Pattern matching device

Abstract

This disclosure proposes a method for significantly reducing the workload associated with data collection and truth generation in pattern matching techniques using machine learning. This method simplifies manual tasks in the learning process of machine learning models (dataset generation, model learning, performance evaluation, etc.). The pattern matching device described in this disclosure performs pattern matching using either a first algorithm without a learner or a second algorithm with a learner. If the previous pattern matching using the first algorithm fails, the previous second algorithm is used to perform the previous pattern matching. If the previous pattern matching using the second algorithm fails, the previous learner is relearned.

Description

[Technical Field]

[0001] This disclosure relates to a technique for performing pattern matching between images. [Previous Technology]

[0002] Apparatus for measuring / inspecting patterns formed on semiconductor wafers often utilizes template matching technology to align the field of view of the apparatus with the desired measurement or inspection position. Template matching is the process of identifying the area that most closely matches a pre-registered template image from the image being explored. Patent Document 1 describes one 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 value representing the correlation between a template image and a target image as a pixel value. This document can perform high-precision matching by learning to absorb the appearance discrepancy between the template image and the target image.

[0004] Patent Document 3 describes a method for selecting the minimum necessary learning data to achieve the target performance of a machine learning model when applying machine learning to a semiconductor inspection apparatus that captures patterns formed on a semiconductor wafer. When applying machine learning to image inspection, the collection of large amounts of image data and the generation of accurate inspection results corresponding to each image data are necessary, which places a significant burden on the device user. Patent Document 3 reduces the burden on the device user by selecting the minimum necessary data for the machine learning model to learn. [Prior Art Documents] [Patent Documents]

[0005] [Patent Document 1] Japanese Patent No. 4218171 (corresponding to US Patent No. 6,627,888) [Patent Document 2] WO2022 / 009357 [Patent Document 3] Japanese Unexamined Patent Publication No. 2020-35282 [Summary of the Invention]

[0006] [Problem to be Solved by the Invention] The pattern matching technique used in semiconductor inspection / testing devices has previously relied on rule-based matching methods that use pre-determined algorithms. General rule-based methods are considered to have high versatility because algorithms can be designed based on various known cases. However, when matching fails due to discrepancies in appearance between the pattern image and the image being investigated, it is not easy to improve the algorithm, and corresponding improvements require time.

[0007] Therefore, a pattern matching method using machine learning with learning capabilities, such as that described in Patent Document 2, has been considered. This method learns to absorb the discrepancies in appearance between a template image and the image being explored, enabling high-precision matching. However, in order to learn a machine learning model with sufficient performance, it is necessary to collect a large amount of learning data, such as images, and to generate correct pattern matching results (hereinafter referred to as ground truth) from the collected learning data. Moreover, generating ground truth from a large amount of data would be time-consuming if done manually, thus placing a heavy burden on the operator. Furthermore, if the types of data contained in the learning data (image shape, image quality, etc.) are insufficient, it will become learning only for specific data, resulting in low versatility of the machine learning model.

[0008] Patent Document 3 describes a method for analyzing semiconductor design data and camera images, and selecting and collecting the minimum necessary learning data to achieve the target performance of a machine learning model. This ensures that the variety of learning data is reduced while simultaneously improving the versatility of the machine learning model and reducing the labor involved in ground truth generation. The document also describes the generation of a ground truth image that simplifies the ground truth generation process for each pixel in the segmentation of a machine learning task. However, it does not eliminate the need for manual operation, leaving room for automation.

[0009] This disclosure was developed in view of the above-mentioned problems. It proposes a method for pattern matching techniques using machine learning that significantly reduces the workload associated with data collection and truth generation, thereby simplifying manual tasks in the learning process of machine learning models (generating learning datasets, learning machine learning models, performance evaluation, etc.). [Means used to solve the problems]

[0010] The pattern matching apparatus disclosed herein performs pattern matching using either a first algorithm without the learner or a second algorithm with the learner; if the previous pattern matching using the first algorithm fails, the previous pattern matching is performed using the second algorithm; if the previous pattern matching using the second algorithm fails, the previous learner is relearned. [Effects of the Invention]

[0011] According to the pattern matching device disclosed herein, the workload associated with collecting learning data and generating truth values ​​during the learning process of a machine learning model can be reduced, thereby simplifying manual tasks in the learning process of the machine learning model (generating the learning dataset, learning the machine learning model, performance evaluation, etc.). This significantly reduces the workload for users of devices equipped with machine learning-based pattern matching.

Implementation Method

[0013] <Embodiment 1> Figure 1 is a structural diagram of the pattern size measurement system described in Embodiment 1 of this disclosure. The pattern size measurement system is a system used to measure the size of a shape pattern formed on a semiconductor wafer. The pattern size measurement system includes: a design data memory 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). The data generated / output by each device constituting the system can be transmitted and received between the devices.

[0014] The matching result of the pattern matching is output as a matching translation amount. The matching translation amount represents the positional translation or positional difference between the design data 101 and the SEM (Scanning Electron Microscope) image 310, and can be represented by a 2D scalar value, such as (translation amount in the X direction, translation amount in the Y direction). Furthermore, the template image is not limited to the SEM image, but can also be other types of images such as the design data 101. Also, the image being explored is not limited to the design data 101, but can also be other types of images, such as the SEM image 310.

[0015] The design data memory unit 100 stores the design data 101 representing the pattern structure of semiconductor elements created by the device user.

[0016] The recipe generation device 200 is equipped with an interface for allowing the device user to input measurement points on the component pattern referenced from the design data 101. Based on the measurement point information input from the interface and the design data 101, the device outputs a measurement recipe 201, which includes the camera conditions in the measurement device 300, the algorithm used in pattern matching based on the rule basis, and the measurement point information.

[0017] The measurement device 300 measures the size of the component pattern based on the measurement formula 201 and outputs the measured value. The measurement device 300 includes: an image camera 301 that captures the component pattern formed on the semiconductor wafer, a regular basic pattern matching unit 302, a machine learning pattern matching unit 303, a matching success or failure determination unit 304, a measurement unit 305, a measurement success or failure determination unit 306, a GUI (Graphical User Interface) 307, and a measurement information memory unit 308.

[0018] The camera unit 301 captures component patterns based on the imaging conditions of the measurement formula 201 and outputs the SEM image 310. The rule-based pattern matching unit 302 uses the algorithm described in the measurement formula 201 to perform pattern matching between the SEM image 310 and the design data 101, and outputs the matching translation amount 314 between the two images. The machine learning pattern matching unit 303 uses the machine learning model 312 to perform pattern matching between the SEM image 310 and the design data 101, and outputs the matching translation amount 314 between the two images. The data used in pattern matching is not limited to the aforementioned combination. As long as data that can identify the consistent location of the field of view between the sample image and the image being explored is used, for example, the SEM image and the design drawing derived 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 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 translation amount 314, and outputs the matching success / failure determination result. The measurement unit 305 measures the size of the component pattern based on the measurement point information recorded in the measurement recipe 201, the design data 101, and the SEM image 310, and outputs the measured value. The measurement success / failure determination unit 306 determines the success or failure of the measurement based on the measured value and outputs the measurement success / failure determination result. If the measurement success / failure determination result is unsuccessful, the device user operates the GUI 307 to input the matching translation amount that indicates the pattern shape between the design data 101 and the SEM image 310 is consistent, and outputs the matching translation amount 314. The measurement information memory unit 308 stores the measurement recipe 201, the SEM image 310, the design data 101, and the machine learning model 312. Machine learning model 312, in its initial state, maintains the basic model learned by machine learning model learning device 400.

