CN115132599A - Defect detection method - Google Patents

Abstract

A defect detection method, comprising: acquiring a defect database, the defect database including several defect data; providing several wafers to be inspected; photographing each of the wafers to be inspected from different angles, and obtaining each of the wafers to be inspected Several pictures to be inspected of the wafer; several initial inspection pictures are obtained from several pictures to be inspected of each of the described wafers to be inspected; several initial inspection pictures of each of the described wafers to be inspected are spliced, forming a detection picture; detecting the detection picture according to the defect detection library to obtain a detection result. By making full use of all the pixel information of the selected image, all the features of the image are completely preserved, so that the final detection accuracy is effectively improved.

Description

Defect detection method

technical field

The invention relates to the technical field of semiconductor manufacturing, and in particular, to a defect detection method.

Background technique

Semiconductor integrated circuit chips are fabricated by batch processing, and a large number of various types of semiconductor devices are formed on the same substrate and interconnected to have complete electronic functions. Among them, defects generated in any step may lead to the failure of circuit fabrication. Therefore, in the fabrication process, it is often necessary to perform defect detection and analysis on the fabrication structure of each step, to find out the cause of the defect and eliminate it. However, with the rapid development of Ultra Large Scale Integration (ULSI, UltraLarge Scale Integration), the integration of chips is getting higher and higher, and the size of devices is getting smaller and smaller. Correspondingly, the production in the process is enough to affect the flat rate of the device. The defect size is also getting smaller and smaller, which puts forward higher requirements for the defect detection of semiconductor devices.

However, the accuracy of existing defect detection methods is low.

SUMMARY OF THE INVENTION

The technical problem solved by the present invention is to provide a defect detection method, which can effectively improve the accuracy of defect detection.

In order to solve the above problems, the present invention provides a defect detection method, which includes: acquiring a defect database, the defect database includes a plurality of defect data; providing a plurality of wafers to be inspected; photographing each wafer to be inspected from different angles , obtain a number of pictures to be inspected of each of the described wafers to be inspected; obtain a number of initial inspection pictures from a number of pictures to be inspected of each of the described wafers to be inspected; The initial inspection images are spliced to form inspection images; the inspection images are inspected according to the defect inspection library to obtain inspection results.

Optionally, the method for acquiring the defect database includes: providing several training wafers, each of which has several defects; performing training and learning on each of the training wafers, and acquiring an initial defect database, so that the The initial defect database includes several initial defect data; provides several groups of test wafer groups, each group of the test wafer groups includes several test wafers; tests the initial defect database according to the several test wafer groups Learning, to obtain the defect database.

Optionally, performing training and learning on each of the training wafers, and the method for obtaining an initial defect database includes: photographing each of the training wafers from different angles, and obtaining a number of training wafers to be trained for each of the training wafers. Picture; obtain several initial training pictures from several described pictures to be trained of each described training wafer; splicing some initial training pictures of each described training wafer to form training pictures; for each described training The image is processed by image recognition to obtain initial defect data; the initial defect database is composed of a plurality of the initial defect data.

Optionally, the method for obtaining a plurality of initial training pictures from a plurality of the pictures to be trained in each of the training wafers includes: obtaining the clearest shooting angle of the defective images in the training wafers of the same type according to an empirical rule; The to-be-trained picture taken at the shooting angle is used as the initial training picture.

Optionally, the initial defect database is tested and learned according to several test wafer groups, and the method for acquiring the defect database includes: photographing each of the test wafers from different angles, and obtaining each of the test wafers. Several pictures to be tested of the test wafer; several initial test pictures are obtained from several pictures to be tested of each of the test wafers; several initial test pictures of each of the test wafers are spliced to form a test picture Carry out image recognition processing to each described test picture, obtain test data; Obtain the real data of each described test picture; According to described initial defect database, described test data is detected, from some described initial defect data Obtaining judgment data; when the judgment data and the real data exceed the preset range, the corresponding test wafers are re-trained and learned; when a group of the test wafer groups is tested, calculate the group of the test wafers. The test accuracy rate of the test wafer group; when the test accuracy rate of the test wafer group of this group is lower than the set threshold, then the test learning of the next group of test wafer groups is performed until the test of the group The defect database is obtained until the test accuracy rate of the wafer group reaches a set threshold, and the learning model with the best accuracy rate in the test learning process is retained.

