Pattern matching device

Abstract

This disclosure proposes a method that significantly reduces the workload associated with data collection and truth-based methods in pattern matching using machine learning, thereby simplifying manual tasks in the learning process of the machine learning model (creation of the learning dataset, learning of the machine learning model, performance evaluation, etc.). The pattern matching apparatus of this disclosure performs pattern matching without using a first machine learning algorithm. If the pattern matching based on the first algorithm is successful, the result is supplied to the learner as learning data for relearning (see Figure 2).

Description

Technical Field

[0001] This disclosure relates to techniques for implementing pattern matching between images. Background Technology

[0002] Apparatus for measuring and inspecting patterns formed on semiconductor wafers mostly utilizes template matching technology to match the field of view of the apparatus with the desired measurement or inspection location. Template matching is the process of finding the region in the searched image that best matches a pre-registered template image. Patent Document 1 describes an example of such template matching.

[0003] Patent Document 2 describes a template matching device with a learning function that estimates a related image having a pixel value representing the correlation between a template image and a searched image. This device can learn to absorb differences in appearance between the template image and the searched image, enabling high-precision matching.

[0004] Patent Document 3 describes a method for selecting the minimum necessary training data to achieve the target performance of a machine learning model, based on semiconductor design data, when applying machine learning to a semiconductor inspection apparatus that captures patterns formed on a semiconductor wafer. In flexibly applying machine learning to image inspection, collecting large amounts of image data and producing accurate inspection results corresponding to each image data point places a significant burden on the apparatus user. Patent Document 3 reduces this burden by selecting the minimum necessary data during the learning process of the machine learning model.

[0005] Existing technical documents

[0006] Patent documents

[0007] Patent Document 1: Japanese Patent No. 4218171 (corresponding to US Patent No. 6627888)

[0008] Patent Document 2: WO2022 / 009357

[0009] Patent Document 3: Japanese Patent Application Publication No. 2020-35282 Summary of the Invention

[0010] The problem that the invention aims to solve

[0011] As a template matching technique used in semiconductor inspection and measurement equipment, conventionally, rule-based pattern matching methods have been employed, where matching is performed according to a predetermined algorithm. Generally, rule-based methods can be designed to handle a wide range of known scenarios, thus considered highly versatile. However, when matching fails due to discrepancies in appearance between the template image and the searched image, the algorithm cannot be easily improved, and the process is time-consuming.

[0012] Therefore, a pattern matching method using machine learning with learning capabilities, as described in Patent Document 2, was proposed. This method learns to absorb the appearance discrepancies between the template image and the searched image, enabling high-precision matching. However, to learn a machine learning model with sufficient performance, a large amount of learning data, such as images, needs to be collected, and correct pattern matching results (hereinafter referred to as ground truth) need to be generated based on the collected learning data. Furthermore, generating ground truths from a large amount of data requires manual work and time, thus becoming a significant burden for the operator. Moreover, when the types of data included in the learning data (image shape, image quality, etc.) are insufficient, the learning becomes only relevant to specific data, reducing the versatility of the machine learning model.

[0013] Patent Document 3 describes a method for analyzing semiconductor design data and captured images to select and collect the minimum necessary learning data to achieve the target performance of a machine learning model. This allows for minimization while ensuring the variety of learning data, thus improving the versatility of the machine learning model while reducing the effort required for ground truth creation. The document also describes creating a ground truth image in segmentation, a machine learning task, to simplify the ground truth creation process for each pixel. However, it does not eliminate the need for manual intervention, leaving room for automation.

[0014] This disclosure was made in view of the above-mentioned issues and proposes a method that significantly reduces the workload associated with data collection and truth-making in pattern matching methods using machine learning, thereby simplifying manual tasks in the learning process of machine learning models (such as creating learning datasets, learning machine learning models, and performance evaluation).

[0015] Methods for solving problems

[0016] The pattern matching apparatus disclosed herein performs pattern matching without using a first algorithm for machine learning, and when the pattern matching based on the first algorithm is successful, the result is supplied to the learner as learning data for the learner to perform relearning.

[0017] Invention Effects

[0018] According to the pattern matching apparatus disclosed herein, the workload associated with collecting learning data and creating truth tables used in machine learning models can be reduced, making manual tasks in the machine learning model learning process (creation of learning datasets, machine learning model learning, performance evaluation, etc.) less labor-intensive. Therefore, the burden on users of devices equipped with machine learning-based pattern matching can be significantly reduced. Attached Figure Description

[0019] Figure 1 This is a structural diagram of the pattern size measurement system according to Embodiment 1.

[0020] Figure 2 This is a flowchart illustrating the operation of the pattern size measurement system.

[0021] Figure 3 This is a flowchart explaining the specific actions of S206~S207.

[0022] Figure 4 This is a flowchart illustrating the specific actions of S214.

[0023] Figure 5 This is a flowchart illustrating the specific actions of S213.

[0024] Figure 6 This represents a data example that protects temporary storage data 406 and learning dataset 407.

[0025] Figure 7 Example data representing learning logs.

[0026] Figure 8 This section illustrates a structural example of a measuring device 300 that includes a SEM600 and a GUI307.

[0027] Figure 9 This is a flowchart illustrating the operation of the pattern size measurement system in Embodiment 3.

[0028] Figure 10 This is a block diagram of the process production apparatus 200. Detailed Implementation

[0029] <Implementation Method 1>

[0030] Figure 1 This is a structural diagram of the pattern size measurement system according to Embodiment 1 of this disclosure. The pattern size measurement system is a system for measuring the dimensions of a shape pattern formed on a semiconductor wafer. The pattern size measurement system includes: a design data storage unit 100, a process generation apparatus 200, a measurement apparatus 300 (pattern matching apparatus), and a machine learning model learning apparatus 400 (pattern matching apparatus). Data generated and output by each apparatus constituting the system can be transmitted and received between the apparatuses.

[0031] The pattern matching result is output as a matching offset. The matching offset represents the positional offset or difference between the design data 101 and the SEM (Scanning Electron Microscope) image 310, and can be represented by a two-dimensional scalar value, such as (offset in the X direction, offset in the Y direction). Furthermore, the template image is not limited to an SEM image; it can also be other types of images, such as the design data 101. Additionally, the searched image is not limited to the design data 101; it can also be other types of images, such as the SEM image 310.

[0032] The design data storage unit 100 holds design data 101 of the pattern structure of the semiconductor device created by the user of the display device.

[0033] The process generation apparatus 200 has an interface for the apparatus user to input measurement points on a device pattern referenced from the design data 101. Based on the measurement point information input from the interface and the design data 101, it outputs a measurement process 201 containing the shooting conditions in the measurement apparatus 300, the algorithm used in the rule-based pattern matching, measurement point information, etc.

[0034] The measuring device 300 measures the size of the device pattern according to the measurement process 201 and outputs the measurement value. The measuring device 300 includes: an imaging unit 301 for capturing the device pattern formed on the semiconductor wafer, a pattern matching unit 302 according to rules, a machine learning pattern matching unit 303, a matching success determination unit 304, a measurement unit 305, a measurement success determination unit 306, a GUI (Graphical User Interface) 307, and a measurement information storage unit 308.