[0019] The measuring device 300 outputs the measuring formula 201, design data 101, SEM image 310, and matching translation amount 314 to the machine learning model learning device 400 after the component pattern size measurement is completed. At this time, the output design data 101 and SEM image 310 can be used as learning data for the machine learning model, and the matching translation amount 314 is the value obtained from the normal implementation of the component pattern size measurement, which can be used as the true value for the machine learning model. Therefore, the manual work in collecting learning data and generating true values ​​can be almost eliminated.

[0020] 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 memory unit 404, and a machine learning model memory unit 405. The learning data memory unit 404 stores temporary memory data 406 and a learning dataset 407. The temporary memory data 406 includes: a calculation formula storage folder 4042, a design data storage folder 4043, and a SEM image storage folder 4044. The temporary memory 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 additional learning data table 4048. Specific examples of these data structures will be described later.

[0021] The machine learning model learning device 400 stores the measurement recipe 201, design data 101, and SEM image 310 collected 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 memory data 406, respectively. At this time, the machine learning model learning device 400 compiles any data ID, the file names of the measurement recipe 201, design data 101, and SEM image 310, and the value of the matching translation amount 314 into a single record to create a data collection table 4049, and stores the created data table in the temporary memory data 406. Data collection form 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 the combination of measurement formula 201, design data 101, SEM image 310, matching translation amount 314, etc. can be used. For example, it can also be managed by folder structure and file name.

[0022] The learning data analysis unit 401 reads in each data (measurement formula 201, design data 101, SEM image 310) based on the existing data tables (e.g., basic model learning data table 4045) in the collection 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 is described in Figure 5). Furthermore, the learning data analysis unit 401 selects records from each data table in a way that ensures there is no bias in the extracted features (the selection method that ensures there is no bias is described in Figure 5). The data (measurement formula 201, design data 101, SEM image 310) contained in the records selected from the collection data table 4049 are stored in the measurement formula storage folder 4042, the design data storage folder 4043, and the SEM image storage folder 4044. Furthermore, records selected from each data table are treated as newly added data tables, given unique names, and saved. This ensures the diversity of data contained in the learning dataset 407, guaranteeing both versatility and improved machine learning model performance. Moreover, since previously used learning datasets can be maintained while new learning datasets are created, managing multiple learning datasets becomes easier. During data parsing in the learning dataset parsing unit 401, errors in the matching shift amount 314 can be manually corrected using the GUI 403 described later.

[0023] The machine learning model learning unit 402 reads in the data tables (e.g., supplementary learning data table 4048) of the learning dataset 407, and inputs the necessary data (design data 101, SEM image 310, matching translation amount 314, etc.) for learning, and performs learning based on preset learning conditions (details of the machine learning model learning method are described later using Figures 3 and 4). To evaluate the performance of the learned machine learning model, the machine learning model learning unit 402 reads in the test data table 4046 stored in the learning dataset 407, and inputs the necessary data (design data 101, SEM image 310, matching translation amount 314, etc.) for inference, and evaluates the performance of the learned model. If the target performance is not achieved, the learning conditions can be changed to perform relearning. In this way, a machine learning model with the necessary performance can be obtained. If the target performance is achieved, the learning log and the data archive of the learned machine learning model 408 are given unique names and stored in the machine learning model 312. At this point, the machine learning models that have been learned in the past are preserved.

[0024] If the matching translation amount 314 is incorrect and the correct matching translation amount needs to be taught, the device user operates the GUI403 to teach and output the correct matching translation amount that matches the pattern shape of the design data 101 and the SEM image 310.

[0025] Figure 2 is a flowchart illustrating the operation of the pattern size measurement system. The steps in Figure 2 are explained below.

[0026] (Figure 2: S200) The recipe generation device 200 generates a measurement recipe 201 based on the measurement point information of the component pattern input by the device user through the interface of the recipe generation device 200 and the design data 101 obtained from the design data memory unit 100. The recipe 201 includes the camera conditions in the measurement device 300 (which may include camera position information, etc.), the algorithm used in pattern matching based on the rule basis, measurement point information, etc.

[0027] (Figure 2: S201) The camera unit 301 of the measuring device 300 captures the component pattern based on the imaging conditions of the measuring formula 201 and outputs the SEM image 310.

[0028] (Figure 2: S202) The rule-based pattern matching unit 302 calculates the matching translation amount 314 between the design data 101 and the SEM image 310. The data used in pattern matching is not limited to this combination; any data that can identify the consistent locations of the view between the template image and the image being explored can be used. For example, the SEM image 310 and the design drawing derived from the design data 101 can be used, as well as the point group information of the SEM image 310 and the design data 101 can be used. The matching translation amount 314 is calculated using a matching algorithm determined based on the measurement formula 201. The matching algorithm can also be, for example, an algorithm that calculates the similarity between the design data 101 and the SEM image 310 at each candidate position using the mean square error, and outputs the matching translation amount of the position with the highest similarity. Matching algorithms are not limited to this method. Any algorithm that extracts each image feature based on a predetermined process and calculates the matching shift based on the extracted image features is acceptable.

[0029] (Figure 2: S203) The matching success or failure determination unit 304 determines whether a match is successful based on the design data 101, the SEM image 310, and the matching translation amount 314. One example of a matching success or failure determination method is to calculate the mean square error representing the similarity between the design data 101 and the SEM image 310 based on the matching translation amount 314, and determine success if the error is above a threshold. The matching success or failure determination method is not limited to this method; it is acceptable as long as it can determine whether the position based on the matching translation amount 314 on the design data 101 corresponds to the field of view position of the SEM image 310. If the matching success or failure determination result is successful, proceed to S208. If the matching success or failure determination result is unsuccessful, proceed to S204.

[0030] (Figure 2: S204) The machine learning pattern matching unit 303 uses a machine learning model that, through learning, estimates the matching translation amount 314 based on the design data 101 and the SEM image 310, and calculates the matching translation amount 314. The data used in pattern matching is not limited to the aforementioned combination; it can use data that can identify the consistent locations of the view between the template image and the image being explored. For example, it can use the SEM image and the design drawing derived from the design data 101, or it can use the point group information of the SEM image and the design data 101. As the machine learning model, it reads a specified machine learning model (the highest-performing machine learning model or the latest machine learning model, etc.) from the machine learning model 312 and uses the read-in machine learning model. As an example, the machine learning-based pattern matching method can be implemented by estimating the correlation between images and calculating the matching translation amount from the correlation images, as in Patent Document 2. Pattern matching methods derived from machine learning are not limited to this, as long as they can pinpoint the points of visual agreement between the template image and the image being explored. For example, methods that directly estimate the matching translation between images using a machine learning model are also possible.

