatial information of the kth die on the test wafer according to the quality prediction result data set of the die to be tested and the test result data set , using formula (1) , thereby obtaining the local spatial information of all the dies on the test wafer : (1) In equation (1), v represents the number of grains surrounding any given grain; Indicates the first The mass of the q-th grain surrounding a given grain, if the q-th grain is the grain to be measured, then let... , Indicates the quality predictor of the test item. For the The quality prediction results of the q-th grain; if the q-th grain To sample grains, let , Indicates the first Quality test results for each grain; According to local spatial information of the crystal grain Classifying the to-be-tested crystal grain; let be the upper and lower boundary classification indexes, respectively, when , it indicates that the grade of the th crystal grain is easy to pass; when , it indicates that the grade of the th crystal grain is easy to fail; otherwise, it indicates that the grade of the th crystal grain is quality ambiguous; Step 4.2: Screen the different levels of grains in the test item subset B by using the RFECV algorithm to obtain the test item subset of easy passing level , the test item subset of easy failure level , and the test item subset of quality ambiguous level ; Step 4.3: Subsets of test items of three levels are created The test result dataset is inputted together with BNR as input data As labels, the test result dataset is inputted together into the random forest model for training to obtain the spatial quality predictor ; Utilizing spatial quality predictor Quality prediction is performed on different levels of the grains under test, resulting in quality prediction results wherein, indicates the prediction result of the first grain under test by the spatial quality predictor​ Step 5: utilizing and the total mass prediction result of the to-be-tested die performing collaborative decision-making to obtain a classification result of the to-be-tested die; Step 5.1: Use the DBSCAN clustering algorithm to process the total quality prediction result R of the test grain to obtain the cluster to which the quality prediction result of the test grain belongs. ,in, Indicates the first One cluster, Indicates the number of clusters; determining whether there is a die in the to-be-tested dies that does not belong to any cluster, if the prediction result of the corresponding to-be-tested die is an outlier, and adding the corresponding to-be-tested die to the outlier die set wherein, represents an outlier die whose prediction result is an outlier of the i-th die in the to-be-tested dies Step 5.2: Calculate the barycentric position of the first cluster of points in the Cartesian two-dimensional coordinate system using formula (2) ):​​ (2) In equation (2), The first Cluster The x and y coordinates of the u-th grain in the middle, Indicates the first Cluster The number of grains in the middle; Step 5.3: Calculate the distance between the center of gravity of all clusters and the ideal points (0, 0) and (1, 1) respectively; and the cluster closest to the distance (0, 0) as the cluster of the failure standard class, recorded as , whose center of gravity is recorded as ); the cluster closest to (1, 1) as the cluster of the qualified standard class, recorded as , whose center of gravity is recorded as ); The cluster of the standard class and the cluster of the failure standard class are denoted as a standard class cluster, and a cluster other than the standard class cluster is denoted as an edge class cluster wherein, denotes the th edge class cluster, denotes the number of edge class clusters, and ; The edge class cluster is recorded as ; wherein represents the barycentric position of the th edge class cluster; Step 5.4: Calculate the distance of the barycentric position of the individual edge class cluster from the barycentric position of the standard class cluster respectively Step 5.5: Calculate the distance of the barycentric position of the individual edge class cluster from the barycentric position of the standard class cluster (3) In formula (3), is the first edge class cluster is the first edge class cluster is the distance to the qualified standard class, is the first edge class cluster is the first edge class cluster is the distance to the failure standard class; Step 5.5: Use the formula (4) to judge the quality of the first edge class cluster ​ (4) In formula (4), denotes the edge class cluster is a qualified die cluster, denotes the edge class cluster is a failed die cluster; Step 5.6: For outliers, calculate the r-th outlier. center of gravity Distance from all cluster centroid locations ,in, express To the Cluster center of gravity The distance of ); and the r-th outlier grain The outlier is assigned to the cluster corresponding to the minimum distance, and its mass is used as the mass of the cluster corresponding to the minimum distance. The quality.

2. An electronic device comprising a memory and a processor, characterized in that The memory is configured to store a program supporting the processor to execute the wafer-level adaptive test method of claim 1.

3. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program, when executed by the processor, performs the steps of the wafer-level adaptive test method of claim 1.

Images