[0035] The imaging unit 301 captures a device pattern according to the imaging conditions of the measurement process 201 and outputs an SEM image 310. The rule-based pattern matching unit 302 uses the algorithm described in the measurement process 201 to perform pattern matching between the SEM image 310 and the design data 101, and outputs a matching offset 314 between the two images. The machine learning pattern matching unit 303 uses a machine learning model 312 to perform pattern matching between the SEM image 310 and the design data 101, and outputs a matching offset 314 between the two images. The data used in pattern matching is not limited to the aforementioned combination; any data capable of determining the consistent area of ​​the field of view between the template image and the searched image can be used. For example, an SEM image and a design drawing obtained from the design data 101 can be used, or point group information from the SEM image and the design data 101 can be used. The matching success / failure determination unit 304 determines whether the matching is successful or not based on the design data 101, the SEM image 310, and the matching offset 314, and outputs a matching success / failure determination result. The measurement unit 305 measures the dimensions of the device pattern based on the measurement point information, design data 101, and SEM image 310 recorded in the measurement process 201, and outputs the measurement value. The measurement success / failure determination unit 306 determines the success or failure of the measurement based on the measurement value and outputs the measurement success / failure determination result. If the measurement success / failure determination result fails, the device user operates the GUI 307 to input the correct matching offset between the design data 101 and the SEM image 310 to ensure the pattern shape matches correctly, and outputs the matching offset 314. The measurement information storage unit 308 stores the measurement process 201, SEM image 310, design data 101, and machine learning model 312. The machine learning model 312 maintains the basic model learned by the machine learning model learning device 400 in the initial state.

[0036] After the measurement device 300 completes the dimensional measurement of the device pattern, it outputs the measurement process 201, design data 101, SEM image 310, and matching offset 314 to the machine learning model learning device 400. The output design data 101 and SEM image 310 can be used as learning data for the machine learning model, and the matching offset 314, being a value that enables normal dimensional measurement of the device pattern, can be used as a ground truth for the machine learning model. Therefore, manual work in collecting learning data and creating ground truth values ​​is almost eliminated.

[0037] The machine learning model learning device 400 includes: a learning data parsing unit 401, a machine learning model learning unit 402, a GUI 403, a learning data storage unit 404, and a machine learning model storage unit 405. The learning data storage unit 404 stores temporary storage data 406 and a learning dataset 407. The temporary storage data 406 includes: a measurement process storage folder 4042, a design data storage folder 4043, and a SEM image storage folder 4044. The temporary storage data 406 also includes a temporary data table 4047 and a collected data table 4049. The learning dataset 407 includes: a basic model learning data table 4045, a testing data table 4046, and an append-learning data table 4048. Specific examples of these data structures will be described later.

[0038] The machine learning model learning device 400 stores the measurement process 201, design data 101, and SEM image 310 received from the measurement device 300 into the measurement process storage folder 4042, design data storage folder 4043, and SEM image storage folder 4044 of the temporary storage data 406, respectively. At this time, the machine learning model learning device 400 creates a collection data table 4049 that summarizes arbitrary data IDs, the filenames of measurement process 201, design data 101, and SEM image 310, and the value of the matching offset 314 into a single record, and stores the created data table in the temporary storage data 406. The collection data table 4049 is created to manage the collected data, but the method of managing the collected data is not limited to this; any method that can manage combinations of measurement process 201, design data 101, SEM image 310, matching offset 314, etc., is acceptable. For example, it can also be managed through folder structure and filenames.

[0039] The learning data parsing unit 401 reads in each record from the existing data tables (e.g., basic model learning data table 4045) in the collected data table 4049 and the learning dataset 407, and extracts features such as pattern shape features, SEM image features, and shooting conditions (regarding the extraction method). Figure 5 (To be described later). Furthermore, the learning data analysis unit 401 selects records from each data table in a way that ensures the extracted features are unbiased (the method for selecting in a way that ensures unbiased selection will use...). Figure 5(As will be described later), the data (measurement process 201, design data 101, SEM image 310) selected from the data collection table 4049 are stored in the measurement process storage folder 4042, the design data storage folder 4043, and the SEM image storage folder 4044. Additionally, the records selected from each data table are saved as new data tables with unique names. This ensures a sufficient variety of data types in the learning dataset 407, improving the performance of the machine learning model while maintaining its versatility. Furthermore, new learning datasets can be created while retaining previously used learning datasets, making it easy to manage multiple learning datasets. When parsing data in the learning data parsing unit 401, errors in the matching offset 314 can be manually corrected using the GUI 403, as described later.

[0040] The machine learning model learning unit 402 reads in the data required for learning (design data 101, SEM image 310, matching offset 314, etc.) from the data table of the learning dataset 407 (e.g., append learning data table 4048, etc.), and learns the machine learning model according to the default learning conditions (details of the machine learning model's learning method can be found using...). Figure 3 , Figure 4 (To be described later). To evaluate the performance of the learned machine learning model, the machine learning model learning unit 402 reads the data required for inference (design data 101, SEM image 310, matching offset 314, etc.) from the test data table 4046 stored in the learning dataset 407, and evaluates the performance of the learned model. Even if the target performance is not achieved, the learning conditions can be changed and relearning can be performed. Thus, a machine learning model with the desired performance can be obtained. If the target performance is achieved, the data files for the learning log and the learned machine learning model 408 are appended with unique names and stored in the machine learning model 312. At this point, the previously learned machine learning model is retained.

[0041] If an error exists in the matching offset 314 and the correct matching offset needs to be taught, the device user operates the GUI 403 to teach and output the correct matching offset that matches the pattern shape of the design data 101 with that of the SEM image 310.

[0042] Figure 2 This is a flowchart illustrating the operation of the pattern size measurement system. The following is... Figure 2 The steps are explained below.

[0043] ( Figure 2 (S200)

[0044] The process generation apparatus 200 generates a measurement process 201 based on the measurement point information of the device pattern input by the device user from the interface of the process generation apparatus 200 and the design data 101 obtained from the design data storage unit 100. This process includes the shooting conditions in the measurement device 300 (sometimes including shooting position information, etc.), the algorithm used in the pattern matching according to the rules, the measurement point information, etc.

[0045] ( Figure 2 (S201)

[0046] The imaging unit 301 of the measuring device 300 captures the device pattern according to the imaging conditions of the measurement process 201 and outputs a SEM image 310.

[0047] ( Figure 2 (S202)

[0048] The matching offset 314 between the design data 101 and the SEM image 310 is calculated by the pattern matching unit 302. The data used in pattern matching is not limited to this combination; any data capable of determining the consistent areas of the field of view between the template image and the searched image can be used. For example, the SEM image 310 and the design drawing obtained from the design data 101 can be used, or the point group information of the SEM image 310 and the design data 101 can be used. The matching offset 314 is calculated using a matching algorithm determined by the measurement process 201. The matching algorithm can be, for example, an algorithm that calculates the similarity between each candidate position of the design data 101 and the SEM image 310 using mean square error, and outputs the matching offset of the position with the highest similarity. The matching algorithm is not limited to this method; any algorithm that extracts image features according to a predetermined process and calculates the matching offset based on the extracted image features is acceptable.