[0031] (Figure 2: S205) The matching success determination unit 304 determines whether a match is successful based on the matching translation amount 314. If the matching success determination result is successful, S208 is implemented. If the matching success determination result is unsuccessful, the design data 101 and the SEM image 310 are stored in the temporary memory data 406 of the machine learning model learning device 400, and S206 is implemented for the purpose of learning the machine learning model. At this time, the machine learning model learning device 400 creates a temporary data table 4047 that records the file names of the received design data 101 and SEM image 310, assigns a unique name to it, and stores it in the temporary memory data 406.

[0032] (Figure 2: S206) The device user uses GUI403 to read in design data 101 and SEM image 310 based on temporary data table 4047. The device user operates GUI403 to teach the matching translation amount of the correct solution where the pattern shapes of the two images are consistent. The method of teaching the matching translation amount of the correct solution is not limited to the method using GUI403, as long as it is a method that can teach the matching translation amount of the correct solution. For example, pattern matching algorithms based on multiple rules can also be used to perform pattern matching on design data 101 and SEM image 310, and the most similar matching translation amount can be used as the matching translation amount of the correct solution. The machine learning model learning device 400 creates a data table 4041 containing the matching translation amount of the correct solution, design data 101, and SEM image 310 for learning failed patterns. Details are described later using Figure 3.

[0033] (Figure 2: S207) 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 memory unit 405. Furthermore, the learned machine learning model 408 is stored in the machine learning model 312 of the measuring device 300. Details are shown in Figure 3 and will be described later. S204 is then implemented again using the learned machine learning model 408.

[0034] (Figure 2: S208) The measurement unit 305 measures the size of the component pattern based on the measurement formula 201, design data 101, SEM image 310 and matching translation amount 314, and outputs the measured values.

[0035] (Figure 2: S209) The measurement success / failure determination unit 306 determines the success or failure of the measurement based on the measurement formula 201, design data 101, SEM image 310, and measurement values, and outputs the success / failure determination result. If the measurement success / failure determination result is successful, proceed to S212. If the measurement success / failure determination result is unsuccessful, proceed to S210 to implement manual assistance.

[0036] (Figure 2: S210) The device user uses GUI307 to input a matching translation amount 314, based on design data 101 and SEM image 310, to ensure that the patterns of the two are consistent. As a method for inputting the matching translation amount, the numerical value of the matching translation amount can be directly input, or the matching translation amount that will become the correct solution can be calculated while confirming the image formed by the overlap of design data 101 and SEM image 310.

[0037] (Figure 2: S211) The measurement unit 305 measures the size of the component pattern based on the measurement formula 201, design data 101, SEM image 310 and matching translation amount 314, and outputs the measured values.

[0038] (Figure 2: S212) After the measurement of the dimensions of the component pattern is completed, the measurement device 300 outputs the measurement recipe 201, design data 101, SEM image 310, and matching translation 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 collected 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 memory data 406, respectively. At this time, the machine learning model learning device 400 records the file names of the collected measurement recipe 201, design data 101, and SEM image 310, as well as the matching translation amount 314, in the data collection table 4049.

[0039] (Figure 2: S213) The learning data analysis unit 401 analyzes the data stored in the temporary memory data 406 to create a new learning data set. Details are shown in Figure 5 and will be described later. S213 can be implemented each time a pattern is measured, or it can be implemented after a certain amount of data has been collected.

[0040] (Figure 2: S214) The machine learning model learning unit 402 learns the machine learning model 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 memory unit 405. Furthermore, the learned machine learning model 408 is stored in the machine learning model 312 of the measurement device 300. Details are shown in Figure 4 and will be described later. S214 can be implemented each time a pattern is measured, or it can be implemented after a certain amount of data has been collected.

[0041] Figure 3 is a flowchart illustrating the specific actions of S206 to S207. The steps in Figure 3 are explained below. S206 to S207 can be executed immediately when pattern matching fails due to machine learning, or they can be executed later after the semiconductor wafer has been unloaded from the measuring device 300.

[0042] (Figure 3: S300) The device user uses GUI 403 to read in design data 101 and SEM image 310 based on temporary data table 4047, and inputs the matching translation amount for the correct solution where the pattern shapes of the two images are consistent. Then, a failure pattern learning data table 4041 containing the input matching translation amount for the correct solution, design data 101, and SEM image 310 is created. At this time, all or part of the records in the basic model learning data table 4045 or the supplementary learning data table 4048 can be added to the failure pattern learning data table 4041. In this way, the general performance of the machine learning model can be ensured while also addressing failure patterns.

[0043] (Figure 3: S301) The machine learning model learning unit 402 implements the learning of 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 number of periods) or the learning rate, etc., which are arbitrarily set values ​​of the machine learning model (hereinafter referred to as hyperparameters), and the presence or absence of the model as the basis for learning. The internal parameters of the machine learning model can be updated by updating all the parameters of the model as the basis for learning, or only a part of them. The predetermined learning conditions are ideal values ​​obtained through prior experiments and empirical rules, but are not limited to this. For example, the number of periods can also be used to confirm the progress of learning while achieving a certain performance. Ideally, the model used as the foundation for learning is a machine learning model learned from the data table 4045 using the base model, or the latest machine learning model stored in the machine learning model memory unit 405. However, it is not limited to these; for example, a new machine learning model can also be learned. Alternatively, the learned machine learning model 408 can be used as the foundation model for future learning.

[0044] (Figure 3: S302) The machine learning model learning unit 402 takes the design data 101 (which taught the correct matching translation amount in S300) and the SEM image 310 as input to the learned machine learning model 408 and performs inference 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 basic model learning data table 4045 or the supplementary learning data table 4048 can be deleted from the failure pattern learning data table 4041. In this way, although the generality of the machine learning model will be reduced, it can be specialized for failure cases, and there is a possibility that the matching will turn into a success. The learned machine learning model 408 can correspond to failure patterns, so it can respond quickly to matching failures.

[0045] (Figure 3: S303) The machine learning model learning unit 402 stores the learned machine learning model 408 and the learning log in the machine learning model memory unit 405.

[0046] (Figure 3: S304) The learned machine learning model 408 is stored in the machine learning model 312.

[0047] Figure 4 is a flowchart illustrating the specific actions of S214. The steps in Figure 4 are explained below.

