Computer system, charged particle beam device, and training method for learning model

Abstract

A trained model 313 for improving the quality of images acquired by a charged particle beam device 1 is a learning model trained using teacher data comprising pairs of pseudo low-quality image data and high-quality image data. Here, an image of each item of pseudo low-quality image data is an image generated by approximating the image characteristics of each region included in an image of the high-quality image data to the image characteristics of a region, included in an image of low-quality image data, having the same association with a pattern on a sample captured in the image. As a result, the estimation accuracy of high-quality images using the trained model can be improved, and the throughput of in-line inspection or measurement can be improved.

Description

Computer system, charged particle beam device and method for training learning model

[0001] The present disclosure relates to a computer system, a charged particle beam device, and a method for training a learning model.

[0002] To improve or maintain high yields in semiconductor mass production processes, various inspection and measurement devices are installed on the production line to inspect and measure semiconductor wafers in-line during the manufacturing process. Known inspection and measurement devices use charged particle beams to generate images of samples and then perform inspection and measurement based on the generated images. Examples include critical dimension scanning electron microscopes (CD-SEMs), which perform dimensional control during the process, and review SEMs, which review and classify various defects that occur on semiconductor wafers on the production line.

[0003] Inspection and measurement of semiconductor wafers requires that charged particle beam systems generate high-quality images so that patterns on the sample can be clearly identified. However, generating high-quality images increases the time required to acquire a single image. Furthermore, the number of inspection points on semiconductor wafers is increasing as semiconductor devices become smaller and more complex, and as semiconductor wafers become larger in diameter. As a result, in-line inspection and measurement throughput decreases, which has a significant impact on the production efficiency of the production line.

[0004] To address these issues, technologies have been developed, such as those disclosed in Patent Documents 1 and 2, in which low-quality images are generated by a charged particle beam device, a high-quality image is estimated from the low-quality image by image processing, particularly image processing using deep learning, and the estimated high-quality image is used for inspection and measurement.

[0005] Patent Document 1 discloses a technology for obtaining a pair of images, a degraded image and a high-quality image, by changing the imaging conditions for the same location on a sample, learning the correspondence between the two using machine learning techniques, and estimating the high-quality image from the degraded image.

[0006] Patent Document 2 also discloses a technique for estimating a high-quality image from a low-quality image by using a learning device that has been trained using a data set of high-quality images and low-quality images as training data.

[0007] JP 2018-137275 A JP 2022-135215 A

[0008] It is actually difficult to obtain a pair of high-quality and low-quality images suitable as teacher images using a charged particle beam device. For example, in the case of Patent Document 1, the low-quality image is acquired before the stage settles, so even though the images are captured at the same location, there is a misalignment between the high-quality and low-quality images. Although alignment is performed to correct this misalignment, precise alignment is difficult. Furthermore, due to the effects of charging and damage caused by charged particle beam irradiation, differences occur between the pattern shape in the low-quality image and the pattern shape in the low-quality image. For example, it is known that when the pattern on the sample is a resist pattern made of an organic material, it shrinks due to electron beam irradiation.

[0009] In Patent Document 2, in order to eliminate the effects of shrinkage, charging, and the like caused by electron beam irradiation, low-quality images are selected such that the relationship between at least one of the state of the high-quality image and the feature quantity extracted from the high-quality image and the state of the low-quality image and at least one of the feature quantities extracted from the low-quality image satisfies a predetermined condition, and a data set of the selected low-quality images and high-quality images is used as training data. This makes it possible to approximate the features of the low-quality image and the high-quality image. However, even with the technology disclosed in Patent Document 2, it is difficult to obtain a pair of high-quality images and low-quality images in which the features of the low-quality image and the high-quality image match, i.e., a pair of high-quality images and low-quality images that are not affected by shrinkage, charging, and the like.

[0010] When training a learning model using a data set that includes effects such as misalignment, damage, and charging as training data, these effects are also learned. For high-quality images for inspection and measurement, it is desirable to convert only the image quality to a clear state without affecting the pattern shapes captured in low-quality images. To achieve this, it is desirable for the training data for the learning model to consist of pairs of high-quality and low-quality images that differ only in image quality and have no difference in the pattern shapes or positions in the images.

[0011] A computer system according to an embodiment of the present disclosure is a computer system for training a learning model that estimates image data of a higher quality than a first image quality and of an image quality that allows inspection or measurement, from image data of a first image quality acquired under first imaging conditions by a charged particle beam device that performs inspection or measurement based on images generated by irradiating a charged particle beam on a sample on which a predetermined pattern is formed at multiple positions. The computer system includes a data storage unit, an image region identification unit that divides an image of the image data into multiple regions and performs region identification that identifies a correspondence between each of the multiple divided 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 storage unit performs region identification on first image data obtained by capturing the predetermined pattern at any of multiple positions by the charged particle beam device under the first imaging conditions, and stores image characteristic data in which image characteristics calculated for each region included in the image of the first image data are registered, and second image data obtained by capturing the predetermined pattern at any of the multiple positions by the charged particle beam device with an image quality that allows inspection or measurement. The image region identification unit performs region identification on the second image data. The teacher data generation unit generates pseudo-low-quality image data that approximates the image characteristics of an area included in the image of the second image data to the image characteristics of an area included in the image of the first image data that has the same correspondence with a specified pattern, and generates a pair of the pseudo-low-quality image data and the second image data as teacher data.

