A method and system for handling semiconductor defects

Abstract

The application discloses a kind of processing method and system of semiconductor blemish product, the method includes the following steps: after receiving semiconductor blemish product, corresponding first particle MAP chart of semiconductor blemish product is acquired from different angles respectively;First flaw detection model is obtained, and according to first particle MAP chart, semiconductor blemish product is carried out flaw detection using first flaw detection model, to obtain the flaw form and flaw position of semiconductor blemish product;Semiconductor blemish product is automatically distributed to corresponding processing flow according to flaw form and flaw position, and based on processing flow, semiconductor blemish product is reprocessed, to obtain secondary processing product;Flaw detection is carried out again to secondary processing product, and whether secondary processing product is scrapped according to detection result is judged. It can be automatically distributed to corresponding processing flow in blemish product and carry out secondary processing treatment, so as to reduce material loss, improve yield.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor defect processing, and particularly relates to a semiconductor defect processing method and system. BACKGROUND

[0002] In the semiconductor production process, usually needs to go through grinding, polishing and other processes, wherein, each process may leave marks on the semiconductor product due to operation errors or equipment aging, thereby causing the semiconductor product to be a defective product.

[0003] In the related art, in order to ensure product quality, a defect detection algorithm is usually used to detect defects of the produced semiconductor product, so as to select good products for use, and to concentrate on the defective product for scrap processing, which not only causes material waste, but also greatly affects the production yield. SUMMARY

[0004] The present application provides a semiconductor defect processing method, which can automatically detect defects of the semiconductor defective product, and automatically distribute the defective product to the corresponding processing flow according to the detected defect shape and defect position for secondary processing, thereby reducing material loss and greatly improving the yield.

[0005] The technical scheme adopted by the present application is as follows:

[0006] A semiconductor defect processing method comprises the following steps: S1, after receiving the semiconductor defective product, acquiring a first particle MAP image corresponding to the semiconductor defective product from different angles respectively; S2, acquiring a first defect detection model, and using the first defect detection model to detect defects of the semiconductor defective product according to the first particle MAP image, to acquire the defect shape and defect position of the semiconductor defective product; S3, automatically distributing the semiconductor defective product to the corresponding processing flow according to the defect shape and defect position, and reprocessing the semiconductor defective product based on the processing flow, to obtain a secondary processing product; S4, detecting defects of the secondary processing product again, and judging whether the secondary processing product is processed for scrap according to the detection result.

[0007] In an embodiment of the present application, step S2 comprises the following steps: S21, obtaining a plurality of unlabeled second particle MAP graphs from a database, and obtaining third particle MAP graphs with resolution level labels in a splicing manner based on the second particle MAP graph samples, wherein each unlabeled second particle MAP graph contains at least one morphological defect, and each third particle MAP graph contains a plurality of morphological defects; S22, establishing a defect detection network, wherein the defect detection network comprises an image compression unit and an image decompression unit; S23, performing upstream training on the defect detection network based on a self-generated learning manner according to the plurality of unlabeled second particle MAP graphs, to obtain a second defect detection model; and S24, performing downstream training on the second defect detection model according to the plurality of third particle MAP graphs with resolution level labels, to obtain the first defect detection model.

[0008] In an embodiment of the present application, step S23 specifically comprises the following steps: S231, performing image set expansion on the second particle MAP graph using different image processing strategies to obtain a fourth particle MAP graph; S232, inputting the fourth particle MAP graph into the image compression unit for image compression processing to obtain compressed image features, and calculating a first generation value based on a full sequence comparison algorithm according to the compressed image features; S233, after the compressed image features are subjected to image decompression processing by the image decompression unit, calculating a corresponding second generation value based on a fragment comparison algorithm; and S234, adjusting and optimizing the correlation coefficients in the image compression unit and the image decompression unit according to the first generation value and the second generation value.

[0009] In an embodiment of the present application, step S232 specifically comprises: S2321, projecting the compressed image features into a high-dimensional feature space through a feature space to obtain high-dimensional feature descriptors of the compressed image features in the high-dimensional feature space; and S2322, calculating a first generation value between the high-dimensional feature descriptors using a first cost function.