[0049] ( Figure 2 (S203)

[0050] The matching success / failure determination unit 304 determines whether the matching is successful or not based on the design data 101, the SEM image 310, and the matching offset 314. As an example of a method for determining matching success / failure, consider the following: Calculate the mean square error (MSE) representing the similarity between the design data 101 and the SEM image 310 based on the matching offset 314; if the value is above a threshold, the matching is considered successful. The method for determining matching success / failure is not limited to this method; it is acceptable as long as it can be determined whether the position of the matching offset 314 on the design data 101 corresponds to the field of view of the SEM image 310. If the matching success / failure determination result is successful, skip step S208. If the matching success / failure determination result is unsuccessful, proceed to step S204.

[0051] ( Figure 2(S204)

[0052] The machine learning pattern matching unit 303 calculates the matching offset 314 using a machine learning model that has been learned to estimate the matching offset 314 based on the design data 101 and the SEM image 310. The data used in pattern matching is not limited to the aforementioned combination; any data capable of determining the consistent area of ​​the field of view between the template image and the searched image can be used. For example, the SEM image and the design drawing obtained from the design data 101 can be used, or the point group information of the SEM image and the design data 101 can be used. The machine learning model is a machine learning model read from the machine learning model 312, specifically a machine learning model (such as the highest-performing or the latest model). As an example, the machine learning-based pattern matching method can be implemented by estimating the related images between images, as in Patent Document 2, and calculating the matching offset based on the related images. The machine learning-based pattern matching method is not limited to this; any method capable of determining the consistent area of ​​the field of view between the template image and the searched image can be used. For example, a method that directly estimates the matching offset between images using a machine learning model can be employed.

[0053] ( Figure 2 (S205)

[0054] The matching success / failure determination unit 304 determines whether the matching is successful or not based on the matching offset 314. If the matching success / failure determination result is successful, S208 is implemented. If the matching success / failure determination result is unsuccessful, the design data 101 and the SEM image 310 are stored in the temporary storage data 406 of the machine learning model learning device 400, and S206 is implemented for learning the machine learning model. At this time, the machine learning model learning device 400 creates a temporary data table 4047 that records the filenames of the received design data 101 and SEM image 310, appends the inherent names, and stores it in the temporary storage data 406.

[0055] ( Figure 2 (S206)

[0056] The user uses GUI 403 to read in design data 101 and SEM image 310 according to temporary data table 4047. The user operates GUI 403 to teach the correct matching offset for patterns in the two images to match. The method for teaching the correct matching offset is not limited to the method using GUI 403; any method capable of teaching the correct matching offset is acceptable. For example, multiple rule-based pattern matching algorithms can be used to perform pattern matching on design data 101 and SEM image 310, and the most similar matching offset can be taken as the correct matching offset. The machine learning model learning device 400 generates a data table 4041 containing the generated correct matching offset, design data 101, and failed pattern learning data for SEM image 310. For details, please refer to... Figure 3 The details will be provided later.

[0057] ( Figure 2 (S207)

[0058] The machine learning model learning unit 402 uses the failure pattern learning data table 4041 to implement the learning of the machine learning model. The learned machine learning model 408 and the learning log are stored in the machine learning model storage unit 405. Additionally, the learned machine learning model 408 is stored in the machine learning model 312 of the measurement device 300. For more details... Figure 3 This will be described later. S204 is then implemented again using the learned machine learning model 408.

[0059] ( Figure 2 (S208)

[0060] The measurement unit 305 outputs its measurement values ​​based on the measurement process 201, design data 101, SEM image 310, matching offset 314, and the size of the measurement device pattern.

[0061] ( Figure 2 (S209)

[0062] The measurement success / failure determination unit 306 determines the success or failure of the measurement based on the measurement process 201, design data 101, SEM image 310, and measured values, and outputs the success / failure determination result. If the measurement success / failure determination result is successful, S212 is skipped. If the measurement success / failure determination result is unsuccessful, S210 is performed to implement manual assistance.

[0063] ( Figure 2 (S210)

[0064] The user of the device uses GUI307 to input a matching offset 314, based on the design data 101 and the SEM image 310, to ensure that the patterns of the two are consistent. The matching offset can be input directly as a numerical value, or it can be calculated while simultaneously verifying the image that makes the design data 101 and the SEM image 310 coincide, to determine the correct matching position.

[0065] ( Figure 2 (S211)

[0066] The measurement unit 305 outputs its measurement values ​​based on the measurement process 201, design data 101, SEM image 310, matching offset 314, and the size of the measurement device pattern.

[0067] ( Figure 2 (S212)

[0068] After measuring the dimensions of the device pattern, the measuring device 300 outputs the measurement process 201, design data 101, SEM image 310, and matching offset 314 to the machine learning model learning device 400. The machine learning model learning device 400 stores the measurement process 201, design data 101, and SEM image 310 received from the measuring device 300 into the measurement process storage folder 4042, design data storage folder 4043, and SEM image storage folder 4044 of the temporary storage data 406, respectively. At this time, the machine learning model learning device 400 records the filenames of the received measurement process 201, design data 101, and SEM image 310, as well as the matching offset 314, in the collected data table 4049.

[0069] ( Figure 2 (S213)

[0070] The learning data parsing unit 401 parses the data stored in temporary storage 406 to create a new learning dataset. For details, please refer to [link / reference]. Figure 5 S213 can be performed each time the pattern is measured, or it can be performed after a certain amount of data has been collected.

[0071] ( Figure 2 (S214)

[0072] The machine learning model learning unit 402 performs machine learning model training based on the new learning dataset created in S213, and saves the learning log and the learned machine learning model 408 in the machine learning model storage unit 405. Additionally, the learned machine learning model 408 is stored in the machine learning model 312 of the measurement device 300. For more details, please refer to... Figure 4S214 can be performed each time a pattern is measured, or it can be performed after a certain amount of data has been collected.

[0073] Figure 3 This is a flowchart explaining the specific actions of S206~S207. The following is... Figure 3 The steps are explained below. S206 to S207 can be executed immediately if the pattern matching based on machine learning fails, or they can be executed later after the semiconductor wafer is unloaded from the measuring device 300.

[0074] ( Figure 3 (S300)

[0075] The user uses GUI 403 to read design data 101 and SEM image 310 from temporary data table 4047, and inputs the correct matching offsets for the pattern shapes of the two images to be consistent. Then, a failure pattern learning data table 4041 is created, containing the input correct matching offsets, design data 101, and SEM image 310. At this time, all or part of the records from the base model learning data table 4045 or the supplementary learning data table 4048 can be appended to the failure pattern learning data table 4041. This ensures the general performance of the machine learning model while corresponding to failure patterns.