[0012] The present disclosure can improve the accuracy of estimating high-quality images using a trained model and improve the throughput of in-line inspection or measurement. Other objects and novel features will become apparent from the description of this specification and the accompanying drawings.

[0013] 1 is a schematic configuration diagram of a measurement system; FIG. 2 is an example of the hardware configuration of a computer system; FIG. 3 is a measurement flow in the measurement system; FIG. 4 is a functional block diagram of a first computer system; FIG. 5 is a schematic diagram of a semiconductor wafer; FIG. 6 is a functional block diagram of a second computer system; FIG. 7 is a flowchart showing details of step S02 in the measurement flow; FIG. 8 is a region identification image of a high-quality image and a low-quality image; FIG. 9 is an example of a display screen for region correspondence; FIG. 10 is a diagram showing the data structure of image characteristic data; FIG. 11 is an example of a display screen for image quality adjustment of a pseudo-low-quality image; FIG. 12 is a functional block diagram of a model learning unit; FIG. 13 is an example of a region identification image of a high-quality image acquired in Modification Example 1;

[0014] FIG. 1 shows a schematic diagram of a measurement system using a scanning electron microscope (SEM) as an example of a charged particle beam device. In the following, the charged particle beam device 1 will be described as an example of a critical dimension SEM that measures the dimensions of a pattern formed on a sample such as a semiconductor wafer. Note that FIG. 1 only shows some of the components, and is not intended to be limiting to including all of the components shown. Furthermore, the positional relationships of the components are shown as an example, and arrangements different from those shown in FIG. 1 are also possible.

[0015] The charged particle beam instrument 1 mainly comprises an imaging unit (SEM main body) 100 and a first computer system 200 that controls the imaging unit 100 to acquire an image of a sample (e.g., a semiconductor device) and measures the dimensions of a pattern formed on the semiconductor device from the image. An input device 231 and a display device 232 are connected to the first computer system 200. The charged particle beam instrument 1 is further 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 instrument 1 for dimension measurement and provides the trained model to the charged particle beam instrument 1. An input device 331 and a display device 332 are connected to the second computer system 300.

[0016] 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 objective lens 114 controlled by an objective lens control unit 124, and irradiated onto a sample 104. Due to the irradiation of the primary electron beam 101, secondary electrons (SE) 102, which are signal electrons, and backscattered electrons (BSE) 103 are emitted from the sample 104. The SE 102 are detected by an SE detector 116, and an SE detection signal processing unit 126 outputs an output signal corresponding to the SE detection intensity. Furthermore, BSE 103 is detected by a BSE detector 117, and a BSE detection signal processing unit 127 outputs an output signal corresponding to the BSE detection intensity.

[0017] The scan deflector 113 controlled by the scan control unit 123 can deflect the primary electron beam 101 to scan the surface of the sample 104. Although not shown in FIG. 1 , an image shift deflector that moves the irradiation position of the primary electron beam 101 may also be provided. The sample 104 is held by a stage 115, and a desired location on the surface of the sample 104 can be observed by moving the stage 115 using the stage control unit 125. Note that the control units that control the respective elements included in the imaging unit 100 as described above may also be collectively referred to as element control unit 12x.

[0018] The image generator 131 generates SE image data by synchronizing the output signal from the SE detection signal processing unit 126 with a control signal from the scan control unit 123 and storing it in a frame memory. Also, it generates BSE image data by synchronizing the output signal from the BSE detection signal processing unit 127 with a control signal from the scan control unit 123 and storing it in a frame memory. SE image data and BSE image data may be collectively referred to as scanned image data. When storing the output signal in the frame memory, signal profile data (one-dimensional information) and scanned image data (two-dimensional information) can be generated by storing the detection signal at a position corresponding to the scanning position of the frame memory.

[0019] The first computer system 200 controls the generation of scanned image data by the imaging unit 100, performs image processing on the scanned image data generated by the image generator 131 to generate SEM image data, and performs dimensional measurement based on the SEM image data, etc. Details of this will be described later.

[0020] The first computer system 200 or the second computer system 300 includes, as its main components, 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 FIG. 2 . The processor 401 functions as a functional unit that provides a predetermined function by executing processing in accordance with a program loaded into the memory 402. The storage device 403 stores data and programs used by the functional unit. The input I / F 404 is connected to input devices such as a keyboard, pointing device, and operation panel, and the output I / F 405 is connected to a display device. The communication device I / F 406 enables communication with other computer systems via a network. These are connected to each other via the bus 407 so that they can communicate with each other.

