Training methods for computer systems, charged particle beam devices, and learning models

Abstract

The pre-trained model (313) used to improve the quality of images acquired by the charged particle beam device (1) is defined as a learning model trained using a pair of simulated low-quality image data and high-quality image data, i.e., training data. Here, the simulated low-quality image data is an image generated by making the image characteristics of each region contained in the high-quality image data approximate the image characteristics of the regions contained in the low-quality image data that are related to the pattern on the sample captured in the image. In this way, the estimation accuracy of the high-quality images produced by the pre-trained model can be improved, thereby improving the output of online inspection or measurement.

Description

[Technical Field]

[0002] This disclosure relates to computer systems, charged particle beam devices, and training methods for learning models. [Previous Technology]

[0004] To improve or maintain high yield in semiconductor mass production processes, various inspection or measurement devices are installed in the manufacturing line to inspect and measure semiconductor wafers during the manufacturing process online. Among such inspection or measurement devices, those that generate images by irradiating a sample with a beam of charged particles, and then perform inspection or measurement based on these images, are known. Examples include critical dimension scanning electron microscopy (CD-SEM) for process dimension management, and review SEM for re-inspecting or classifying various defects occurring on semiconductor wafers in the manufacturing line.

[0005] To inspect and measure semiconductor wafers, charged particle beam devices must generate high-quality images so that patterns on the sample can be clearly identified from the images. However, generating a high-quality image takes longer than the time required to obtain a single image. Furthermore, with the miniaturization of manufactured semiconductor devices or the increasing complexity of their structures, or the enlargement of semiconductor wafers, the number of inspection points on the semiconductor wafer also increases. As a result, the output of on-line inspection or measurement decreases, which has a significant impact on the manufacturing efficiency of the production line.

[0006] For such unsolved problems, some people have developed techniques such as those in Patent Documents 1 and 2, that is, to make the image generated by the charged particle beam device a low-quality image, and to use image processing, especially image processing using deep learning, to infer a high-quality image from the low-quality image, and to use the inferred high-quality image for inspection and measurement.

[0007] Patent document 1 discloses a technology that obtains image pairs of degraded and high-quality images by changing the shooting conditions at the same location on the sample, and learns the correspondence between the two by machine learning, thereby inferring the high-quality image from the degraded image.

[0008] Patent Document 2 also discloses a technique using a learner, which involves training a dataset of high-quality and low-quality images to infer high-quality images from low-quality images. (Prior Art Documents, Patent Documents)

[0009] Patent Document 1: Japanese Patent Application Publication No. 2018-137275; Patent Document 2: Japanese Patent Application Publication No. 2022-135215 [Summary of the Invention]

[0010] The problem that the invention is intended to solve

[0011] Obtaining a pair of high-quality and low-quality images suitable for training images using a charged particle beam device is practically difficult. For example, in the case of Patent Document 1, the low-quality image was obtained before the platform settled, so even if the images were taken at the same location, there would still be a positional deviation between the high-quality and low-quality images. Although alignment is performed to correct this positional deviation, precise alignment is difficult. Furthermore, due to the effects of charging or damage accompanying charged particle beam irradiation, differences will occur between the pattern shapes in the low-quality images. As an example, it is known that when the pattern on the sample is an organic material, i.e., a resist pattern, shrinkage will occur due to electron beam irradiation.

[0012] In Patent Document 2, in order to remove the effects of shrinkage or charging caused by electron beam irradiation, a low-quality image is selected, and a dataset of the selected low-quality image and a high-quality image is compiled as training data. The low-quality image is provided such that the relationship between at least one of the state of the high-quality image and at least one of the feature quantities extracted from the high-quality image satisfies a predetermined condition. This allows the low-quality image and the high-quality image to have similar features. However, even according to the technology disclosed in Patent Document 2, it is still difficult to obtain a pair of high-quality and low-quality images with identical features, i.e., eliminating the effects of shrinkage or charging.

[0013] When a dataset containing effects such as positional deviations, damage, or electrical charge is used as training data for a learning model, these effects are also learned. Ideally, high-quality images used for inspection or measurement should not affect the patterns and shapes captured in low-quality images, but only improve image quality. Therefore, the training data used to train the learning model should ideally be a pair of high-quality and low-quality images where the patterns and shapes are identical, differing only in image quality. Technical means to solve the problem.

