CN116342422A

CN116342422A - Defect identification method based on wafer map denoising

Info

Publication number: CN116342422A
Application number: CN202310321620.4A
Authority: CN
Inventors: 穆晓宝; 易丛文; 夏敏
Original assignee: Shenzhen Zhixian Future Industrial Software Co ltd
Current assignee: Shenzhen Zhixian Future Industrial Software Co ltd
Priority date: 2023-03-29
Filing date: 2023-03-29
Publication date: 2023-06-27

Abstract

The invention relates to a defect identification method based on wafer map denoising, which comprises the following steps: obtaining wafer map data, wherein the wafer map data comprises a defective wafer map and a non-defective wafer map; based on connectivity of defect points on each wafer map, polymerizing the defect points to obtain a plurality of defect clusters; removing noise clusters on the wafer map, wherein the noise clusters are defect clusters with defect areas smaller than a preset first threshold value in the plurality of defect clusters; determining a wafer map still containing defect clusters as a defective wafer map; and performing defect detection on the defective wafer map to obtain a corresponding defect type.

Description

Defect identification method based on wafer map denoising

Technical Field

The invention relates to the field of semiconductor manufacturing, in particular to a defect identification method based on wafer map denoising.

Background

In the manufacture of semiconductors, identifying defects in wafer maps is an important step. Before identifying the defective pattern of the wafer map, the wafer map is usually subjected to denoising treatment, so that noise points irrelevant to defects on the pattern are removed, and then the process of dividing and identifying the defective pattern can be performed.

In the existing scheme, a DBSCAN algorithm is generally adopted, and defect points in a wafer map are clustered based on two core parameters of a preset neighborhood radius and a minimum clustering point number, so that a defect pattern and a noise point are distinguished. However, the DBSCAN algorithm has many problems in terms of denoising the wafer map, mainly including: the algorithm cannot deal with the situation that the defect pattern is disconnected, when the defect pattern is disconnected, the part with smaller area can be judged as noise point and can be removed by mistake; the algorithm core parameters are not easy to determine, the small change of the core parameters can have great influence on the final denoising result, and meanwhile, even if a group of more reasonable core parameters can be determined, the situation of broken defect patterns can not be processed.

Disclosure of Invention

One or more embodiments of the present disclosure describe a method for identifying defects based on denoising a wafer map, which does not use density-based clustering such as DBSCAN algorithm, but clusters interconnected defect points into the same cluster based on connectivity of the defect points, so as to more accurately distinguish defect patterns from noise points.

The specification provides a defect identification method based on wafer map denoising, which comprises the following steps:

obtaining wafer map data, wherein the wafer map data comprises a defective wafer map and a non-defective wafer map;

based on connectivity of defect points on each wafer map, polymerizing the defect points to obtain a plurality of defect clusters;

removing noise clusters on the wafer map, wherein the noise clusters are defect clusters with defect areas smaller than a preset first threshold value in the plurality of defect clusters;

determining a wafer map still containing defect clusters as a defective wafer map;

and performing defect detection on the defective wafer map to obtain a corresponding defect type.

In one possible implementation manner, based on connectivity of defect points on each wafer map, the defect points are aggregated to obtain a plurality of defect clusters, including:

setting the defect points on the wafer map to be in an unmarked state, and carrying out multi-round clustering operation on the defect points until all the defect points are marked; wherein, any round of clustering operation includes:

selecting one defect point from untagged defect points, determining the defect point as a target point and marking the target point;

searching for unmarked communication points of the target point, if the target point has unmarked communication points, determining all unmarked communication points as new target points and marking, and repeating the steps;

if all the target points have no unmarked connected points, determining all marked defect points in the current round of clustering operation as a defect cluster, and carrying out the next round of clustering operation.

In one possible embodiment, the communication point of the target point is a defect point located at eight directions adjacent to each other, i.e., above, below, left, right, above-left, below-left, above-right, and below-right of the target point.

In one possible embodiment, before performing defect detection on the defective wafer map, the method further includes:

and performing size transformation on the defective wafer map to a preset size.

In one possible implementation, the first threshold is determined based on area statistics of defect clusters in a sample wafer map that are known to be defect free in advance.

In one possible implementation manner, performing defect detection on the defective wafer map to obtain a corresponding defect type, including:

dividing the wafer map by using a U-Net image dividing algorithm according to the pixel array corresponding to the wafer map to obtain a plurality of defect subgraphs containing single wafer defects;

inputting target defect subgraphs in the plurality of defect subgraphs into a defect detection model to obtain corresponding defect types.

