Collaborative intelligence based semiconductor wafer semi-automated labeling system

Abstract

A collaborative intelligence-based semi-automated semiconductor wafer labeling system according to an embodiment of the present invention comprises: a first database providing at least one set of WBM image data that has been labeled; a second database providing a plurality of test data set control groups composed of a plurality of unlabeled WBM image data; a learning model comprising a first process for receiving WBM image data from the first database and learning defect patterns and normal patterns, a second process for predicting the defect probability of WBM image data received from the second database based on the learned pattern information, a third process for calculating an uncertainty value for each data based on the predicted defect probability value, and a fourth process for determining whether to perform an auto-labeling operation using a labeling model or an engineer labeling operation by a worker on the WBM image data according to the result of comparing the calculated uncertainty value for each data with a preset threshold value; and a first labeling unit that inputs auto-labeled WBM image data into the first database according to the determination result of the learning model and provides it as training data for the learning model. It may include a second labeling unit that inputs engineer-labeled WBM image data into a first database and provides it as training data for the learning model based on the judgment result of the learning model.

Description

Technology Field

[0001] The present invention relates to a collaborative intelligence-based semi-automated semiconductor wafer labeling system and a semiconductor wafer labeling method using the same. More specifically, it relates to a collaborative intelligence-based semi-automated semiconductor wafer labeling system and a semiconductor wafer labeling method using the same, which enables the performance of high-accuracy labeling tasks by providing a labeling system that combines experts and artificial intelligence. Background Technology

[0002] Generally, the wafer-based semiconductor manufacturing process consists of eight processes: photolithography, etching, diffusion, chemical vapor deposition, metallization, chemical mechanical polishing, cleaning, and ion implantation. If defective products are produced during this process, product quality and yield deteriorate. Therefore, it is important to rapidly identify defective dies on the wafer in the semiconductor manufacturing process.

[0003] Once the manufacturing process is complete, defects are identified through EDS (Electrical Die Sorting) inspection. EDS is a quality inspection method that classifies whether individual semiconductor wafer chips are functioning normally by verifying whether each die of the semiconductor wafer meets the desired quality level through electrical characteristic inspection.

[0004] As a result of the EDS inspection, a Wafer Bin Map (WBM) image is generated that may include defective die patterns.

[0005] WBM images show the location information of defective chips on the wafer and include specific defect patterns. Since these patterns suggest the cause of the defects, engineering experts are performing labeling tasks in the field to manually classify each pattern.

[0006] The labeled datasets for each die pattern obtained through such labeling operations serve as essential data for pattern analysis using machine learning techniques, such as image mining algorithms, or AI-based approaches to process large volumes of wafers; however, there is a problem in that the labeled WBM images represent only a tiny fraction of the total volume due to the limited number of experts available to process the massive datasets generated by EDS, and each of these approaches is subject to limitations such as labor costs, time consumption, and accuracy.

[0007] Therefore, there is a need to develop a collaborative intelligent system that labels each defect pattern by combining semi-supervised machine learning and human intelligence to address the limitations of existing methodologies. The problem to be solved

[0008] The technical problem that the present invention aims to solve is to provide a collaborative intelligence-based semi-automated semiconductor wafer labeling system and a semiconductor wafer labeling method using the same, which enables high-accuracy labeling by providing a labeling system combining experts and artificial intelligence, improves the performance and accuracy of the learning model as the number of labeling repetitions increases, thereby reducing the need for expert labeling work and enabling highly efficient labeling, and allows for flexible response to the occurrence of new defect patterns by performing labeling based on uncertainty.