[0021] In the following description, when describing processing by a program, the program or functional units may be described as the main components, but the main hardware component of these components is a processor or a computer system including the processor. The computer system executes processing according to a program loaded into memory using resources such as memory and a communication interface as appropriate. While FIG. 2 shows an example of a CPU as the processor, a GPU (Graphical Processing Unit) or the like may also be used. Furthermore, processing to realize a function is not limited to software program processing, and can also be implemented using a dedicated circuit. Examples of the dedicated circuit include a field programmable gate array (FPGA) and an application specific integrated circuit (ASIC).

[0022] (Measurement Flow) Figure 3 shows the measurement flow in the measurement system. First, the charged particle beam device 1 acquires high-quality images and low-quality images (S01). Next, the second computer system 300 generates training data using the acquired high-quality images and low-quality images (S02), and trains the learning model using the generated training data (S03). After that, in the charged particle beam device 1, the first computer system 200 receives the trained model from the second computer system 300 and measures a sample (e.g., a semiconductor wafer) using the trained model (S04). Steps S01 to S03 are also referred to as the preparation phase, and step S04 is also referred to as the operation phase.

[0023] FIG. 4 shows a functional block diagram of a first computer system 200 according to this embodiment. The first computer system 200 includes functional units, namely, an imaging condition setting unit 201, an image processing unit 202, a high-quality image estimation unit 203, and a measurement unit 204, as well as a data storage unit 210 that stores data used by these functional units. The imaging condition setting unit 201 generates scanned image data by controlling each element of the imaging unit 100 via the element control unit 12x (see FIG. 1) in accordance with imaging conditions set by a user via an input device 231. 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 referred to as SEM image data.

[0024] In the preparation phase, the image processing unit 202 stores SEM image data acquired under high-quality image shooting conditions as high-quality image data 211 and SEM image data acquired under low-quality image shooting conditions as low-quality image data 212 in the data storage unit 210. On the other hand, in the operation phase, the image processing unit 202 outputs SEM image data acquired under measurement shooting conditions to the high-quality image estimation unit 203.

[0025] A trained model 313 for estimating a high-quality image from SEM image data acquired under measurement and imaging conditions is registered in advance in the data storage unit 210. As will be described later, the trained model 313 is a learning model trained in the second computer system 300. The high-quality image estimation unit 203 estimates a high-quality image from the SEM image data output from the image processing unit 202 using the trained model 313. The measurement unit 204 measures the dimensions of the pattern using the estimated high-quality image and stores the measured data 213. Each step will be described below.

[0026] (Preparation Phase S01) In step S01, the charged particle beam device 1 acquires high-quality images and low-quality images for generating teacher data. Here, the image quality of each of the high-quality images and low-quality images is determined by the user. Specifically, the image quality of the high-quality images can be determined as an image quality that allows dimension measurement by the measurement unit 204. This is because this embodiment is premised on a charged particle beam device that measures the dimensions of a pattern. In the case of a charged particle beam device (review SEM) that reviews and classifies defects, the image quality can be determined as an image quality that achieves that purpose. On the other hand, it is desirable to set the imaging conditions for the low-quality images to be the same as the measurement imaging conditions set in the imaging condition setting unit 201 in the operation phase. This is for the following reason.

[0027] To improve the throughput of the charged particle beam device 1 during operation, the time required to acquire one SEM image data frame must be reduced. The time required to acquire one SEM image data frame mainly consists of the time required for the stage 115 to move the observation position on the sample to the irradiation position of the primary electron beam 101 (stage movement time), the time required for the objective lens 114 to focus the primary electron beam 101 on the sample surface (autofocus time), and the time required to acquire multiple scanned image data to be integrated (imaging time). Although the term "low-quality image" is generally used, different characteristics emerge depending on the time reduction method. For example, if the stage movement time is shortened and SEM image data is acquired before the sample has completely settled, deformation may appear in the SEM image. If the magnification is reduced and the focal depth is made shallow to shorten the autofocus time, the resolution of the SEM image may decrease. Similarly, if a certain degree of defocus is allowed to shorten the autofocus time, blurring may occur in the SEM image. Furthermore, in order to improve the quality of the image, the sample is generally scanned repeatedly with the primary electron beam 101 and the obtained multiple scanned image data are accumulated. However, if the number of accumulations is reduced to shorten the imaging time, the amount of noise contained in the SEM image will increase.

[0028] As described above, different degradation modes appear in low-quality images depending on the time reduction method. Therefore, if the imaging conditions for low-quality images are set as measurement imaging conditions in the imaging condition setting unit 201 in the operation phase, it becomes possible to train the learning model to improve the quality of the SEM images acquired in the operation phase in accordance with the degradation modes contained in the images. For this reason, the imaging conditions for low-quality images are stored in the data storage unit 210 in association with the low-quality images. The imaging conditions include the acceleration voltage, probe current, magnification, number of pixels, and the like, as well as image processing algorithms such as noise removal and contrast adjustment, and the order of these processes.

