Pattern matching device

The pattern matching device addresses the challenge of high workload in semiconductor inspection by switching between rule-based and machine learning algorithms and automating retraining, improving model versatility and reducing human effort.

KR1020260162297APending Publication Date: 2026-11-10HITACHI HIGH TECH CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
KR1020267029333
Authority / Receiving Office
KR · KR
Patent Type
Applications
Current Assignee / Owner
Filing Date
2024-04-02
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Conventional pattern matching techniques in semiconductor inspection and measurement devices face challenges in handling apparent discrepancies between template and target images, requiring significant human effort for data collection and true value generation, leading to limited generality and increased workload.

Method used

A pattern matching device that switches between rule-based and machine learning algorithms, with automatic retraining upon failure, reducing the need for manual data collection and true value generation by using successful measurement results for learning.

Benefits of technology

Significantly reduces the workload associated with training data collection and true value generation, enhancing the versatility and performance of the machine learning model while minimizing human intervention.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure P1020267029333_ABST
    Figure P1020267029333_ABST
Patent Text Reader

Abstract

The present disclosure proposes a pattern matching technique using machine learning that significantly reduces the workload related to data collection and true value generation, and reduces human labor in the learning process of a machine learning model (creation of a learning data set, training of a machine learning model, performance evaluation, etc.). A pattern matching device according to the present disclosure performs pattern matching using either a first algorithm that does not use a learner or a second algorithm that uses said learner, and if the pattern matching using said first algorithm fails, performs the pattern matching using said second algorithm, and if the pattern matching using said second algorithm fails, performs retraining of said learner (see FIG. 9).
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present disclosure relates to a technique for performing pattern matching between images. Background Technology

[0002] Devices for measuring and inspecting patterns formed on a semiconductor wafer often use template matching technology to align the device's field of view with a desired measurement or inspection location. Template matching is a process of finding the region that most closely matches a pre-registered template image from the image to be searched. Patent Document 1 describes an example of such template matching.

[0003] Patent Document 2 describes a template matching device having a learning function that estimates a correlation image having a numeric value representing the correlation between a template image and a search image as a pixel value. The said document can absorb the apparent discrepancy between the images of the template image and the search image through learning, thereby enabling high-precision matching.

[0004] Patent Document 3 describes a technique for selecting the minimum necessary training data capable of achieving the target performance of a machine learning model based on semiconductor design data when applying machine learning to a semiconductor inspection device that captures a pattern formed on a semiconductor wafer. When machine learning is utilized for image inspection, it is necessary to collect a large amount of image data and generate correct inspection results corresponding to each image data, and this task places a heavy load on the device user. Patent Document 3 makes it possible to reduce the load on the device user by selecting the minimum necessary data when training the machine learning model. Prior art literature

[0005] Japanese Patent Publication No. 4218171 (corresponding specification of U.S. Patent No. 6,627,888) WO2022 / 009357 Japanese Patent Publication No. 2020-35282 The problem to be solved

[0006] Conventionally, as a template matching technology used in semiconductor inspection and measurement devices, a rule-based pattern matching technique that performs matching based on a predetermined algorithm has been used. Generally, rule-based techniques can be considered to have high versatility because it is possible to design algorithms by assuming various known cases. However, when matching fails due to apparent discrepancies between the template image and the target image, it is difficult to easily improve the algorithm, and it takes time to respond.

[0007] Therefore, a pattern matching technique using machine learning with a learning function such as that described in Patent Document 2 has been devised. This technique enables high-precision matching by absorbing the apparent discrepancy between a template image and a target image through learning. However, in order to train a machine learning model with sufficient performance, it is necessary to collect a large amount of training data, such as images, and generate correct pattern matching results (hereinafter referred to as true values) for the collected training data. Furthermore, the task of generating true values ​​for a large amount of data requires human labor and time, thus placing a heavy burden on the operator. In addition, if the types of data included in the training data (image shape, quality, etc.) are insufficient, the training becomes limited to specific data, and the generality of the machine learning model is reduced.

[0008] Patent Document 3 describes a technique for analyzing semiconductor design data and captured images, and for selecting and collecting the minimum necessary training data capable of achieving the target performance of a machine learning model. By doing so, the variety of training data can be secured while reducing the volume, thereby improving the versatility of the machine learning model and reducing the effort involved in generating true values. The document also describes the generation of an estimated true value image that simplifies the task of generating true values ​​for each pixel in segmentation, which is one of the machine learning tasks. However, it did not eliminate the need for manual labor, leaving room for automation.

[0009] The present disclosure is made in consideration of the above-mentioned problems and proposes a pattern matching technique using machine learning that significantly reduces the workload related to data collection and true value generation, and reduces the workload performed by human personnel in the machine learning model training process (creation of training data sets, training of machine learning models, performance evaluation, etc.). means of solving the problem

[0010] A pattern matching device according to the present disclosure performs pattern matching using either a first algorithm that does not use a learner or a second algorithm that uses said learner, and if the pattern matching using said first algorithm fails, performs the pattern matching using said second algorithm, and if the pattern matching using said second algorithm fails, performs retraining of said learner. Effects of the invention

[0011] According to the pattern matching device of the present disclosure, the workload related to the collection of training data and the creation of true values ​​used for training a machine learning model is reduced, thereby reducing the workload performed by human personnel in the machine learning model training process (creation of training data sets, training of the machine learning model, performance evaluation, etc.). As a result, the load on the user of a device equipped with pattern matching by machine learning can be significantly reduced. Brief explanation of the drawing

[0012] FIG. 1 is a configuration diagram of a pattern dimension measurement system according to embodiment 1. Figure 2 is a flowchart explaining the operation of a pattern dimension measurement system. Figure 3 is a flowchart explaining the specific operations of S206 to S207. Figure 4 is a flowchart explaining the specific operation of S214. Figure 5 is a flowchart explaining the specific operation of S213. FIG. 6 shows an example of data held by temporary memory data (406) and a learning data set (407). Figure 7 shows an example of data from a learning log. FIG. 8 shows an example configuration of a measuring device (300) including an SEM (600) and a GUI (307). FIG. 9 is a flowchart illustrating the operation of a pattern dimension measurement system according to embodiment 3. FIG. 10 is a block diagram of a recipe generating device (200). Specific details for implementing the invention

[0013] <Embodiment 1>

[0014] FIG. 1 is a configuration diagram of a pattern dimension measurement system according to Embodiment 1 of the present disclosure. The pattern dimension measurement system is a system for measuring the size of a shape pattern formed on a semiconductor wafer. The pattern dimension measurement system has a design data storage unit (100), a recipe generation device (200), a measurement device (300) (pattern matching device), and a machine learning model learning device (400) (pattern matching device). Data generated and output by each device constituting the system can be transmitted and received between devices.

[0015] The matching result of pattern matching is output as a matching shift amount. The matching shift amount represents the positional shift amount and positional difference between the design data (101) and the SEM (Scanning Electron Microscope) image (310), and can be represented by a two-dimensional scalar value, for example (shift amount in the X direction, shift amount in the Y direction). In addition, the template image is not limited to the SEM image, but may be of another type of image, for example, the design data (101). In addition, the image to be searched is not limited to the design data (101), but may be of another type of image, for example, the SEM image (310).

[0016] The design data storage unit (100) holds design data (101) representing the pattern structure of a semiconductor device created by the device user.