[0009] The technical problems of the present invention are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art from the description below. means of solving the problem

[0010] A collaborative intelligence-based semiconductor wafer semi-automated labeling system according to one embodiment of the present invention for solving the above technical problem is,

[0011] A first database providing at least one set of labeled WBM image data;

[0012] A second database providing multiple test data set control groups consisting of multiple unlabeled WBM image data;

[0013] A first process for learning defect patterns and normal patterns by receiving WBM image data from a first database, and

[0014] A second process for predicting the defect probability of WBM image data received from a second database based on the learned pattern information above, and

[0015] A third process for calculating data-specific uncertainty values ​​based on the above-mentioned predicted defect probability values, and

[0016] A learning model comprising a fourth process for determining whether to perform auto-labeling work using a labeling model or engineer labeling work by a worker on the corresponding WBM image data based on the result of comparing the above-mentioned data-specific uncertainty value with a preset threshold value;

[0017] A first labeling unit that inputs auto-labeled WBM image data into a first database according to the judgment result of a learning model and provides it as training data for a learning model; and

[0018] It may include a second labeling unit that inputs engineer-labeled WBM image data into a first database based on the judgment result of the learning model and provides it as training data for the learning model.

[0019] According to one embodiment of the present invention, the learning model can predict the defect probability of WBM image data provided in a second database based on information learned for at least eight defect patterns including Center, Donut, Edge-Loc, Edge-Ring, Loc, Scratch, Random, and Near-full, and one normal pattern.

[0020] According to one embodiment of the present invention, the number of WBM images in the test data set provided to the learning model can be provided in an equal number to the number of WBM images in the initial training data set.

[0021] According to one embodiment of the present invention, the third process can calculate three uncertainty values ​​for each WBM image data through: a Minimum Confidence technique that selects the lowest defect probability value among the defect probability values ​​predicted by the learning model for eight defect patterns; a Minimum Margin technique that calculates the difference between the second highest defect probability value and the highest defect probability value predicted by the learning model for eight defect patterns; and an Entropy technique that calculates the entropy of the distribution of defect probability values ​​predicted by the learning model for eight defect patterns.

[0022] According to one embodiment of the present invention, when all three uncertainty values ​​are below a threshold value, an auto-labeling operation is performed, and when any one of the three uncertainty values ​​is above a threshold value, an engineer-labeling operation can be performed.

[0023] According to one embodiment of the present invention, the learning model can predict a defect probability for one unlabeled test data set among the control groups of the second database based on information learned from training data added through the first labeling work unit and the second labeling work unit, calculate an uncertainty value for each data based on the predicted defect probability value, and determine whether to perform an auto-labeling work using the labeling model or an engineer labeling work by a worker on the corresponding WBM image data based on the result of comparing the calculated uncertainty value with a preset threshold value, and repeat the work until all the control groups of the second database are labeled.

[0024] According to one embodiment of the present invention, as the labeling work for the control groups of the second database is repeated, the ratio of engineer labeling work may gradually decrease, and the ratio of auto-labeling work may gradually increase.

[0025] A semiconductor wafer labeling method using a collaborative intelligence-based semiconductor wafer semi-automated labeling system according to another embodiment of the present invention is,

[0026] Step (S10) of preparing a training data set consisting of labeled WBM image data through a first database;

[0027] A step (S20) of providing the above training data set to a learning model to learn defect patterns and normal patterns;

[0028] Step (S30) of preparing multiple test data set control groups consisting of multiple unlabeled WBM image data through a second database;

[0029] One test data set among the above control groups is input into a learning model, and the learning model predicts the defect probability of the test data set based on previously learned pattern information (S40);

[0030] A step (S50) of calculating uncertainty values ​​for each data based on defect probability values ​​predicted through a learning model;

[0031] A step (S60) of determining whether to perform auto-labeling work using a labeling model or engineer labeling work by a worker on the WBM image data based on the result of comparing the calculated uncertainty value with a preset threshold value;

[0032] Step (S70) of performing an auto-labeling task or an engineer-labeling task based on the judgment result;

[0033] Step (S80) of adding training data by inputting labeled WBM image data into the first database;

[0034] A step (S90) of providing the added training data to a learning model to learn the added defect pattern;

[0035] A step (S100) of inputting one unlabeled test data set among the control groups of the second database into a learning model, and the learning model predicting the defect probability of the test data set based on previously learned defect pattern information; and

[0036] It may include a step (S110) of repeating steps S50 through S100 until all control groups of the second database are labeled.

[0037] According to another embodiment of the present invention, the S40 step can predict defect probabilities for at least eight defect patterns including Center, Donut, Edge-Loc, Edge-Ring, Loc, Scratch, Random, and Near-full, and one normal pattern.