[0076] ( Figure 3 (S301)

[0077] The machine learning model learning unit 402 learns the machine learning model based on the failure pattern learning data table 4041 and predetermined learning conditions, and outputs the learned machine learning model 408. Generally, the learning conditions of the machine learning model include the number of times each record in the data table is learned (hereinafter referred to as the epoch number), the learning rate and other values ​​that the machine learning model can arbitrarily set (hereinafter referred to as hyperparameters), and whether there is a model as the basis for learning. The internal parameters of the machine learning model can be updated with all the content of the model as the basis for learning, or only a part of it can be updated. The predetermined learning conditions are preferably the optimal values ​​obtained empirically through prior experiments, etc., but are not limited to this. For example, the epoch number can also be set until a certain performance is achieved while confirming the progress of learning. The model as the basis for learning is preferably the machine learning model that has been learned from the base model learning data table 4045, or the latest machine learning model stored in the machine learning model storage unit 405, but is not limited to this. For example, a new machine learning model can also be learned. The learned machine learning model 408 can also be used as the basis for future learning.

[0078] ( Figure 3 (S302)

[0079] The machine learning model learning unit 402, having been taught in S300 to use the design data 101 with the correct matching offset and the SEM image 310 as input, performs inference on the learned machine learning model 408 to evaluate whether the matching is successful. If the matching is successful, S303 is executed. If it fails, S301 is executed again. At this time, records selected from the base model learning data table 4045 or the supplementary learning data table 4048 in the failure pattern learning data table 4041 can also be deleted. Thus, although the generality of the machine learning model is reduced, it can be specialized for failure situations, and there is a possibility that the matching will turn into success. The learned machine learning model 408 can correspond to the failure pattern, therefore, it can quickly respond to matching failures.

[0080] ( Figure 3 (S303)

[0081] The machine learning model learning unit 402 stores the learned machine learning model 408 and the learning log in the machine learning model storage unit 405.

[0082] ( Figure 3 (S304)

[0083] The learned machine learning model 408 is stored in machine learning model 312.

[0084] Figure 4 This is a flowchart illustrating the specific actions of S214. The following is... Figure 4 The steps are explained below.

[0085] ( Figure 4 (S400)

[0086] The machine learning model learning unit 402 uses the supplementary learning data table 4048 to implement the learning of the machine learning model according to predetermined learning conditions. The predetermined learning conditions are preferably optimal values ​​obtained empirically through prior experiments, but are not limited to this; for example, the number of epochs can be set until a certain performance is achieved while confirming the progress of learning. Furthermore, the model used as the basis for learning is preferably a machine learning model that has been learned using the base model learning data table 4045, or the latest machine learning model stored in the machine learning model storage unit 405 based on the learning log, but is not limited to this; for example, a newly learned machine learning model can also be used. The method for selecting the model as the basis for learning is not limited to referring to the learning log; any model that can be selected as the optimal basis for learning is acceptable. For example, the latest machine learning model can be determined based on the machine learning model's save date and time, or it can be determined based on the machine learning model's filename.

[0087] ( Figure 4(S401)

[0088] The machine learning model learning unit 402 uses data from the test data table 4046 (with known truth values) as input to perform inference on the learned machine learning model 408 and evaluate its performance. Regarding the test data table 4046, to evaluate the general performance of the machine learning model, data with various shape features can be kept in the records as other data; for example, data that the device user is expected to select for successful matching can also be kept. As a performance evaluation method, it can be set to an accuracy rate above a threshold, or other methods, such as ensuring the difference between the inferred matching offset and the true matching offset is sufficiently small, or a combination of these methods. If sufficient performance is obtained, step S403 is implemented. If performance is insufficient, step S402 is implemented.

[0089] ( Figure 4 (S402)

[0090] The machine learning model learning unit 402 can also verify the data included in the supplementary learning data table 4048 and delete data that may adversely affect learning (e.g., data where the template image does not exist within the searched image's field of view due to field of view shift, data where the shape of the device pattern formed on the wafer differs significantly from the design data due to deformation, data where image quality or other characteristics differ significantly from other data within the same data table, etc.) (adverse data types). Additionally, learning conditions can be adjusted. Methods for adjusting learning conditions include changing the model used as the basis for learning, increasing the number of epochs if learning is insufficient, etc. Afterwards, S400 is executed again. This step is optional and may not be performed.

[0091] ( Figure 4 (S403)

[0092] The machine learning model learning unit 402 stores the learned machine learning model 408 and the learning log in the machine learning model storage unit 405.

[0093] ( Figure 4 (S404)

[0094] The learned machine learning model 408 is stored in machine learning model 312.

[0095] Figure 5 This is a flowchart explaining the specific actions of S213. The following is... Figure 5 The steps are explained below.

[0096] ( Figure 5 (S500)

[0097] The learning data analysis unit 401 reads the design data, SEM images, and measurement processes contained in the records from the collected data table 4049, the failed pattern learning data table 4041, the latest supplementary learning data table 4048, and the basic model learning data table 4045. All four data tables can be used, or only a portion of them can be used. The learning data analysis unit 401 analyzes the read data, extracting features such as pattern shape characteristics, SEM image features, and shooting conditions. Specific examples of each feature are as follows: Pattern shape features include pattern type (line, spaced pattern, hole array pattern, etc.), pattern density, and constituent materials. SEM image features include the presence or absence of transmission of the underlying pattern, deformation, and image quality. Shooting conditions include accelerating voltage, detector type, shooting magnification, cumulative frame rate, and high-quality processing. Specific extraction methods for pattern shape features include, for example, inferences based on information obtained from image edge and corner detection, and image frequency analysis. Alternatively, the process maker can include this information in the measurement process itself during the measurement process fabrication. Specific methods for extracting features from SEM images include extracting the outline of the pattern area from both the SEM image and the design data, comparing the extracted outlines, and thereby inferring the presence or absence of transmission of the underlying pattern, the amount of pattern deformation, etc. Furthermore, the process maker can also include this information in the measurement process during fabrication. The imaging conditions can directly use the information recorded in the measurement process, or some preprocessing can be performed. The extracted features are not limited to these; they can also include the model number of the measuring device 300, the identification number of the same type of device, the process name (photolithography process, etching process, etc.) recorded by the process maker in the measurement process during fabrication.

[0098] ( Figure 5 (S501)

[0099] The learning data parsing unit 401 selects records from each data table in a way that ensures the features extracted in S500 are unbiased, and creates a new supplementary learning data table. For example, a method for selecting records may involve clustering based on the characteristics of the records, selecting records from each cluster in a way that eliminates bias in each cluster. As a specific clustering method, it is possible to apply methods such as k-means clustering, where each record is configured in a space with a dimension corresponding to the number of extracted features, but this method is not limited to any particular method; any method that can cluster similar records into a single class based on the number of extracted features is acceptable.