[0017] The recipe generation device (200) has an interface for a device user to input a measurement point on a device pattern referenced from design data (101), and outputs a measurement recipe (201) including imaging conditions in a measurement device (300), an algorithm used in pattern matching by a rule base, measurement point information, etc. based on the measurement point information input from the interface and the design data (101).

[0018] The measuring device (300) measures the dimensions of a device pattern based on a measuring recipe (201) and outputs the measured value. The measuring device (300) is equipped with an imaging unit (301) for capturing a device pattern 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 measuring unit (305), a measuring success / failure determination unit (306), a GUI (Graphical User Interface) (307), and a measuring information storage unit (308).

[0019] The imaging unit (301) captures a device pattern based on the imaging conditions of the measurement recipe (201) and outputs an SEM image (310). The rule-based pattern matching unit (302) performs pattern matching between the SEM image (310) and design data (101) using an algorithm described in the measurement recipe (201) and outputs a matching shift amount (314) between the two images. The machine learning pattern matching unit (303) performs pattern matching between the SEM image (310) and design data (101) using a machine learning model (312) and outputs a matching shift amount (314) between the two images. The data used for pattern matching is not limited to the above combinations, but may use data in which the matching points of the field of view between the template image and the image to be searched can be specified. For example, a design drawing obtained from the SEM image and the design data (101) may be used, or point group information of the SEM image and the design data (101) may be used. The matching success / failure determination unit (304) performs a matching success / failure determination based on the design data (101), the SEM image (310), and the matching shift amount (314), and outputs the matching success / failure determination result. The measurement unit (305) measures the dimensions of the device pattern based on the measurement point information listed in the measurement recipe (201), the design data (101), and the SEM image (310), and outputs the measurement value. The measurement success / failure determination unit (306) determines the success / failure of the measurement based on the measurement value and outputs the measurement success / failure determination result. If the measurement success / failure judgment result is a failure, the device user operates the GUI (307) to input a matching shift amount of the correct solution in which the pattern shape between the design data (101) and the SEM image (310) matches, and outputs a matching shift amount (314). The measurement information storage unit (308) stores the measurement recipe (201), the SEM image (310), the design data (101), and the machine learning model (312).The machine learning model (312) has a base model learned by the machine learning model learning device (400) in an initial state.

[0020] After the dimensional measurement of the device pattern is completed, the measurement device (300) outputs the measurement recipe (201), design data (101), SEM image (310), and matching shift amount (314) to the machine learning model learning device (400). At this time, the design data (101) and SEM image (310) being output can be used as training data for the machine learning model, and the matching shift amount (314) is a value that was successfully used to measure the dimensions of the device pattern and can be used as a true value for the machine learning model. Therefore, manual labor in the collection of training data and the creation of true values ​​can be made almost unnecessary.

[0021] A machine learning model learning device (400) comprises a learning data interpretation 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 data set (407). The temporary storage data (406) includes a measurement recipe storage folder (4042), a design data storage folder (4043), and an 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 data set (407) includes a base model learning data table (4045), a test data table (4046), and an additional learning data table (4048). Specific examples of the data configurations will be described later.

[0022] The machine learning model learning device (400) stores the measurement recipe (201), design data (101), and SEM image (310) received from the measurement device (300) in the measurement recipe 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 an integrated collected data table (4049) in which an arbitrary data ID, the filenames of the measurement recipe (201), design data (101), and SEM image (310), and the numerical value of the matching shift amount (314) are organized into a single row of records, and stores the created data table in the temporary storage data (406). The collected data table (4049) is created to manage collected data, but the method of managing collected data is not limited to this; any method that allows for the management of combinations such as measurement recipes (201), design data (101), SEM images (310), and matching shift amounts (314) may be used, and for example, it may be managed by folder organization or file names.

[0023] The learning data interpretation unit (401) reads each data (measurement recipe (201), design data (101), SEM image (310)) based on each record of the existing data table (e.g., base model training data table (4045)) of the collected data table (4049) and the learning data set (407), and extracts features such as pattern shape features, SEM image features, and imaging conditions (the extraction method will be described later using FIG. 5). In addition, the learning data interpretation unit (401) selects records from each data table so that there is no bias in the extracted features (the technique for selecting without bias will be described later using FIG. 5), and stores each data (measurement recipe (201), design data (101), SEM image (310)) included in the records selected from the collected data table (4049) in the measurement recipe storage folder (4042), design data storage folder (4043), and SEM image storage folder (4044). In addition, records selected from each data table are preserved by giving them a unique name as a new data table. By doing so, a sufficient variety of data types can be secured for the training data set (407), thereby ensuring versatility and improving the performance of the machine learning model. Furthermore, since a new training data set can be created while retaining the training data set used in the past, the management of multiple training data sets can be facilitated. When interpreting data in the training data interpretation unit (401), if there is an error in the matching shift amount (314), it can be manually corrected by the GUI (403) described later.

[0024] The machine learning model training unit (402) reads data required for training (design data (101), SEM image (310), matching shift amount (314), etc.) based on a data table of the training data set (407) (e.g., additional training data table (4048), etc.) and trains the machine learning model based on default training conditions (details of the machine learning model training method will be described later using FIGS. 3 and FIGS. 4). To evaluate the performance of the machine learning model after training, the machine learning model training unit (402) reads data required for inference (design data (101), SEM image (310), matching shift amount (314), etc.) from a test data table (4046) stored in the training data set (407) and evaluates the performance of the model after training. At this time, if the target performance is not met, the training conditions may be changed and retrained. By doing so, a machine learning model with the required performance can be obtained. When the target performance is achieved, the training log and the data file of the machine learning model (408) after training are given unique names and stored in the machine learning model (312). At this time, the machine learning model trained in the past is retained.

[0025] If there is an error in the matching shift amount (314) and it is necessary to teach the correct matching shift amount, the device user operates the GUI (403) to teach and output a fixed matching shift amount in which the pattern shape of the design data (101) and the SEM image (310) match.

[0026] FIG. 2 is a flowchart illustrating the operation of a pattern dimension measurement system. Each step of FIG. 2 will be explained below.

[0027] (Fig. 2: S200)

[0028] The recipe generation device (200) generates a measurement recipe (201) including measurement point information of a device pattern input by a device user from the interface of the recipe generation device (200) and design data (101) obtained from the design data storage unit (100), imaging conditions in the measurement device (300) (which may include imaging location information, etc.), an algorithm used in pattern matching by a rule base, measurement point information, etc.

[0029] (Fig. 2: S201)

[0030] The imaging unit (301) of the measuring device (300) captures the device pattern based on the imaging conditions of the measuring recipe (201) and outputs an SEM image (310).

[0031] (Fig. 2: S202)

[0032] The rule-based pattern matching unit (302) calculates a matching shift amount (314) between the design data (101) and the SEM image (310). The data used for pattern matching is not limited to this combination, but may use data in which the matching location of the field of view between the template image and the image to be searched can be specified. For example, a design drawing obtained from the SEM image (310) and the design data (101) may be used, or point cloud information of the SEM image (310) and the design data (101) may be used. The matching shift amount (314) is calculated by a matching algorithm determined based on the measurement recipe (201). The matching algorithm may be, for example, an algorithm that calculates the similarity at each candidate location of the design data (101) and the SEM image (310) as a mean squared error and outputs the matching shift amount of the location with the highest similarity. The matching algorithm is not limited to this technique; any algorithm that extracts each image feature based on a predetermined process and calculates a matching shift amount based on the extracted image features is sufficient.