In a possible implementation manner, before inputting the target defect subgraph of the plurality of defect subgraphs into the defect detection model, the method further includes:

extracting first characteristics of any defect subgraph to obtain subgraph characteristics;

searching and comparing in a preset feature set according to the subgraph features, and determining whether the defect subgraph belongs to a known defect type or not; the predetermined feature set is composed of map features of known defect types;

in case of belonging to a known defect type, the defect subgraph is determined as the target defect subgraph.

In a possible implementation manner, the searching and comparing in a predetermined feature set according to the sub-graph features to determine whether the defect sub-graph belongs to a known defect type includes:

searching K candidate samples with highest similarity with the sub-graph features in the preset feature set;

and determining whether the defect subgraph belongs to a known defect type according to the similarity of the subgraph characteristics and the K candidate samples.

In a possible implementation manner, inputting the target defect subgraph of the plurality of defect subgraphs into a defect detection model to obtain a corresponding defect type includes:

performing second feature extraction on the target defect subgraph to obtain target features;

inputting the target characteristics into a full-connection layer, and inputting the output of the full-connection layer into a Softmax function to obtain probabilities corresponding to various known defect types;

and selecting the known defect type with the highest probability as the defect type.

In a possible implementation manner, the second feature extraction includes: and performing feature extraction on the target defect subgraph by using a convolutional neural network CNN or ResNet model.

One or more embodiments of the present disclosure describe a method for identifying defects based on denoising a wafer map, which does not use density-based clustering such as DBSCAN algorithm, but clusters interconnected defect points into the same cluster based on connectivity of the defect points, so as to more accurately distinguish defect patterns from noise points. Meanwhile, the wafer map with the defect points on the wafer map after denoising can be regarded as a defect-free wafer map in the denoising stage, so that the defect pattern of the wafer map can be directly processed without considering the defect-free situation when the defect pattern of the wafer map is detected and classified later, and the whole process of defect pattern recognition is simplified.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments disclosed in the present specification, the drawings required for the description of the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only examples of the embodiments disclosed in the present specification, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a block diagram of a method for identifying defects based on wafer map denoising according to an embodiment of the present invention;

FIG. 2 is a flow chart of a method for identifying defects based on wafer map denoising according to an embodiment of the present invention;

FIG. 3 is a flow chart of a method for aggregating defective points on a wafer map according to an embodiment of the present invention;

fig. 4 is a flowchart of a method for performing defect detection on a defective wafer map according to an embodiment of the present invention.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Fig. 1 illustrates a frame diagram of a method for identifying defects based on wafer map denoising, according to one embodiment. As shown in fig. 1, after the original wafer map is obtained, connectivity noise point removal based on the defect points is performed first, after the noise points are removed, the wafer map without the defect points on the surface of the pattern is a non-defective wafer map, and the processing is not continued in the subsequent steps; and (3) firstly adjusting the sizes of the defective wafer maps with the defect points on the surface of the pattern to be uniform, then dividing the defect pattern on each wafer map into a plurality of wafer maps with single defects, and then carrying out defect type identification on the single defect wafer map to obtain the types of the defects on the wafer map.

The following description will proceed with reference being made to the drawings, which are not intended to limit the scope of embodiments of the invention.

Fig. 2 is a flowchart of a method for identifying defects based on denoising a wafer map according to an embodiment of the present invention. As shown in fig. 2, the method at least includes: step 201, obtaining wafer map data, wherein the wafer map data comprises a defective wafer map and a non-defective wafer map; step 202, based on connectivity of defect points on each wafer map, aggregating the defect points to obtain a plurality of defect clusters; step 203, removing noise clusters on the wafer map, wherein the noise clusters are defect clusters with defect areas smaller than a preset first threshold value in the plurality of defect clusters; step 204, determining the wafer map still containing the defect clusters as a defective wafer map; and 206, performing defect detection on the defective wafer map to obtain a corresponding defect type.

In step 201, wafer map data is obtained, wherein the wafer map data includes a defective wafer map and a non-defective wafer map.

The acquisition mode may be various, for example, wafer map data may be directly acquired from an acquisition device of each machine in the semiconductor manufacturing process, or wafer map data may be acquired by calling from a database, where the wafer map data generated by each machine in the semiconductor manufacturing process is stored.