[0100] ( Figure 5 (S502)

[0101] The device user confirms the matching offset in the new supplementary learning data table in GUI403, and manually corrects the incorrect matching offset if errors exist. The method for correcting the matching offset is not limited to the method used in GUI403; any method capable of correcting matching errors is acceptable. For example, multiple rule-based pattern matching algorithms can be used to perform pattern matching on the design data 101 and the SEM image, and the most similar matching offset can be taken as the correct matching offset. Alternatively, as the object data for confirming the matching offset, only data with a high probability of matching offset errors can be selected beforehand, and their matching offsets can be confirmed. For example, a method could be to perform pattern matching on the design data 101 and the SEM image using multiple rule-based pattern matching algorithms, and select data from which similar matching offsets are not obtained in each algorithm. S502 may not be necessary. This is because the collected data is obtained from correct measurements, therefore, the probability of correct matching is high. This step is used to reliably provide correct data.

[0102] ( Figure 5 (S503)

[0103] The learning data parsing unit 401 stores the data (measurement process 201, design data 101, SEM image 310) contained in the temporary storage data 406 within the new supplementary learning data table 4048 in the measurement process storage folder 4042, the design data storage folder 4043, and the SEM image storage folder 4044. The new supplementary learning data table 4048 is given a unique name and stored in the learning dataset 407.

[0104] Figure 6 This represents the data instance that is temporarily stored in data 406 and maintained in the learning dataset 407. For example... Figure 6 As shown, temporary storage data 406 and learning dataset 407 contain multiple data tables for managing various data, a measurement process storage folder 4042, a design data storage folder 4043, and a SEM image storage folder 4044. Temporary storage data 406 and learning dataset 407 maintain three folders, respectively storing measurement processes, design data, and SEM images. However, all data files can also be stored at the same level as the data tables, in a single folder, or in multiple folders.

[0105] The learning dataset 407 is kept in its initial state as the base model for learning the machine learning model. The learning dataset 4045 contains the base model data, and the test dataset 4046 contains the test dataset for evaluating the performance of the machine learning model. The base model refers to the machine learning model that has been learned using patterns with high measurement frequency or general characteristics.

[0106] like Figure 6 As shown, each data table stores the data ID, measurement process, design data, SEM image, and matching offset in a single record. By maintaining multiple such data tables, multiple datasets can be managed as data tables. The dataset management method is not limited to table-based management; any method capable of managing combinations of measurement processes, design data, SEM images, and matching offsets is acceptable. For example, management can also be achieved through folder structures, filenames, etc. Furthermore, regarding each data table, while storing the data ID, measurement process, design data, SEM image, and matching offset information in a record, the amount of information stored can vary, as long as it contains the information necessary for data management. For example, the data ID may not be stored, or other information such as the registration date and time may be retained.

[0107] Figure 7 Example data representing learning logs. For example... Figure 7 As shown, the learning log can be recorded in a data table format, with each row containing the update date, machine learning model name, data table used, and learning conditions. This simplifies the management of machine learning models that have undergone learning and allows for the retrieval of past machine learning models as needed. Figure 7 In the learning log, the update date, machine learning model name, data table used, and learning conditions are recorded. However, the amount of information recorded can vary, as long as it is sufficient to manage the machine learning model after learning. For example, other information such as the machine learning model ID can also be recorded.

[0108] Figure 8 This section illustrates a structural example of a measuring device 300 including an SEM 600 and a GUI 307. The SEM 600 can function as an imaging unit 301. The control unit 614 can function as a rule-based pattern matching unit 302, a machine learning pattern matching unit 303, a matching success / failure determination unit 304, a measurement unit 305, and a measurement success / failure determination unit 306. The image memory 615 can function as a measurement information storage unit 308. The SEM 600, for example, measures the pattern dimensions of semiconductor devices formed on a semiconductor wafer. The arithmetic processing unit or computer system in the measuring device 300 can be configured as, for example, the control unit 614. The control unit 614 has a processing unit (e.g., a CPU / GPU 616) and a storage device (e.g., a memory including the image memory 615). Information can be stored in the storage device, such as a program related to pattern matching processing. The CPU / GPU 616 executes this program, and the process is then executed. Figure 2The pattern matching process is shown. That is, the control unit 614 functions as a pattern matching unit. In other words, this program enables the computer system to function as a processing unit included in the pattern matching unit, executing... Figure 2 The pattern matching process shown.

[0109] The SEM 600 generates an electron beam from the electron gun 601. The deflector 604 and objective lens 605 are controlled to focus the electron beam and irradiate any position on a semiconductor wafer 603, which serves as a sample placed on the stage 602. Secondary electrons are emitted from the irradiated semiconductor wafer 603 and detected by the secondary electron detector 606. The detected secondary electrons are converted into digital signals by an A / D converter 607. The image represented by the digital signal is stored in the image memory 615 within the control unit 614. This image is used, for example, as the SEM image 310, based on which pattern matching processing is performed by the control unit 614 or the CPU / GPU 616. The necessary settings and display of the processing results can be performed on the GUI 307. An optical camera 611 can also be used in alignments using an optical camera with a lower magnification than the SEM. The signal obtained by the optical camera 611 from the semiconductor wafer 603 is converted into a digital signal by the A / D converter 612 (the A / D converter 612 is unnecessary if the signal from the optical camera 611 is a digital signal). The image represented by the digital signal is stored in the image memory 615 in the control unit 614, and the CPU / GPU 616 performs image processing corresponding to the purpose.

[0110] The SEM 600 may also include a reflection electron detector 608. When the reflection electron detector 608 is included, reflected electrons emitted from the semiconductor wafer 603 are detected by the detector, and the detected reflected electrons are converted into digital signals by an A / D converter 609 or 610. The image represented by the digital signal is stored in the image memory 615 in the control unit 614, and the CPU / GPU 616 performs image processing corresponding to the purpose. The control unit 614 can control the stage 602 via the stage controller 630, or control the objective lens 605, etc., via the deflection control unit 631.

[0111] <Implementation Method 1: Summary>

[0112] The pattern size measurement system of Implementation 1 uses the successful results of rule-based matching or machine learning matching to implement the learning of the learner (S214). These successful results can be obtained each time pattern matching and pattern-based measurement are performed, and are therefore suitable as learning data. Thus, the workload of collecting learning data and providing correct solutions can be reduced.

[0113] <Implementation Method 2>

[0114] Embodiment 2 of this disclosure describes a structural example of a learner for selectively using matching objects suitable for specific characteristics. Embodiment 2, like Embodiment 1, is constructed by... Figure 1 The system comprises a pattern size measurement system. Its basic operation is the same as in Embodiment 1; the following mainly describes the differences in operation compared to Embodiment 1.

[0115] The design data storage unit 100, similar to Embodiment 1, stores design data 101 representing the pattern structure of a semiconductor device created by the device user.

[0116] In addition to the interface, the process generation apparatus 200 also includes a pattern feature analysis unit. This unit analyzes the design data 101 obtained from the design data storage unit 100 and the measurement process 201, extracting pattern shape features and imaging conditions. Specific examples of each feature are as follows: Pattern shape features include pattern type (lines, spacing patterns, hole array patterns, etc.), pattern density, and constituent materials. Imaging conditions include accelerating voltage, detector type, magnification, cumulative frame rate, and high-quality processing. Specific extraction methods for pattern shape features can include estimation based on information obtained from image edge and corner detection, image frequency analysis, etc. Alternatively, the user can input pattern type information through the interface. Imaging conditions can directly use information recorded in the measurement process, or some preprocessing can be performed. The extracted features are not limited to these; they can also include the model number of the measuring device, the identification number of the same type of device, and the process name (photolithography process, etching process, etc.) recorded by the process manufacturer during the measurement process.