[0029] The high-quality image and low-quality image acquired in step S01 are images of the pattern to be measured, each acquired at one or more points. FIG. 5 schematically illustrates a semiconductor wafer 500. The semiconductor wafer 500 includes multiple chip regions 501. Predetermined patterns of semiconductor layers, insulator layers, conductor layers, etc. are stacked in each of the multiple chip regions 501, and finally, multiple semiconductor chips are obtained by dicing. Here, the pattern to be measured is assumed to be a line-and-space (L / S) pattern formed in a certain process. For example, assume an L / S pattern in which a resist 511 is formed on an underlying conductor layer 512, and measure the width W of the resist 511. In this case, the semiconductor wafer 500 includes L / S patterns to be measured, which are formed in the same process within the same chip region or another chip region. Therefore, observation regions 502 and 503 are set to include the L / S patterns formed in the same process, and images are acquired. As will be described later, in this embodiment, the number of high-quality images corresponds to the number of training data, so it is desirable to acquire a large number of images. On the other hand, it is sufficient to acquire at least one low-quality image.

[0030] The pattern to be measured is not limited to an L / S pattern, but may be a pattern or via having a specific shape. Furthermore, any pattern formed in a different process may be used as long as it can be considered identical to the pattern to be measured. For example, while FIG. 5 illustrates an L / S pattern extending in the Y direction, in such cases, an L / S pattern extending in the X direction is often formed in the upper layer. If it is determined that there is no problem in using a mixture of images of an L / S pattern in an upper layer and an L / S pattern in a lower layer as training data, high-quality images may be acquired that include patterns formed in such different processes.

[0031] (Preparation Phase S02) Prior to describing the processing of this step, the functions of the second computer system 300 according to this embodiment will be described. FIG. 6 shows a functional block diagram of the second computer system 300. The second computer system 300 includes functional units, namely, an image region identification unit 301, an image characteristic calculation unit 302, a training data generation unit 303, and a model learning unit 304, as well as a data storage unit 310 that stores data used by these functional units. The image region identification unit 301 divides the high-quality image and the low-quality image acquired in step S01 into regions and performs region identification to identify the correspondence between each region and the corresponding part of the pattern to be measured. The image characteristic calculation unit 302 calculates the image characteristics of each region included in the low-quality image. The training data generation unit 303 generates training data using the high-quality image data and the image characteristic data of each region calculated for the low-quality image. The model learning unit 304 trains the learning model using the generated training data.

[0032] The data storage unit 310 stores the high-quality image data 211 and low-quality image data 212 acquired in the preparation phase S01, the image characteristic data 311 calculated by the image characteristic calculation unit 302, the teacher data 312 generated by the teacher data generation unit 303, and the trained model 313 trained and trained by the model training unit 304. The high-quality image data 211 and low-quality image data 212 may be transferred from the first computer system 200 to the second computer system 300, or these data may be stored in a data server accessible to both the first computer system 200 and the second computer system 300.

[0033] FIG. 7 is a flowchart showing the details of step S02.

[0034] (Step S11) The image region identification unit 301 divides each of the high-quality image and the low-quality image acquired in step S01 into regions and identifies the correspondence between each divided region and the pattern to be measured (region identification). Any known method can be applied to this region identification. For example, it can be performed manually, or an image processing method such as the k-means method, image recognition AI, or a combination thereof can be applied.

[0035] In this step, the required range of region identification differs between high-quality images and low-quality images. For high-quality images, the image must be divided so that the entire image is included in one of the regions. For low-quality images, however, it is sufficient that regions are identified so that at least a portion of the image corresponds to a predetermined pattern. FIG. 8 shows the results (region identification images) of region identification performed on high-quality and low-quality images when the pattern to be measured is an L / S pattern. In the region identification image 601 of the high-quality image, the entire image is divided into two types of regions: a line region 602 and a space region 603. In contrast, in the region identification image 611 of the low-quality image, a portion of the region included in the line pattern is identified as a line region 612, and a portion of the region included in the space pattern is identified as a space region 613.

[0036] High-quality images are acquired in the order of tens to hundreds of images and have clear image quality, so it is desirable to automatically segment the regions using image processing or AI. Regarding the correspondence between the segmented regions and the pattern to be measured, it is possible to identify, for example, regions with higher brightness as line regions and regions with lower brightness as space regions based on feature quantities such as brightness in the segmented regions. Alternatively, the high-quality images may be compared with design information (layout diagram) of the coordinates from which they were acquired to determine whether the segmented regions correspond to line regions or space regions.

[0037] On the other hand, it is sufficient to acquire one or a few low-quality images, and since the image quality is unclear, they may be identified manually. In this case, areas that are clearly line patterns when visually inspected may be identified as line areas, and areas that are clearly space patterns may be identified as space areas. Of course, the same techniques as for high-quality images may be used to divide and identify the areas.