Optionally, the method for acquiring several initial test pictures from several of the to-be-tested pictures of each of the test wafers includes: according to an empirical rule, acquiring the shooting angle with the clearest defect image in the test wafer of the same type. ; Use the picture to be tested taken at the shooting angle as the initial test picture.

Optionally, the method for acquiring several initial inspection pictures from several of the to-be-detected pictures includes: according to empirical rules, acquiring the clearest shooting angle of the defect images in the same type of the to-be-detected wafer; The obtained picture to be detected is used as the initial detected picture.

Optionally, the same type includes: one or more of the same batch, the same process step and the same wafer.

Optionally, the detection picture is detected according to the defect detection library, and the method for obtaining the detection result includes: performing image recognition processing on each of the detection pictures to obtain detection data; comparing the detection data with the defect. Several defect data in the database are compared separately; within a certain comparison error range, when the detection data is the same as any of the defect data, the corresponding defect data is obtained, and the defect data is used as the detection result .

Optionally, the method for forming a detection picture by splicing a plurality of the initial detection pictures includes: docking a plurality of the initial detection pictures along a first direction or a second direction to form the detection picture, the first direction perpendicular to the second direction.

Optionally, the learning model used in the training learning and the testing learning process includes a convolutional neural network model.

Compared with the prior art, the technical solution of the present invention has the following advantages:

In the detection method of the technical solution of the present invention, a number of initial detection pictures are obtained from a plurality of the pictures to be detected of each of the wafers to be detected; Splicing to form a detection picture; detecting the detection picture according to the defect detection library to obtain a detection result. By making full use of all the pixel information of the selected image, all the features of the image are completely preserved, so that the final detection accuracy is effectively improved.

Further, the method for obtaining a number of initial detection pictures from a number of the pictures to be detected includes: according to empirical rules, acquiring a shooting angle with the clearest defect image in the wafer to be detected of the same type; The to-be-detected picture is used as the initial detection picture, which reduces pseudo-pixel interference of other to-be-detected pictures, so that the final detection accuracy is effectively improved.

Description of drawings

1 is a schematic diagram of a picture processing structure of a defect detection method;

2 is a schematic diagram of a picture processing structure of another defect detection method;

3 to 15 are schematic structural diagrams of each step of an embodiment of the defect detection method of the present invention.

Detailed ways

As mentioned in the background art, the existing defect detection methods have low accuracy. The following will be described in detail with reference to the accompanying drawings.

The existing wafer defect detection process includes two methods. One method is: taking several pictures to be inspected obtained from multi-angle shooting, mixing the pictures according to a random pixel ratio, and then obtaining the inspection pictures (as shown in FIG. 1 ). Although this method makes full use of all the pixel information, it will introduce some very unnatural pseudo-pixel information, which will affect the final detection judgment and reduce the detection accuracy.

Another method is: splicing several pictures to be detected obtained by shooting at multiple angles according to the method of randomly intercepting local parts, and then obtaining the detected pictures (as shown in FIG. 2 ). This method cannot guarantee that all the defective positions of the picture will be intercepted, which will also affect the final detection judgment and reduce the detection accuracy.

On this basis, the present invention provides a defect detection method, which obtains a number of initial detection pictures from a number of the pictures to be detected; splices a plurality of the initial detection pictures to form a detection picture; The detection picture is used for detection, and the detection result is obtained. By making full use of all the pixel information of the selected image, all the features of the image are completely preserved, so that the final detection accuracy is effectively improved.

In order to make the above objects, features and advantages of the present invention more clearly understood, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

3 to 15 are schematic structural diagrams of a process of forming a semiconductor structure according to an embodiment of the present invention.

In this embodiment, before the defect detection of the wafer is performed, a defect database needs to be obtained first. The defect database includes several defect data, and the defect of the wafer is judged according to the defect data stored in the defect database, and then the final result is given. test results. For the specific acquisition process of the defect database, please refer to FIG. 3 to FIG. 11 .

In this embodiment, acquiring the defect database is divided into two stages. The first stage is to form an initial defect database through training and learning; the second stage is to further test and verify the initial defect database to obtain the defect database. For the training and learning process, please refer to FIG. 3 to FIG. 6 .

Referring to FIG. 3 , several training wafers 100 are provided, and each of the training wafers 100 has several defects 101 .

In this embodiment, samples are provided for defect data in the defect database through defects 101 on the training wafer 100 .