[0117] The process generation apparatus 200 uses the inherent feature group IDs of pre-created feature groups (details of which will be described later) to select a machine learning model that has been trained on data from feature groups that classify the extracted features. The method for selecting the machine learning model is not limited to this method; any machine learning model that has been trained on data similar to the extracted features can be selected. For example, the apparatus user can also use the interface to select the optimal machine learning model for the design data.

[0118] The process generation apparatus 200 outputs a measurement process 201 based on the measurement point information input from the interface, the design data 101, and the selected machine learning model. This output includes the shooting conditions in the measurement device 300, the algorithm used in rule-based pattern matching, the machine learning model used in machine learning-based pattern matching, and the measurement point information. The selected machine learning model is pre-learned to specialize in specific features, resulting in high matching performance for similar data. Therefore, it can reduce matching failures in the machine learning pattern matching unit 303.

[0119] The measurement apparatus 300 includes: an imaging unit 301 for capturing device patterns formed on a semiconductor wafer, a rule-based pattern matching unit 302, a machine learning pattern matching unit 303, a matching success / failure determination unit 304, a measurement unit 305, a measurement success / failure determination unit 306, a GUI 307, and a measurement information storage unit 308. The machine learning pattern matching unit 303 retrieves a machine learning model used for pattern matching from the machine learning model 312 according to the measurement process 201. Alternatively, the machine learning pattern matching unit 303 receives a machine learning model used for pattern matching from the machine learning model storage unit 405 of the machine learning model learning device 400 according to the measurement process 201. The machine learning pattern matching unit 303 uses the acquired machine learning model to perform pattern matching.

[0120] The machine learning model learning device 400 includes: a learning data parsing unit 401, a machine learning model learning unit 402, a GUI 403, a learning data storage unit 404, and a machine learning model storage unit 405. The learning dataset 407 of the learning data storage unit 404 contains a pre-prepared dataset for additional learning for each feature group. The method for creating the dataset for each feature group will be described later. The machine learning model storage unit 405 stores the machine learning model for each feature group. The learning / management method for the machine learning model for each feature group will be described later.

[0121] The machine learning model learning device 400 stores the measurement process 201, design data 101, and SEM image 310 received from the measurement device 300 into the measurement process storage folder 4042, design data storage folder 4043, and SEM image storage folder 4044 of the temporary storage data 406, respectively. At this time, the machine learning model learning device 400 creates a collection data table 4049 that summarizes arbitrary data IDs, filenames of the measurement process 201, design data 101, and SEM image 310, as well as the values ​​of the matching offset 314 and feature group ID into a single record. This data table is then stored in the temporary storage data 406. The collection data table 4049 is created to manage the collected data, but the management method for the collected data is not limited to the method based on the collection data table 4049. Any method that can manage combinations of measurement processes, design data, SEM images, matching offsets, and feature group IDs is acceptable; for example, it can also be managed through folder structure and filenames.

[0122] Implementation method 2 is similar to implementation method 1, and can be achieved through... Figure 2 The flowchart shown illustrates the actions. Here, we will mainly focus on the second embodiment. Figure 2 The differences between the steps in this embodiment and those in Embodiment 1 will be described. Steps not described are the same as in Embodiment 1.

[0123] ( Figure 2 (S200)

[0124] The process generation apparatus 200 generates a measurement process 201 based on the measurement point information of the device pattern input by the device user from the interface, the design data 101 obtained from the design data storage unit 100, and the feature group obtained from the pattern feature analysis unit. This process includes the shooting conditions (sometimes including shooting position information, etc.) in the measurement apparatus 300, the algorithm used in rule-based pattern matching, the machine learning model used in machine learning-based pattern matching, and the measurement point information.

[0125] ( Figure 2 (S204)

[0126] The machine learning pattern matching unit 303 retrieves the machine learning model used for pattern matching from the machine learning model 312 according to the measurement process 201. Alternatively, it receives it from the machine learning model storage unit 405 of the machine learning model learning device 400. Then, the machine learning pattern matching unit 303 uses the acquired machine learning model to perform pattern matching.

[0127] ( Figure 2 (S206)

[0128] Using the same method as in Embodiment 1, the device user operates the GUI 403 to teach the correct matching offset. Afterwards, the machine learning model learning device 400 creates a failure pattern learning data table 4041 containing the input correct matching offset, design data 101, and SEM image 310. When creating the failure pattern learning data table 4041, all or part of the dataset used when learning the machine learning model obtained in S204 can be appended to the failure pattern learning data table 4041. Thus, it is possible to ensure the performance of the machine learning model for a specific pattern while corresponding to the failure pattern.

[0129] ( Figure 2 (S207)

[0130] The machine learning model learning unit 402 performs machine learning model learning using the same method as in Embodiment 1, and outputs the learned machine learning model 408. The model used as the basis for learning is preferably the machine learning model obtained in S204, but it is not limited to this; for example, a new machine learning model can also be learned. The learned machine learning model 408 is stored in the machine learning model 312 of the measuring device 300.

[0131] ( Figure 2 (S212)

[0132] After measuring the dimensions of the device pattern, etc., the measuring device 300 outputs the measurement process 201, design data 101, SEM image 310, and matching offset 314 to the machine learning model learning device 400. The machine learning model learning device 400 stores the measurement process 201, design data 101, and SEM image 310 received from the measuring device 300 into the measurement process storage folder 4042, design data storage folder 4043, and SEM image storage folder 4044 of the temporary storage data 406, respectively. At this time, the machine learning model learning device 400 records the filenames of the received measurement process 201, design data 101, and SEM image 310, as well as the values ​​of the matching offset 314 and the pattern feature group ID, in the collected data table 4049.

[0133] ( Figure 2 (S213)

[0134] The learning data parsing unit 401 extracts feature group ID information from the failure pattern learning data table 4041 and the collection data table 4049. Records from each data table are appended to the supplementary learning data table 4048 for each feature group, using the same feature group ID information. This allows for the collection and accumulation of learning data by feature group. Therefore, for specific features, a higher-performance machine learning model can be learned.

[0135] ( Figure 2 (S214)

[0136] The machine learning model learning unit 402 uses the supplementary learning data table 4048 created in S213 for each feature group, and implements the learning of machine learning models adapted to each feature group using the same learning method as in Embodiment 1. Each learned machine learning model, with its current learning logs and inherent name appended, is stored in the machine learning model storage unit 405. In addition to the information shown in Embodiment 1, feature group IDs can also be assigned to the learning logs. This allows for the association between each feature group and each machine learning model. The learned machine learning models are stored in the machine learning model 312 of the measurement device 300.

[0137] As a method for creating feature groups in Implementation Method 2, a clustering-based method is considered, for example. The method for creating feature groups based on the clustering method will be described below.