[0038] According to another embodiment of the present invention, the number of WBM images in the test data set provided to the learning model in step S40 may be provided in an equal number to the number of WBM images in the training data set in step S10.

[0039] According to another embodiment of the present invention, the calculation of the uncertainty value in step S50 is a Minimum Confidence technique that selects the lowest defect probability value among the defect probability values ​​predicted by the learning model for eight defect patterns;

[0040] A minimum margin technique that calculates the difference between the second-highest defect probability value and the highest defect probability value predicted by the learning model for eight defect patterns; and

[0041] Three uncertainty values ​​can be calculated for each WBM image data through an entropy technique that calculates the entropy of the distribution of defect probability values ​​predicted by the learning model for eight defect patterns.

[0042] According to another embodiment of the present invention, when all three uncertainty values ​​are below a threshold value, an auto-labeling operation may be performed, and when any one of the three uncertainty values ​​is above a threshold value, an engineer-labeling operation may be performed.

[0043] According to another embodiment of the present invention, as the labeling work for the control groups of the second database is repeated, the ratio of engineer labeling work may gradually decrease, and the ratio of auto-labeling work may gradually increase. Effects of the invention

[0044] As described above, the collaborative intelligence-based semi-automated semiconductor wafer labeling system according to the embodiment of the present invention provides a labeling system that combines experts and artificial intelligence, thereby enabling the performance of high-accuracy labeling tasks. Furthermore, as the number of labeling repetitions increases, the performance and accuracy of the learning model improve, which reduces the need for expert labeling tasks and enables highly efficient labeling. Additionally, since labeling is performed based on uncertainty, it has the effect of flexibly responding to the occurrence of new defect patterns. Brief explanation of the drawing

[0045] FIG. 1 is a conceptual diagram illustrating a collaborative intelligence-based semiconductor wafer semi-automated labeling system according to one embodiment of the present invention. FIG. 2 is a conceptual diagram of learning training data using a learning model according to one embodiment of the present invention. FIG. 3 is a conceptual diagram of predicting the probability of defects in test data using a learning model according to one embodiment of the present invention. FIG. 4 is a conceptual diagram of calculating uncertainty values ​​based on defect probability values ​​using a learning model according to an embodiment of the present invention. FIG. 5 is a conceptual diagram for determining whether to perform an auto-labeling task or an engineer-labeling task based on the result of comparing an uncertainty value calculated according to one embodiment of the present invention with a preset threshold value. FIG. 6 is a conceptual diagram of adding WBM image data, which has been auto-labeled or engineer-labeled according to one embodiment of the present invention, as training data. FIG. 7 is a conceptual diagram of a method for repeatedly performing a labeling operation using a labeled defect pattern as training data according to an embodiment of the present invention. FIGS. 8a to 8h are exemplary diagrams of eight defect patterns provided to a learning model according to one embodiment of the present invention. FIG. 9 is a flowchart illustrating a semiconductor wafer labeling method using a collaborative intelligence-based semiconductor wafer semi-automated labeling system according to another embodiment of the present invention. Specific details for implementing the invention

[0046] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims. Throughout the specification, the same reference numerals refer to the same components.

[0047] The terms used herein are for describing embodiments and are not intended to limit the invention. In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. As used herein, "comprising" and / or "comprising" does not exclude the presence or addition of one or more other components, steps, actions, and / or elements to the mentioned components, steps, actions, and / or elements.

[0048] Furthermore, throughout the specification, when a part is described as being "connected" to another part, this includes not only cases where they are "directly connected," but also cases where they are "indirectly" or "electrically connected" with other members or elements in between.

[0049] Additionally, throughout the specification, the description that each layer (film), region, pattern, or structure is formed "on" or "under" the substrate, each layer (film), region, pad, or pattern includes both direct formation and formation through another layer. The criteria for "on" or "under" each layer are described based on the drawings.

[0050] Furthermore, expressions such as 'first, second,' etc., are used solely to distinguish multiple compositions and do not limit the order or other characteristics between the compositions.