[0010] In an embodiment of the present application, the first cost function is generated by the following formula:

[0011] ,

[0012] wherein, represents a positive class feature representation, represents a negative class feature representation, represents a negative class feature representation set, represents a randomness adjustment coefficient, represents a generation value output by the first cost function.

[0013] In one embodiment of the present application, step S233 specifically comprises the following steps: S2331, inputting the fourth particle MAP graph into a convolution operation layer with a kernel size same as the size of the patch selection range, calculating the weight of each patch selection range by moving a window, and selecting a target patch selection range according to the weight; wherein, each time the selection is performed, the selected target patch selection range is removed from each patch selection range; S2332, obtaining a patch selection feature map corresponding to the selected target patch selection range in the decompressed MAP graph, and performing average down-sampling on the patch selection feature map to obtain a corresponding patch selection feature descriptor; S2333, calculating the second generation value between the patch selection feature descriptors by using the first cost function.

[0014] In one embodiment of the present application, step S3 comprises the following steps: S31, obtaining a plurality of groups of historical processing information corresponding to the processing procedure; wherein, each group of historical processing information comprises historical processing data and corresponding historical removal amount, and the historical processing data comprises applied pressure, processing time and processing rate of the semiconductor surface; S32, obtaining a removal amount evaluation model corresponding to the processing procedure based on the historical processing data and the corresponding historical removal amount; S33, obtaining optimal processing data according to the removal amount evaluation model; wherein, the deviation between the removal amount evaluation value output by the removal amount evaluation model corresponding to the optimal processing data and the ideal removal amount is the smallest; S34, reprocessing the semiconductor defective product according to the optimal processing data.

[0015] In one embodiment of the present application, step S32 comprises the following steps: S321, generating an initial estimation model according to the historical processing data and the corresponding historical removal amount by the following formula:

[0016] ,

[0017] wherein, represents the removal amount evaluation value, represents the intercept term, represents the first-order term skewness coefficient, represents the second-order term skewness coefficient, represents the jth processing data in the historical processing data, represents the interference term; S322, establishing a second cost function corresponding to the initial estimation model by the following formula:

[0018] ,

[0019] wherein, represents the output value of the second cost function, represents the historical removal amount corresponding to the ith group of historical processing data, represents the removal amount evaluation value corresponding to the i-th group of historical processing data, represents the number of groups of the historical processing data; S323, respectively input each group of historical processing data and the corresponding historical removal amount into the second cost function, and obtain the intercept term, the first-order term skewness coefficient and the second-order term skewness coefficient in the initialization estimation model by using the least square method to minimize the second cost function, so as to generate the removal amount evaluation model.

[0020] A semiconductor defective product processing system comprises a first acquisition module, a second acquisition module, an allocation module and a judgment module. The first acquisition module is configured to acquire a first particle MAP image corresponding to the semiconductor defective product from different angles after receiving the semiconductor defective product. The second acquisition module is configured to acquire a first defect detection model, and perform defect detection on the semiconductor defective product according to the first particle MAP image by using the first defect detection model, so as to acquire a defect shape and a defect position of the semiconductor defective product. The allocation module is configured to automatically allocate the semiconductor defective product to a corresponding processing flow according to the defect shape and the defect position, and perform reprocessing on the semiconductor defective product based on the processing flow, so as to acquire a secondary processing product. The judgment module is configured to perform defect detection on the secondary processing product again, and determine whether the secondary processing product is processed for scrapping according to a detection result.

[0021] The present application has the following advantages:

[0022] The present application can automatically perform defect detection on a semiconductor defective product, and automatically allocate the defective product to a corresponding processing flow according to a detected defect shape and a defect position, so as to perform secondary processing on the defective product, thereby reducing material loss and greatly improving the yield. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A flowchart of a semiconductor defective product processing method according to an embodiment of the present application;

[0024] Figure 2 A block schematic diagram of a semiconductor defective product processing system according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0026] Figure 1 A flowchart of the raw material automatic supervision method based on the industrial Internet of Things according to an embodiment of the present application.