[0138] The design data and measurement process contained in the records of the pre-obtained data collection table 4049 and the basic model learning data table 4045 are read in. Both data tables can be used, or only one can be used. The read data is parsed and compared with implementation method 1 ( Figure 5 The same method is used to extract features such as pattern shape and shooting conditions. Records are then configured in a space with the number of extracted features, and clustering is performed on this space. Specific clustering methods can include, but are not limited to, k-means, etc., as long as they can cluster similar records into one class based on the number of extracted features. Each class obtained above is assigned an arbitrary feature group ID, and the arbitrary feature group IDs assigned to each record are recorded. Afterwards, data tables are created separately according to the feature group IDs.

[0139] So far, a method for creating feature groups based on clustering has been described. However, the method for creating feature groups is not limited to this. For example, it can also be done through a classification task based on feature group IDs assigned by the device user. Specifically, for some or all of the records obtained from the pre-obtained collection data table 4049 and the basic model learning data table 4045, the device user assigns feature group IDs based on pattern shape features, SEM image features, shooting conditions, and the device user's requirements for the machine learning model. The device user's requirements for the machine learning model may include, for example, "I want to prepare a machine learning model according to pattern type," "I want to classify the machine learning model according to the presence or absence of transmission in the lower layers of the SEM image," or "I want to prepare a machine learning model according to the measuring device." Then, features such as pattern shape features and shooting conditions are extracted from each record. Each record is configured in a space with the number of extracted features, and classification is performed on this space with the feature group ID assigned by the device user as the ground truth. As a specific classification method, support vector machines, which are a general method in classification tasks, are considered, but are not limited to this method, as long as the feature group ID can be inferred from features such as pattern shape features and shooting conditions. The resulting feature group classification model can estimate the feature group IDs that reflect the user's requirements based on the input design data and measurement process information. The feature group IDs estimated by the feature group classification model are then recorded in each record. Afterwards, data tables are created separately for each feature group ID.

[0140] The records in the data tables for each feature group obtained through the two methods described above have similar characteristics. Therefore, by using these records in append-only learning, machine learning models specialized for specific features can be obtained. Preferably, the machine learning model for each feature group is learned using the machine learning model created from data table 4045 as the base model. However, this is not a limitation; other methods may be used, such as learning a new machine learning model without using the base model. The machine learning model that has undergone learning for each feature group is given a unique name and saved along with the learning log.

[0141] Feature-group-specific machine learning models can identify corresponding feature groups and model versions based on learning logs, etc. This information can be used to select machine learning models for pattern matching and as the foundation for learning. The method for identifying machine learning models is not limited to this one; any method that selects the optimal model is acceptable. For example, the optimal machine learning model can also be selected based on its filename, save date, etc.

[0142] The pattern feature analysis unit of the process generation apparatus 200 uses the class information or feature group classification model of each feature group obtained by the above method to select the best machine learning model for the input data. The specific process is as follows.

[0143] (1) Extract pattern shape features, shooting conditions, etc. from design data 101 and measurement process 201.

[0144] (2) Use the class information or feature group classification model of each feature group to infer the feature group ID for the features extracted in (1).

[0145] (3) The machine learning model learned using the data of the feature group estimated in (2) is used as the machine learning model used in pattern matching and is recorded in the measurement process 201.

[0146] <Implementation Method 3>

[0147] Figure 9 This is a flowchart illustrating the operation of the pattern size measurement system according to Embodiment 3 of this disclosure. In Embodiment 1, rule-based pattern matching is first performed, and if that fails, machine learning-based pattern matching is further performed. Alternatively, the algorithm initially implemented may be suitable for either rule-based or machine learning-based algorithms. Specifically, between S201 and S202, S901 to S902, as described below, are performed. Everything else is the same as in Embodiments 1 and 2.

[0148] ( Figure 9 (S901)

[0149] The process generation apparatus 200 analyzes the measurement process 201, and based on information from past measurement processes 201 and their corresponding matching success or failure results, refers to the matching success or failure results of similar measurement processes 201, determining whether rule-based pattern matching or machine learning-based pattern matching is more appropriate. It can be considered that cases where past matching was successful (or with a higher success rate) are more appropriate. If no measurement process 201 with similar information exists, or if no past matching success or failure results exist, rule-based pattern matching is selected. In other words, at the time of generating the measurement process 201, the matching algorithm corresponding to that measurement process 201 is also determined. Regarding the specific operations of the process generation apparatus 200, [further details are needed]. Figure 10 The details will be provided later.

[0150] ( Figure 9 (S902)

[0151] The measuring device 300 switches the pattern matching algorithm based on the result of S901. If rule-based pattern matching is specified, S202 is performed. If a machine learning-based pattern matching algorithm is specified, S204 is skipped. This step allows the selection of the optimal pattern matching method before performing pattern matching processing, thus reducing unnecessary processing and improving the matching success rate.

[0152] Figure 10 This is a block diagram of the process generation apparatus 200. The process generation apparatus 200 analyzes and measures the process 201, selecting the optimal pattern matching method. The measurement process 201 includes pattern type, imaging conditions, and apparatus number. Pattern type includes pattern shape information, pattern density information, semiconductor pattern constituent materials, whether the underlying semiconductor pattern is transparent, and the amount of deformation generated. Imaging conditions include the accelerating voltage of the electron microscope used for imaging, detector type, imaging magnification, the number of image frames accumulated to reduce noise, and the presence or absence of pre-processing for high-quality images based on AI technology and image processing technology. Apparatus number includes apparatus model and identification number for similar apparatuses.

[0153] The process generation unit 210 creates the measurement process 201. The process analysis unit 220 analyzes all or part of the information of the measurement process 201 and extracts its features. The process analysis unit 220 refers to the matching success or failure determination result associated with the measurement process 201 that has features most similar to the extracted features from the matching result storage unit 230, and selects a method with a high probability of successful pattern matching. The matching result storage unit 230 is a database containing information such as the pattern type, shooting conditions, and device number of the measurement process 201, as well as keywords of the current matching success or failure result, and stores the information of each keyword in the form of text, numerical values, etc.

[0154] As a method for process analysis, classification models such as random forests based on machine learning can be cited as examples. In this learning process, the input is set to the data corresponding to the measurement process 201 in the matching result storage unit 230, and the teacher label is set to the matching success or failure judgment result corresponding to the measurement process 201. Thus, the machine learning model can estimate the optimal matching method corresponding to the measurement process. When performing the estimation based on the machine learning model, the input is set to the information of the measurement process 201, and the output is set to the optimal pattern matching method. As another method, multiple columns of items that are most consistent with the measurement process 201 can be referenced from the database of the matching result storage unit 230, and the matching success or failure result at this time can be referenced. Through these methods, the optimal pattern matching method can be selected based on the results of the pattern matching process performed in the past. The method for process analysis is not limited to these methods, as long as the optimal pattern matching method can be selected.