[0033] (Fig. 2: S203)

[0034] The matching success / failure determination unit (304) determines the success or failure of the matching based on the design data (101), the SEM image (310), and the matching shift amount (314). As an example of a matching success / failure determination method, a method can be considered in which a mean squared error representing the similarity between the design data (101) and the SEM image (310) is calculated based on the matching shift amount (314), and if the value is greater than or equal to a threshold, it is determined to be successful. The matching success / failure determination method is not limited to this method, and it is sufficient if it can determine whether the position based on the matching shift amount (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 to S208. If the matching success / failure determination result is a failure, proceed to S204.

[0035] (Fig. 2: S204)

[0036] The machine learning pattern matching unit (303) calculates the matching shift amount (314) by using a machine learning model trained to estimate the matching shift amount (314) based on design data (101) and SEM images (310). The data used for pattern matching is not limited to the above combination, but may use data in which the matching location of the field of view between the template image and the image to be searched can be specified. For example, a design drawing obtained from the SEM image and design data (101) may be used, or point cloud information of the SEM image and design data (101) may be used. As for the machine learning model, a designated machine learning model (such as the machine learning model with the highest performance or the latest machine learning model) is read from the machine learning model (312), and the read machine learning model is used. The pattern matching method by machine learning can be implemented, as an example, by estimating the correlation image between images and calculating the matching shift amount from the correlation image, as in Patent Document 2. The pattern matching method based on machine learning is not limited to this; it is sufficient if it can identify the matching points in the field of view between the template image and the image to be searched. For example, it may be a method in which the amount of matching shift between images is directly estimated by a machine learning model.

[0037] (Fig. 2: S205)

[0038] The matching success / failure determination unit (304) determines the success or failure of the matching based on the matching shift amount (314). If the result of the matching success / failure determination is successful, S208 is performed. If the result of the matching success / failure determination is a failure, 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 performed to perform machine learning model learning. At this time, the machine learning model learning device (400) creates a temporary data table (4047) having the filenames of the received design data (101) and SEM image (310) as records, assigns a unique name, and stores it in the temporary storage data (406).

[0039] (Fig. 2: S206)

[0040] The device user reads the design data (101) and the SEM image (310) based on the time data table (4047) using the GUI (403). The device user operates the GUI (403) to teach the matching shift amount of the predetermined solution, where the pattern shapes of the two images match. The method of teaching the matching shift amount of the predetermined solution is not limited to the method of using the GUI (403), but any method capable of teaching the matching shift amount of the predetermined solution is acceptable. For example, pattern matching may be performed on the design data (101) and the SEM image (310) by a pattern matching algorithm based on multiple rule bases, and the most similar matching shift amount may be set as the matching shift amount of the predetermined solution. The machine learning model learning device (400) creates a data table (4041) for learning failure patterns, including the created matching shift amount of the predetermined solution, the design data (101), and the SEM image (310). Details will be described later using FIG. 3.

[0041] (Fig. 2: S207)

[0042] The machine learning model learning unit (402) performs machine learning of the machine learning model using a data table (4041) for learning failure patterns. The machine learning model (408) and the learning log after learning are stored in the machine learning model memory unit (405). Additionally, the machine learning model (408) after learning is stored in the machine learning model (312) of the measuring device (300). Details will be described later using FIG. 3. S204 is performed again using this machine learning model (408) after learning.

[0043] (Fig. 2: S208)

[0044] The measurement unit (305) measures the dimensions of the device pattern, etc., based on the measurement recipe (201), design data (101), SEM image (310), and matching shift amount (314), and outputs the measured value.

[0045] (Fig. 2: S209)

[0046] The measurement success / failure determination unit (306) determines the success or failure of the measurement based on the measurement recipe (201), design data (101), SEM image (310), and measurement value, and outputs the success / failure determination result. If the measurement success / failure determination result is successful, skip to S212. If the measurement success / failure determination result is a failure, perform S210 to perform manual assistance.

[0047] (Fig. 2: S210)

[0048] The device user inputs a matching shift amount (314) that matches the pattern shapes of the two based on the design data (101) and the SEM image (310) using the GUI (307). As a method for inputting the matching shift amount, the numerical value of the matching shift amount may be entered directly, or the matching shift amount that becomes the matching position may be calculated by checking an image in which the design data (101) and the SEM image (310) are superimposed.

[0049] (Fig. 2: S211)

[0050] The measurement unit (305) measures the dimensions of the device pattern, etc., based on the measurement recipe (201), design data (101), SEM image (310), and matching shift amount (314), and outputs the measured value.

[0051] (Fig. 2: S212)

[0052] After the measurement of dimensions of the device pattern, etc., is completed, the measurement device (300) outputs the measurement recipe (201), design data (101), SEM image (310), and matching shift amount (314) to the machine learning model learning device (400). The machine learning model learning device (400) stores the measurement recipe (201), design data (101), and SEM image (310) received from the measurement device (300) in the measurement recipe 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) adds the filenames of the received measurement recipe (201), design data (101), and SEM image (310), and the matching shift amount (314) as records to the collected data table (4049).

[0053] (Fig. 2: S213)

[0054] The learning data interpretation unit (401) interprets the data stored in the temporary memory data (406) and creates a new learning data set. Details will be described later using FIG. 5. S213 may be performed whenever a pattern is measured, or it may be performed after a certain amount of data has been collected.

[0055] (Fig. 2: S214)

[0056] The machine learning model learning unit (402) performs machine learning of the machine learning model based on the new learning data set created in S213, and stores the learning log and the machine learning model (408) after learning in the machine learning model memory unit (405). In addition, the machine learning model (408) after learning is stored in the machine learning model (312) of the measurement device (300). Details will be described later using FIG. 4. S214 may be performed every time a pattern is measured, or it may be performed after a certain amount of data has been collected.

[0057] FIG. 3 is a flowchart illustrating the specific operations of S206 to S207. Each step of FIG. 3 is described below. S206 to S207 may be executed immediately upon failure of pattern matching by machine learning, or may be executed later after unloading the semiconductor wafer from the measuring device (300).

[0058] (Fig. 3: S300)

[0059] The device user uses the GUI (403) to read design data (101) and SEM images (310) based on the time data table (4047), and inputs a matching shift amount for a predetermined solution in which the pattern shapes of the two images match. Then, a failure pattern learning data table (4041) is created including the input matching shift amount, design data (101), and SEM images (310). At this time, all or part of the records of the base model learning data table (4045) or the additional learning data table (4048) may be added to the failure pattern learning data table (4041). By doing so, the machine learning model can be made to correspond to failure patterns while ensuring general-purpose performance.

[0060] (Fig. 3: S301)

[0061] The machine learning model training unit (402) performs training of the machine learning model based on the failure pattern learning data table (4041) and the predetermined training conditions, and outputs the machine learning model (408) after training. Generally, the training conditions of the machine learning model include values ​​that can be arbitrarily set by the machine learning model, such as the number of training sessions per record of the data table (hereinafter referred to as the number of epochs) and the learning rate (hereinafter referred to as hyperparameters), and the presence or absence of a model that serves as the basis for training. The internal parameters of the machine learning model may be updated entirely with the values ​​possessed by the model that serves as the basis for training, or only partially. The predetermined training conditions are preferably optimal values ​​obtained empirically through prior experiments, but are not limited to this; for example, the number of epochs may be set until a certain performance is achieved while checking the trend of training. The model serving as the basis for learning is preferably a machine learning model that has been completed using the base model learning data table (4045), or the latest machine learning model held by the machine learning model memory unit (405), but is not limited to this, and, for example, a new machine learning model may be learned. It is also possible to use the machine learning model (408) after learning as a model serving as the basis for future learning.