[0027] As shown in Figure 1 The semiconductor defective product processing method according to an embodiment of the present application can include the following steps:

[0028] S1, after receiving the semiconductor defective product, a first particle MAP image corresponding to the semiconductor defective product is obtained from different angles respectively.

[0029] Specifically, as one possible implementation, an ellipsoidal mirror can be used to focus a laser beam to scan the semiconductor surface, and a photodetector can be used to collect the scattered light signal, and finally generate the corresponding particle MAP image.

[0030] S2, a first defect detection model is obtained, and the first defect detection model is used to detect defects of the semiconductor defective product according to the first particle MAP image, so as to obtain the defect morphology and the defect position of the semiconductor defective product.

[0031] In one embodiment of the present application, step S2 includes the following steps:

[0032] S21, a plurality of unlabeled second particle MAP images are obtained from the database, and a third particle MAP image with resolution level annotation is obtained based on the second particle MAP image sample in a splicing manner. Each unlabeled second particle MAP image contains at least one type of defect, and each third particle MAP image contains multiple types of defects.

[0033] Specifically, the database can be pre-stored with particle MAP images of defects of various types, i.e. second particle MAP images, wherein each unlabeled second particle MAP image contains at least one type of defect. The types of defects that can occur during semiconductor manufacturing include polishing contamination, edge aggregation, concrete anti-corrosion and anti-permeability protective agent back splash, wheel grinding marks, incoming material scratches, scratches, cutting marks, wheel grinding chuck marks, and polishing liquid residues.

[0034] Wherein, after obtaining a plurality of unannotated second particle MAP graphs from the database, a second particle MAP graph containing only one morphological defect is selected, and the defect range is extracted. The defect range corresponding to different morphological defects is modified to different levels of resolution, thereby obtaining the resolution level annotation of the second particle MAP graph containing only one morphological defect, and then, based on the corresponding resolution level, the second particle MAP graph containing only one morphological defect with resolution level annotation is spliced, wherein the resolution is used as the image splicing weight, and the higher the resolution, the greater the image processing weight. When different resolution images intersect, the resolution with the maximum image processing weight is selected. Thus, the third particle MAP graph with resolution level annotation can be obtained.

[0035] S22, establishing a defect detection network. Wherein, the defect detection network includes an image compression unit and an image decompression unit.

[0036] Wherein, the image compression unit includes an enhanced architecture with a residual neural network as the backbone network, and the image decompression unit is used for feature recovery of the compressed image features to generate a feature map with the same size as the input image of the image compression unit.

[0037] S23, upstream training of the defect detection network based on self-generated learning mode according to a plurality of unannotated second particle MAP graphs to obtain a second defect detection model.

[0038] In an embodiment of the present application, step S23 specifically includes the following steps:

[0039] S231, image set expansion of the second particle MAP graph is performed using different image processing strategies to obtain a fourth particle MAP graph.

[0040] Wherein, the image processing strategy can include image deformation, image fusion, adding noise and the like to realize image expansion.

[0041] S232, the fourth particle MAP graph is input into the image compression unit for image compression processing to obtain compressed image features, and the first generation value is calculated based on the full sequence comparison algorithm according to the compressed image features.

[0042] In an embodiment of the present application, step S232 specifically includes the following steps:

[0043] S2321, the compressed image features are projected into a high-dimensional feature space through a feature space to obtain a high-dimensional feature descriptor of the compressed image features in the high-dimensional feature space.

[0044] Wherein, the high-dimensional feature descriptor can include positive class feature representation and negative class feature representation.

[0045] S2322, the first cost function is used to calculate the first cost value between the high-dimensional feature descriptors.

[0046] In an embodiment of the present application, the first cost function is generated by the following formula:

[0047] ,

[0048] wherein, represents the positive class feature representation, represents the negative class feature representation, represents the set of negative class feature representations, represents a randomness adjustment coefficient, represents the cost value output by the first cost function.

[0049] S233, after the compressed image features are image decompressed by the image decompression unit, the corresponding second cost value is calculated based on the segment comparison algorithm.