[0051] Furthermore, the flowcharts illustrated in the drawings are merely illustrative steps to obtain the most desirable results in carrying out the present invention, and it is obvious that other steps may be added or some steps may be deleted.

[0052] Unless otherwise defined, all terms used in this specification (including technical and scientific terms) may be used in a meaning commonly understood by those skilled in the art to which the present invention pertains. Additionally, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0053] Hereinafter, preferred embodiments of the present invention will be described with reference to the attached drawings.

[0054] FIG. 1 is a conceptual diagram illustrating a collaborative intelligence-based semi-automated semiconductor wafer labeling system according to an embodiment of the present invention; FIG. 2 is a conceptual diagram illustrating training data using a learning model according to an embodiment of the present invention; FIG. 3 is a conceptual diagram illustrating predicting the defect probability of test data using a learning model according to an embodiment of the present invention; FIG. 4 is a conceptual diagram illustrating calculating an uncertainty value based on a defect probability value using a learning model according to an embodiment of the present invention; FIG. 5 is a conceptual diagram illustrating determining an auto-labeling operation or an engineer-labeling operation based on the result of comparing the calculated uncertainty value with a preset threshold value according to an embodiment of the present invention; FIG. 6 is a conceptual diagram illustrating adding WBM image data that has undergone an auto-labeling operation or an engineer-labeling operation as training data according to an embodiment of the present invention; FIG. 7 is a conceptual diagram illustrating repeatedly performing a labeling operation using a labeled defect pattern as training data according to an embodiment of the present invention; and FIG. 8a to 8h are example diagrams of eight defect patterns provided to a learning model according to an embodiment of the present invention.

[0055] Referring to FIG. 1, the present invention may largely include a first database (110); a second database (120); a learning model (130); a first labeling unit (140); and a second labeling unit (150).

[0056] The first database (110) above can provide at least one set of labeled wafer bin map (WBM) image data (hereinafter referred to as WBM image).

[0057] The second database (120) above may provide multiple test data set control groups consisting of multiple unlabeled WBM image data.

[0058] The above learning model (130) may include a first process (131) for receiving WBM image data from a first database (110) and learning defect patterns and normal patterns, a second process (132) for predicting the defect probability of WBM image data received from a second database (120) based on the learned pattern information, a third process (133) for calculating an uncertainty value for each data based on the predicted defect probability value, and a fourth process (134) for determining whether to perform auto-labeling work using a labeling model or engineer labeling work by a worker on the WBM image data based on the result of comparing the calculated uncertainty value for each data with a preset threshold value.

[0059] A learning model (130) according to one embodiment of the present invention used ResNet50 (residual neural network 50: CNN-based engine) and used WM-811K, an open source dataset provided by TSMC (Taiwan semiconductor manufacturing company).

[0060] In an experiment according to one embodiment of the present invention, 172,950 WBM images include normal patterns and 8 other defect patterns, 10% of the total data is used for model training, and the remaining 90% of the data set is delabeled for model testing.

[0061] The first labeling unit (140) can input the auto-labeled WBM image data into the first database (110) according to the judgment result of the learning model (130) and provide it as training data for the learning model (130).

[0062] Additionally, the second labeling unit (150) can input engineer-labeled WBM image data into the first database (110) according to the judgment result of the learning model (130) and provide it as training data for the learning model (130).

[0063] In addition, the above learning model (130) can predict the defect probability of WBM image data provided from the second database (120) based on information learned for at least 8 defect patterns including Center, Donut, Edge-Loc, Edge-Ring, Loc, Scratch, Random, and Near-full as seen in FIGS. 8a to 8h, and 1 normal pattern.

[0064] In addition, the number of WBM images in the test data set provided to the above learning model (130) can be provided in an equal number to the number of WBM images in the initial training data set.

[0065] Figure 1 illustrates an example where the number of WBM images in the initial training data set is 17,295. At this time, the number of WBM images in the initial test data set can be set to 17,295.