In step 202, based on connectivity of defect points on each wafer map, the defect points are aggregated to obtain a plurality of defect clusters.

Specifically, an embodiment of step 202 is shown in FIG. 3. Fig. 3 is a flowchart of a method for aggregating defective points on a wafer map according to an embodiment of the present invention, and as shown in fig. 3, the method includes a step 310 and sub-steps 311 to 313 thereof.

In step 310, the defective points on the wafer map are set to be in an unmarked state, and a plurality of rounds of clustering operation are performed on the defective points until all the defective points are marked; wherein, any round of clustering operation includes steps 311 to 313.

In step 311, one defect point is selected among the untagged defect points to be determined as a target point and marked.

There are various ways of selection, for example, a defective point may be randomly selected as the target point among the defective points that are not marked; the first unlabeled defect point may also be selected as the target point in a sequential order, e.g., from top left to bottom right. The description is not intended to be limiting.

In step 312, the unlabeled connectivity points of the target point are found, if there are connectivity points of the target point that are not labeled, all connectivity points that are not labeled are determined as new target points and labeled, and then step 312 is repeated.

Specifically, the points on the wafer map are distributed in a two-dimensional plane lattice, and the connection points of one defect point are defect points at eight adjacent positions in the directions of up, down, left, right, left up, left down, right up and right down of the defect point.

In step 313, if all the target points have no unlabeled connected points, all the labeled defective points in the current round of clustering operation are determined as one defective cluster, and the next round of clustering operation is performed.

In one embodiment, steps 311 through 313 may be implemented based on a stack. First, in step 311, a defective point is selected from the defective points that are not marked, and the defective points are pressed onto a stack after marking. Then, in step 312, when the stack is not empty, a defect point is popped from the stack top, all the unmarked connection points of the defect point are found, all the unmarked connection points are marked and pushed into the stack, then step 312 is repeated, and when the stack is empty, the process goes to step 313. Finally, when the stack is empty, all marked defect points in the current round of clustering operations are determined as one defect cluster, and the next round of clustering operations is performed in step 313.

Returning to fig. 2, in step 203, the noise clusters on the wafer map are removed, where the noise clusters are defect clusters with a defect area smaller than a preset first threshold value in the plurality of defect clusters.

After the noise-removed clusters, the rest of the wafer map is true defect clusters.

In one embodiment, the first threshold is determined based on area statistics of defect clusters in a sample wafer map that are known to be defect free in advance.

In step 204, the wafer map that still contains defective clusters is determined to be a defective wafer map.

Through step 204, the wafer map to be inspected can be divided into a defective wafer map and a non-defective wafer map at the denoising stage of the pretreatment, and then in step 206, only the defective wafer map is inspected, and the situation of no defect is not considered any more, so that the whole process of defect inspection is simplified.

In step 206, defect detection is performed on the defective wafer map, so as to obtain a corresponding defect type.

Specifically, an embodiment of step 206 is shown in FIG. 4. Fig. 4 is a flowchart of a method for performing defect detection on a defective wafer map according to an embodiment of the present invention. As shown in fig. 4, the method includes steps 401 to 405.

In step 401, the wafer map is segmented by using a U-Net image segmentation algorithm according to the pixel array corresponding to the wafer map, so as to obtain a plurality of defect subgraphs including single wafer defects.

When the wafer map only contains a single defect, the algorithm directly outputs the pattern of the defect, the coordinate information of the wafer defect and the number of pixel points; when the wafer map includes a plurality of defects (overlapping defects), the algorithm divides the plurality of defects into a plurality of individual defects, and outputs the pattern of each wafer defect, and the coordinate information and the number of pixels of the wafer defect. The number of pixels of a wafer defect may be used to calculate the area of the wafer defect. It will be appreciated that the shape and size of each defect map is the same as the wafer map before singulation, except that the defect in the pattern becomes a sub-defect of the defect in the original wafer map.

As will be appreciated by those skilled in the art, the U-Net image segmentation algorithm is one algorithm that can implement pixel-level image segmentation.

In step 402, a first feature extraction is performed on any defect subgraph, resulting in subgraph features.

In a specific embodiment, a convolutional neural network CNN is used to perform a first feature extraction on the arbitrary defect subgraph, so as to obtain subgraph features.