[0050] In an embodiment of the present application, step S233 specifically includes the following steps:

[0051] S2331, the fourth particle MAP image is input into a convolution operation layer with the same kernel size as the segment selection range size, the weight of each segment selection range is calculated by moving the window, and the target segment selection range is selected according to the weight. Wherein, in each selection, the selected target segment selection range is removed from each segment selection range.

[0052] S2332, the segment selection feature map corresponding to the selected target segment selection range is obtained in the decompressed MAP image, and the segment selection feature map is averaged and down-sampled to obtain the corresponding segment selection feature descriptor.

[0053] Wherein, the segment selection feature descriptor can include the positive class feature representation and the negative class feature representation.

[0054] S2333, the first cost function is used to calculate the second cost value between the segment selection feature descriptors.

[0055] Wherein, the way of calculating the second cost value between the segment selection feature descriptors is similar to the way of calculating the first cost value, which will not be described here in detail.

[0056] S234, the related coefficients in the image compression unit and the image decompression unit are adjusted and optimized according to the first cost value and the second cost value.

[0057] S24, the second defect detection model is trained downstream according to a plurality of third particle MAP images with resolution level labels to obtain the first defect detection model.

[0058] In an embodiment of the present application, the second defect detection model can be trained downstream based on a plurality of third particle MAP images with resolution level annotations according to a third cost function, wherein the third cost function can be generated by the following formula:

[0059]

[0060] wherein, represents the output value of the third cost function; represents the number of samples; represents the number of defect types; represents the true label of the i-th sample, which is 1 if the sample is a positive class, and 0 if the sample is a negative class; represents the probability that the i-th sample is the s-th defect type.

[0061] S3, automatically assigning the semiconductor defective product to the corresponding processing flow according to the defect type and the defect position, and reprocessing the semiconductor defective product based on the processing flow to obtain a secondary processing product.

[0062] Specifically, different processing flows can be used to reprocess the semiconductor defective product to obtain a secondary processing product according to different defect types detected for different defect positions. Specifically, a statistical table of defect position-defect type-processing flow can be generated in advance, and when step S3 is performed, the statistical table can be directly called to automatically assign the semiconductor defective product to the corresponding processing flow. For example, in an embodiment of the present application, the statistical table of defect position-defect type-processing flow can be as shown in the following table:

[0063]

[0064] In an embodiment of the present application, step S3 includes the following steps:

[0065] S31, obtaining a plurality of sets of historical processing information corresponding to the processing flow. Each set of historical processing information includes historical processing data and corresponding historical removal amount, and the historical processing data includes the applied pressure on the surface of the semiconductor, the processing time and the processing rate.

[0066] S32, obtaining a removal amount evaluation model corresponding to the processing flow based on the historical processing data and the corresponding historical removal amount.

[0067] In an embodiment of the present application, step S32 includes the following steps:

[0068] S321, generating an initial estimation model according to the historical processing data and the corresponding historical removal amount by the following formula:

[0069] ,

[0070] wherein, represents a removal amount evaluation value, represents an intercept term, represents a linear term skewness coefficient, represents a quadratic term skewness coefficient, represents the jth processing data in the historical processing data, represents an interference term,

[0071] S322, a second cost function corresponding to the initialization estimation model is established by the following formula:

[0072] ,

[0073] wherein, represents an output value of the second cost function, represents a historical removal amount corresponding to the ith group of historical processing data, represents a removal amount evaluation value corresponding to the ith group of historical processing data, represents the number of groups of historical processing data,

[0074] S323, each group of historical processing data and the corresponding historical removal amount are respectively input into the second cost function, and the intercept term, the linear term skewness coefficient and the quadratic term skewness coefficient in the initialization estimation model are obtained by using the least square method to minimize the second cost function, so as to generate a removal amount evaluation model.

[0075] S33, the optimal processing data is obtained according to the removal amount evaluation model. The deviation between the removal amount evaluation value output by the removal amount evaluation model corresponding to the optimal processing data and the ideal removal amount is the smallest. S34, the semiconductor defective product is reprocessed according to the optimal processing data.