[0014] One embodiment of the computer system disclosed herein is a computer system for training a learning model. The learning model is used to estimate image data of higher quality than the first image quality that can be inspected or measured from first-quality image data. The first-quality image data is obtained by a charged particle beam device that inspects or measures images generated by irradiating a sample with a predetermined pattern formed at a plurality of positions with charged particle beams under first photographic conditions. The system includes: a data memory unit; an image region recognition unit that divides the image data into a plurality of regions and performs region recognition, i.e., identifies the correspondence between each of the divided plurality of regions and the predetermined pattern; a training data generation unit that generates training data; and a model learning unit that trains the learning model using the training data generated by the training data generation unit. The data memory unit stores image characteristic data, which is obtained by performing region recognition on first image data obtained by the charged particle beam device at one of a plurality of locations using a first photographic condition to capture a predetermined pattern, and calculates the region characteristics based on each region contained in the first image data; and second image data, obtained by the charged particle beam device at one of a plurality of locations using an image quality capable of performing the aforementioned checks or measurements to capture a predetermined pattern; the image region recognition unit performs region recognition on the second image data; and the training data generation unit generates simulated low-quality image data, and generates a pair of simulated low-quality image data and second image data as training data, wherein the simulated low-quality image data is obtained by making the image characteristics of regions contained in the second image data approximate the image characteristics of regions contained in the first image data that are associated with the same predetermined pattern. Effects of the Invention

[0015] This disclosure enables the improvement of the estimation accuracy of high-quality images produced by pre-trained models, thereby increasing the output of online inspections or measurements. Other unresolved issues and novel features will become apparent from the description in this specification and the accompanying drawings.

Implementation Method

[0019] Figure 1 shows a schematic diagram of a measurement system using a scanning electron microscope (SEM) as an example of a charged particle beam device. Hereinafter, the charged particle beam device 1 will be described using a length-measuring SEM to measure the dimensions of patterns formed on a sample such as a semiconductor wafer. Furthermore, Figure 1 only shows a portion of the components and is not limited to including all the components shown. The positional relationship of the components is also shown as an example; configurations different from those in Figure 1 are also possible.

[0020] The charged particle beam device 1, as its main component, includes: an imaging unit (SEM body) 100; and a first computer system 200, which controls the imaging unit 100 to acquire an image of a sample (e.g., a semiconductor element) and measures the size of a pattern formed on the semiconductor element from the image. The first computer system 200 is connected to an input device 231 and a display device 232. Furthermore, the charged particle beam device 1 is connected to a second computer system 300 via a network 330. The second computer system 300 trains a learning model used by the charged particle beam device 1 in size measurement and provides the pre-trained model to the charged particle beam device 1. The second computer system 300 is also connected to an input device 331 and a display device 332.

[0021] In the imaging unit 100, a primary electron beam 101 emitted from an electron source 111 controlled by an electron source control unit 121 is accelerated to a desired acceleration voltage value by an acceleration electrode 112 controlled by an acceleration electrode control unit 122, and then focused by an object lens 114 controlled by an object lens control unit 124, irradiating the sample 104. Due to the irradiation by the primary electron beam 101, signal electrons, namely secondary electrons (SE) 102 and backscattered electrons (BSE) 103, are emitted from the sample 104. SE 102 is detected by an SE detector 116, and an SE detection signal processing unit 126 outputs an output signal based on the SE detection intensity. BSE 103 is detected by a BSE detector 117, and a BSE detection signal processing unit 127 outputs an output signal based on the BSE detection intensity.

[0022] The sweep deflector 113, controlled by the sweep control unit 123, can deflect the primary electron beam 101 so that it sweeps across the surface of the sample 104. Although omitted in Figure 1, an image translation deflector that moves the irradiation position of the primary electron beam 101 may also be included. The sample 104 is held by the platform 115, and the platform 115 is moved by the platform control unit 125, thereby enabling observation of desired areas on the surface of the sample 104. Furthermore, the control units that control the various elements included in the imaging unit 100 as described above are sometimes collectively referred to as element control units 12x.

[0023] The image generator 131 synchronizes the output signal from the SE detection signal processing unit 126 and the control signal from the sweep control unit 123 and stores them in the image frame memory, thereby generating SE image data. Similarly, it synchronizes the output signal from the BSE detection signal processing unit 127 and the control signal from the sweep control unit 123 and stores them in the image frame memory, thereby generating BSE image data. Sometimes, SE image data and BSE image data are collectively referred to as scan image data. When the output signal is stored in the image frame memory, the detection signal is stored in the image frame memory at the position corresponding to the scan position, thereby generating signal profile data (1D information) and scan image data (2D information).

[0024] The first computer system 200 controls the generation of scanned image data by the imaging unit 100, or performs image processing on the scanned image data generated by the image generator 131 to generate SEM image data, and performs dimensional measurements based on the SEM image data. Details will be described later.

[0025] The first computer system 200 or the second computer system 300, as its main components, includes a processor (Central Processing Unit: CPU) 401, memory 402, storage device 403, input interface (I / F) 404, output I / F 405, communication I / F 406, and bus 407, as shown in FIG2. The processor 401 performs processing according to the program loaded into the memory 402, thereby becoming a functional unit that provides the prescribed functions. The storage device 403 stores the data or programs used in the functional unit. The input I / F 404 is connected to input devices such as keyboards, pointing elements, and operation panels, and the output I / F 405 is connected to a display device. The communication device I / F 406 can communicate with other computer systems via a network. They are communicatively connected to each other via bus 407.