[0048] (Figure 4: S400) The machine learning model learning unit 402 uses the supplementary learning data table 4048 to implement machine learning model learning based on predetermined learning conditions. The predetermined learning conditions are ideally the best values ​​obtained through prior experiments and according to empirical rules, but are not limited to this; for example, the number of periods can also be used, to confirm the progress of learning while achieving a certain performance. Furthermore, the model used as the basis for learning is ideally the machine learning model learned using the basic model learning data table 4045, or the latest machine learning model kept in 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 can also be learned. The method for selecting the model as the basis for learning is not limited to referring to the learning log; as long as the most suitable model as the basis for learning can be selected, for example, the latest machine learning model can be specified based on the saving date and time of the machine learning model, or it can be specified based on the file name of the machine learning model.

[0049] (Figure 4: S401) The machine learning model learning unit 402 performs inference on the learned machine learning model 408 using data from the test data table 4046, whose true values ​​are known, as input, and evaluates its performance. The test data table 4046 is used to evaluate the general performance of the machine learning model. It can also record data with various shape features, such as data that the device user wants the selected match to succeed. As a performance evaluation method, it can be, for example, a correct answer rate above a threshold, or other methods, such as the difference between the inferred match shift and the true match shift being very small, or a method combining these. If sufficient performance is obtained, S403 is implemented. If the performance is insufficient, S402 is implemented.

[0050] (Figure 4: S402) The machine learning model learning unit 402 can also identify and delete data in the additional learning data table 4048 that may have an adverse effect on learning (e.g., data where the template image is not present in the field of view due to field of view shift, data that becomes larger than the design data due to deformation of the component pattern formed on the wafer, data with significantly different image quality and other characteristics compared to other data in the same data table, etc.) (adverse data type). Furthermore, the learning conditions can also be adjusted. Methods for adjusting learning conditions include changing the model that forms the basis of learning, and increasing the number of periods if learning is insufficient. Afterwards, S400 is performed again. This step is optional and not mandatory.

[0051] (Figure 4: S403) The machine learning model learning unit 402 stores the learned machine learning model 408 and the learning log in the machine learning model memory unit 405.

[0052] (Figure 4: S404) The learned machine learning model 408 is stored in the machine learning model 312.

[0053] Figure 5 is a flowchart illustrating the specific actions of S213. The steps in Figure 5 are explained below.

[0054] (Figure 5: S500) The learning data analysis unit 401 reads in the design data, SEM images, and measurement formulas contained in the records of the collected data sheet 4049, the failed pattern learning data sheet 4041, the latest supplementary learning data sheet 4048, and the basic model learning data sheet 4045. All four data sheets can be used, or only one can be used. The learning data analysis unit 401 analyzes the read data and extracts features such as pattern shape characteristics, SEM image characteristics, and imaging conditions. Specific examples of each feature are as follows: Pattern shape characteristics include pattern type (line width / spacing pattern or hole array pattern, etc.), pattern density, and constituent materials. SEM image characteristics include the presence or absence of penetration of the underlying pattern, deformation amount, and image quality. Imaging conditions include acceleration voltage, detector type, imaging magnification, image quality integrator frame count, and high-definition processing. Specific methods for extracting pattern shape features include, for example, inferences derived from information obtained from edge and corner detection of the image, frequency analysis of the image, etc. Alternatively, this information can be included in the measurement formula by the formula maker during the formulation process. Specific methods for extracting SEM image features include, for example, extracting the outline of the pattern portion from the SEM image and design data respectively, and by comparing the extracted outlines, inferring the presence or absence of penetration of the underlying pattern or the amount of pattern deformation. Again, this information can be included in the measurement formula by the formula maker during the formulation process. The imaging conditions can also directly use the information recorded in the measurement formula, or perform certain preprocessing. The extracted features are not limited to these; they may also include, during the formulation process, the type of the measurement device 300, the identification number of the same type of device, the project name (photolithography or etching, etc.) recorded by the formula maker in the measurement formula.

[0055] (Figure 5: S501) The learning data analysis unit 401 selects records from each data table in a way that ensures the features extracted in S500 are unbiased, creating a new supplementary learning data table. Methods for selecting records include, for example, clustering based on the characteristics of the records, selecting records from each category in a way that ensures each category is unbiased. Specific clustering methods include k-means clustering, where each record is arranged in a space with the same dimension as the number of extracted features, but this method is not limited to any specific method; any method that can cluster similar records into one category based on the extracted features is acceptable.

[0056] (Figure 5: S502) The device user confirms the matching shift amount in the newly added supplementary learning data table in GUI403, and manually corrects the erroneous matching shift amount if there is an error. The method for correcting the matching shift amount is not limited to the method using GUI403; any method that can correct the matching error is acceptable. For example, pattern matching algorithms based on multiple rules can be used to perform pattern matching on design data 101 and SEM images, and the most similar matching shift amount can be taken as the correct matching shift amount. Furthermore, as the object data for which the matching shift amount is to be confirmed, only data with a high probability of matching shift amount errors can be selected in advance, and the matching shift amount can be confirmed for those data. As a selection method, for example, pattern matching algorithms based on multiple rules can be used to perform pattern matching on design data 101 and SEM images, and data for which no similar matching shift amount is obtained in each algorithm can be selected. S502 is not necessarily to be implemented. This is because the collected data is data that indicates the correct solution, so the probability of a correct match is very high. This step is to reliably provide the data needed for the correct solution.

[0057] (Figure 5: S503) The learning data analysis unit 401 stores the data (measurement formula 201, design data 101, SEM image 310) contained in the temporary memory data 406 in the newly added supplementary learning data table 4048 into the measurement formula storage folder 4042, the design data storage folder 4043, and the SEM image storage folder 4044. The newly added supplementary learning data table 4048 is given a unique name and stored in the learning data set 407.

[0058] Figure 6 shows an example of the data stored in temporary memory data 406 and learning data set 407. Temporary memory data 406 and learning data set 407, as shown in Figure 6, contain: multiple data tables for managing various data, a calculation formula storage folder 4042, a design data storage folder 4043, and a SEM image storage folder 4044. Temporary memory data 406 and learning data set 407 are stored in three folders and respectively divided into calculation formula, design data, and SEM images. However, all data files can also be stored at the same level as the data tables, all data files can be stored in one folder, or they can be divided into multiple folders.

[0059] The learning data set 407 contains: a basic model learning data sheet 4045, which is required for learning the basic machine learning model in the initial stage; and a test data sheet 4046, which is required for evaluating the performance of the machine learning model. The data files recorded in these two data sheets are as follows. The basic model refers to a machine learning model that has been trained using patterns with high measurement frequency or general patterns.

[0060] Each data table, as shown in Figure 6, maintains the data ID, measurement formula, design data, SEM image, and matching shift amount as a single row of records. By maintaining multiple such data tables, multiple datasets can be managed in a table-based manner. The data set management method is not limited to the management method derived from data tables; any method that can manage combinations of measurement formulas, design data, SEM images, and matching shift amounts is acceptable. For example, it can also be managed through folder structures, file names, etc. Furthermore, each data table maintains the information of data ID, measurement formula, design data, SEM image, and matching shift amount in the record, but the amount of information maintained can be more or less, as long as the information necessary for data management is maintained. For example, the data ID may not be maintained, but other information such as the login date and time may be maintained instead.