[0076] It can be understood that in the semiconductor product processing process (grinding, fine polishing, rough polishing, etc.), the removal amount corresponds to the processing effect, and the closer the removal amount is to the ideal removal amount, the closer the processing effect is to the ideal processing effect. Therefore, the removal amount evaluation model is used to estimate the removal amount evaluation value corresponding to each processing data, and the removal amount evaluation value with the smallest deviation from the ideal removal amount is selected, and the corresponding processing data is obtained as the optimal processing data. Then, the semiconductor defective product is reprocessed according to the optimal processing data. Thus, the processing data can be corrected in real time to ensure that the processing is performed by the optimal processing data, thereby greatly improving the processing quality of the product.

[0077] S4, the secondary processing product is detected again for defects, and whether the secondary processing product is processed for scrap is judged according to the detection result.

[0078] In an embodiment of the present application, when the secondary processing product is subjected to flaw detection again, the first flaw detection model in the above embodiment can be used for flaw detection; in other embodiments of the present application, the flaw detection method in the prior art can also be used to detect whether the secondary processing product has flaws, for example, based on a traditional deep learning model for flaw detection. Thus, it is determined whether the secondary processing product still has flaws, and if it is determined that the secondary processing product still has flaws, the secondary processing product is subjected to scrap processing.

[0079] Of course, in other embodiments of the present application, the above steps can be repeated multiple times for multiple reprocessing, and after a preset number of times, flaw detection is performed again, and if flaws still exist, scrap processing is performed.

[0080] Thus, the present application can automatically perform flaw detection on the semiconductor flawed product, and automatically distribute the flawed product to the corresponding processing flow according to the detected flaw shape and flaw position for secondary processing, thereby not only reducing material loss, but also greatly improving the yield.

[0081] In summary, according to the semiconductor flawed product processing method according to the embodiments of the present application, after receiving the semiconductor flawed product, the first particle MAP image corresponding to the semiconductor flawed product is obtained from different angles, and the first flaw detection model is obtained, and the first flaw detection model is used to perform flaw detection on the semiconductor flawed product according to the first particle MAP image to obtain the flaw shape and flaw position of the semiconductor flawed product, and the semiconductor flawed product is automatically distributed to the corresponding processing flow according to the flaw shape and flaw position, and the semiconductor flawed product is reprocessed based on the processing flow to obtain a secondary processing product, and the secondary processing product is subjected to flaw detection again, and whether the secondary processing product is subjected to scrap processing is determined according to the detection result. Thus, the semiconductor flawed product can be automatically subjected to flaw detection, and the flawed product can be automatically distributed to the corresponding processing flow for secondary processing according to the detected flaw shape and flaw position, thereby not only reducing material loss, but also greatly improving the yield.

[0082] Corresponding to the semiconductor flawed product processing method of the above embodiment, the present application further proposes a semiconductor flawed product processing system.

[0083] As shown in Figure 2 The semiconductor flawed product processing system according to the embodiments of the present application can include a first acquisition module 100, a second acquisition module 200, a distribution module 300, and a determination module 400.

[0084] The first acquisition module 100 is configured to acquire a first particle MAP image corresponding to the semiconductor defective product from different angles after receiving the semiconductor defective product; the second acquisition module 200 is configured to acquire a first defect detection model, and perform defect detection on the semiconductor defective product according to the first particle MAP image by using the first defect detection model, so as to acquire a defect shape and a defect position of the semiconductor defective product; the distribution module 300 is configured to automatically distribute the semiconductor defective product to a corresponding processing flow according to the defect shape and the defect position, and perform reprocessing on the semiconductor defective product based on the processing flow, so as to obtain a secondary processing product; and the judgment module 400 is configured to perform defect detection on the secondary processing product again, and determine whether the secondary processing product is subjected to a scrapping process according to a detection result.