[0062] (Fig. 3: S302)

[0063] The machine learning model training unit (402) performs inference by inputting the design data (101) and SEM image (310), which were taught the matching shift amount in S300, to the machine learning model (408) after training, and evaluates whether the matching is successful. If the matching is successful, S303 is performed. If it fails, S301 is performed again. At this time, records selected from the base model training data table (4045) or the additional training data table (4048) among the failure pattern training data table (4041) may be deleted. By doing so, the generality of the machine learning model is reduced, but it can be specialized for failure cases, and there is a possibility that the matching will turn to success. Since the machine learning model (408) after training is capable of responding to failure patterns, it is possible to respond quickly to matching failures.

[0064] (Fig. 3: S303)

[0065] The machine learning model learning unit (402) stores the machine learning model (408) and the learning log after learning in the machine learning model memory unit (405).

[0066] (Fig. 3: S304)

[0067] The machine learning model (408) after learning is stored in the machine learning model (312).

[0068] FIG. 4 is a flowchart illustrating the specific operation of S214. Each step of FIG. 4 is described below.

[0069] (Fig. 4: S400)

[0070] The machine learning model learning unit (402) uses an additional learning data table (4048) and performs learning of the machine learning model based on 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 may be set until a certain performance is achieved while checking the progress of learning. In addition, the model that serves as the basis for learning is preferably a machine learning model that has been completed using a base model learning data table (4045), or the latest machine learning model held by the machine learning model memory unit (405) obtained by referring to the learning log, but is not limited to this; for example, a new machine learning model may be learned. The method of selecting the model that serves as the basis for learning is not limited to the method of referring to the learning log; it is sufficient to select the model that serves as the basis for optimal learning. For example, the latest machine learning model may be identified from the storage date of the machine learning model, or it may be identified from the filename of the machine learning model.

[0071] (Fig. 4: S401)

[0072] The machine learning model training unit (402) performs inference on the machine learning model (408) after training by inputting data from a test data table (4046) in which the true value is known, and evaluates the performance. The test data table (4046) may contain data with various shape features as records to evaluate the general performance of the machine learning model, or other data, such as data selected by the device user that is desired to succeed in matching. As a method for evaluating performance, for example, the accuracy rate may be above a threshold, or other methods may be used, such as the difference between the inferred matching shift amount and the true value's matching shift amount being sufficiently small, or a combination of these methods may be used. If sufficient performance is obtained, S403 is performed. If performance is insufficient, S402 is performed.

[0073] (Fig. 4: S402)

[0074] The machine learning model training unit (402) checks the data included in the additional training data table (4048) and may delete data that may have a negative impact on training (e.g., data where the template image does not exist within the field of view of the image to be searched due to field of view misalignment, data where the shape of the device pattern formed on the wafer is significantly different from the design data due to deformation, data where the characteristics such as image quality are significantly different from other data within the same data table, etc.) (types of negative impact data). In addition, training conditions may be adjusted. As a method for adjusting training conditions, adjustment methods such as changing the model that serves as the basis for training or increasing the number of epochs if training is insufficient may be used. After that, S400 is performed again. This step is optional and does not necessarily have to be performed.

[0075] (Fig. 4: S403)

[0076] The machine learning model learning unit (402) stores the machine learning model (408) and the learning log after learning in the machine learning model memory unit (405).

[0077] (Fig. 4: S404)

[0078] The machine learning model (408) after learning is stored in the machine learning model (312).

[0079] FIG. 5 is a flowchart illustrating the specific operation of S213. Each step of FIG. 5 will be explained below.

[0080] (Fig. 5: S500)

[0081] The learning data analysis unit (401) reads design data, SEM images, and measurement recipes included in their records from the collected data table (4049), the failure pattern learning data table (4041), the latest additional learning data table (4048), and the base model learning data table (4045). All four data tables may be used, or some of them may be used. The learning data analysis unit (401) analyzes the read data and extracts features such as pattern shape features, SEM image features, and imaging conditions. Specific examples of each feature are as follows. As pattern shape features, the pattern type (line-space pattern or hole array pattern, etc.), pattern density, constituent material, etc. As SEM image features, the presence or absence of transmission of the lower layer pattern, deformation amount, image quality, etc. As imaging conditions, acceleration voltage, detector type, imaging magnification, number of integrated frames for image quality, high-definition processing, etc. As a specific method for extracting pattern shape features, for example, estimation can be performed using information obtained from image edge and corner detection, image frequency analysis, etc. Alternatively, the recipe creator may include such information in the measurement recipe when creating the measurement recipe. As a specific method for extracting SEM image features, for example, the contour of the pattern part can be extracted from the SEM image and the design data, respectively, and by comparing the extracted contours, the presence or absence of transmission of the lower layer pattern or the amount of deformation of the pattern can be estimated. In addition, the recipe creator may include such information in the measurement recipe when creating the measurement recipe. The imaging conditions may use the information listed in the measurement recipe as is, or any preprocessing may be performed. The features to be extracted are not limited to these, and when creating the measurement recipe, the recipe creator may include the type of the measurement device (300), the identification number of the identical device, the process name (photolithography process, etching process, etc.) listed in the measurement recipe.

[0082] (Fig. 5: S501)

[0083] The learning data interpretation unit (401) selects records from each data table so that there is no bias in the features extracted in S500, and creates a new additional learning data table. As a method for selecting records, for example, there is a method of clustering based on the features of the records and selecting records from each class so that there is no bias in each class. As a specific method for clustering, the k-means method, which places each record in a space having a dimension equal to the number of extracted features and performs clustering on this space, can be applied, but it is not limited to this method, and any technique capable of clustering similar records into one class based on the amount of extracted features is sufficient.

[0084] (Fig. 5: S502)

[0085] The device user checks the matching shift amount in the new additional learning data table in the GUI (403), and if there is an error, manually corrects the incorrect matching shift amount. The method of correcting the matching shift amount is not limited to the method using the GUI (403), but any method capable of correcting the matching error is sufficient. For example, pattern matching may be performed on the design data (101) and the SEM image according to a pattern matching algorithm based on multiple rule bases, and the most similar matching shift amount may be determined as the matching shift amount. Additionally, as the target data for checking the matching shift amount, only data with a high probability of being incorrect in the matching shift amount may be selected in advance, and the matching shift amount may be checked against this. As a method of selection, for example, a method of performing pattern matching on the design data (101) and the SEM image according to a pattern matching algorithm based on multiple rule bases, and selecting data for which a similar matching shift amount was not obtained by each algorithm. S502 does not necessarily need to be performed. This is because the collected data is the data determined by measurement, so the matching is likely to be correct. This step is intended to provide clearly defined data.

