Defect classification device and method
By introducing a two-layer neural network architecture into the defect automatic classification system, and using the second neural network to extract feature information that cannot be observed in the defect classification model, the problem of low accuracy of automatic classification of defects in the prior art is solved, and accurate identification and automatic classification of multiple defect types are achieved.
Patent Information
- Application Number
- CN202510214416.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-25
- Publication Date
- 2025-06-06
AI Technical Summary
The existing automatic defect classification method is difficult to improve the accuracy rate, especially the inability to effectively classify the graphics of small defects, resulting in the phenomenon of "the part visible to the human eye cannot be seen by the machine".
A two-layer neural network architecture is adopted, the first neural network is used as the defect classification model, and the second neural network is used as the defect detection model. By extracting feature information that cannot be observed in the defect classification model, the accuracy of defect classification is improved.
It has achieved the accuracy of automatic defect classification, and can effectively identify various defect types including small defects, which has improved the accuracy and efficiency of automatic classification.
Smart Images
Figure CN120107682A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of semiconductor integrated circuit manufacturing, and in particular to a defect classification device; the present invention also relates to a defect classification method. Background Art
[0002] Currently, defect classification in wafer manufacturing plants (fabs) is mainly completed by manual classification by yield engineers (YE). Although the automatic defect classification (ADC) of the machine itself does basic classification, automatic classification cannot be achieved due to its low accuracy.
[0003] There is an existing method of automatic defect classification, which is to use AMTD combined with artificial intelligence to achieve automatic defect classification. The method is to apply the classification method of neural network and extract features from images to achieve automatic defect classification. However, in practice, it is found that the defect classification accuracy of the existing automatic defect classification method cannot be improved, especially for small defect graphics, which cannot be clearly distinguished from graphics, resulting in the phenomenon of "parts visible to the human eye cannot be seen by the machine". Summary of the invention
[0004] The technical problem to be solved by the present invention is to provide a defect classification device which can realize automatic defect classification and improve the accuracy. To this end, the present invention also provides a defect classification method.
[0005] In order to solve the above technical problems, the defect classification device provided by the present invention comprises:
[0006] The first neural network is a neural network for a defect classification model.
[0007] The second neural network is a neural network of a defect detection model. The wafer image is input into the input end of the second neural network. The defect detection model detects a first defect on the wafer image and extracts first feature information of the first defect.
[0008] The first feature information is also input into the input end of the first neural network, and the defect classification model also obtains the defect category of the first defect according to the first feature information.
[0009] A further improvement is that the input end of the first neural network also includes directly inputting the wafer image, and the defect classification model directly classifies the defects on the wafer image and outputs the corresponding defect category at the output end.
[0010] A further improvement is that the first defects at least include defects that cannot be observed by the defect classification model.
[0011] A further improvement is that the first characteristic information includes: position, size or shape.
[0012] A further improvement is that the first feature information also includes: height or boundary patch.
[0013] A further improvement is that a reference image is also input into the input end of the second neural network, and the defect detection model compares the wafer image with the reference image to obtain the first feature information.
[0014] A further improvement is that the first neural network includes a deep neural network, and the first neural network obtains the parameter values of the defect classification model through learning and training.
[0015] A further improvement is that the second neural network includes a deep neural network, and the second neural network obtains the parameter values of the defect detection model through learning and training.
[0016] In order to solve the above technical problems, the defect classification method provided by the present invention comprises the following steps:
[0017] The wafer image is input into a second neural network, where the second neural network is a neural network of a defect detection model. The defect detection model is used to detect a first defect on the wafer image and extract first feature information of the first defect.
[0018] The second neural network inputs the first feature information to the input end of the first neural network, the first neural network is a neural network of a defect classification model, and the defect classification model also obtains the defect category of the first defect according to the first feature information.
[0019] Further improvements include:
[0020] The wafer image is directly input into the input end of the first neural network, and the defect classification model directly classifies the defects on the wafer image and outputs the corresponding defect category at the output end.
[0021] A further improvement is that the first defects at least include defects that cannot be observed by the defect classification model.
[0022] A further improvement is that the first characteristic information includes: position, size or shape.
[0023] A further improvement is that the first feature information also includes: height or boundary patch.
[0024] A further improvement is that a reference image is also input into the input end of the second neural network, and the defect detection model compares the wafer image with the reference image to obtain the first feature information.
[0025] A further improvement is that the first neural network includes a deep neural network, and the first neural network obtains the parameter values of the defect classification model through learning and training.
[0026] A further improvement is that the second neural network includes a deep neural network, and the second neural network obtains the parameter values of the defect detection model through learning and training.