Referring to FIG. 4 , each of the training wafers 100 is photographed from different angles to obtain a plurality of pictures 102 to be trained for each of the training wafers 100 .

Since one shooting angle cannot fully reflect the defect 101 characteristics of the training wafer 100, in this embodiment, each of the training wafers 100 is shot from different angles to ensure a more accurate and comprehensive picture. Obtain defect information on the training wafer 100 .

In this embodiment, the shooting angles include vertically downward, 45° downward from the upper left, and 45° downward from the upper right.

Referring to FIG. 5, several initial training pictures 103 are obtained from several of the to-be-trained pictures 102 of each piece of the training wafer 100; several initial training pictures 103 of each piece of the training wafer 100 are spliced to form a training Picture 104.

In this embodiment, the method for acquiring several initial training pictures 103 from several of the to-be-trained pictures 102 of each of the training wafers 100 includes: acquiring the same type of defect images in the training wafer 100 according to an empirical rule The clearest shooting angle; the to-be-trained picture 102 shot at the shooting angle is used as the initial training picture 103 .

In this embodiment, the same type includes one or more of the same batch, the same process step, and the same wafer.

Since some of the shooting angles in each shooting angle cannot clearly reflect the defect information of the training wafer 100, if these images to be trained 102 that cannot clearly reflect the defect information are also selected, a lot of pseudo pixels will be added. interference, which in turn affects the final detection accuracy.

The staff knows the best shooting angles corresponding to different types of training wafers 100 through the experience rules accumulated over a long period of time. The pictures taken at these shooting angles can clearly reflect the defect information. 102 is used as the initial training picture 103, which can reduce the pseudo-pixel interference of the remaining pictures 102 to be trained, so that the final detection accuracy can be effectively improved.

In this embodiment, the method for forming a training picture 104 by splicing several initial training pictures 103 of each of the training wafers 100 includes: forming the training picture 104 by splicing a plurality of the initial training pictures 103 along the first direction X Picture 104; In other embodiments, a method for splicing several initial training pictures of each of the training wafers to form a training picture includes: forming the training picture by splicing a plurality of the initial training pictures along the second direction Y , the first direction X is perpendicular to the second direction Y.

In this embodiment, each of the training wafers 100 shoots three pictures 102 to be trained, and selects two of them as the initial training pictures 103 as an example.

In this embodiment, by making full use of all the pixel information of the selected picture, all the features of the picture are completely preserved, so that the final detection accuracy is effectively improved.

Referring to FIG. 6 , image recognition processing is performed on each of the training pictures 104 to obtain initial defect data; the initial defect database is composed of a plurality of the initial defect data.

In this embodiment, the image of the training picture 104 is scanned by image recognition processing, the defects in the training picture 104 are identified, and the defects are converted into corresponding defect data for storage.

In this embodiment, the defect data corresponding to each training wafer 100 further includes type information of the training wafer 100 .

In this embodiment, the learning model used in the training and learning process is a convolutional neural network model.

In this embodiment, the initial defect database is obtained through training and learning, but the reliability of the initial defect database has not been verified. Therefore, after the training and learning are performed, the initial defect database needs to be tested and learned, and then the defect database is acquired, so as to improve the reliability of the initial defect data. Please refer to FIG. 7 to FIG. 11 for the process of the test learning.

Referring to FIG. 7 , several sets of test wafer sets are provided, and each set of the test wafer sets includes several test wafers 200 .

In this embodiment, the purpose of providing several groups of test wafer groups is that after completing the test learning of a group of the test wafer groups, the test accuracy rate of the initial defect database needs to be calculated. If the test accuracy rate reaches a threshold value , then the test learning is completed; when the test accuracy rate does not reach the threshold, the test learning is performed through the next group of the test wafer groups until the test accuracy rate reaches the threshold.

The following will take a group of the test wafer groups as an example to describe the process of the test learning.

Referring to FIG. 8 , each of the test wafers 200 is photographed from different angles to obtain several pictures 201 to be tested of each of the test wafers 200 .

In this embodiment, the acquisition process of the picture to be tested 201 is the same as the acquisition process of the picture to be trained 102 , which will not be repeated here.

Referring to FIG. 9, several initial test images 202 are obtained from several of the to-be-tested images 201 of each of the test wafers 200; several initial test images 202 of each of the test wafers 200 are spliced to form a test Picture 203.