[0038] (Step S12) The image region identification unit 301 associates regions having the same identified pattern between the high-quality image and the low-quality image. In the example of Fig. 8, in step S11, line regions and space regions are identified for each of the region identification image of the high-quality image and the region identification image of the low-quality image. In this case, by identifying region 602 of the high-quality image and region 612 of the low-quality image as the same line region, and region 603 of the high-quality image and region 613 of the low-quality image as the same space region, it can be said that steps S11 and S12 were performed simultaneously.

[0039] On the other hand, when it is difficult to associate regions using region classification processing alone, or when the user wishes to confirm the region classification results, it is preferable to display a GUI for performing region association on the display device 332. Fig. 9 shows an example of a display screen for region association. The region association screen 700 includes region identification image display sections 701 and 702 for the high-quality image and the low-quality image, association setting sections 703 and 704 for each region included in the high-quality image, and an association confirmation button 705.

[0040] The region identification image display units 701 and 702 display region identification images of the high-quality image and the low-quality image, respectively, for which region identification is to be performed. Instead of the region identification images, the high-quality image or the low-quality image itself may be displayed so that the region identification results are clear. Regarding the correspondence setting units 703 and 704, since the high-quality image in this example has two regions, i.e., region A (line region) and region B (space region), for each region, the correspondence between the two regions in the low-quality image, i.e., region a and region b, can be selected. If a correspondence has already been determined as a result of region identification in step S11, the determined correspondence is displayed in a shaded area. The user can confirm the correspondence, and if they determine that the correspondence is incorrect, they can correct the correspondence by reselecting the corresponding region (shaded area) in the correspondence setting units 703 and 704. The user can confirm the correspondence between the regions by pressing the correspondence confirmation button 705.

[0041] (Step S13) The image characteristic calculation unit 302 calculates image characteristics for each region of the low-quality image acquired in step S11. FIG. 10 shows the data structure of the image characteristic data 311 acquired in this step. The ID 801 is an ID that uniquely identifies the low-quality image. The shooting condition data 802 registers the shooting conditions when the low-quality image was acquired. The image characteristic data 803 registers the image characteristics for each region. For example, in the case of the low-quality image shown in FIG. 9, the image characteristic data 803 is registered for region a and region b.

[0042] Any image characteristic may be registered as image characteristic data 803, and here noise characteristics, brightness characteristics, shape characteristics, and degree of blur are exemplified, but it is not necessary to register all of these, and image characteristics other than those exemplified may also be registered. Noise characteristics can be quantitatively determined as the variation in brightness in the region, brightness characteristics as the brightness distribution in the region, shape characteristics as the unevenness of the edge in the region, and degree of blur as the sharpness of the region.

[0043] In addition, if there are multiple low-quality images acquired under the same shooting conditions, a representative value such as the average or median of the image characteristic values ​​of those multiple low-quality images may be used as the image characteristic of the low-quality image.

[0044] (Step S14) The teacher data generation unit 303 generates a pseudo low-quality image from the high-quality image acquired in step S11 using the image characteristic data 803 calculated in step S13. A pseudo low-quality image is an image in which the image characteristics of each region of the high-quality image are approximated to the image characteristics of the corresponding region of the low-quality image. For example, in the example of Figure 9, the pseudo low-quality image can be generated by converting the image characteristics of region A of the high-quality image to approximate the image characteristics of region a of the low-quality image, and converting the image characteristics of region B of the high-quality image to approximate the image characteristics of region b of the low-quality image.

[0045] (Steps S15 to S17) The teacher data generation unit 303 determines whether the pseudo-low-quality image generated in step S14 is actually similar to the low-quality image with similar image characteristics. In addition to the user's visual inspection, a similarity index is also calculated to enable a quantitative determination. Indices that can be used as the similarity index include PSNR (Peak Signal to Noise Ratio), SSIM (Structural Similarity), and MSE (Mean Squared Error). At this time, both a similarity index between corresponding regions and a similarity index for the entire image are calculated. If it is determined that the pseudo-low-quality image and the low-quality image are similar (Yes in S15), a pair of the generated pseudo-low-quality image and its original high-quality image is registered as teacher data 312 (S17).

[0046] On the other hand, if the pseudo low-quality image and the low-quality image are not determined to be similar (No in S15), the parameters of the pseudo low-quality image are adjusted to adjust the image quality of the pseudo low-quality image (S16), and if a pseudo low-quality image similar to the low-quality image is obtained, the pair of the parameter-adjusted pseudo low-quality image and the high-quality image is registered as training data 312 (S17).

[0047] 11 shows an example of a GUI (pseudo-low-quality image adjustment screen 900) for adjusting the image quality of a pseudo-low-quality image displayed on the display device 332. The pseudo-low-quality image and the low-quality image are displayed in image display sections 901 and 902, respectively, so that the user can visually compare the two. The similarity index display section 903 displays the calculated similarity index for each region and the similarity index for the entire image.