[0086] (Fig. 5: S503)

[0087] The learning data interpretation unit (401) stores each data (measurement recipe (201), design data (101), SEM image (310)) in the temporary storage data (406) included in the new additional learning data table (4048) in the measurement recipe storage folder (4042), design data storage folder (4043), and SEM image storage folder (4044). The new additional learning data table (4048) is given a unique name and stored in the learning data set (407).

[0088] FIG. 6 illustrates an example of data held by temporary storage data (406) and a training data set (407). As shown in FIG. 6, the temporary storage data (406) and the training data set (407) include a plurality of data tables for managing each data, a measurement recipe storage folder (4042), a design data storage folder (4043), and a SEM image storage folder (4044). Although the temporary storage data (406) and the training data set (407) have three folders and store measurement recipes, design data, and SEM images separately in each, all data files may be stored in the same layer as the data tables, all data files may be stored in a single folder, or divided and stored in multiple folders.

[0089] The training data set (407) contains a base model training data table (4045) for training a machine learning model that serves as a base in an initial state, a test data table (4046) for evaluating the performance of the machine learning model, and each data file listed in these two data tables. A base model refers to a machine learning model that has been trained with a high measurement frequency or a general pattern.

[0090] Each data table holds a data ID, measurement recipe, design data, SEM image, and matching shift amount as a set of records, as shown in FIG. 6. By holding multiple such data tables, multiple data sets can be managed as data tables. The method of managing data sets is not limited to management by data tables; any method that allows for the management of combinations such as measurement recipes, design data, SEM images, and matching shift amounts is acceptable, for example, by folder organization or file names. Furthermore, although each data table holds information such as the data ID, measurement recipe, design data, SEM image, and matching shift amount as records, the amount of information held may be greater or less than this, and it is sufficient to hold information necessary for data management. For example, it is not necessary to hold the data ID, and other information such as the registration date and time may be held.

[0091] Figure 7 shows an example of data for a training log. As shown in Figure 7, the training log can be recorded in the format of a data table in which the update date, machine learning model name, data table used, and training conditions are recorded as a single row of records. This facilitates the management of trained machine learning models and allows for the retrieval of past machine learning models if necessary. In Figure 7, the training log retains information such as the update date, machine learning model name, data table used, and training conditions as records, but the amount of information retained may be greater or less than this, as long as it allows for the management of machine learning models after training. For example, other information such as a machine learning model ID may also be retained.

[0092] FIG. 8 shows an example configuration of a measuring device (300) including a SEM (600) and a GUI (307). The SEM (600) can be operated as an imaging unit (301). The control unit (614) can be operated as a rule-based pattern matching unit (302), a machine learning pattern matching unit (303), a matching success / failure determination unit (304), a measuring unit (305), and a measuring success / failure determination unit (306). The image memory (615) can be operated as a measuring information storage unit (308). The SEM (600) measures, for example, the pattern dimensions of a semiconductor device formed on a semiconductor wafer. The computational processing unit or computer system in the measuring device (300) can be configured, for example, as a control unit (614). The control unit (614) is equipped with a computational means (e.g., a CPU / GPU (616)) and a memory device (e.g., a memory including an image memory (615)). Information can be stored in the memory device, for example, a program regarding pattern matching processing is stored. By the CPU / GPU (616) executing this program, the pattern matching processing shown in FIG. 2 is executed. That is, the control unit (614) functions as each pattern matching unit. In other words, this program causes the computer system to function as an arithmetic processing unit included in the pattern matching unit, thereby executing the pattern matching processing shown in FIG. 2.

[0093] The SEM (600) generates an electron beam from an electron gun (601). A deflector (604) and an objective lens (605) are controlled so that the electron beam is focused and irradiated at any position on a semiconductor wafer (603), which is a sample placed on a stage (602). Secondary electrons are emitted from the semiconductor wafer (603) irradiated by the electron beam and are detected by a secondary electron detector (606). The detected secondary electrons are converted into a digital signal by an A / D converter (607). An image displayed by the digital signal is stored in an image memory (615) within a control unit (614). This image is used, for example, as an SEM image (310), and based on this image, pattern matching processing is performed by the control unit (614) or a CPU / GPU (616). The setting processing required for these processes and the display of the processing results can be performed on a GUI (307). For alignment using an optical camera with a lower magnification than SEM, an optical camera (611) may be used. The signal obtained by the optical camera (611) capturing the semiconductor wafer (603) is converted into a digital signal by an A / D converter (612) (if the signal from the optical camera (611) is a digital signal, the A / D converter (612) is unnecessary), the image displayed by the digital signal is stored in an image memory (615) within the control unit (614), and the CPU / GPU (616) performs image processing according to the purpose.

[0094] The SEM (600) may be equipped with a reflection electron detector (608). When the reflection electron detector (608) is equipped, reflection electrons emitted from the semiconductor wafer (603) are detected by the reflection electron detector (608), and the detected reflection electrons are converted into a digital signal by an A / D converter (609 or 610). An image displayed by the digital signal is stored in an image memory (615) within the control unit (614), and the CPU / GPU (616) performs image processing according to the purpose. The control unit (614) may control the stage (602) through the stage controller (630) and may control the objective lens (605), etc., through the deflection control unit (631).

[0095] <Embodiment 1: Arrangement>

[0096] The pattern dimension measurement system according to Embodiment 1 performs learning of a learner using the success results of rule-based matching or machine learning matching (S214). These success results can be obtained whenever pattern matching and measurement based thereon are performed, and since they are results of successful measurement, they are suitable as learning data. Therefore, the workload for collecting and determining learning data can be suppressed.

[0097] <Embodiment 2>

[0098] Embodiment 2 of the present disclosure describes an example of a configuration for selectively using a learning device suitable for a matching target having specific features. Embodiment 2 is configured by the pattern dimension measurement system of FIG. 1, just like Embodiment 1. The basic operation is the same as Embodiment 1, and the differences in operation from Embodiment 1 are described below.

[0099] The design data storage unit (100) holds design data (101) representing the pattern structure of a semiconductor device created by a device user, similar to embodiment 1.

[0100] The recipe generation device (200) is equipped with a pattern feature analysis unit in addition to an interface. The pattern feature analysis unit analyzes the design data (101) obtained from the design data storage unit (100) and the measurement recipe (201), and extracts the pattern shape features and imaging conditions. Specific examples of each feature are as follows. As pattern shape features, the pattern type (line-space pattern or hole array pattern, etc.), the density of the pattern, the constituent material, etc. As imaging conditions, the acceleration voltage, the type of detector, the imaging magnification, the number of frames accumulated for image quality, the high-definition processing, etc. As a specific method for extracting pattern shape features, for example, estimation based on information obtained from image edge and corner detection, image frequency analysis, etc. Alternatively, the device user may input pattern type information, etc. by operating the interface. As imaging conditions, the information listed in the measurement recipe may be used as is, or any preprocessing may be performed. The features to be extracted are not limited to these, and may include the type of measuring device, identification number of a similar device, process name (photolithography process or etching process, etc.) that the recipe creator has listed in the measuring recipe when creating the measurement recipe.

[0101] The recipe generation device (200) selects a machine learning model trained using data from a feature group in which the extracted feature is classified, by using a unique feature group ID of each feature group (details of the feature group will be described later) that was created in advance. The method of selecting the machine learning model is not limited to this method, but can select a machine learning model trained on data similar to the extracted feature. For example, the device user may select the optimal machine learning model for the design data by operating an interface.