In this embodiment, the process of acquiring a plurality of the initial test pictures 202 and forming the test pictures 203 is the same as the process of acquiring a plurality of the initial training pictures 103 and forming the training pictures 104 , which will not be repeated here. .

It should be noted that the shooting angles of the test wafer 200 and the training wafer 100 of the same type should be consistent, that is, the shooting angle corresponding to the initial test picture 202 is consistent with the shooting angle corresponding to the initial training picture 103 .

Referring to FIG. 10 , image recognition processing is performed on each of the test pictures 203 to obtain test data.

In this embodiment, the process of performing image recognition processing on the test picture 203 is the same as the process of performing image recognition processing on the training picture 104, which will not be repeated here.

Please refer to FIG. 11 to obtain the real data of each of the test pictures 203; the test data is detected according to the initial defect database, and judgment data is obtained from a number of the initial defect data; When the real data exceeds the preset range, the real data is replaced with the corresponding initial defect data, and is updated to the initial defect database.

Since it is a process of testing and learning, the real data is the result of human judgment, and the judgment of the initial defect database is verified through human judgment, thereby verifying the reliability of the initial defect data.

When the human judgment is inconsistent with the judgment of the initial defect database, it is considered that the judgment of the initial defect database is wrong, and then the corresponding test wafer 200 is re-trained and learned, so as to improve the test accuracy of the initial defect database.

After completing the test learning of a group of the test wafer groups, the test accuracy rate of the defect database is calculated, if the test accuracy rate reaches the threshold, then the test learning is completed; when the test accuracy rate does not reach the threshold, This is tested through the next set of test wafer sets until the test accuracy reaches a threshold. So far, the process of acquiring the defect database is over, and the model with the best accuracy rate in the test learning process is retained.

In this embodiment, the learning model used in the test learning process is a convolutional neural network model.

Referring to FIG. 12 , a number of wafers 300 to be inspected are provided.

In this embodiment, it is unknown whether the wafer 300 to be inspected has defects, so it is necessary to judge the wafer 300 to be inspected through the acquired defect database, and then obtain the inspection result.

Referring to FIG. 13 , each of the wafers 300 to be inspected is photographed from different angles, and several images 301 to be inspected of each wafer 300 to be inspected are obtained.

In this embodiment, the acquisition process of the picture to be detected 301 is the same as the acquisition process of the picture to be trained 102 , which will not be repeated here.

Referring to FIG. 14 , a plurality of initial inspection pictures 302 are obtained from a plurality of the to-be-inspected pictures 301 of each of the to-be-inspected wafers 300 ; Splicing to form a detection picture 303 .

In this embodiment, the process of acquiring a plurality of the initial detection pictures 302 and forming the detection pictures 303 is the same as the process of acquiring a plurality of the initial training pictures 103 and forming the training pictures 104 , which will not be repeated here. .

It should be noted that the selection of the shooting angles of the same type of wafers 300 to be tested and the training wafers 100 should be consistent, that is, the shooting angles corresponding to the initial testing pictures 302 are consistent with the shooting angles corresponding to the initial training pictures 103 . In this way, subsequent detections can be performed under the same pixel conditions.

Referring to FIG. 15, the inspection picture 303 is inspected according to the defect inspection library to obtain inspection results.

In this embodiment, the detection picture 303 is detected according to the defect detection library, and the method for obtaining the detection result includes: performing image recognition processing on each of the detection pictures 303 to obtain detection data; Compare with several defect data in the defect database; within a certain comparison error range, when the detection data is the same as any of the defect data, obtain the corresponding defect data, and use the defect data as the test results.

In this embodiment, the process of performing image recognition processing on the detection picture 303 is the same as the process of performing image recognition processing on the training picture 104 , which will not be repeated here.

In this embodiment, the inspection data corresponding to each wafer 300 to be inspected further includes type information of the wafer to be inspected 300 .

It should be noted that, before the detection data is compared with the defect data, it is necessary to ensure that the type information corresponding to the detection data is consistent with the type information corresponding to the defect data.

In this embodiment, several initial inspection pictures 302 are obtained from a plurality of the to-be-inspected pictures 301 of each of the to-be-inspected wafers 300 ; 302 is spliced to form a detection picture 303; the detection picture 303 is detected according to the defect detection library, and a detection result is obtained. By making full use of all the pixel information of the selected image, all the features of the image are completely preserved, so that the final detection accuracy is effectively improved.