[0066] Additionally, the third process (133) can calculate three uncertainty values ​​for each WBM image data through: a Minimum Confidence technique that selects the lowest defect probability value among the defect probability values ​​predicted by the learning model (130) for eight defect patterns; a Minimum Margin technique that calculates the difference between the second highest defect probability value and the highest defect probability value predicted by the learning model (130) for eight defect patterns; and an Entropy technique that calculates the entropy of the distribution of defect probability values ​​predicted by the learning model (130) for eight defect patterns.

[0067] The calculation formulas for each technique are as follows.

[0068] Minimum Confidence Technique:

[0069] Minimum Margin Technique:

[0070] Entropy Technique:

[0071] At this time, when all three of the above uncertainty values ​​are below a threshold, an auto-labeling operation is performed, and when any one of the three uncertainty values ​​is above a threshold, an engineer-labeling operation can be performed.

[0072] Additionally, the learning model (130) can predict the defect probability for one unlabeled test data set among the control groups of the second database (120) based on information learned from the training data added through the first labeling work unit (140) and the second labeling work unit (150), calculate the uncertainty value for each data based on the predicted defect probability value, and determine whether to perform auto-labeling work using the labeling model or engineer labeling work by a worker on the WBM image data based on the result of comparing the calculated uncertainty value with a preset threshold value, and repeat this process until all the control groups of the second database (120) are labeled.

[0073] In this invention, as the labeling work for the control groups of the second database (120) is repeated, the ratio of engineer labeling work and the ratio of auto-labeling work can gradually decrease and increase.

[0074] The labeling system according to the present invention can predict the label of test data using a learning model and calculate uncertainty using three different strategies: minimum confidence, minimum margin, and entropy. Additionally, based on each uncertainty result, the test data can be automatically labeled or engineer-labeled according to a threshold and added to the training data.

[0075] In addition, all processes can be repeated until all labels are assigned. Test results according to one embodiment of the present invention showed that the labeling accuracy for each uncertainty was 99.72% (minimum confidence), 99.78% (minimum margin), and 99.78% (entropy), respectively, and the engineer labeling data ratios for each uncertainty were 18.84%, 24.96%, and 21.77%, respectively. Through this, it can be seen that the accuracy of the labeling system according to the present invention is very high.

[0076] In addition, the ratio of engineer-labeled data according to each uncertainty strategy was found to be 18.84% (minimum confidence), 24.96% (minimum margin), and 21.77% (entropy), respectively. This also indicates that the labeling system according to the present invention is effective in reducing the burden on engineers.

[0077] FIG. 9 is a flowchart illustrating a semiconductor wafer labeling method using a collaborative intelligence-based semiconductor wafer semi-automated labeling system according to another embodiment of the present invention.

[0078] Referring to FIG. 9, the semiconductor wafer labeling method according to the present invention first prepares a training data set consisting of labeled WBM image data through a first database (110) (S10), and then provides the training data set to a learning model (130) to learn defect patterns and normal patterns (S20).

[0079] Subsequently, multiple test data set control groups consisting of multiple unlabeled WBM image data are prepared through the second database (120) (S30), one of the test data set control groups is input into the learning model (130), and the learning model (130) predicts the defect probability of the test data set based on previously learned pattern information (S40).

[0080] Afterwards, an uncertainty value for each data is calculated based on the defect probability value predicted through the learning model (130) (S50), and based on the result of comparing the calculated uncertainty value with a preset threshold value, it is determined whether to perform an auto-labeling operation using a labeling model or an engineer labeling operation by a worker on the corresponding WBM image data (S60).

[0081] Afterwards, an auto-labeling operation or an engineer-labeling operation is performed according to the judgment result (S70), and the labeled WBM image data is entered into the first database (110) to add training data (S80).

[0082] Afterwards, the added training data is provided to the learning model (130) to learn the added defect pattern (S90), and one unlabeled test data set among the control groups of the second database (120) is input into the learning model (130), and the learning model (130) predicts the defect probability of the test data set based on the previously learned defect pattern information (S100).

[0083] At this time, steps S50 through S100 can be repeated until all control groups of the second database (120) are labeled (S110).