[0027] The present invention adds a defect detection model on the basis of the defect classification model. The defect detection model can extract feature information of defects that cannot be observed by the defect classification model. The first feature information extracted by the defect detection model will be input into the first neural network of the defect classification model to obtain the corresponding defect category. Since the defect detection model and the defect classification model are both implemented through neural networks, they can both run automatically without manual classification. Therefore, the present invention can realize automatic defect classification and improve the accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:
[0029] Figure 1 It is a structural block diagram of an existing defect classification device;
[0030] Figure 2 is a structural block diagram of a defect classification device according to an embodiment of the present invention;
[0031] Figure 3 is a schematic structural diagram of an example of a defect classification device according to an embodiment of the present invention;
[0032] Figure 4 It is a diagram of defect classification results achieved by the defect classification device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0033] like Figure 2 , which is a structural block diagram of a defect classification device according to an embodiment of the present invention; the defect classification device according to an embodiment of the present invention comprises:
[0034] The first neural network 201 is a neural network of a defect classification model, namely a defect classification network.
[0035] In the embodiment of the present invention, the first neural network 201 includes a deep neural network, and the first neural network 201 obtains the parameter values of the defect classification model through learning and training.
[0036] The second neural network 202 is a neural network of a defect detection model, i.e., a defect detection network. The wafer image 203 is input into the input end of the second neural network 202. The defect detection model detects a first defect on the wafer image 203 and extracts first feature information 205 of the first defect.
[0037] In the embodiment of the present invention, the first defects at least include defects that cannot be observed by the defect classification model.
[0038] The first feature information 205 includes: position, size or shape.
[0039] The first feature information 205 also includes: height or boundary patch.
[0040] The first feature information 205 is also input to the input end of the first neural network 201 , and the defect classification model also obtains the defect category of the first defect according to the first feature information 205 .
[0041] The reference image 204 is also input to the input end of the second neural network 202 , and the defect detection model compares the wafer image 203 with the reference image 204 to obtain the first feature information 205 .
[0042] In the embodiment of the present invention, the second neural network 202 includes a deep neural network, and the second neural network 202 obtains the parameter values of the defect detection model through learning and training.
[0043] In the embodiment of the present invention, the input end of the first neural network 201 also includes directly inputting the wafer image 203, and the defect classification model directly classifies the defects on the wafer image 203 and outputs the corresponding defect category at the output end.
[0044] The first neural network 201 can simultaneously process the first feature information 205 and the information from the wafer image 203, so it can achieve mixed processing of multiple different types of information. Among them, the first feature information 205 can make up for the defects in the wafer image 203203 that the defect classification model cannot identify, such as smaller defects, so it can finally improve the accuracy of defect classification.
[0045] The embodiment of the present invention adds a defect detection model on the basis of the defect classification model. The defect detection model can extract feature information of defects that cannot be observed by the defect classification model. The first feature information 205 extracted by the defect detection model will be input into the first neural network 201 of the defect classification model to obtain the corresponding defect category. Since the defect detection model and the defect classification model are both implemented through neural networks, they can both run automatically without manual classification. Therefore, the embodiment of the present invention can realize automatic defect classification and improve the accuracy. Similarly, since the embodiment of the present invention does not require manual classification, it can also improve efficiency.
[0046] The embodiment of the present invention modifies the network structure and adds a defect detection model on the basis of the classification model. By comparing the detection with the reference graph, the coordinates and morphological information of the defect are given to the defect classification neural network, thereby increasing the accuracy of defect detection.
[0047] The defect classification method according to the embodiment of the present invention comprises the following steps:
[0048] The wafer image 203 is input into the second neural network 202, which is a neural network of a defect detection model. The defect detection model is used to detect a first defect on the wafer image 203 and extract first feature information 205 of the first defect.
[0049] The second neural network 202 inputs the first feature information 205 to the input end of the first neural network 201 . The first neural network 201 is a neural network of a defect classification model. The defect classification model also obtains the defect category of the first defect according to the first feature information 205 .
[0050] The method of the embodiment of the present invention also includes:
[0051] The wafer image 203 is directly input to the input end of the first neural network 201, and the defect classification model directly classifies the defects on the wafer image 203 and outputs the corresponding defect category at the output end.
[0052] The first defects include at least defects that cannot be observed by the defect classification model.
[0053] The first feature information 205 includes: position, size or shape.
[0054] The first feature information 205 also includes: height or boundary patch.
[0055] The reference image 204 is also input to the input end of the second neural network 202 , and the defect detection model compares the wafer image 203 with the reference image 204 to obtain the first feature information 205 .
[0056] The first neural network 201 includes a deep neural network, and the first neural network 201 obtains parameter values of the defect classification model through learning and training.
[0057] The second neural network 202 includes a deep neural network, and the second neural network 202 obtains parameter values of the defect detection model through learning and training.
[0058] like Figure 3 FIG. 1 is a schematic diagram of a structure of an example of a defect classification device according to an embodiment of the present invention; in order to more clearly understand the embodiment of the present invention, the following is combined with Figure 3 The example is further explained as follows:
[0059] Figure 3 In the figure, the first neural network is marked solely with label 201a, the second neural network is marked solely with label 202a, the wafer image is marked solely with label 203a, the reference image is marked solely with label 204a, the first feature information is marked solely with label 205a, and the defect classification result is represented solely with label 206a.
[0060] An efficient hybrid encoder (Efficient Hybrid Encoder) 301 is adopted in the first neural network 201a, and the defect classification model adopted is the RT-DETR model.