[0102] The recipe generation device (200) outputs a measurement recipe (201) including measurement point information input from an interface, design data (101), and a selected machine learning model, based on the imaging conditions in the measurement device (300), an algorithm used for pattern matching by rule base, a machine learning model used for pattern matching by machine learning, and measurement point information. The selected machine learning model has been trained in advance to be specialized for specific features, and since the matching performance for similar data is high, the number of matching failure cases in the machine learning pattern matching unit (303) can be reduced.

[0103] The measurement device (300) comprises an imaging unit (301) for capturing a device pattern 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) extracts a machine learning model used for pattern matching from a machine learning model (312) based on a measurement recipe (201). Alternatively, the machine learning pattern matching unit (303) receives a machine learning model used for pattern matching from a machine learning model storage unit (405) of a machine learning model learning device (400) based on a measurement recipe (201). The machine learning pattern matching unit (303) performs pattern matching using the acquired machine learning model.

[0104] A machine learning model learning device (400) comprises a learning data interpretation 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 set (407) of the learning data storage unit (404) has an additional learning data set for each feature group that has been created in advance. The method for creating the data set for each feature group will be described later. The machine learning model storage unit (405) holds a machine learning model for each feature group. The method for learning and managing the machine learning model for each feature group will be described later.

[0105] The machine learning model learning device (400) stores the measurement recipe (201), design data (101), and SEM image (310) received from the measurement device (300) in the measurement recipe 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 collected data table (4049) in which an arbitrary data ID, the filenames of the measurement recipe (201), design data (101), and SEM image (310), and the numeric values ​​of the matching shift amount (314) and feature group ID are organized into a single row of records, and stores the created data table in the temporary storage data (406). The collected data table (4049) is created to manage collected data, but the method of managing the collected data is not limited to the method of managing the collected data table (4049). Any method that allows for the management of combinations such as measurement recipes, design data, SEM images, matching shift amounts, feature group IDs, etc., may be used. For example, it may be managed by folder organization or file names.

[0106] Embodiment 2, like Embodiment 1, can be explained by the flowchart shown in FIG. 2. Here, mainly, the processing of each step in FIG. 2 in Embodiment 2 that differs from Embodiment 1 is described. Steps not described are the same as in Embodiment 1.

[0107] (Fig. 2: S200)

[0108] The recipe generation device (200) generates a measurement recipe (201) including measurement point information of a device pattern input by a device user from an interface, design data (101) obtained from a design data storage unit (100), and a feature group obtained in a pattern feature analysis unit, based on imaging conditions (which may include imaging location information, etc.) in a measurement device (300), an algorithm used in pattern matching by a rule base, a machine learning model used in pattern matching by machine learning, measurement point information, etc.

[0109] (Fig. 2: S204)

[0110] The machine learning pattern matching unit (303) retrieves a machine learning model used for pattern matching from the machine learning model (312) based on the measurement recipe (201), or receives it from the machine learning model memory unit (405) of the machine learning model learning device (400). After that, the machine learning pattern matching unit (303) performs pattern matching using the acquired machine learning model.

[0111] (Fig. 2: S206)

[0112] In the same manner as in Embodiment 1, the device user operates the GUI (403) to teach the matching shift amount of the predetermined solution. After that, the machine learning model learning device (400) creates a failure pattern learning data table (4041) including the input matching shift amount of the predetermined solution, design data (101), and SEM image (310). When creating the failure pattern learning data table (4041), all or part of the data set used when learning the machine learning model acquired in S204 may be added to the failure pattern learning data table (4041). By doing so, the machine learning model can be matched to a failure pattern while ensuring performance on a specific pattern.

[0113] (Fig. 2: S207)

[0114] The machine learning model learning unit (402) performs machine learning of the machine learning model in the same manner as in embodiment 1 and outputs the machine learning model (408) after learning. The machine learning model that serves as the basis for learning is preferably the machine learning model acquired in S204, but is not limited to this, and, for example, a new machine learning model may be learned. The machine learning model (408) after learning is stored in the machine learning model (312) of the measuring device (300).

[0115] (Fig. 2: S212)

[0116] After the measurement device (300) finishes measuring the dimensions of the device pattern, etc., it outputs the measurement recipe (201), design data (101), SEM image (310), and matching shift amount (314) to the machine learning model learning device (400). The machine learning model learning device (400) stores the measurement recipe (201), design data (101), and SEM image (310) received from the measurement device (300) in the measurement recipe 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) adds the filenames of the received measurement recipe (201), design data (101), and SEM image (310), and the numeric values ​​of the matching shift amount (314) and pattern feature group ID as records to the collected data table (4049).

[0117] (Fig. 2: S213)

[0118] The learning data interpretation unit (401) extracts information on feature group IDs from the failure pattern learning data table (4041) and the collected data table (4049). Records from each data table are added to an additional learning data table (4048) for each feature group so that the information on the feature group IDs matches. By doing so, learning data can be collected and accumulated for each feature group. Accordingly, a machine learning model with higher performance can be trained for a specific feature.

[0119] (Fig. 2: S214)

[0120] The machine learning model learning unit (402) uses an additional learning data table (4048) for each feature group created in S213, and performs learning of the machine learning model adapted to each feature group by the same learning method as in Embodiment 1, and stores each learning log and each machine learning model after learning, which has a unique name, in the machine learning model memory unit (405). In addition to the information shown in Embodiment 1, a feature group ID may be assigned to the learning log. By doing so, it becomes possible to link each feature group with each machine learning model. The machine learning model after learning is stored in the machine learning model (312) of the measuring device (300).

[0121] As a method for creating feature groups in Embodiment 2, for example, a method by clustering can be considered. A method for creating feature groups by a clustering technique is described below.

[0122] Design data and measurement recipes included in their records are read from the previously obtained collected data table (4049) and base model training data table (4045). Both data tables may be used, or only one may be used. The read data is interpreted, and features such as pattern shape features and imaging conditions are extracted in the same manner as in Embodiment 1 (Fig. 5). Each record is placed in a space having a number of dimensions equal to the number of extracted features, and clustering is performed on this space. As a specific method for clustering, the k-means method may be applied, but it is not limited to this method; any technique capable of clustering similar records into a single class based on the amount of extracted features is sufficient. For each class obtained above, an arbitrary feature group ID is assigned, and an arbitrary feature group ID assigned to each record is added. After that, data tables are created by dividing them according to feature group ID.

[0123] Up until now, a method for creating feature groups by clustering has been described, but the method for creating feature groups is not limited to this; for example, they can also be created by a classification task based on a feature group ID assigned by the device user. Specifically, for some or all of the records obtained from the previously obtained collection data table (4049) and base model training data table (4045), the device user assigns a feature group ID based on pattern shape features, SEM image features, imaging conditions, and the device user's requirements for the machine learning model. The device user's requirements for the machine learning model can be, for example, "I want to prepare one machine learning model for each pattern type," "I want to divide machine learning models based on whether the lower layer is transparent in the SEM image," or "I want to prepare one machine learning model for each measurement device." After that, features such as pattern shape features and imaging conditions are extracted from each record. Each record is placed in a space having a number of dimensions equal to the number of extracted features, and classification is performed on this space using the feature group ID assigned by the device user as the true value. As a specific method for classification, one can consider methods such as Support Vector Machines, which are common for classification tasks; however, it is not limited to this technique, and it is sufficient if a feature group ID can be estimated from features such as pattern shape features and imaging conditions. The feature group classification model obtained in this way can estimate a feature group ID that reflects the device user's requirements based on input design data or information from measurement recipes. The feature group IDs estimated by the feature group classification model are appended to each record. Subsequently, data tables are created by dividing them according to feature group ID.