[0026] In the following description, when describing the processing performed by a program, the description may sometimes focus on the program or functional unit, but the hardware component that constitutes the main body of these components is the processor or the computer system comprising the processor. A computer system executes processing according to a program read from memory by utilizing resources such as memory or communication interfaces appropriately through a processor. While Figure 2 illustrates an example of a CPU as a processor, a GPU (Graphical Processing Unit) or similar device can also be used. Furthermore, the processing used to implement functions is not limited to software program processing; it can also be implemented in dedicated circuits. Dedicated circuits can include FPGAs (Field Programmable Gate Arrays), ASICs (Application Specific Integrated Circuits), etc.

[0027] (Measurement Flow) Figure 3 illustrates the measurement flow in the measurement system. First, in the charged particle beam device 1, high-quality and low-quality images are acquired (S01). Next, in the second computer system 300, training data is generated using the acquired high-quality and low-quality images (S02), and the generated training data is used to train the learning model (S03). Subsequently, in the charged particle beam device 1, the first computer system 200 receives the pre-trained model from the second computer system 300 and uses the pre-trained model to perform measurement on the sample (e.g., a semiconductor wafer) (S04). Steps S01 to S03 are also referred to as the preparation stage, and step S04 is also referred to as the application stage.

[0028] Figure 4 shows a functional block diagram of the first computer system 200 of this embodiment. The first computer system 200 includes a functional unit, namely a photography condition setting unit 201, an image processing unit 202, a high-quality image estimation unit 203, a measurement unit 204, and a data memory unit 210 for storing data used in these functional units. The photography condition setting unit 201 controls the various elements of the shooting unit 100 through the element control unit 12x (see Figure 1) according to the photography conditions set by the user through the input device 231, thereby generating scanned image data. The image processing unit 202 performs image processing on the scanned image data output from the image generator 131. The image data generated by the image processing unit 202 is called SEM image data.

[0029] During the preparation stage, the image processing unit 202 designates the SEM image data obtained under high-quality image photography conditions as high-quality image data 211, and the SEM image data obtained under low-quality image photography conditions as low-quality image data 212, and stores them in the data memory unit 210. On the other hand, during the application stage, the image processing unit 202 outputs the SEM image data obtained under measured photography conditions to the high-quality image estimation unit 203.

[0030] In the data memory unit 210, a pre-trained model 313 is pre-registered for estimating high-quality images from SEM image data obtained by measuring photographic conditions. As described later, the pre-trained model 313 is a learning model already trained in the second computer system 300. The high-quality image estimation unit 203 uses the pre-trained model 313 to estimate high-quality images from the SEM image data output from the image processing unit 202. The measurement unit 204 uses the estimated high-quality image to measure the size of the pattern and accumulates it as measurement data 213. The following describes each step.

[0031] (Preparation Stage S01) In step S01, high-quality and low-quality images for generating training data are obtained using the charged particle beam device 1. Here, the image quality of the high-quality and low-quality images is determined by the user. Specifically, the image quality of the high-quality image can be set to a quality suitable for dimensional measurement by the measurement unit 204. This is because this embodiment is based on a charged particle beam device for measuring the size of a pattern. If it is a charged particle beam device for defect re-inspection or classification (re-inspection SEM), it is sufficient to set the image quality to a level that can achieve that purpose. On the other hand, the photographic conditions for the low-quality image are ideally the same as the measurement photographic conditions set in the photographic condition setting unit 201 during the setting and application stage. This is based on the following reasons.

[0032] In order to improve the output of the charged particle beam device 1 during the application phase, it is necessary to shorten the time required to acquire one SEM image. The time required to acquire one SEM image mainly consists of the following: the time for the platform 115 to move the irradiation position of the primary electron beam 101 to the observation position on the sample (platform movement time); the time for the primary electron beam 101 to be focused on the sample surface by the objective lens 114 (autofocus time); and the time for acquiring a cumulative plurality of scanned image data (photographing time). Even though they are all called low-quality images, they will exhibit different characteristics depending on the time reduction method. For example, if the platform movement time is shortened and the SEM image is acquired before the sample is completely settled, there is a concern that the SEM image may be distorted. If the magnification is reduced and the depth of focus is shallowed in order to shorten the autofocus time, there is a concern that the resolution of the SEM image may be reduced. Similarly, if a certain degree of focus deviation is allowed in order to shorten the autofocus time, there is a concern that the SEM image may be out of focus. Furthermore, in order to improve the quality of the images, it is generally necessary to perform one electron beam 101 scan on the sample and accumulate multiple scan images. However, if you want to reduce the number of scans and shorten the imaging time, the noise in the SEM images will increase.

[0033] As such, depending on the time reduction method, low-quality images may exhibit different degradation patterns. Therefore, if the photographic conditions of the low-quality image are set to the measurement photographic conditions set in the photographic condition setting unit 201 during the application phase, a learning model can be trained to improve the image quality based on the degradation patterns contained in the SEM images obtained during the application phase. Therefore, the data memory unit 210 pre-memorizes the photographic conditions of the low-quality image and the low-quality image. The photographic conditions include, in addition to, acceleration voltage, probe current, magnification, and number of pixels, image processing algorithms such as noise removal and contrast adjustment, as well as the order of these processing steps.