[0084] At this time, the above S40 step can predict defect probabilities for at least 8 defect patterns including Center, Donut, Edge-Loc, Edge-Ring, Loc, Scratch, Random, and Near-full, and 1 normal pattern.

[0085] In addition, the number of WBM images in the test data set provided to the learning model (130) in step S40 can be provided in the same number as the number of WBM images in the training data set in step S10.

[0086] Additionally, the calculation of the uncertainty value in step S50 above can calculate three uncertainty values ​​for each WBM image data through: a Minimum Confidence technique that selects the lowest defect probability value among the defect probability values ​​predicted by the learning model (130) for eight defect patterns; a Minimum Margin technique that calculates the difference between the second highest defect probability value and the highest defect probability value predicted by the learning model (130) for eight defect patterns; and an Entropy technique that calculates the entropy of the defect probability value distribution predicted by the learning model (130) for eight defect patterns.

[0087] At this time, when all three of the above uncertainty values ​​are below a threshold, an auto-labeling operation is performed, and when any one of the three uncertainty values ​​is above a threshold, an engineer-labeling operation can be performed.

[0088] Additionally, as the labeling work for the control groups of the second database (120) is repeated, the ratio of engineer labeling work may gradually decrease, and the ratio of auto-labeling work may gradually increase.

[0089] The collaborative intelligence-based semi-automated semiconductor wafer labeling system according to an embodiment of the present invention provides a labeling system that combines experts and artificial intelligence, thereby enabling high-accuracy labeling work. As the number of labeling repetitions increases, the performance and accuracy of the learning model improve, which reduces the need for expert labeling work and enables highly efficient labeling. Furthermore, since labeling is performed based on uncertainty, it has the effect of flexibly responding to the occurrence of new defect patterns.

[0090] Although the present invention has been described above, those skilled in the art will recognize that the invention may be implemented in other forms while maintaining the technical concept and essential features of the invention.

[0091] The scope of the rights of the present invention shall be determined primarily by the patent claims; however, configurations directly derived from the descriptions in the patent claims, as well as all modifications or variations derived from configurations equivalent thereto, shall be interpreted as being included within the scope of the rights of the present invention. Explanation of the symbols

[0092] 110: 1st Database 120: 2nd Database 130: Learning Model 131: First Process Learning 132: Second process 133: Third process 134: 4th Process 140: 1st Labeling Work Unit 150: 2nd Labeling Work Unit