[0124] The records in the data table for each feature group obtained by the two methods above have similar features to each other, so by using them for additional learning, a machine learning model specialized in specific features can be obtained. It is preferable to train the machine learning model for each feature group by training the machine learning model created by training the base model training data table (4045) as the model that serves as the basis for training, but it is not limited to this method; for example, other methods may be used, such as training a new machine learning model without using the model that serves as the basis for training. The machine learning model trained for each feature group is given a unique name and is saved along with the training log.

[0125] Machine learning models specialized for each feature group can identify the corresponding feature group or model version from sources such as training logs. Using this information, it is possible to select the machine learning model to be used for pattern matching or to choose the model that serves as the basis for training. The method for identifying machine learning models is not limited to this approach; it is sufficient to be able to select the optimal machine learning model. For example, it is also possible to select the optimal machine learning model based on its filename or save date.

[0126] The pattern feature analysis unit of the recipe generation device (200) selects an optimal machine learning model for the input data using class information of each feature group or a feature group classification model obtained by the above method. The specific procedure is as follows.

[0127] (1) From the design data (101) and measurement recipe (201), pattern shape features or imaging conditions are extracted.

[0128] (2) Using class information or a feature group classification model for each feature group, the feature group ID is estimated for the features extracted in (1).

[0129] (3) A machine learning model trained using the data of the feature group estimated in (2) is described in the measurement recipe (201) as a machine learning model used in pattern matching.

[0130] <Embodiment 3>

[0131] FIG. 9 is a flowchart illustrating the operation of a pattern dimension measurement system according to Embodiment 3 of the present disclosure. In Embodiment 1, rule-based pattern matching is performed first, and if it fails, pattern matching by machine learning is additionally performed. Alternatively, the algorithm performed initially may be suitable for either rule-based or machine learning. Specifically, between S201 and S202, steps S901 and S902 described below are performed. Otherwise, it is the same as Embodiments 1 and 2.

[0132] (Fig. 9: S901)

[0133] The recipe generation device (200) interprets the measurement recipe (201) and, based on the information of past measurement recipes (201) and the matching success / failure results corresponding thereto, refers to the matching success / failure results having information of similar measurement recipes (201) to determine which method of pattern matching is appropriate to perform between rule-based pattern matching and machine learning pattern matching. It may be considered that the one that succeeded in past matching (or has a higher success rate) is more appropriate. If there is no measurement recipe (201) having similar information or if there are no past matching success / failure results, rule-based pattern matching is selected. In other words, at the time the measurement recipe (201) is generated, the matching algorithm corresponding to that measurement recipe (201) is also determined. The specific operation of the recipe generation device (200) will be described later using FIG. 10.

[0134] (Fig. 9: S902)

[0135] The measuring device (300) switches the pattern matching algorithm based on the result of S901. If a rule-based pattern matching is specified, S202 is performed. If a machine learning-based pattern matching device is specified, it skips to S204. By this step, the optimal pattern matching technique can be selected before executing the pattern matching process, thereby reducing unnecessary processing and improving the matching success rate.

[0136] FIG. 10 is a block diagram of a recipe generation device (200). The recipe generation device (200) analyzes a measurement recipe (201) and selects an optimal pattern matching technique. The measurement recipe (201) includes a pattern type, imaging conditions, a device number, etc. The pattern type includes shape information of the pattern, information on the density of the pattern, constituent material of the semiconductor pattern, whether or not there is transmission of the semiconductor pattern in the lower layer, and the amount of deformation that occurred. The imaging conditions include the acceleration voltage of the electron microscope used for imaging, the type of detector, the imaging magnification, the number of image frames of image integration performed for noise reduction, and whether or not prior high-definition processing by AI technology or image processing technology has been performed. The device number includes the type of device, the identification number of a device of the same type, etc.

[0137] The recipe generation unit (210) creates a measurement recipe (201). The recipe interpretation unit (220) interprets all or part of the information of the measurement recipe (201) and extracts the features. The recipe interpretation unit (220) refers to the matching success / failure judgment result linked to the measurement recipe (201) having the feature most similar to the extracted feature from the matching result storage unit (230) and selects a technique that is highly likely to succeed in pattern matching processing. The matching result storage unit (230) is a database that has each information of the pattern type, imaging condition, and device number of the measurement recipe (201) and the key of the matching success / failure result at that time, and holds the information of each key in the form of characters or numbers.

[0138] As a technique for interpreting recipes, for example, a classification model such as a random forest based on machine learning can be used. In this learning process, the input is set to each data corresponding to the measurement recipe (201) in the matching result storage unit (230), and the teacher label is set to the matching success / failure judgment result corresponding to the measurement recipe (201). By doing so, the machine learning model can estimate the optimal matching technique corresponding to the measurement recipe. When performing the estimation by the machine learning model, the input is set to each information of the measurement recipe (201), and the output is set to the optimal pattern matching technique. As another method, it is also possible to refer to the column with the most items matching the measurement recipe (201) in the database of the matching result storage unit (230), and refer to the matching success / failure result at that time. By these methods, the optimal pattern matching technique can be selected from the results of pattern matching processing performed in the past. The technique for interpreting recipes is not limited to these techniques, and any method capable of selecting the optimal pattern matching technique is sufficient.

[0139] <Embodiment 3: Arrangement>

[0140] The pattern dimension measurement system according to Embodiment 3 performs either pattern matching by a rule-based algorithm or pattern matching by machine learning, and if rule-based matching fails, performs machine learning matching, and if machine learning matching fails, performs retraining. Retraining of the learner is performed only when machine learning matching fails, and it is also possible to perform it automatically (see S206). Therefore, in Embodiment 3, it is possible to suppress the workload for learning.

[0141] <Regarding variations of the present disclosure>

[0142] The present disclosure is not limited to the embodiments described above and includes various variations. For example, the embodiments described above are detailed for the purpose of making the present disclosure easier to understand and are not necessarily limited to having all the described configurations. Furthermore, it is possible to substitute a part of the configuration of one embodiment with a part of another embodiment, and it is also possible to add a part of another embodiment to a part of an embodiment. Additionally, it is possible to add, delete, or substitute other configurations for a part of the configuration of each embodiment.

[0143] In the above embodiments, a scanning electron microscope was described as an example of a measuring device (300), but the present disclosure is not limited thereto and can be applied to any measuring device or inspection device that acquires an image and performs pattern matching.

[0144] In the above embodiments, the design data storage unit (100), the recipe generation device (200), the measuring device (300), and the machine learning model learning device (400) are shown as separate devices, but they may be implemented as one or more devices capable of realizing similar processing (a pattern matching device configured by one or more computer systems). For example, the system may be configured with a single device having the functions of the design data storage unit (100), the measuring device (300), and the machine learning model learning device (400), and the recipe generation device (200).

[0145] In Embodiment 2, the following may be used as the feature quantity used to perform clustering: (a) image feature quantity: image luminance distribution, image frequency components, (b) shape pattern feature quantity: shape pattern type, shape pattern planar density, (c) imaging condition feature quantity: optical conditions and electrical conditions of the imaging device acquiring the image.