[0034] The high-quality and low-quality images obtained in step S01 are images obtained by acquiring the pattern designated as the measurement object from more than one point each. Figure 5 schematically illustrates a semiconductor wafer 500. The semiconductor wafer 500 includes a plurality of wafer regions 501. Semiconductor layers, insulating layers, or conductor layers of a predetermined pattern are gradually deposited in each of the plurality of wafer regions 501, and finally separated by dicing, thereby obtaining a plurality of semiconductor wafers. Here, it is assumed that the pattern designated as the measurement object is a line and spacing (L / S) pattern formed in a certain process. For example, it is assumed that an L / S pattern of resist 511 is formed on a conductor layer 512 as a substrate, and the width W of the resist 511 is measured. In this case, the semiconductor wafer 500 contains L / S patterns formed by the same process as the measurement object in the same wafer region or in other wafer regions, so observation regions 502 and 503 containing L / S patterns formed by the same process are set to obtain images. As will be described later, in this embodiment, the number of high-quality images will be the number of training data, so ideally, a large number of images should be obtained. On the other hand, for low-quality images, only at least one image needs to be obtained.

[0035] Furthermore, the pattern designated as the measurement object is not limited to L / S patterns, but may also be a pattern with a specific shape or a via. In addition, if it is a pattern that can be regarded as the same as the pattern designated as the measurement object, it may also be a pattern formed by a different process. For example, although Figure 5 uses an L / S pattern extending in the Y direction as an example, in such a case, an L / S pattern extending in the X direction is mostly formed on top of it. When it is determined that there is no problem in using the image of the upper L / S pattern and the image of the lower L / S pattern as training data, it may also include patterns formed by a different process to obtain high-quality images.

[0036] (Preparation Stage S02) Before explaining the processing of this step, the functions of the second computer system 300 in this embodiment will be explained. Figure 6 shows a functional block diagram of the second computer system 300. The second computer system 300 includes a functional unit, namely an image region recognition unit 301, an image feature calculation unit 302, a training data generation unit 303, a model learning unit 304, and a data memory unit 310 for storing the data used by these functional units. The image region recognition unit 301 segments the high-quality image and low-quality image obtained in step S01 into regions, and performs region recognition, that is, identifies which part of the pattern designated as the measurement object belongs to each region. The image feature calculation unit 302 calculates the image features of each region contained in the low-quality image. The training data generation unit 303 generates training data using the high-quality image data and the image feature data of each region calculated for the low-quality image. In the model learning unit 304, the generated training data is used to train the learning model.

[0037] In the data memory unit 310, the high-quality image data 211 and low-quality image data 212 obtained in the preparation stage S01 are stored, as well as the image feature data 311 calculated by the image feature calculation unit 302, the training data 312 generated by the training data generation unit 303, and the pre-trained model 313 that has been trained by the model learning unit 304. Furthermore, the high-quality image data 211 and low-quality image data 212 can also be transferred from the first computer system 200 to the second computer system 300, or these data can be stored on a data server accessible to both the first computer system 200 and the second computer system 300.

[0038] Figure 7 shows a flowchart of the details of step S02.

[0039] (Step S11) The image region recognition unit 301 segments the high-quality image and low-quality image obtained in step S01 into regions, and identifies the correspondence between each segmented region and the pattern designated as the measurement object (region recognition). This region recognition can be performed using any known method. For example, it can be performed using human methods, image processing methods such as k-means, image recognition AI, or a combination thereof.

[0040] Regarding the processing in this step, the necessary scope of region identification differs between high-quality and low-quality images. In the case of a high-quality image, the entire image must be segmented in such a way that it is contained within a region. In contrast, in the case of a low-quality image, it is sufficient to identify regions by associating at least a portion of the image with a defined pattern. Figure 8 illustrates the results of region identification (region identification images) for both high-quality and low-quality images when the pattern to be measured is an L / S pattern. Region identification image 601 of a high-quality image divides the entire image into two types of regions: line regions 602 and interval regions 603. In contrast, region identification image 611 of a low-quality image identifies a portion of the region contained in the line pattern as a line region 612 and a portion of the region contained in the interval pattern as an interval region 613.

[0041] High-quality images are acquired at a scale of tens to hundreds of frames and possess vivid image quality. Therefore, it is ideal to automatically perform region segmentation using image processing or AI. The correspondence between the segmented regions and the pattern designated as the measurement object can be determined by characteristic quantities such as brightness within the segmented regions, for example, identifying higher-brightness regions as line regions and lower-brightness regions as interval regions. Alternatively, it can be designed to compare with the coordinate design information (layout diagram) of the acquired high-quality images to determine whether the segmented regions belong to line regions or interval regions.

[0042] On the other hand, low-quality images only require one or a few images, resulting in a lack of clarity, and can therefore be identified manually. In this case, areas that are clearly line patterns can be identified as line areas, and areas that are clearly spaced patterns can be identified as spaced areas. Of course, the same method used for high-quality images can also be used for region segmentation and identification.