[0085] In an embodiment of the present application, the second acquisition module 200 is specifically configured to: acquire a plurality of unlabeled second particle MAP images from a database, and acquire a third particle MAP image with resolution level annotation in a splicing manner based on the second particle MAP image sample, wherein each unlabeled second particle MAP image contains at least one defect shape, and each third particle MAP image contains a plurality of defect shapes; establish a defect detection network, wherein the defect detection network includes an image compression unit and an image decompression unit; perform upstream training on the defect detection network based on a self-generated learning manner according to the plurality of unlabeled second particle MAP images, so as to obtain a second defect detection model; and perform downstream training on the second defect detection model according to the plurality of third particle MAP images with resolution level annotation, so as to obtain the first defect detection model.

[0086] In an embodiment of the present application, the second acquisition module 200 is specifically configured to: perform image set expansion on the second particle MAP image by using different image processing strategies, so as to obtain a fourth particle MAP image; input the fourth particle MAP image into the image compression unit to perform image compression processing, so as to obtain compressed image features, and calculate a first generation value based on the compressed image features according to a full sequence comparison algorithm; after the compressed image features are subjected to image decompression processing by the image decompression unit, calculate a corresponding second generation value based on a fragment comparison algorithm; and adjust and optimize the correlation coefficients in the image compression unit and the image decompression unit according to the first generation value and the second generation value.

[0087] In an embodiment of the present application, the second acquisition module 200 is specifically configured to: project the compressed image features into a high-dimensional feature space through a feature space, so as to obtain a high-dimensional feature descriptor of the compressed image features in the high-dimensional feature space; and calculate a first generation value between the high-dimensional feature descriptors by using a first cost function.

[0088] In an embodiment of the present application, the first cost function is generated by the following formula:

[0089] ,

[0090] wherein, represents a positive class feature representation, represents a negative class feature representation, represents a negative class feature representation set, represents a randomness adjustment coefficient, represents a value of a first cost function output.

[0091] In an embodiment of the present application, the second acquisition module 200 is specifically configured to: input the fourth particle MAP graph into a convolution operation layer with a kernel size same as a patch selection range size, calculate a weight value of each patch selection range through window moving, and select a target patch selection range according to the weight value; wherein, in each selection, the selected target patch selection range is removed from each patch selection range; a patch selection feature map corresponding to the selected target patch selection range is acquired in the decompressed MAP graph, and the patch selection feature map is averaged and down-sampled to obtain a corresponding patch selection feature descriptor; and a second value is calculated between the patch selection feature descriptors by using a first cost function.

[0092] In an embodiment of the present application, the distribution module 300 is specifically configured to: acquire a plurality of groups of historical processing information corresponding to the processing procedure; wherein, each group of historical processing information comprises historical processing data and corresponding historical removal amount, and the historical processing data comprises an applied pressure on the semiconductor surface, a processing time and a processing rate; acquire a removal amount evaluation model corresponding to the processing procedure based on the historical processing data and the corresponding historical removal amount; acquire optimal processing data according to the removal amount evaluation model; wherein, a deviation between a removal amount evaluation value output by the removal amount evaluation model corresponding to the optimal processing data and an ideal removal amount is minimum; and reprocess the semiconductor defective product according to the optimal processing data.

[0093] In an embodiment of the present application, the distribution module 300 is specifically configured to: generate an initial estimation model according to the historical processing data and the corresponding historical removal amount through the following formula:

[0094] ,

[0095] wherein, represents a removal amount evaluation value, represents an intercept term, represents a first-order term skewness coefficient, represents a second-order term skewness coefficient, represents a jth processing data in the historical processing data, represents an interference term; and a second cost function corresponding to the initial estimation model is established through the following formula:

[0096] ,

[0097] wherein, represents an output value of the second cost function, represents a historical removal amount corresponding to the i-th set of historical processing data, represents a removal amount evaluation value corresponding to the i-th set of historical processing data, represents a number of sets of historical processing data; each set of historical processing data and the corresponding historical removal amount are input into the second cost function, and the intercept term, the linear term skewness coefficient and the quadratic term skewness coefficient in the initial estimation model are obtained by using the least square method to minimize the second cost function, so as to generate the removal amount evaluation model.

[0098] It should be noted that details of the semiconductor defective product processing system in the embodiments of the present application are not disclosed, and please refer to the details disclosed in the semiconductor defective product processing method described above, which will not be described in detail here.