[0061] Figure 7 illustrates an example of a learning log. As shown in Figure 7, the learning log records the update date, machine learning model name, data table used, and learning conditions as a single row in a data table format. This simplifies the management of learned machine learning models, allowing for the retrieval of past models if necessary. In Figure 7, the learning log maintains information such as the update date, machine learning model name, data table used, and learning conditions. However, the amount of information maintained can vary, as long as it is sufficient to manage the learned machine learning models. For example, other information such as the machine learning model ID can also be maintained.

[0062] Figure 8 shows an example of the configuration of a measurement device 300 containing an SEM 600 and a GUI 307. The SEM 600 can be operated as a camera 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 measurement unit 305, and a measurement success / failure determination unit 306. The image memory 615 can be operated as a measurement information memory unit 308. The SEM 600 measures, for example, the pattern dimensions of semiconductor elements formed on a semiconductor wafer. The calculation processing device or computer system in the measurement device 300 can be configured as, for example, the control unit 614. The control unit 614 has a calculation means (e.g., a CPU / GPU 616) and a memory device (e.g., a memory including the image memory 615). The memory device can store information, such as a program related to pattern matching processing. The pattern matching process shown in Figure 2 is executed by the CPU / GPU 616. That is, the control unit 614 functions as the pattern matching unit. In other words, the program enables the computer system to function as the calculation and processing device contained in the pattern matching unit to execute the pattern matching process shown in Figure 2.

[0063] The SEM 600 generates an electron beam from the electron gun 601. The deflector 604 and the object lens 605 are controlled to focus the electron beam at any point on the sample, i.e., the semiconductor wafer 603, placed on the platform 602. The semiconductor wafer 603, irradiated by the electron beam, emits secondary electrons, which are detected by the secondary electron detector 606. The detected secondary electrons are converted into digital signals by the A / D converter 607. The image displayed as a digital signal is stored in the image memory 615 within the control unit 614. This image is used as, for example, an SEM image 310, and pattern matching processing is performed based on this image by the control unit 614 or the CPU / GPU 616. The necessary settings and display of the processing results for these processes can be implemented on the GUI 307. In alignment and positioning using an optical camera with a lower magnification than SEM, the optical camera 611 can also be used. The signal captured by the optical camera 611 on 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, then the A / D converter 612 is not needed). The image displayed 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.

[0064] The SEM600 series 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 reflection electron detector 608, and the detected reflected electrons are converted into digital signals by an A / D converter 609 or 610. The image displayed by the digital signal is stored in the image memory 615 within the control unit 614, and the CPU / GPU 616 performs image processing corresponding to the purpose. The control unit 614 can also control the platform 602 via the platform controller 630, and can also control the object lens 605, etc., via the bias control unit 631.

[0065] <Implementation Version 1: Summary> The pattern size measurement system described in Implementation Version 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 by performing pattern matching and measurement based on it, and since they are the results of successful measurement, they are suitable as learning data. Therefore, the collection of learning data and the workload required to provide the correct solution can be reduced.

[0066] <Embodiment 2> This disclosure describes an embodiment of a configuration required for selectively using a learner suitable for matching objects with specific characteristics. Embodiment 2, like Embodiment 1, is configured using the pattern size measurement system of FIG1. ​​The basic operation is the same as in Embodiment 1; the differences in operation from Embodiment 1 will be explained below.

[0067] The design data memory unit 100, similar to Embodiment 1, stores the design data 101 representing the pattern structure of the semiconductor element created by the device user.

[0068] The recipe generation device 200, in addition to its interface, also includes a pattern feature analysis unit. The pattern feature analysis unit analyzes the design data 101 and the measured recipe 201 obtained from the design data memory unit 100, extracting their pattern shape features and imaging conditions. Specific examples of each feature are as follows: Pattern shape features include pattern type (line width / spacing pattern or hole array pattern, etc.), pattern density, and constituent materials. Imaging conditions include acceleration voltage, detector type, camera magnification, image quality integrator frame rate, and high-definition processing. Specific extraction methods for pattern shape features can include, for example, estimations based on information obtained from image edge and corner detection, image frequency analysis, etc. Alternatively, the device user can operate the interface and input pattern type information. Imaging conditions can also be obtained by directly using information recorded in the measured recipe, or by performing certain preprocessing. The extracted features are not limited to these, and may also include the type of measuring device, the identification number of the same type of device, the name of the project (photolithography or etching, etc.) recorded by the formulator in the measuring formula when the measuring formula is made.

[0069] The recipe generation device 200 uses the inherent feature group IDs of each feature group (details of which will be described later) created in advance to select a machine learning model that has been trained on data from feature groups that will be classified using the extracted features. The method for selecting the machine learning model is not limited to this method, as long as it is possible to select a machine learning model that has been trained on data similar to the extracted features. For example, the user of the device can also operate the interface to select the machine learning model that is most suitable for the design data.

[0070] The recipe generation device 200 outputs a measurement recipe 201, which includes the camera 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, based on the measurement point information, design data 101, and a selected machine learning model input from the interface. The selected machine learning model has been pre-trained to specialize in learning specific features, and because it has high matching performance for similar data, it can reduce the number of matching failure cases in the machine learning pattern matching unit 303.

[0071] The measurement apparatus 300 includes: an imaging unit 301 for capturing images of element 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 memory unit 308. The machine learning pattern matching unit 303 retrieves the machine learning model used in pattern matching from the machine learning model 312 based on the measurement recipe 201. Alternatively, the machine learning pattern matching unit 303 retrieves the machine learning model used in pattern matching from the machine learning model memory unit 405 of the machine learning model learning device 400 based on the measurement recipe 201. The machine learning pattern matching unit 303 performs pattern matching using the acquired machine learning model.

[0072] 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 memory unit 404, and a machine learning model memory unit 405. The learning data set 407 of the learning data memory unit 404 contains a pre-created additional learning data set for each feature group. The method for creating the data set for each feature group will be described later. The machine learning model memory unit 405 stores the machine learning model for each feature group. The learning / management method for the machine learning model of each feature group will be described later.

[0073] The machine learning model learning device 400 stores the measurement recipe 201, design data 101, and SEM image 310 collected 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 memory data 406, respectively. At this time, the machine learning model learning device 400 compiles any data ID, the file names of the measurement recipe 201, design data 101, and SEM image 310, as well as the values ​​of the matching translation amount 314 and feature group ID into a single record to create a data collection table 4049, and stores the created data table in the temporary memory data 406. Data collection form 4049 was created to manage the collected data, but the management method of the collected data is not limited to the management method of data collection form 4049. Any method that can manage the combination of measurement formulas, design data, SEM images, matching translation amounts, feature group IDs, etc. can be used. For example, it can also be managed by folder structure and file name.