[0043] (Step S12) The image region recognition unit 301 associates the patterns identified between the high-quality image and the low-quality image as the same region. In the example of FIG8, in step S11, line regions and interval regions are identified for each of the region recognition images of the high-quality image and the low-quality image. In this case, it can be said that by identifying the region 602 of the high-quality image and the region 612 of the low-quality image as the same line region, and identifying the region 603 of the high-quality image and the region 613 of the low-quality image as the same interval region, steps S11 and S12 are performed simultaneously.

[0044] On the other hand, when it is difficult to associate regions with each other using only region recognition processing, or when it is necessary to confirm the region recognition results, a GUI for associating regions can be designed to be displayed on the display device 332. Figure 9 illustrates an example of a display screen for region association. The region association screen 700 includes region recognition image display units 701 and 702 for each of high-quality and low-quality images, association setting units 703 and 704 for each region containing high-quality images, and an association confirmation button 705.

[0045] The region recognition image display units 701 and 702 respectively display region recognition images of the high-quality image and the low-quality image, which are associated with the regions. Alternatively, instead of region recognition images, the high-quality image or the low-quality image itself can be displayed in a way that clearly shows the region recognition results. Regarding the association setting units 703 and 704, the high-quality image in this example has two regions, namely region A (line region) and region B (interval region). Therefore, for each region, the correspondence with the two regions of the low-quality image, namely region a and region b, can be selected. The result of region recognition in step S11, when the correspondence has been identified, displays the identified correspondence with a halftone background. The user confirms the correspondence, and if the association is judged to be incorrect, the user reselects the associated region (the halftone region) in the association setting units 703 and 704 to correct the association. By pressing the association confirmation button 705, the user can confirm the association between the regions.

[0046] (Step S13) The image feature calculation unit 302 calculates the image features of each region for the low-quality image obtained in step S11. Figure 10 illustrates the data structure of the image feature data 311 obtained in this step. ID 801 is an ID that uniquely identifies the low-quality image. Photography condition data 802 records the photography conditions when the low-quality image was obtained. Image feature data 803 records the image features of each region. For example, in the case of the low-quality image shown in Figure 9, image feature data 803 is recorded for regions a and b.

[0047] The image characteristics recorded as image characteristic data 803 are arbitrary. Examples here include noise characteristics, brightness characteristics, shape characteristics, and defocus degree, but it is not necessary to record all of them. In addition, image characteristics other than those listed in the examples can also be recorded. Noise characteristics are defined as the variation in brightness in the area, brightness characteristics are defined as the brightness distribution in the area, shape characteristics are defined as the unevenness of the edges of the area, and defocus degree is defined as the sharpness of the area, which can be quantitatively determined.

[0048] Alternatively, when there are multiple low-quality images obtained under the same photographic conditions, it is also possible to design the average or median of the image characteristic values ​​of these multiple low-quality images as the representative value of the low-quality image.

[0049] (Step S14) The training data generation unit 303 uses the image characteristic data 803 calculated in step S13 to generate a pseudo-low-quality image from the high-quality image obtained in step S11. A pseudo-low-quality image refers to an image whose image characteristics of each region of the high-quality image are approximated to the image characteristics of a region of the associated low-quality image. For example, in the case of Figure 9, the image characteristics of region A of the high-quality image are transformed to approximate the image characteristics of region a of the low-quality image, and the image characteristics of region B of the high-quality image are transformed to approximate the image characteristics of region b of the low-quality image, thereby generating a pseudo-low-quality image.

[0050] (Steps S15-S17) The training data generation unit 303 determines whether the simulated low-quality image generated in step S14 is actually similar to the low-quality image whose characteristics are approximated. In addition to the user's visual perception, a similarity index is also calculated to make a quantitative judgment. As a similarity index, for example, indices such as PSNR (Peak Signal to Noise Ratio), SSIM (Structural Similarity), and MSE (Mean Squared Error) can be used. At this time, both the similarity index between corresponding regions and the similarity index of the entire image are calculated. When it is determined that the simulated low-quality image is similar to the low-quality image (Yes in S15), the generated simulated low-quality image and its source high-quality image are registered as training data 312 (S17).

[0051] In contrast, when it is determined that the simulated low-quality image is not similar to the low-quality image (No in S15), the parameters of the simulated low-quality image are adjusted to adjust the image quality of the simulated low-quality image (S16). If a simulated low-quality image similar to the low-quality image can be obtained, the simulated low-quality image with adjusted parameters and the high-quality image are registered as training data 312 (S17).

[0052] Figure 11 shows an example of a GUI (Simulated Low-Quality Image Adjustment Screen 900) for adjusting the image quality of a simulated low-quality image displayed in the display device 332. The image display units 901 and 902 display a simulated low-quality image and a low-quality image, respectively, allowing the user to visually compare the two. The similarity index display unit 903 displays the calculated similarity index for each region and the similarity index for the entire image.