In step 403, searching and comparing in a predetermined feature set according to the sub-graph features, and determining whether the defect sub-graph belongs to a known defect type; the predetermined feature set is made up of map features of known defect types.

In one embodiment, in the preset feature set, K candidate samples with highest similarity with the sub-graph features are searched out; and determining whether the defect subgraph belongs to a known defect type according to the similarity of the subgraph characteristics and the K candidate samples.

Specifically, after searching and comparing the sub-graph features and the graph features of known defect types in the preset feature set, outputting the first K candidate search samples with the highest similarity, and judging whether the defect sub-graph belongs to the known defect types according to the similarity. For example, a similarity threshold may be set, and when the similarity is greater than or equal to a preset similarity threshold, the defect is considered to be a known defect type; and when the similarity is lower than a preset similarity threshold, the defect is considered to be of an unknown defect type.

In step 404, the defect subgraph is determined as the target defect subgraph in case it belongs to a known defect type.

If the defect subgraph belongs to an unknown defect type, the existing information is directly output, and the defect subgraph is sent to a related engineer for manual judgment, so that the follow-up steps are not performed. After the manual judgment, if the system is found to be misjudged, namely the defect actually belongs to the known defect, the defect is added into a preset characteristic set formed by the known defect; if the defect is found to belong to a true unknown defect, the defect is updated to a new type of defect and added into a predetermined feature set.

In step 405, a target defect subgraph of the plurality of defect subgraphs is input into a defect detection model, and a corresponding defect type is obtained.

In one embodiment, performing second feature extraction on the target defect subgraph to obtain target features; inputting the target characteristics into a full-connection layer, and inputting the output of the full-connection layer into a Softmax function to obtain probabilities corresponding to various known defect types; and selecting the known defect type with the highest probability as the defect type. Wherein the second feature extraction comprises: and performing feature extraction on the target defect subgraph by using a convolutional neural network CNN or ResNet model.

In another embodiment, a Detection Transformer model (hereinafter simply referred to as DETR model) is used as the defect detection model. As will be appreciated by those skilled in the art, the DETR model regards the problem of defect recognition of a wafer map as a problem of target detection in the field of computer vision, specifically, the input of the model firstly extracts features through a backhaul module, then spreads the features for position coding, and transmits the features into an Encoder of a transducer, and then obtains the defect type and detection frame of a target defect subgraph after processing and decoding by a Decoder.

Returning to fig. 2, in some possible embodiments, prior to step 206, the method further comprises: and 205, performing size transformation on the defective wafer map to a preset size.

Step 205 is used to unify wafer maps of different sizes from different tools to facilitate processing in subsequent steps.

It is noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises an element.

It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program for instructing relevant hardware, and the program may be stored in a computer readable storage medium, where the storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.

The foregoing description of the embodiments has been provided for the purpose of illustrating the general principles of the invention, and is not meant to limit the scope of the invention, but to limit the invention to the particular embodiments, and any modifications, equivalents, improvements, etc. that fall within the spirit and principles of the invention are intended to be included within the scope of the invention.

Claims

1. A defect identification method based on wafer map denoising comprises the following steps:

2. The method of claim 1, wherein aggregating defective points on each wafer map based on connectivity of the defective points to obtain a plurality of defective clusters, comprises:

3. The method according to claim 2, wherein the communication point of the target point is a defect point located adjacent to eight directions of up, down, left, right, up left, down left, up right, down right of the target point.

4. The method of claim 1, wherein prior to defect inspection of the defective wafer map, the method further comprises:

and performing size transformation on the defective wafer map to a preset size.

5. The method of claim 1, wherein the first threshold is determined based on prior statistics of areas of defect clusters in the sample wafer map that are known to be defect free.

6. The method of claim 1, wherein performing defect inspection on the defective wafer map to obtain a corresponding defect type comprises:

7. The method of claim 6, wherein prior to inputting a target defect subgraph of the number of defect subgraphs into a defect detection model, the method further comprises:

8. The method of claim 7, wherein determining whether the defect sub-graph is of a known defect type based on the sub-graph features being retrieved and compared in a predetermined feature set, comprises:

9. The method of claim 6, wherein inputting the target defect subgraph of the plurality of defect subgraphs into a defect detection model results in a corresponding defect type, comprising:

10. The method of claim 9, wherein the second feature extraction comprises: and performing feature extraction on the target defect subgraph by using a convolutional neural network CNN or ResNet model.