[0048] If the user determines that the similarity between the pseudo-low-quality image and the low-quality image is insufficient, the parameters of the pseudo-low-quality image are adjusted. In this example, an adjustment selection unit 905 is provided to allow the user to easily adjust the parameters in a desired direction. The user sets the adjustment direction in the adjustment selection unit 905 and adjusts the image quality of the pseudo-low-quality image by pressing a parameter adjustment button 906. The parameter-adjusted pseudo-low-quality image is displayed in the image display unit 901. The user then presses a similarity calculation button 904, which calculates a similarity index between the parameter-adjusted pseudo-low-quality image and the low-quality image, and updates the value in the similarity index display unit 903. If the user determines that the parameter-adjusted pseudo-low-quality image is appropriate as training data, the user presses a registration button 907, which registers the parameter-adjusted pseudo-low-quality image as training data.

[0049] By performing the above process on the high-quality image acquired in step S11, training data 312 is obtained (S02).

[0050] 12 shows a functional block diagram of the model learning unit 304. The learning model 1000 is, for example, a neural network that performs deep learning, and includes an input layer 1011, an output layer 1012, and multiple intermediate layers 1013. The model learning unit 304 adjusts parameters such as weights and biases in the intermediate layers 1013 so that an image (inferred image) output from the output layer 1012 for an image input to the input layer 1011 is sufficiently similar to a desired image.

[0051] 12 shows the training of the learning model 1000, i.e., the adjustment of parameters, using a pair of pseudo low-quality image 1021 and high-quality image 1022 included in the training data 312. Specifically, an inferred image 1023 is obtained by inputting the pseudo low-quality image 1021 into the learning model 1000. An error calculator 1001 compares the inferred image 1023 with the high-quality image 1022 to calculate the error, and a parameter adjuster 1002 adjusts the parameters of the learning model 1000 to reduce the error. When the error becomes sufficiently small, the training of the learning model 1000 is terminated, and the parameters of the learning model 1000 are registered as a trained model 313.

[0052] In this way, in this embodiment, a pair of pseudo low-quality images and high-quality images is used as training data. Because the pseudo low-quality images are generated based on the high-quality images, the regions, i.e., the patterns in the high-quality images, do not change from the patterns in the pseudo low-quality images. Therefore, it is possible to obtain a trained learning model that improves image quality without learning the effects of misalignment, damage, charging, etc., as in conventional techniques.

[0053] (Operation Phase S04) The trained model 313 obtained in step S03 is transferred to the first computer system 200 and registered in the data storage unit 210.

[0054] When SEM image data acquired under measurement shooting conditions (operational shooting conditions) is input, the high-quality image estimation unit 203 calls the trained model 313 and estimates a high-quality image from the SEM image data. If multiple trained models are stored in the data storage unit 210 as the trained models 313, the high-quality image estimation unit 203 compares the shooting conditions of the low-quality images used to generate the pseudo-low-quality images that make up the training data with the operational shooting conditions, and selects a trained model whose shooting conditions are the same as those of the low-quality images. The measurement unit 204 performs dimensional measurement of the estimated high-quality image.

[0055] A modification of this embodiment will now be described.

[0056] (Variation 1) In the embodiment, an example has been described in which high-quality images and low-quality images are captured by the charged particle beam device 1 to generate teacher data. In Variation 1, existing calculated image characteristic data 311 is used, and only high-quality images are captured by the charged particle beam device 1 without newly acquiring low-quality images, to generate teacher data. The method of generating teacher data in this case (preparation phase S02) will be described, focusing on the differences from the embodiment.

[0057] 13 shows an example of a region identification image 1300 of a high-quality image acquired in Modification 1. For such a high-quality image, a pseudo low-quality image is generated using the image characteristic data 311 of the low-quality image described in the embodiment.

[0058] In the region identification image 1300, region 1301 is assumed to be an image of the resist, and region 1302 is assumed to be 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 FIG. 7 ). This process can be performed in the same manner as in the embodiment, and thereby, a pseudo-low-quality image of a high-quality image can be generated using the image characteristic data 311.

[0059] In the embodiment, the pseudo-low-quality image was evaluated by evaluating the similarity between the pseudo-low-quality image and the low-quality image. In contrast, in Modification Example 1, the pseudo-low-quality image and the low-quality image are different images, so the similarity between the image characteristics of the associated regions is evaluated as a similarity index. Specifically, for the obtained pseudo-low-quality image, the image characteristics of the region 1301 and the image characteristics of the region 1302 are obtained in the same way as for the low-quality image, and the similarity between the image characteristics of the associated region 1301 and the image characteristics of the region a, and the similarity between the image characteristics of the region 1302 and the image characteristics of the region b are calculated and evaluated. In this case, the Euclidean distance or Manhattan distance between the image characteristics of the regions may be used as the similarity index.

[0060] Although the first modification has been described by taking as an example a case where different patterns are measured as measurement targets, the same applies when the patterns to be measured are the same patterns formed in the same process.