[0053] When the user determines that the similarity between the simulated low-quality image and the low-quality image is insufficient, the parameters of the simulated low-quality image are adjusted. In this example, an adjustment selection unit 905 is provided so that the user can easily adjust the parameters in the desired direction. The user sets the adjustment direction in the adjustment selection unit 905 and presses the parameter adjustment button 906 to adjust the image quality of the simulated low-quality image. The simulated low-quality image after parameter adjustment is displayed on the image display unit 901. Then, the user presses the similarity calculation button 904, which calculates the similarity index between the simulated low-quality image after parameter adjustment and the low-quality image, and the value of the similarity index display unit 903 is updated. If the user determines that the simulated low-quality image after parameter adjustment is suitable as training data, the user presses the registration button 907, thereby registering the simulated low-quality image after parameter adjustment as training data.

[0054] Perform the above processing on the high-quality image obtained in step S11 to obtain training data 312 (S02).

[0055] (Preparation Stage S03) Figure 12 shows a functional block diagram of the model learning unit 304. The learning model 1000 is, for example, a neural network for deep learning, including an input layer 1011, an output layer 1012, and a plurality of intermediate layers 1013. In the model learning unit 304, parameters such as weights or biases in the intermediate layers 1013 are adjusted so that the image output from the output layer 1012 (inference image) is sufficiently close to the desired image relative to the image input to the input layer 1011.

[0056] Figure 12 illustrates the training of the learning model 1000, i.e., parameter adjustment, using a pair of simulated low-quality images 1021 and high-quality images 1022 contained in the training data 312. Specifically, the simulated low-quality image 1021 is input to the learning model 1000 to obtain the inference image 1023. The error calculator 1001 compares the inference image 1023 with the high-quality image 1022 to calculate the error, and the parameter adjuster 1002 adjusts the parameters of the learning model 1000 to reduce the error. If the error is small enough, the training of the learning model 1000 ends, and the parameters of the learning model 1000 are registered as the pre-trained model 313.

[0057] In this embodiment, a pair of simulated low-quality images and high-quality images is used as training data. The simulated low-quality image is generated based on the high-quality image, so there is no change between the region, i.e., the pattern captured in the high-quality image, and the pattern captured in the simulated low-quality image. Therefore, unlike conventional techniques, it does not learn the effects of positional deviations, damage, or electrical charge, and a pre-trained learning model that improves image quality can be obtained.

[0058] (Application stage S04) The pre-trained model 313 obtained in step S03 is transferred to the first computer system 200 and logged into the data memory unit 210.

[0059] The high-quality image estimation unit 203, once inputting SEM image data obtained from measurement and photography conditions (application photography conditions), calls up the pre-trained model 313 and estimates a high-quality image from the SEM image data. When the data memory unit 210 stores a plurality of learning models as pre-trained models 313, it compares the photography conditions of the low-quality images used in generating the simulated low-quality images constituting the training data with the application photography conditions, and selects a pre-trained model whose low-quality image photography conditions and application photography conditions are the same. The measurement unit 204 performs size measurement on the estimated high-quality image.

[0060] The following describes a variation of this embodiment.

[0061] (Modification 1) In the embodiment, an example of generating training data by capturing high-quality and low-quality images using the charged particle beam device 1 is described. In Modification 1, using existing calculated image characteristic data 311, training data is generated by capturing only high-quality images using the charged particle beam device 1 without prior acquisition of low-quality images. The method for generating training data in this case (preparation stage S02) will be described focusing on the differences from the embodiment.

[0062] Figure 13 shows an example of region recognition image 1300 of a high-quality image obtained in Modification 1. For such a high-quality image, a pseudo-low-quality image is generated using image characteristic data 311 of the low-quality image described in the embodiment.

[0063] Regarding region identification image 1300, assuming region 1301 is an image of the resist and region 1302 is an image of the conductor layer, the user associates region 1301 with region a of the low-quality image and region 1302 with region b of the low-quality image (step S12, see Figure 7). This process is performed as in the embodiment, thereby enabling the generation of a simulated low-quality image of a high-quality image using image feature data 311.

[0064] In this embodiment, a simulated low-quality image is evaluated, therefore the similarity between the simulated low-quality image and the low-quality image is evaluated. In contrast, in Modification 1, the simulated low-quality image and the low-quality image are dissimilar images, therefore, the similarity index is used to evaluate the similarity between the image characteristics of the associated regions. Specifically, for the obtained simulated low-quality image, the image characteristics of region 1301 and region 1302 are calculated as in the case of the low-quality image, and the similarity between the image characteristics of region 1301 and the image characteristics of region a, and the similarity between the image characteristics of region 1302 and the image characteristics of region b are calculated and evaluated. As a similarity index in this case, the Euclidean distance or Manhattan distance between the image characteristics of the regions can be used.

[0065] Furthermore, although the modified example 1 was explained using the case where the pattern to be measured is different, the same applies when the pattern to be measured is the same pattern formed by the same project.

[0066] (Modification 2) In this embodiment, in principle, a pre-trained model is obtained by acquiring multiple high-quality images of a pattern formed by the same process as the measurement object. In this case, a pre-trained model is created for each pattern as the measurement object. When there are multiple patterns as measurement objects for a single semiconductor wafer, multiple pre-trained models are necessary. In this case, the principle can be relaxed, for example, training data generated from high-quality images of patterns such as those shown in Figure 13 can be mixed to train the pre-trained model.