[0074] Embodiment 2, similar to Embodiment 1, can be described using the flowchart shown in FIG2. Here, the differences between Embodiment 2 and Embodiment 1 are mainly explained in the processing of each step in FIG2. Steps not described are the same as in Embodiment 1.

[0075] (Figure 2: S200) The recipe generation device 200 generates a measurement recipe 201 based on the measurement point information of the component pattern input by the device user through the interface, the design data 101 obtained from the design data memory unit 101, and the feature groups obtained from the pattern feature analysis unit. The recipe 201 includes the camera conditions in the measurement device 300 (which may include camera position information, etc.), the algorithm used in pattern matching based on rules, the machine learning model used in pattern matching based on machine learning, and the measurement point information.

[0076] (Figure 2: S204) The machine learning pattern matching unit 303 retrieves the machine learning model used in pattern matching from the machine learning model 312 based on the measurement formula 201. Alternatively, it retrieves it from the machine learning model memory unit 405 of the machine learning model learning device 400. Thereafter, the machine learning pattern matching unit 303 performs pattern matching using the acquired machine learning model.

[0077] (Figure 2: S206) Using the same method as in Embodiment 1, the device user operates the GUI 403 to teach the matching translation amount of the correct solution. Subsequently, the machine learning model learning device 400 creates a failure pattern learning data table 4041 containing the input matching translation amount of the correct solution, design data 101, and SEM image 310. During the creation of the failure pattern learning data table 4041, all or part of the dataset used by the machine learning model during learning, obtained in S204, can also be added to the failure pattern learning data table 4041. This ensures both the performance of the machine learning model for a specific pattern and its ability to handle failure patterns.

[0078] (Figure 2: S207) The machine learning model learning unit 402 learns the machine learning model in the same way as in Embodiment 1, and outputs the learned machine learning model 408. The model used as the basis for learning is ideally 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.

[0079] (Figure 2: S212) After the measurement device 300 has completed the measurement of the dimensions of the component pattern, etc., it outputs the measurement recipe 201, design data 101, SEM image 310, and matching translation 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 collected 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 memory data 406, respectively. At this time, the machine learning model learning device 400 records the file names of the collected measurement recipe 201, design data 101, and SEM image 310, as well as the values ​​of matching translation amount 314 and pattern feature group ID in the data collection table 4049.

[0080] (Figure 2: S213) The learning data parsing unit 401 extracts the feature group ID information from the failure pattern learning data table 4041 and the collection data table 4049. The records from each data table are appended to the supplementary learning data table 4048 for each feature group, ensuring consistency in the feature group ID information. In this way, learning data can be collected and accumulated for each feature group. Therefore, a machine learning model with high performance for specific features can be learned.

[0081] (Figure 2: S214) The machine learning model learning unit 402 uses the additional learning data table 4048 created in S213 for each feature group, and performs learning of machine learning models adapted to each feature group using the same learning method as in Embodiment 1. The learning logs and the learned machine learning models, each with a unique name, are stored in the machine learning model memory unit 405. In addition to the information shown in Embodiment 1, the learning logs can also be assigned feature group IDs. This allows each feature group to be bound to each machine learning model. The learned machine learning models are stored in the machine learning model 312 of the measuring device 300.

[0082] As a method for creating the feature groups in Embodiment 2, a method resulting from clustering can be considered, for example. The method for creating feature groups resulting from clustering will be described below.

[0083] The collected data table 4049 obtained in advance, and the data table 4045 for learning from the basic model, are used to read the design data and measurement formula contained in their records. The data table can be used from both sides or only from one side. The read data is parsed, and features such as pattern shape features and imaging conditions are extracted in the same way as in Embodiment 1 (Fig. 5). Each record is arranged in a space with a dimension equal to the number of extracted features, and clustering is performed on this space. As a specific clustering method, k-means method can be used, but it is not limited to this method. Any method that can cluster similar records into one category from the extracted feature quantity is acceptable. Each category obtained by the above is assigned an arbitrary feature group ID, and the assigned arbitrary feature group ID is recorded for each record. Then, a data table is created for each feature group ID.

[0084] So far, a method for creating feature groups resulting from clustering has been described. However, the method for creating feature groups is not limited to this. For example, it can be created by performing 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 data collection table 4049 and the data table 4045 for basic model learning, the device user assigns feature group IDs based on pattern shape features, SEM image features, camera conditions, and the device user's requirements for the machine learning model. The device user's requirements for the machine learning model can be conceived as, for example, "I want to prepare a machine learning model for each pattern type," "I want to distinguish machine learning models based on the presence or absence of penetration in the lower layers of the SEM image," "I want to prepare a machine learning model for each measurement device," etc. Subsequently, features such as pattern shape features and camera conditions are extracted from each record. Each record is arranged in a space with a dimension equal to the number of extracted features, and classification is performed on this space using the feature group IDs assigned by the device user as the ground truth. As a specific method for classification, common methods in classification tasks, such as support vector machines, can be considered, but this approach is not limited to it. Any method that can infer feature group IDs from features such as pattern shape characteristics and camera conditions is acceptable. The resulting feature group classification model can infer feature group IDs that reflect the user's requirements based on the input design data or measured formula information. The feature group IDs inferred by the feature group classification model are then recorded in each record. Subsequently, a data table is created for each feature group ID.

[0085] The records in the data tables of each feature group obtained by the two methods described above have similar features. Therefore, by using them for supplementary learning, a machine learning model specialized for a specific feature can be obtained. Ideally, the machine learning model for each feature group should use the machine learning model created by learning from the base model using data table 4045 as the foundation model for learning. However, this is not limited to this method; other methods can also be used, such as learning a new machine learning model without using the foundation model. The machine learning model learned for each feature group is given a unique name and saved along with the learning log.

[0086] For machine learning models that have been specialized for each feature group, the corresponding feature group or model version can be determined from learning logs, etc. This information can be used to select the machine learning model used in pattern matching caused by machine learning, or to select the model as the basis for learning. The method for determining the machine learning model is not limited to this method, as long as the most suitable machine learning model can be selected. For example, the most suitable machine learning model can also be selected based on the file name or the date and time of saving the machine learning model.

[0087] The pattern feature analysis unit of the recipe generation device 200 uses the category information of each feature group or the feature group classification model obtained by the above method to select the most suitable machine learning model for the input data. The specific procedure is as follows.

[0088] (1) Extract pattern shape features or imaging conditions from design data 101 and measurement formula 201. (2) Use the category information of each feature group or the feature group classification model to estimate the feature group ID of the features extracted in (1). (3) Record the machine learning model that has been learned using the data of the feature groups estimated in (2) as the machine learning model used in pattern matching in measurement formula 201.