[0099] According to the semiconductor defective product processing system in the embodiments of the present application, after receiving the semiconductor defective product, the first acquisition module acquires the first particle MAP image corresponding to the semiconductor defective product from different angles, the second acquisition module acquires the first defect detection model, and the first defect detection model is used to detect the defect of the semiconductor defective product according to the first particle MAP image to obtain the defect shape and the defect position of the semiconductor defective product, the distribution module automatically distributes the semiconductor defective product to the corresponding processing flow according to the defect shape and the defect position, the semiconductor defective product is reprocessed based on the processing flow to obtain the secondary processing product, and the judgment module detects the defect of the secondary processing product again and judges whether the secondary processing product is scrapped according to the detection result. Therefore, the semiconductor defective product can be automatically detected, and the defective product can be automatically distributed to the corresponding processing flow for secondary processing according to the detected defect shape and defect position, so as to not only reduce material loss, but also greatly improve the yield.

[0100] The semiconductor defective product processing system,

[0101] In the description of the present application, the terms "first", "second" are only used for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features with "first", "second" can explicitly or implicitly include one or more features. The meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0102] In the present application, unless specifically defined otherwise and limited, the terms "mount", "connected", "connection", "fixed", and the like should be construed as being broadly, for example, it can be fixed connection, or detachable connection, or integrated; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, or it can be the internal communication of two elements or the interaction relationship between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0103] In the present application, unless specifically defined otherwise and limited, the first feature is "on" or "under" the second feature. The first and second features can be in direct contact, or the first and second features can be indirectly in contact through an intermediate medium. Moreover, the first feature "above", "over" and "on" the second feature can be the first feature directly above or obliquely above the second feature, or only indicates that the horizontal height of the first feature is higher than that of the second feature. The first feature "below", "under" and "under" the second feature can be the first feature directly below or obliquely below the second feature, or only indicates that the horizontal height of the first feature is less than that of the second feature.

[0104] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or features of different embodiments or examples described in the present application without contradiction.

[0105] In addition, each functional unit in each embodiment of the present application can be integrated in one processing module, or each unit can be physically present separately, or two or more units can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of a software function module. The integrated module, if realized in the form of a software function module and sold or used as an independent product, can also be stored in a computer readable storage medium.

[0106] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and cannot be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A method for processing defective semiconductor products, characterized in that, Includes the following steps: S1, after receiving the defective semiconductor product, the first particle MAP image corresponding to the defective semiconductor product is obtained from different angles; S2, obtain a first defect detection model, and use the first defect detection model to perform defect detection on the semiconductor defective product according to the first particle MAP map, so as to obtain the defect morphology and defect location of the semiconductor defective product; wherein, step S2 includes the following steps: S21, retrieve multiple unlabeled second particle MAP images from the database, and obtain a third particle MAP image with resolution level labeling by stitching together the second particle MAP image samples. Each unlabeled second particle MAP image contains at least one morphological defect, and each third particle MAP image contains multiple morphological defects. S22, establish a defect detection network, wherein the defect detection network includes an image compression unit and an image decompression unit; S23, the defect detection network is trained upstream based on multiple unlabeled second particle MAP maps using a self-generating learning method to obtain a second defect detection model; S24, the second defect detection model is trained downstream based on multiple third particle MAP maps with resolution level annotations to obtain the first defect detection model; S3, the semiconductor defective product is automatically assigned to the corresponding processing flow according to the defect morphology and the defect location, and the semiconductor defective product is reprocessed based on the processing flow to obtain a secondary processed product; S4. Perform defect detection on the secondary processed product again, and determine whether the secondary processed product should be scrapped based on the detection results.

2. The method for processing defective semiconductor products according to claim 1, characterized in that, Step S23 specifically includes the following steps: S231, the second particle MAP map is expanded by using different image processing strategies to obtain a fourth particle MAP map; S232, the fourth particle MAP map is input into the image compression unit for image compression processing to obtain compressed image features, and the first generation value is calculated based on the compressed image features using a full sequence alignment algorithm; S233, after the compressed image features are decompressed by the image decompression unit, the corresponding second-generation value is calculated based on the fragment comparison algorithm; S234, The correlation coefficients in the image compression unit and the image decompression unit are adjusted and optimized based on the first generation value and the second generation value.