[0067] In order to avoid the degradation of high-quality images inferred by pre-trained models, it is necessary to limit the mixed patterns to patterns that are similar to a certain extent, which makes the management of pre-trained models easier.

[0068] This disclosure is not limited to the embodiments described above, but includes various modifications. For example, the embodiments described above are explained in detail for the purpose of simply illustrating this disclosure, and are not limited to having all the described configurations. Furthermore, a part of the configuration of a certain embodiment or modification can be replaced with the configuration of other embodiments or modifications, and the configuration of other embodiments or modifications can also be added to the configuration of a certain embodiment or modification. In addition, for a part of the configuration of each embodiment or modification, other configurations can be added, deleted, or replaced. [Simplified Explanation of the Diagram]

[0017] [Figure 1] Schematic diagram of the measurement system. [Figure 2] Example of the hardware configuration of the computer system. [Figure 3] Measurement process in the measurement system. [Figure 4] Functional block diagram of the first computer system. [Figure 5] Model diagram of a semiconductor wafer. [Figure 6] Functional block diagram of the second computer system. [Figure 7] Flowchart illustrating the details of step S02 in the measurement process. [Figure 8] Region recognition images of high-quality and low-quality images. [Figure 9] Example of a display screen for region association. [Figure 10] Diagram illustrating the data structure of image characteristic data. [Figure 11] Example of a display screen for quality adjustment of simulated low-quality images. [Figure 12] Functional block diagram of the model learning unit. [Figure 13] Example of a region recognition image of a high-quality image obtained in Modified Example 1.

Claims

1. A computer system for training a learning model, wherein the learning model is derived from first-quality image data to determine higher-quality image data that is inspectable or measurable, the first-quality image data being obtained by a charged particle beam device for inspecting or measuring an image generated by irradiating a sample with a predetermined pattern formed at a plurality of positions with a charged particle beam under first photographic conditions, the system comprising: a data memory unit; an image region recognition unit for dividing the image data into a plurality of regions and performing region recognition, i.e., identifying the correspondence between each of the divided plurality of regions and the predetermined pattern; a training data generation unit for generating training data; and a model learning unit for training the learning model using the training data generated by the training data generation unit. The aforementioned data memory unit stores: image characteristic data, which is image characteristics obtained by the aforementioned charged particle beam device for region recognition of a first image data obtained at one of the aforementioned plurality of locations by photographing the aforementioned predetermined pattern under the aforementioned first photographic conditions, and is calculated based on each region contained in the image of the aforementioned first image data; and second image data obtained by the aforementioned charged particle beam device for photographing the aforementioned predetermined pattern at one of the aforementioned plurality of locations using high-quality images capable of performing the aforementioned inspection or measurement; the aforementioned image region recognition unit performs region recognition on the aforementioned second image data; the aforementioned training data generation unit generates simulated low-quality image data, and generates a pair of the aforementioned simulated low-quality image data and the aforementioned second image data as the aforementioned training data, wherein the simulated low-quality image data is obtained by making the image characteristics of the regions contained in the image of the aforementioned second image data approximate the image characteristics of the regions contained in the image of the aforementioned first image data that are associated with the aforementioned predetermined pattern.

2. The computer system as described in claim 1, wherein, It further includes: an image feature calculation unit that calculates image features based on each region contained in the image data; a data memory unit that stores the first image data; an image region recognition unit that performs region recognition on the first image data; and an image feature calculation unit that calculates image features based on each region contained in the first image data and stores them in the data memory unit as the image feature data.

3. The computer system as described in claim 2, wherein, The aforementioned image region recognition unit divides the entire second image data into regions, and identifies regions of the first image data by associating at least a portion of the image with the aforementioned defined pattern.

4. The computer system as described in claim 2, wherein, The aforementioned image region recognition unit displays a region association screen on the display device. This region association screen selects and associates with one of the regions contained in the image of the aforementioned second image data, based on the regions contained in the image of the aforementioned first image data.

5. The computer system as described in claim 4, wherein, The aforementioned image region recognition unit, for each region contained in the image of the aforementioned second image data, selects a region in the image of the aforementioned first image data that is associated with the aforementioned predetermined pattern, and displays the aforementioned region association screen on the aforementioned display device.

6. The computer system as described in claim 2, wherein, The aforementioned training data generation unit calculates the similarity index between the aforementioned first image data and the aforementioned simulated low-quality image data.

7. The computer system as described in claim 6, wherein, The aforementioned training data generation unit calculates the aforementioned similarity index based on the image characteristics of the regions contained in the image of the aforementioned second image data that are associated with the aforementioned pattern and the image characteristics of the regions contained in the image of the aforementioned first image data.

8. The computer system as described in claim 1, wherein, The aforementioned pattern is a pattern formed on the aforementioned sample through the same process.

9. The computer system as described in claim 1, wherein, The aforementioned first photographic condition refers to the time required for the aforementioned charged particle beam device to acquire image data under the aforementioned first photographic condition being set to be shorter than the time required to acquire a single high-quality image data that can be used for the aforementioned inspection or measurement.