[0089] <Embodiment 3> Figure 9 is a flowchart illustrating the operation of the pattern size measurement system described in Embodiment 3 of this disclosure. As explained in Embodiment 1, rule-based pattern matching is performed first, and only if this fails is machine learning-based pattern matching performed. Alternatively, the initially implemented algorithm can be either rule-based or machine learning-based. Specifically, between S201 and S202, S901 to S902 as described below are implemented. The rest is the same as in Embodiments 1 and 2.

[0090] (Figure 9: S901) The recipe generation device 200 analyzes the measured recipe 201, and determines which is more suitable for pattern matching processing: rule-based pattern matching or machine learning pattern matching, by referring to the matching results of similar measured recipe 201 information and comparing them with the matching results of similar measured recipe 201 information. A pattern matching with a successful history (or a higher success rate) is considered more suitable. If a measured recipe 201 with similar information does not exist, or if the past matching results do not exist, then rule-based pattern matching is selected. In other words, at the point of generation of the measured recipe 201, the matching algorithm corresponding to that measured recipe 201 is also determined. The specific operation of the recipe generation device 200 will be described later using Figure 10.

[0091] (Figure 9: S902) The measuring device 300 switches the pattern matching algorithm based on the result of S901. If the pattern matching based on rules has already been specified, S202 is executed. If the pattern matching algorithm based on machine learning has already been specified, the process is skipped to S204. Through this step, the most suitable pattern matching method can be selected before the pattern matching process is executed, thus reducing unnecessary processing and improving the matching success rate.

[0092] Figure 10 is a block diagram of the recipe generation device 200. The recipe generation device 200 analyzes the measured recipe 201 and selects the most suitable pattern matching method. The measured recipe 201 contains pattern type, imaging conditions, device number, etc. The pattern type includes: pattern shape information, pattern density information, semiconductor pattern constituent material, presence or absence of penetration of the underlying semiconductor pattern, and amount of deformation, etc. The imaging conditions include: the accelerating voltage of the electron microscope used for imaging, the type of detector, the imaging magnification, the number of image frames for image integration to reduce noise, and the presence or absence of pre-processing for high-quality images due to AI technology or image processing technology, etc. The device number includes the device type and the identification number of similar devices, etc.

[0093] The formula generation unit 210 generates the measured formula 201. The formula analysis unit 220 analyzes all or part of the information of the measured formula 201 and extracts its features. The formula analysis unit 220 refers to the matching result of the measured formula 201 with the feature most similar to the extracted features from the matching result memory unit 230, and selects the method with a higher probability of success in the pattern matching process. The matching result memory unit 230 is a database containing information such as the pattern type, imaging conditions, and device number of the measured formula 201, as well as keywords of the current matching result, and stores the information of each keyword in the form of text or numbers.

[0094] As a method for formula analysis, classification models such as random forests derived from machine learning can be used. In this learning process, the input is set to the data of the measured formula 201 in the matching result memory unit 230, and the teacher label is set to the matching result of the measured formula 201. In this way, the machine learning model can estimate the most suitable matching method for the measured formula. When the estimation is performed by the machine learning model, the input is set to the information of the measured formula 201, and the output is set to the most suitable pattern matching method. As another method, the database of the matching result memory unit 230 can be referenced to the column with the most items that are most consistent with the measured formula 201, and the matching result at that time can be referenced. By these methods, the most suitable pattern matching method can be selected from the results of the pattern matching process performed in the past. As a method for formula analysis, it is not limited to these methods, as long as the most suitable pattern matching method can be selected.

[0095] <Implementation Version 3: Summary> The pattern size measurement system described in Implementation Version 3 implements either pattern matching based on a rule-based algorithm or pattern matching based on machine learning. If rule-based matching fails, machine learning matching is implemented; if machine learning matching fails, relearning is implemented. The relearning of the learner only needs to be implemented when machine learning matching fails, and it can also be implemented automatically (see S206). Therefore, in Implementation Version 3, the workload required for learning can also be suppressed.

[0096] <Variations on this Disclosure> This disclosure is not limited to the embodiments described above and may include various variations. For example, the embodiments described above are detailed explanations provided to illustrate this disclosure in an easily understandable manner, and are not limited to possessing all the described components. Furthermore, a part of the components of a certain embodiment may be replaced by the components of other embodiments, and the components of other embodiments may be added to the components of a certain embodiment. Moreover, for a part of the components of each embodiment, other components may be added, deleted, or replaced.

[0097] In the above embodiments, the example of the measuring device 300 is a scanning electron microscope, but this disclosure is not limited to this. This disclosure can be applied to any measuring device or inspection device that acquires images and performs pattern matching.

[0098] In the above embodiments, the design data memory unit 100, the recipe generation device 200, the measurement device 300, and the machine learning model learning device 400 are illustrated as individual devices, but they can also be implemented as one or more devices (a pattern matching device composed of one or more computer systems) capable of performing the same processing. For example, a system can be constructed using a single device having the functions of the design data memory unit 100, the measurement device 300, and the machine learning model learning device 400, and the recipe generation device 200.

[0099] In Embodiment 2, the feature quantities used for implementing clustering may be, for example, as shown below. (a) Image feature quantities: brightness distribution of the image, frequency components of the image, (b) Shape and pattern feature quantities: type of shape and pattern, planar density of shape and pattern, (c) Image condition feature quantities: optical conditions and electrical conditions of the image acquisition device.

[0100] As explained in Implementation Example 3, after implementing S200 to S207, the measurement is determined to be correct (S209) in the same manner as in Implementation Example 1, and the successful result is used to implement learning (S210 to S214). By implementing S209, learning data can be automatically collected in the same way as in Implementation Example 1, but these steps are not necessarily required. That is, Implementation Example 3 is more suitable for implementing machine learning or rule-based learning, so as long as this can be achieved, the general purpose can be achieved. Of course, implementing S209 will be even more effective.

[0101] In the above embodiments, the functional units of each device (e.g., recipe generation unit 210, recipe analysis unit 220, rule-based pattern matching unit 302, machine learning pattern matching unit 303, matching success determination unit 304, measurement unit 305, measurement success 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 implement these functions, or by executing software that implements these functions using a computing device such as a CPU (Central Processing Unit).

[0102] In the above embodiments, the rule-based matching algorithm can be described as a specific construct formed by the internal logic of the algorithm, which specifically constructs the correspondence between image feature quantities and the positional translation quantities between two images corresponding to those image feature quantities. On the other hand, the machine learning matching algorithm, in addition to implementing the internal logic of matching processing, also has a learning model that describes the result of learning the correspondence between image feature quantities and positional translation quantities. This is different from the rule-based algorithm, and this point will be further explained.