Although the present invention is disclosed above, the present invention is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention should be based on the scope defined by the claims.

Claims (11)

1. a defect detection method, is characterized in that, comprises: Obtaining a defect database, the defect database includes several defect data; Provide a number of wafers to be inspected; Photograph each of the wafers to be inspected from different angles, and obtain several images to be inspected of each of the wafers to be inspected; Obtain a number of initial detection pictures from a plurality of the to-be-detected pictures of each of the to-be-detected wafers; splicing several of the initial detection pictures of each of the to-be-detected wafers to form a detection picture; The detection picture is detected according to the defect detection library, and a detection result is obtained. 2. The defect detection method according to claim 1, wherein the method for obtaining the defect database comprises: providing several training wafers, each of which has several defects; The training wafer is trained and learned, and an initial defect database is obtained, and the initial defect database includes several initial defect data; several groups of test wafer groups are provided, and each group of the test wafer groups includes several test wafers; The test wafer group performs test learning on the initial defect database to obtain the defect database. 3. The defect detection method according to claim 2, wherein training and learning are performed on each of the training wafers, and the method for obtaining an initial defect database comprises: photographing each of the training wafers from different angles , obtain a number of pictures to be trained for each piece of the training wafer; obtain a number of initial training pictures from a number of pictures to be trained for each piece of the training wafer; Perform splicing to form a training picture; perform image recognition processing on each of the training pictures to obtain initial defect data; and form the initial defect database from a plurality of the initial defect data. 4. The defect detection method according to claim 3, wherein the method for obtaining a plurality of initial training pictures from a plurality of the pictures to be trained in each of the training wafers comprises: obtaining all the same type of pictures according to empirical rules. The most clear shooting angle of the defect image in the training wafer; the picture to be trained taken from the shooting angle is used as the initial training picture. 5 . The defect detection method according to claim 3 , wherein the initial defect database is tested and learned according to a plurality of the test wafer groups, and the method for obtaining the defect database comprises: analyzing each wafer from different angles. 6 . The test wafer is photographed to obtain several pictures to be tested of each of the test wafers; several initial test pictures are obtained from several pictures to be tested of each of the test wafers; Several initial test pictures of the wafer are spliced to form a test picture; image recognition processing is performed on each of the test pictures to obtain test data; the real data of each of the test pictures is obtained; The test data is detected, and judgment data is obtained from a number of the initial defect data; when the judgment data and the real data exceed the preset range, the corresponding test wafer is re-trained and learned; when a group of the After the test of the test wafer group is completed, the test accuracy rate of the test wafer group of this group is calculated; when the test accuracy rate of the test wafer group of this group is lower than the set threshold, then the next group of the test is performed The test learning of the wafer group is performed until the test accuracy rate of the test wafer group in this group reaches the set threshold, so as to obtain the defect database and retain the learning model with the best accuracy rate in the test learning process. 6. The defect detection method according to claim 5, wherein the method for obtaining a plurality of initial test pictures from a plurality of the pictures to be tested of each of the test wafers comprises: according to empirical rules, obtaining the same type of The photographing angle with the clearest defect image in the test wafer; the picture to be tested photographed at the photographing angle is used as the initial test picture. 7. The defect detection method according to claim 1, wherein the method for obtaining a plurality of initial detection pictures from a plurality of the pictures to be detected comprises: according to empirical rules, obtaining defects in the wafers to be detected of the same type The shooting angle with the clearest image; the to-be-detected picture captured by the shooting angle is used as the initial detection picture. 8. The defect detection method according to claim 4, 6 or 7, wherein the same type includes one or more of the same batch, the same process step and the same wafer. 9. The defect detection method according to claim 1, wherein the detection picture is detected according to the defect detection library, and the method for obtaining the detection result comprises: performing image recognition processing on each of the detection pictures, Obtain inspection data; compare the inspection data with several defect data in the defect database respectively; within a certain comparison error range, when the inspection data is the same as any of the defect data, obtain the corresponding defect data, and use the defect data as the detection result. 10. The defect detection method according to claim 1, wherein the method for splicing a plurality of the initial detection pictures to form a detection picture comprises: splicing a plurality of the initial detection pictures along a first direction or a second direction Docking to form the detection picture, the first direction is perpendicular to the second direction. 11 . The defect detection method according to claim 2 , wherein the learning model used in the training learning and the testing learning process comprises a convolutional neural network model. 12 .

Images