10. A charged particle beam device comprising: an imaging unit for irradiating a sample having a predetermined pattern formed at a plurality of positions with a charged particle beam to detect signal electrons emitted from the sample; an image generator for generating scanned image data based on scanned position information of the charged particle beam and detection information of the signal electrons; and a computer system for inspecting or measuring images based on the image data generated from the scanned image data; wherein the charged particle beam device, The aforementioned computer system includes: an image processing unit that acquires image data using photographic conditions; and a high-quality image estimation unit that uses a pre-trained model to estimate higher-quality, high-resolution image data suitable for the aforementioned checks or measurements from the image data acquired using the aforementioned photographic conditions. The aforementioned pre-trained model is a learning model trained using a pair of simulated low-quality image data and a second image data, i.e., training data. The simulated low-quality image data is generated using a first image data obtained by photographing the aforementioned predetermined pattern at one of the aforementioned plurality of locations using the first photographic conditions, and the aforementioned second image data obtained by photographing the aforementioned predetermined pattern at one of the aforementioned plurality of locations using high-resolution images suitable for the aforementioned checks or measurements. The aforementioned first image data and the aforementioned second image data are each divided into a plurality of regions, and the correspondence between each of the aforementioned plurality of regions and the aforementioned prescribed pattern is identified. The aforementioned simulated low-quality image data is generated by making the image characteristics of the regions contained in the image of the aforementioned second image data approximately the same as the image characteristics of the regions contained in the image of the aforementioned first image data that are associated with the aforementioned prescribed pattern. The aforementioned computer system uses the image data estimated by the aforementioned high-quality image estimation unit to perform the aforementioned inspection or measurement.

11. The charged particle beam device as described in claim 10, wherein, The aforementioned computer system stores a plurality of the aforementioned pre-trained models. The aforementioned high-quality image estimation unit selects a pre-trained model with the same first photographic condition and the aforementioned applied photographic condition. The first photographic condition is the aforementioned first photographic condition of the aforementioned first image data used in the generation of the aforementioned simulated low-quality image data of the aforementioned training data used in the training of the aforementioned pre-trained model.

12. The charged particle beam device as described in claim 10, wherein, The aforementioned simulated low-quality image data is used to determine whether it is suitable as training data based on the similarity between the images in the simulated low-quality image data and the images in the first image data.

13. The charged particle beam device as described in claim 10, wherein, The aforementioned pattern is a pattern formed on the aforementioned sample through the same process.

14. The charged particle beam device as described in claim 10, wherein, The aforementioned first photographic condition refers to the time required for the aforementioned charged particle beam device to acquire image data under the aforementioned first photographic condition being set to be shorter than the time required to acquire a single high-quality image data that can be used for the aforementioned inspection or measurement.

15. A training method for a learning model, which is a training method for a learning model using a computer system, wherein the learning model estimates high-quality image data that is of higher quality than the aforementioned first-quality image data and is capable of being inspected or measured, the first-quality image data being obtained by a charged particle beam device for the aforementioned inspection or measurement based on an image generated by irradiating a sample with a predetermined pattern formed at a plurality of positions with a charged particle beam under first photographic conditions, wherein... The aforementioned computer system includes: a data storage unit; an image region recognition unit that segments image data into a plurality of regions and performs region recognition, i.e., identifies the correspondence between each of the segmented plurality of regions and the aforementioned defined pattern; a training data generation unit that generates training data; and a model learning unit that trains the aforementioned learning model; the aforementioned data storage unit stores: image characteristic data, which records image characteristics obtained by the aforementioned charged particle beam device for region recognition of a first image data obtained at one of the aforementioned plurality of locations using the aforementioned first photographic conditions, and calculates the characteristics based on each region contained in the image of the aforementioned first image data; and second image data obtained by the aforementioned charged particle beam device for high-definition photographing of the aforementioned defined pattern at one of the aforementioned plurality of locations, which allows for the aforementioned inspection or measurement; the aforementioned image region recognition unit performs region recognition on the aforementioned second image data. The aforementioned training data generation unit generates simulated low-quality image data, and generates a pair of the aforementioned simulated low-quality image data and the aforementioned second image data as the aforementioned training data. The simulated low-quality image data is obtained by making the image characteristics of the region contained in the image of the aforementioned second image data approximate the image characteristics of the region contained in the image of the aforementioned first image data that are associated with the aforementioned defined pattern. The aforementioned model learning unit uses the aforementioned training data to train the aforementioned learning model.

16. A training method for the learning model as described in claim 15, wherein, The aforementioned computer system further includes: an image feature calculation unit that calculates image features based on each region contained in the image data; a data memory unit that stores the aforementioned first image data; an image region recognition unit that performs region recognition on the aforementioned first image data; and the image feature calculation unit that calculates image features based on each region contained in the image data of the aforementioned first image data and stores them in the aforementioned data memory unit as the aforementioned image feature data.

17. A training method for the learning model as described in claim 15, wherein, The aforementioned pattern is a pattern formed on the aforementioned sample through the same process.

Images