[0061] (Variation 2) In the embodiment, in principle, a trained model is obtained by acquiring multiple high-quality images of patterns to be measured that are formed in the same process. In this case, a trained model is created for each pattern to be measured. If multiple patterns to be measured exist on a single semiconductor wafer, multiple trained models are required. In this case, the principle may be relaxed, and the trained model may be trained by mixing training data generated from high-quality images of patterns such as those shown in FIG. 13 .

[0062] To prevent degradation of the high-quality images inferred by the trained model, it is necessary to narrow down the patterns to be mixed to patterns that are somewhat similar, but this makes it easier to manage the trained model.

[0063] The present disclosure is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments have been described in detail to make the present disclosure easier to understand, and are not necessarily limited to those including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment or modification with the configuration of another embodiment or modification, and it is also possible to add the configuration of another embodiment or modification to the configuration of one embodiment or modification. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment or modification with other configurations.

[0064] 1: charged particle beam device, 100: imaging unit, 101: primary electron beam, 102: secondary electrons, 103: backscattered electrons, 104: specimen, 111: electron source, 112: accelerating electrode, 113: scan deflector, 114: objective lens, 115: stage, 116: SE detector, 117: BSE detector, 121: electron source control unit, 122: accelerating electrode control unit, 123: scan control unit, 124: objective lens control unit, 125: stage control unit, 126: SE detection signal processing unit, 127: BSE detection signal processing unit, 131: image generator, 200: first computer system system, 201: imaging condition setting unit, 202: image processing unit, 203: high-quality image estimation unit, 204: measurement unit, 210: data storage unit, 211: high-quality image data, 212: low-quality image data, 213: measurement data, 231: input device, 232: display device, 300: second computer system, 301: image region identification unit, 302: image characteristic calculation unit, 303: training data generation unit, 304: model learning unit, 310: data storage unit, 311: image characteristic data, 312: training data, 313: trained model, 330: network, 331: input device, 332 : display device, 401: processor (CPU), 402: memory, 403: storage device, 404: input interface, 405: output interface, 406: communication interface, 407: bus, 500: semiconductor wafer, 501: chip area, 502, 503: observation area, 511: resist, 512: conductor layer, 601, 611: area identification image, 602, 612: line area, 603, 613: space area, 700: area association screen, 701, 702: area identification image display section, 703, 704: association setting section, 705: association confirmation button , 801: ID, 802: Shooting condition data, 803: Image characteristic data, 900: Pseudo low-quality image adjustment screen, 901, 902: Image display section, 903: Similarity index display section, 904: Similarity calculation button, 905: Adjustment selection section, 906: Parameter adjustment button, 907: Registration button, 1000: Learning model, 1001: Error calculator, 1002: Parameter adjuster, 1011: Input layer, 1012: Output layer, 1013: Intermediate layer, 1021: Pseudo low-quality image, 1022: High-quality image, 1023: Inference image, 1300: Region identification image, 1301, 1302: Region.

Claims

1. A computer system for training a learning model that estimates image data of a quality higher than the first image quality and enabling the inspection or measurement, from the first image quality image data acquired by a charged particle beam apparatus under first imaging conditions, for performing inspection or measurement based on an image generated by irradiating a sample in which a predetermined pattern is formed at a plurality of positions with a charged particle beam, the computer system comprising: a data storage unit; an image region identification unit that divides an image of the image data into a plurality of regions and performs region identification to identify a correspondence relationship between each of the plurality of divided regions and the predetermined pattern; a teacher data generation unit that generates teacher data; and a model learning unit that trains the learning model using the teacher data generated by the teacher data generation unit, wherein the data storage unit performs region identification on first image data obtained by the charged particle beam apparatus by imaging the predetermined pattern under the first imaging conditions at any one of the plurality of positions, stores image characteristic data in which image characteristics calculated for each region included in the image of the first image data are registered, and stores second image data obtained by the charged particle beam apparatus by imaging the predetermined pattern under an image quality enabling the inspection or measurement at any one of the plurality of positions, the image region identification unit performs region identification on the second image data, and the teacher data generation unit generates pseudo low-quality image data in which the image characteristics of the regions included in the image of the second image data are approximated to the image characteristics of the regions included in the image of the first image data having the same association with the predetermined pattern, and generates a pair of the pseudo low-quality image data and the second image data as the teacher data.

2. The computer system according to claim 1, further comprising an image characteristic calculation unit that calculates image characteristics for each region included in the image of the image data, wherein the data storage unit stores the first image data, the image region identification unit performs region identification on the first image data, and the image characteristic calculation unit calculates image characteristics for each region included in the image of the first image data and stores the image characteristics in the data storage unit as the image characteristic data.

3. The computer system according to claim 2, wherein the image area identification unit divides the second image data so that the entire image is included in any area, and identifies an area of the first image data so that at least a part of the image is associated with the predetermined pattern.

4. The computer system according to claim 2, wherein the image area identification unit displays, on a display device, an area association screen for selecting an association with any area included in the image of the first image data for each area included in the image of the second image data.