[0061] The efficient hybrid encoder 301 includes an attention-based intra-scale feature interaction (AIFI) encoder 302 and a CNN-based cross-scale feature fusion module (CCFM) 303. The CCFM module 303 includes a plurality of fusion units 304.
[0062] The first neural network 201a also includes: an IOU-aware query selection mechanism (Iou-aware query selection) module 305 and a decoding and header (Decoder & header) module 306.
[0063] The second neural network 202a has been trained using data sets including Videomatte 240k, Photomatte 13k, and Distinctions-646.
[0064] The first feature information 205a includes four types, such as size (eg, 0.475 μ), shape, boundary patches, and height (predicted height).
[0065] The defect classification result 206a lists a variety of defects, among which buried particles are individually marked with a mark 206a1 and the photo is enlarged. It can be seen that the embodiment of the present invention can accurately classify the buried particles 206a1.
[0066] By adding a detection network, the embodiment of the present invention can provide the position, size, shape and even height information of the defect to the defect classification neural network, and combine this part of the "artificial labels" that cannot be observed by the machine to improve the overall classification accuracy.
[0067] like Figure 4 , which is a diagram of defect classification results achieved by the defect classification device according to an embodiment of the present invention; Figure 4 The corresponding result is the result of classifying the defects of the metal layer of the back-end-of-line (BEOL) process using the network of the embodiment of the present invention. Figure 4 , nine types of defects are shown, which are marked with labels 402a, 402b, 402c, 402d, 402e, 402f, 402g, 402h and 402i respectively.
[0068] Among them, defect 402a is a buried particle (buried PA), defect 402b is a bump defect, defect 402c is an etch barrier (block ET) defect, defect 402d is a line open defect, defect 402e is a high line open (Hline open) defect, defect 402f is a photoresist bubble (PR bubble) defect, defect 402g is a pattern fail defect, defect 402h is a residue defect, and defect 402i is a surface particle (surface PA).
[0069] Figure 4Table 401 is also provided, which provides data on the number, recall rate and accuracy of the above 9 types of defects. From the data in Table 401, it can be seen that the accuracy of defect classification in the embodiment of the present invention meets the requirements and is much higher than that of the existing ADC system.
[0070] The present invention has been described in detail above through specific embodiments, but these do not constitute limitations of the present invention. Without departing from the principle of the present invention, those skilled in the art may also make many variations and improvements, which should also be considered as the protection scope of the present invention.
Claims
1. A defect classification device, characterized in that: include: The first neural network is a neural network of a defect classification model; A second neural network is a neural network of a defect detection model, wherein an input end of the second neural network inputs a wafer image, and the defect detection model detects a first defect on the wafer image and extracts first feature information of the first defect; The first feature information is also input into the input end of the first neural network, and the defect classification model also obtains the defect category of the first defect according to the first feature information.
2. The defect classification device according to claim 1, characterized in that: The input end of the first neural network also includes directly inputting the wafer image, and the defect classification model directly classifies the defects on the wafer image and outputs the corresponding defect category at the output end.
3. The defect classification device according to claim 2, characterized in that: The first defects include at least defects that cannot be observed by the defect classification model.
4. The defect classification device according to claim 1, characterized in that: The first characteristic information includes: position, size or shape.
5. The defect classification device according to claim 4, characterized in that: The first feature information also includes: height or boundary patch.
6. The defect classification device according to claim 1, characterized in that: The input end of the second neural network also inputs a reference image, and the defect detection model compares the wafer image with the reference image to obtain the first feature information.
7. The defect classification device according to claim 1, characterized in that: The first neural network includes a deep neural network, and the first neural network obtains parameter values of the defect classification model through learning and training.
8. The defect classification device according to claim 1, characterized in that: The second neural network includes a deep neural network, and the second neural network obtains parameter values of the defect detection model through learning and training.
9. A defect classification method, characterized in that: The steps include: Inputting the wafer image into a second neural network, where the second neural network is a neural network of a defect detection model, using the defect detection model to detect a first defect on the wafer image and extract first feature information of the first defect; The second neural network inputs the first feature information to the input end of the first neural network, the first neural network is a neural network of a defect classification model, and the defect classification model also obtains the defect category of the first defect according to the first feature information.
10. The defect classification method according to claim 9, characterized in that: Also includes: The wafer image is directly input into the input end of the first neural network, and the defect classification model directly classifies the defects on the wafer image and outputs the corresponding defect category at the output end.
11. The defect classification method according to claim 10, characterized in that: The first defects include at least defects that cannot be observed by the defect classification model.
12. The defect classification method according to claim 9, characterized in that: The first characteristic information includes: position, size or shape.
13. The defect classification method according to claim 12, characterized in that: The first feature information also includes: height or boundary patch.
14. The defect classification method according to claim 9, characterized in that: The input end of the second neural network also inputs a reference image, and the defect detection model compares the wafer image with the reference image to obtain the first feature information.
15. The defect classification method according to claim 9, characterized in that: The first neural network includes a deep neural network, and the first neural network obtains parameter values of the defect classification model through learning and training.
16. The defect classification method according to claim 9, characterized in that: The second neural network includes a deep neural network, and the second neural network obtains parameter values of the defect detection model through learning and training.