[0103] In the above embodiments, it is explained that S201 is performed after S200. However, the order of each step is not limited to this. Changes may be made as long as they do not contradict the method disclosed herein. For example, S902 may be performed after S200, or S201 may be performed as shown in Figure 9. [Simplified Explanation of the Diagram]

[0012] [Figure 1] A structural diagram of the pattern size measuring system described in Embodiment 1. [Figure 2] A flowchart explaining the operation of the pattern size measuring system. [Figure 3] A flowchart explaining the specific operations of S206 to S207. [Figure 4] A flowchart explaining the specific operations of S214. [Figure 5] A flowchart explaining the specific operations of S213. [Figure 6] Examples of data held by temporary memory data 406 and learning data set 407. [Figure 7] Examples of data in the learning log. [Figure 8] A structural example of the measuring device 300 containing SEM600 and GUI307. [Figure 9] A flowchart explaining the operation of the pattern size measuring system described in Embodiment 3. [Figure 10] A block diagram of the recipe generation device 200.

Claims

1. A pattern matching apparatus for performing pattern matching between images, characterized in that it comprises a computer system for performing prior pattern matching; the prior computer system comprises a learner that outputs the result of prior pattern matching based on machine learning; the prior computer system comprises a first algorithm that does not use the prior learner and a second algorithm that uses the prior learner; the prior computer system performs prior pattern matching using the first prior algorithm, and if prior pattern matching using the first prior algorithm fails, it performs prior pattern matching using the second prior algorithm; if prior pattern matching using the second prior algorithm fails, the prior computer system generates learning data required for the prior learner to perform machine learning, and performs relearning of the prior learner using the prior learning data; the prior computer system performs prior pattern matching again using the prior learner that has undergone relearning.

2. The pattern matching device as described in claim 1, wherein, The pre-learning computer system is configured to supply information containing the positional translation amount between two images for which pre-learning pattern matching is to be performed to the pre-learning learner as pre-learning learning data; the pre-learning computer system accepts specified inputs for specifying the pre-learning positional translation amount through its interface, and generates pre-learning data based on the specified pre-learning inputs.

3. The pattern matching device as described in claim 1, wherein, The pre-learning computer system is configured to supply information containing the positional translation between two images for which pre-learning pattern matching is to be performed to a pre-learning learner as pre-learning learning data. The pre-learning computer system obtains the pre-learning positional translation from each pre-learning first algorithm by implementing a plurality of pre-learning first algorithms, each of which has a different matching procedure. The pre-learning computer system generates pre-learning data by selecting the pre-learning positional translation from each pre-learning first algorithm that yields the best match between the two pre-learning images.

4. The pattern matching device as described in claim 1, wherein, The first algorithm mentioned above is a rule-based algorithm that is configured to specify the correspondence between image feature quantities and the positional translation between two images corresponding to the image feature quantities mentioned above through the internal logic of the first algorithm mentioned above.

5. The pattern matching device as described in claim 1, wherein, The preceding computer system, when the preceding pattern matching is successful due to the preceding first algorithm or the preceding learner, uses the result as learning material for the preceding learner to perform relearning and supplies it to the preceding learner.

6. The pattern matching device as described in claim 5, wherein, The prior computer system evaluates the inference performance of the prior learner after prior relearning. If the prior inference performance does not meet the baseline value, the prior computer system deletes the portion of the learning materials used for prior relearning that is pre-defined as potentially reducing prior inference performance. The prior computer system then uses the learning materials with the previously deleted negatively impactful data types to relearn the prior.

7. The pattern matching device as described in claim 5, wherein, The prior computer system clusters the learning data used for prior relearning according to its features; the prior computer system corrects the configuration of data in each prior category to suppress bias in each category obtained by prior clustering; the prior computer system uses the prior learning data with the prior bias suppressed to perform prior relearning.

8. The pattern matching device as described in claim 1, wherein, The preceding image is an observation image of the sample; the preceding computer system, when the preceding pattern matching is successful, collects the results of the measurement or inspection performed on the preceding sample; if the preceding result indicates that the preceding measurement or inspection has failed, the preceding computer system, through its interface, collects the specified input to specify the positional translation amount between preceding images, and instructs the preceding learner to relearn based on the specified preceding input.

9. The pattern matching device as described in claim 1, wherein, The prior computer system comprises: a plurality of prior learners that have performed machine learning using mutually different learning data; the prior computer system groups the learning data used by each prior learner during machine learning according to the feature quantities of the prior learning data; the prior computer system uses prior learners that have performed machine learning using the prior learning data corresponding to at least one of the feature quantities of the prior image, the feature quantities of the shape and pattern contained in the prior image, or the feature quantities of the imaging conditions of the prior image, to perform prior pattern matching.

10. The pattern matching device as described in claim 9, wherein, The characteristic quantities of the preceding image include at least one of the following: the brightness distribution of the preceding image and the frequency component of the preceding image; the characteristic quantities of the preceding shape pattern include at least one of the following: the type of the preceding shape pattern and the planar density of the preceding shape pattern; the characteristic quantities of the preceding imaging conditions include at least one of the following: the optical conditions and electrical conditions of the imaging device that acquires the preceding image.

11. A pattern matching apparatus for performing pattern matching between images, characterized in that it comprises a computer system for performing prior pattern matching; the prior computer system comprises a learner that outputs the result of prior pattern matching based on machine learning; the prior computer system is configured to perform prior pattern matching by implementing any one of two or more algorithms that have performed prior pattern matching; the prior computer system selects any one of the prior algorithms based on the matching success rate when each prior algorithm has been implemented in the past; the prior computer system performs prior pattern matching by implementing the previously selected prior algorithm.

12. The pattern matching device as described in claim 11, wherein, The pre-learning computer system, which uses two or more pre-learning algorithms, includes: a first algorithm that does not use a pre-learning machine, and a second algorithm that uses a pre-learning machine; if the pre-learning pattern matching fails when using the first pre-learning algorithm, the pre-learning computer system uses the second pre-learning algorithm to perform pre-learning pattern matching; if the pre-learning pattern matching fails when using the second pre-learning algorithm, the pre-learning computer system performs relearning of the pre-learning machine.

13. A pattern matching apparatus for performing pattern matching between images, characterized in that it comprises a computer system for performing prior pattern matching; the prior computer system comprises a learner that outputs the result of prior pattern matching based on machine learning; the prior computer system is configured to perform prior pattern matching by implementing any one of two or more algorithms that have performed prior pattern matching; the prior computer system is configured to perform prior pattern matching according to the conditions described in the measurement recipe; the prior computer system extracts features of the new measurement recipe once a new measurement recipe is obtained; the prior computer system selects any one of the prior algorithms by identifying the feature most similar to the previously extracted feature from measurement recipes used in the past when each prior algorithm was successfully implemented; the prior computer system performs prior pattern matching by implementing the previously selected algorithm.

14. The pattern matching device as described in claim 13, wherein, The pre-learning computer system, which uses two or more pre-learning algorithms, includes: a first algorithm that does not use a pre-learning machine, and a second algorithm that uses a pre-learning machine; if the pre-learning pattern matching fails when using the first pre-learning algorithm, the pre-learning computer system uses the second pre-learning algorithm to perform pre-learning pattern matching; if the pre-learning pattern matching fails when using the second pre-learning algorithm, the pre-learning computer system performs relearning of the pre-learning machine.

Images