[0146] In Embodiment 3, after performing steps S200 to S207, the success or failure of the measurement is determined (S209) as in Embodiment 1, and learning is performed using the successful result (S210 to S214). By performing steps S209 and beyond, learning data can be automatically collected as in Embodiment 1, but these steps do not necessarily have to be performed. That is, since Embodiment 3 performs either machine learning or rule-based learning, the objective can be achieved as long as this is realized. Of course, performing steps S209 and beyond is more effective.

[0147] In the above embodiments, the functional units provided by each device (e.g., recipe generation unit (210), recipe interpretation unit (220), rule-based pattern matching unit (302), machine learning pattern matching unit (303), matching success / failure determination unit (304), measurement unit (305), measurement success / failure determination unit (306), GUI (307), learning data interpretation unit (401), machine learning model learning unit (402), GUI (403), etc.) may be configured by hardware such as a circuit device that implements these functions, or by software that implements these functions that is executed by a computing device such as a CPU (Central Processing Unit).

[0148] In the above embodiments, the rule-based matching algorithm can be described as being configured to determine, by the internal logic of the algorithm, the correspondence relationship between an image feature and the position shift amount between two images corresponding to that image feature. On the other hand, it should be noted that the machine learning matching algorithm differs from the rule-based algorithm in that, outside the internal logic implementing the matching process, there exists a learning model that describes the result of learning the correspondence relationship between the image feature and the position shift amount.

[0149] In the above embodiments, it has been described that S201 is performed after S200, but the order in which each step is performed is not limited to this and can be changed within a range that does not contradict the technique according to the present disclosure. For example, S902 may be performed after S200, or S201 may be performed as described in FIG. 9. Explanation of the symbols

[0150] 100: Design data storage unit 200: Recipe Generator 300: Measuring device 400: Machine learning model learning device

Claims

Claim 1 A pattern matching device for performing pattern matching between images, comprising a computer system for performing pattern matching, wherein the computer system comprises a learning unit that outputs the result of the pattern matching based on machine learning, wherein the computer system performs the pattern matching using either a first algorithm that does not use the learning unit or a second algorithm that uses the learning unit, wherein the computer system performs the pattern matching using the second algorithm when the pattern matching using the first algorithm fails, and wherein the computer system performs re-learning of the learning unit when the pattern matching using the second algorithm fails. Claim 2 A pattern matching device according to claim 1, wherein, when the pattern matching using the second algorithm fails, the computer system generates training data for the learning unit to perform machine learning and performs retraining on the learning unit using the training data, and the computer system performs the pattern matching again using the learning unit that performed the retraining. Claim 3 A pattern matching device according to paragraph 2, wherein the computer system is configured to supply information including a position shift amount between two images performing pattern matching as the learning data to the learning device, and the computer system receives a designated input specifying the position shift amount through an interface provided by the computer system, and generates the learning data based on the designated input. Claim 4 A pattern matching device according to claim 2, wherein the computer system is configured to supply information including a position shift amount between two images performing pattern matching as the learning data to the learning device, and the computer system obtains the position shift amount from each of the first algorithms by each implementing a plurality of first algorithms each having a different matching procedure, and the computer system generates the learning data by using the position shift amount obtained from each of the first algorithms that provides the best matching between the two images. Claim 5 A pattern matching device according to claim 1, characterized in that the first algorithm implements a rule-based algorithm configured to determine, by the internal logic of the first algorithm, the correspondence relationship between an image feature quantity and a position shift quantity between two images corresponding to the image feature quantity. Claim 6 A pattern matching device according to claim 1, wherein the computer system supplies the result to the learner as learning data for the learner to perform re-learning when the pattern matching by the first algorithm or the learner is successful. Claim 7 A pattern matching device according to claim 6, wherein the computer system evaluates the inference performance of the learner that has performed the retraining, and if the inference performance does not satisfy a reference value, the computer system deletes a portion of the training data used for the retraining that belongs to a pre-defined adverse effect data type that is likely to degrade the inference performance, and the computer system performs the retraining again using the training data from which the adverse effect data type has been deleted. Claim 8 A pattern matching device according to claim 6, wherein the computer system clusters the training data used to perform the retraining according to the quantity of features, the computer system modifies the arrangement of data belonging to each class to suppress bias in the data of each class obtained by the clustering, and the computer system performs the retraining using the training data with suppressed bias. Claim 9 A pattern matching device according to claim 1, wherein the image is an observation image of a sample, and the computer system receives the result of measuring or inspecting the sample when the pattern matching is successful, and when the result indicates that the measurement or inspection has failed, the computer system receives a designated input specifying the position shift amount between the images through an interface provided by the computer system, and performs retraining on the learning device based on the designated input. Claim 10 A pattern matching device according to claim 1, wherein the computer system comprises a plurality of learning units that perform machine learning using different learning data, the computer system groups the learning data used by each learning unit when performing machine learning according to the characteristic quantity of the learning data, and the computer system performs pattern matching using the learning unit that performs machine learning using the learning data corresponding to at least one of the characteristic quantity of the image, the characteristic quantity of the shape pattern included in the image, or the characteristic quantity of the image capture condition. Claim 11 A pattern matching device according to claim 10, wherein the characteristic quantity of the image includes at least one of the luminance distribution of the image and the frequency component of the image, the characteristic quantity of the shape pattern includes at least one of the type of the shape pattern and the planar density of the shape pattern, and the characteristic quantity of the imaging condition includes at least one of the optical condition and the electrical condition of the imaging device acquiring the image. Claim 12 A pattern matching device for performing pattern matching between images, comprising a computer system for performing pattern matching, wherein the computer system comprises a learning unit that outputs the result of the pattern matching based on machine learning, wherein the computer system is configured to perform the pattern matching by executing one of two or more algorithms that implement the pattern matching, wherein the computer system selects one of the algorithms based on the matching success rate when each algorithm was previously executed, and wherein the computer system performs the pattern matching by executing the selected algorithm. Claim 13 A pattern matching device according to claim 12, wherein the computer system performs pattern matching using either a first algorithm that does not use the learner or a second algorithm that uses the learner, as two or more algorithms, and if the pattern matching using the first algorithm fails, the computer system performs pattern matching using the second algorithm, and if the pattern matching using the second algorithm fails, the computer system performs retraining of the learner. Claim 14 A pattern matching device for performing pattern matching between images, comprising a computer system for performing pattern matching, wherein the computer system comprises a learning unit that outputs the result of the pattern matching based on machine learning, wherein the computer system is configured to perform the pattern matching by executing any one of two or more algorithms that implement the pattern matching, wherein the computer system is configured to perform the pattern matching according to conditions described in a measurement recipe, wherein when the computer system acquires a new measurement recipe, it extracts features of the new measurement recipe, wherein the computer system selects any one of the algorithms by identifying the one most similar to the extracted features among the measurement recipes used when each algorithm was previously executed successfully, and wherein the computer system performs the pattern matching by executing the selected algorithm. Claim 15 A pattern matching device according to claim 14, wherein the computer system performs pattern matching using either a first algorithm that does not use the learner or a second algorithm that uses the learner, as two or more algorithms, and if the pattern matching using the first algorithm fails, the computer system performs pattern matching using the second algorithm, and if the pattern matching using the second algorithm fails, the computer system performs retraining of the learner.