3. The method for processing defective semiconductor products according to claim 2, characterized in that, Step S232 specifically includes: S2321, The compressed image features are projected from the feature space to the high-dimensional feature space to obtain the high-dimensional feature descriptor of the compressed image features in the high-dimensional feature space; S2322, The first generation value between the high-dimensional feature descriptors is calculated using the first cost function.

4. The method for processing defective semiconductor products according to claim 3, characterized in that, The first cost function is generated using the following formula: ， in, This represents the positive class feature representation. This represents the negative class feature representation. The set represents the negative class features. Represents the randomness adjustment coefficient. This represents the cost value output by the first cost function.

5. The method for processing defective semiconductor products according to claim 4, characterized in that, Step S233 specifically includes the following steps: S2331, the fourth particle MAP map is input into a convolutional operation layer with the same kernel size and chip selection range size. The weight of each chip selection range is calculated by moving the window, and the target chip selection range is selected according to the weight. In each selection, the selected target chip selection range is removed from each chip selection range. S2332, Obtain the chip selection feature map corresponding to the selected target chip selection range in the decompressed MAP map, and perform average downsampling on the chip selection feature map to obtain the corresponding chip selection feature descriptor; S2333, calculate the second generation value between the slice selection feature descriptors using the first cost function.

6. The method for processing defective semiconductor products according to claim 1, characterized in that, Step S3 includes the following steps: S31, acquire multiple sets of historical processing information corresponding to the processing flow; wherein, each set of historical processing information includes historical processing data and corresponding historical removal amount, and the historical processing data includes the applied pressure on the semiconductor surface, processing time and processing rate; S32, Based on the historical processing data and the corresponding historical removal volume, obtain the removal volume evaluation model corresponding to the processing process; S33, Obtain optimal processing data according to the removal amount evaluation model; wherein, the deviation between the removal amount evaluation value output by the removal amount evaluation model corresponding to the optimal processing data and the ideal removal amount is the smallest; S34, The defective semiconductor product is reprocessed according to the optimal processing data.

7. The method for processing defective semiconductor products according to claim 6, characterized in that, Step S32 includes the following steps: S321, an initial estimation model is generated based on the historical processing data and the corresponding historical removal amount using the following formula: ， in, This represents the assessed value of the amount removed. Represents the intercept term. This represents the skewness coefficient of the first-order term. This represents the skewness coefficient of the quadratic term. This represents the j-th processing data in the historical processing data. Indicates distractor items; S322, the second cost function corresponding to the initial estimation model is established using the following formula: ， in, This represents the output value of the second cost function. This represents the historical removal amount corresponding to the i-th set of historical processing data. This represents the estimated removal amount corresponding to the i-th set of historical processing data. This indicates the number of groups of the historical processing data; S323, input each group of historical processing data and the corresponding historical removal amount into the second cost function, with the goal of minimizing the second cost function, and use the least squares method to obtain the intercept term, the first-order skewness coefficient and the second-order skewness coefficient in the initial estimation model, so as to generate the removal amount evaluation model.

8. A system for handling defective semiconductor products, characterized in that, The semiconductor defect processing system performs the semiconductor defect processing method according to any one of claims 1-7, the semiconductor defect processing system comprising: The first acquisition module is used to acquire the first particle MAP image corresponding to the semiconductor defective product from different angles after receiving the semiconductor defective product. The second acquisition module is used to acquire a first defect detection model and use the first defect detection model to perform defect detection on the semiconductor defective product according to the first particle MAP map, so as to obtain the defect morphology and defect location of the semiconductor defective product. The allocation module is used to automatically allocate the semiconductor defective products to the corresponding processing flow according to the defect morphology and the defect location, and to reprocess the semiconductor defective products based on the processing flow to obtain secondary processed products. The judgment module is used to perform defect detection on the secondary processed product again, and determine whether the secondary processed product should be scrapped based on the detection results.

Images