[0155] <Implementation Method 3: Summary>

[0156] The pattern size measurement system in Implementation 3 implements either rule-based pattern matching or machine learning-based pattern matching. If rule-based matching fails, machine learning matching is implemented; if machine learning matching fails, relearning is performed. Relearning of the learner can be performed only when machine learning matching fails, and it can also be performed automatically (see S206). Therefore, in Implementation 3, the learning workload can also be reduced.

[0157] <Variations on this disclosure>

[0158] This disclosure is not limited to the embodiments described above, but includes various modifications. For example, the embodiments described above are those that have been explained in detail for the purpose of easily understanding this disclosure, and are not necessarily limited to having all the structures described. In addition, a part of the structure of a certain embodiment can be replaced with the structure of another embodiment, and it is also possible to add the structure of another embodiment to the structure of a certain embodiment. Furthermore, with respect to a part of the structure of each embodiment, other structures can be added, deleted, or replaced.

[0159] In the above embodiments, a scanning electron microscope was described as an example of a measuring device 300, but this disclosure is not limited thereto and can be applied to any measuring device or inspection device that acquires images and performs pattern matching.

[0160] In the above embodiments, the design data storage unit 100, the process generation device 200, the measurement device 300, and the machine learning model learning device 400 are represented as different devices, but they can also be installed as one or more devices capable of performing the same processing (a pattern matching device composed of one or more computer systems). For example, a system can also be composed of a device having the functions of the design data storage unit 100, the measurement device 300, and the machine learning model learning device 400, and the process generation device 200.

[0161] In Implementation 2, the following features can be used as features for implementing clustering: (a) Image features: brightness distribution of the image, frequency components of the image; (b) Shape and pattern features: category of shape and pattern, planar density of shape and pattern; (c) Shooting condition features: optical conditions and electrical conditions of the shooting device for acquiring the image.

[0162] In Implementation 3, the process of determining whether the measurement is correct (S209) after performing S200-S207, similar to Implementation 1, and then performing learning using the successful result (S210-S214) is described. By performing the steps after S209, learning data can be automatically collected in the same way as in Implementation 1, but these steps may not necessarily be performed. That is, since Implementation 3 performs machine learning or according to either of the suitable rules, as long as these steps are implemented, the general objective can be achieved. Of course, performing the steps after S209 would be more effective.

[0163] In the above embodiments, the functional units of each device (e.g., process generation unit 210, process analysis unit 220, rule-based pattern matching unit 302, machine learning pattern matching unit 303, matching success or failure determination unit 304, measurement unit 305, measurement success or failure determination unit 306, GUI 307, learning data analysis unit 401, machine learning model learning unit 402, GUI 403, etc.) can be constructed by hardware such as circuit devices that have these functions installed, or they can be constructed by executing software with these functions installed by a computing device such as a CPU (Central Processing Unit).

[0164] In the above implementation, the rule-based matching algorithm can be configured to determine the correspondence between an image feature and the positional offset between two images corresponding to that image feature through the algorithm's internal logic. On the other hand, it should be noted that the machine learning matching algorithm differs from the rule-based algorithm in that, in addition to the internal logic for matching processing, it also includes a learning model that describes the result of learning the correspondence between image features and positional offsets.

[0165] In the above embodiments, S201 was described as being performed after S200; however, the order of the steps is not limited to this and can be modified within the scope that does not contradict the method of this disclosure. For example, S902 can be performed after S200, or as follows... Figure 9 S201 shall be implemented as described.

[0166] Explanation of reference numerals in the attached figures

[0167] 100: Design Data Storage Department

[0168] 200: Process Production Equipment

[0169] 300: Measuring device

[0170] 400: Machine learning model learning device.

Claims

1. A pattern matching apparatus for performing pattern matching between images, characterized in that, The pattern matching device includes: a computer system that performs the pattern matching. The computer system includes: a learner that outputs the pattern matching result through machine learning. The computer system performs the pattern matching by using a first algorithm that does not employ machine learning. If the pattern matching based on the first algorithm is successful, the computer system provides the result as learning data for the learner to perform relearning.

2. The pattern matching device according to claim 1, characterized in that, The computer system evaluates the inference performance of the learner after the relearning process. If the inference performance fails to meet the baseline value, the computer system deletes portions of the learning data used for relearning that belong to pre-defined data types that may negatively impact inference performance. The computer system re-implements the relearning using learning data with the adverse data types removed.

3. The pattern matching device according to claim 1, characterized in that, The computer system clusters the learning data used to perform the relearning according to its feature values. The computer system corrects the configuration of data belonging to each of the said classes in order to suppress bias in the data of each class obtained through the clustering. The computer system uses the learning data, which has suppressed the bias, to perform the relearning.

4. The pattern matching device according to claim 1, characterized in that, The image is an observational image of the sample. If the pattern matching is successful, the computer system receives the results of the measurement or inspection performed on the sample. If the result indicates that the measurement or inspection has failed, the computer system receives a specified input specifying the positional offset between the images via an interface of the computer system, and causes the learner to perform relearning based on the specified input.

5. The pattern matching device according to claim 1, characterized in that, If the pattern matching based on the first algorithm fails, the computer system performs the pattern matching by using a second algorithm of the learner.

6. The pattern matching device according to claim 5, characterized in that, If the pattern matching using the second algorithm fails, the computer system generates learning data for the learner to perform machine learning, and uses the learning data to enable the learner to perform relearning. The computer system uses the learner, which has undergone the relearning process, to re-implement the pattern matching.

7. The pattern matching device according to claim 6, characterized in that, The computer system is configured to supply the learner with information including the positional offset between the two images in which the pattern matching is performed. The computer system receives a specified input specifying the position offset via an interface provided by the computer system, and generates the learning data based on the specified input.

8. The pattern matching device according to claim 6, characterized in that, The computer system is configured to supply the learner with information including the positional offset between the two images in which the pattern matching is performed. The computer system obtains the position offset by implementing multiple first algorithms, each with a different matching process, respectively. The computer system generates the learning data using the position offset obtained according to each of the first algorithms that yields the best match between the two images.

9. The pattern matching device according to claim 1, characterized in that, The first algorithm is equipped with a rule-based algorithm, which is configured to determine the correspondence between image feature quantities and positional offsets between two images corresponding to the image feature quantities through the internal logic of the first algorithm.

10. The pattern matching device according to claim 1, characterized in that, The computer system has: multiple learners that perform machine learning using distinct learning data. The computer system groups the learning data used by each learner when performing machine learning according to the feature values ​​of the learning data. The computer system uses the learner to perform the pattern matching, and the learner uses the learning data corresponding to at least one of the feature quantities of the image, the feature quantities of the shape patterns contained in the image, and the feature quantities of the image's shooting conditions to perform machine learning.

11. The pattern matching device according to claim 10, characterized in that, The feature quantities of the image include at least one of the image's brightness distribution and the image's frequency components. The feature quantities of the shape pattern include at least one of the shape pattern category and the shape pattern planar density. The characteristic quantities of the shooting conditions include at least one of the optical and electrical conditions of the shooting device that acquires the image.

12. The pattern matching device according to claim 1, characterized in that, If the pattern matching is successful using the learner, the computer system supplies the result to the learner as learning data for the learner to perform relearning.

Images