5. The computer system according to claim 4, wherein the image area identification unit displays the area association screen on the display device in a state where, for each area included in the image of the second image data, an area having the same association with the predetermined pattern as an area included in the image of the first image data is selected.

6. The computer system according to claim 2, wherein the teacher data generation unit calculates a similarity index between the first image data and the pseudo low-quality image data.

7. The computer system according to claim 6, wherein the teacher data generation unit calculates the similarity index based on image characteristics of an area included in the image of the second image data having the same association with the predetermined pattern and image characteristics of an area included in the image of the first image data.

8. The computer system according to claim 1, wherein the predetermined pattern is a pattern formed on the sample in the same process.

9. The computer system according to claim 1, wherein the first imaging condition is set such that the time required for the charged particle beam device to acquire image data under the first imaging condition is shorter than the time required to acquire one piece of image data with an image quality enabling the inspection or measurement.

10. A charged particle beam device comprising: an imaging unit which irradiates a charged particle beam onto a sample on which a predetermined pattern is formed at a plurality of positions, and detects signal electrons emitted from the sample; an image generator which generates scanned image data based on scanning position information of the charged particle beam and detection information of the signal electrons; and a computer system which performs inspection or measurement based on an image of image data generated based on the scanned image data, wherein the computer system has an image processing unit which acquires image data under operational shooting conditions; and a high-quality image estimation unit which estimates image data of higher quality and image quality capable of the inspection or measurement from the image data acquired under the operational shooting conditions using a trained model, wherein the trained model is a learning model trained using teacher data which is a pair of pseudo-low-quality image data and the second image data, which are generated using first image data obtained by shooting the predetermined pattern at any of the plurality of positions under first shooting conditions, and second image data obtained by shooting the predetermined pattern at any of the plurality of positions with image quality capable of the inspection or measurement, The first image data and the second image data each have their images divided into a plurality of regions, and a correspondence between each of the divided regions and the specified pattern is identified, and the pseudo-low-quality image data is generated by approximating image characteristics of a region included in the image of the second image data to image characteristics of a region included in the image of the first image data that has the same correspondence with the specified pattern, and the computer system is a charged particle beam device that performs the inspection or measurement using an image of the image data estimated by the high-quality image estimation unit.

11. A charged particle beam device as described in claim 10, wherein the computer system stores a plurality of the trained models, and the high-quality image estimation unit selects a trained model whose operational shooting conditions are the same as the first shooting conditions of the first image data used to generate the pseudo-low-quality image data of the teacher data used to train the trained models.

12. The charged particle beam apparatus according to claim 10, wherein the pseudo low-quality image data is determined as to whether it is suitable as the teacher data based on the similarity between the image of the pseudo low-quality image data and the image of the first image data.

13. The charged particle beam apparatus according to claim 10, wherein the predetermined pattern is a pattern formed on the sample in the same process.

14. The charged particle beam apparatus according to claim 10, wherein the first imaging condition is set such that the time required for the charged particle beam apparatus to acquire image data under the first imaging condition is shorter than the time required to acquire one piece of image data with an image quality that enables the inspection or measurement.

15. A method for training a learning model that estimates image data of a quality higher than the first image quality and enabling the inspection or measurement from the first image quality image data acquired by a charged particle beam apparatus under first imaging conditions, based on an image generated by irradiating a sample in which a predetermined pattern is formed at a plurality of positions with a charged particle beam, and performing inspection or measurement. The computer system includes a data storage unit, an image region identification unit that divides an image of the image data into a plurality of regions and performs region identification to identify a correspondence relationship between each of the plurality of divided regions and the predetermined pattern, a teacher data generation unit that generates teacher data, and a model learning unit that trains the learning model. The data storage unit performs region identification on first image data obtained by the charged particle beam apparatus imaging the predetermined pattern under the first imaging conditions at any of the plurality of positions, and stores image characteristic data in which image characteristics calculated for each region included in the image of the first image data are registered, and second image data obtained by the charged particle beam apparatus imaging the predetermined pattern at any of the plurality of positions with an image quality enabling the inspection or measurement. The image region identification unit performs region identification on the second image data. The teacher data generation unit generates pseudo low-quality image data in which the image characteristics of the regions included in the image of the second image data are approximated to the image characteristics of the regions included in the image of the first image data having the same association with the predetermined pattern, and generates a pair of the pseudo low-quality image data and the second image data as the teacher data. The model learning unit is a method for training a learning model that trains the learning model using the teacher data.

16. In claim 15, the computer system further includes an image characteristic calculation unit that calculates image characteristics for each region included in the image of the image data. The data storage unit stores the first image data. The image region identification unit performs region identification on the first image data. The image characteristic calculation unit calculates image characteristics for each region included in the image of the first image data, and stores the image characteristics in the data storage unit as the image characteristic data. A method for training a learning model.

17. The training method of the learning model according to claim 15, wherein the predetermined pattern is a pattern formed on the sample in the same process.

Images