Claims

Claim 1 A first database providing at least one set of labeled WBM image data; a second database providing multiple test data set control groups consisting of multiple unlabeled WBM image data; a first process for receiving WBM image data from the first database and learning defect patterns and normal patterns; a second process for receiving WBM image data from the second database based on the learned pattern information and predicting the defect probability of the WBM image data in the second database; a Minimum Confidence technique for calculating uncertainty values ​​for each data based on the predicted defect probability values, wherein the lowest defect probability value among the defect probability values ​​predicted by the learning model for eight defect patterns is selected; and a Minimum Margin technique for calculating the difference between the second highest defect probability value and the highest defect probability value predicted by the learning model for eight defect patterns. A learning model comprising: a third processor characterized by calculating three uncertainty values ​​for each WBM image data through an entropy technique that calculates the entropy of the distribution of defect probability values ​​predicted by the learning model for eight defect patterns; and a fourth process that determines, based on the result of comparing the calculated uncertainty values ​​for each data with a preset threshold value, to perform an auto-labeling operation using a labeling model on the corresponding WBM image data if all values ​​included in the three uncertainty values ​​are below the threshold value, and to perform an engineer labeling operation by a worker if at least one of the values ​​included in the uncertainty values ​​is above the threshold value; and a first labeling unit that inputs the auto-labeled WBM image data into a first database according to the determination result of the learning model and provides it as training data for the learning model.A collaborative intelligence-based semi-automated semiconductor wafer labeling system comprising: a second labeling unit that inputs engineer-labeled WBM image data into a first database based on the judgment result of a learning model and provides it as training data for a learning model. Claim 2 A collaborative intelligence-based semi-automated semiconductor wafer labeling system according to claim 1, characterized in that the learning model predicts the defect probability of WBM image data provided from a second database based on information learned for at least eight defect patterns including Center, Donut, Edge-Loc, Edge-Ring, Loc, Scratch, Random, and Near-full, and one normal pattern. Claim 3 A collaborative intelligence-based semiconductor wafer semi-automated labeling system according to claim 1, characterized in that the number of WBM images in the test data set provided to the learning model is equal to the number of WBM images in the initial training data set. Claim 4 delete Claim 5 delete Claim 6 A collaborative intelligence-based semiconductor wafer semi-automated labeling system according to claim 1, characterized in that the learning model predicts a defect probability for one unlabeled test data set among the control groups of the second database based on information learned from training data added through the first labeling work unit and the second labeling work unit, calculates an uncertainty value for each data based on the predicted defect probability value, and determines whether to perform auto-labeling work using the labeling model or engineer labeling work by a worker on the corresponding WBM image data according to the result of comparing the calculated uncertainty value with a preset threshold value, and repeats the process until all control groups of the second database are labeled. Claim 7 A collaborative intelligence-based semiconductor wafer semi-automated labeling system according to claim 6, characterized in that as the labeling work for the control groups of the second database is repeated, the ratio of engineer labeling work gradually decreases and the ratio of auto-labeling work gradually increases. Claim 8 A step of preparing a training data set consisting of labeled WBM image data through a first database (S10); a step of providing the training data set to a learning model to learn defect patterns and normal patterns (S20); a step of preparing multiple test data set control groups consisting of multiple unlabeled WBM image data through a second database (S30); a step of inputting one test data set among the control groups into a learning model, and the learning model predicting the defect probability of the test data set based on previously learned pattern information (S40); a step of calculating an uncertainty value for each data based on the defect probability value predicted by the learning model, wherein the uncertainty value for each data is a Minimum Confidence technique that selects the lowest defect probability value among the defect probability values ​​predicted by the learning model for eight defect patterns; and a Minimum Margin technique that calculates the difference between the second highest defect probability value and the highest defect probability value predicted by the learning model for eight defect patterns. A step (S50) comprising three uncertainty values ​​for each WBM image data through an entropy technique that calculates the entropy of the distribution of defect probability values ​​predicted by the learning model for eight defect patterns; a step (S60) determining whether to perform an auto-labeling operation using a labeling model on the corresponding WBM image data if all three uncertainty values ​​are below the threshold value based on the result of comparing the calculated uncertainty values ​​with a preset threshold value, and determining whether to perform an engineer labeling operation by a worker if any one of the three uncertainty values ​​is above the threshold value; a step (S70) performing an auto-labeling operation or an engineer labeling operation according to the determination result; a step (S80) adding training data by inputting the labeled WBM image data into a first database; and a step (S90) providing the added training data to the learning model to learn the added defect patterns.A semiconductor wafer labeling method using a collaborative intelligence-based semi-automated semiconductor wafer labeling system, comprising: a step (S100) of inputting one unlabeled test data set among the control groups of a second database into a learning model, and the learning model predicting the defect probability of the test data set based on previously learned defect pattern information; and a step (S110) of repeating steps S50 through S100 until all control groups of the second database are labeled. Claim 9 A semiconductor wafer labeling method using a collaborative intelligence-based semiconductor wafer semi-automated labeling system, wherein in claim 8, the above-mentioned step S40 predicts defect probabilities for at least 8 defect patterns including Center, Donut, Edge-Loc, Edge-Ring, Loc, Scratch, Random, and Near-full, and 1 normal pattern. Claim 10 A semiconductor wafer labeling method using a collaborative intelligence-based semiconductor wafer semi-automated labeling system, characterized in that, in claim 8, the number of WBM images in the test data set provided to the learning model in step S40 is equal to the number of WBM images in the training data set in step S10. Claim 11 delete Claim 12 delete Claim 13 A semiconductor wafer labeling method using a collaborative intelligence-based semiconductor wafer semi-automated labeling system, characterized in that, in claim 8, as the labeling work for the control groups of the second database is repeated, the ratio of engineer labeling work gradually decreases and the ratio of auto-labeling work gradually increases.

Images