Dimension measurement device, dimension measurement method, and semiconductor device manufacturing system

Abstract

An objective of the present invention is to provide a method whereby the dimensions of a pattern prior to deformation can be measured for a cross-sectional image having a structure in which pattern deformation has occurred.　With regard to a dimension measurement device according to the present invention, the dimensions of a desired portion of a pattern are measured using a cross-sectional image including an etched pattern, and when the pattern is in contact with another pattern, the contacting portion of the pattern is cut, and then a segmentation image of the pattern in which the contact portion is cut is obtained using a semantic segmentation model, contour line data of the pattern is acquired using the segmentation image, the position of the pattern prior to contact is estimated using the contour line data, the amount of displacement between the estimated position and a contour line is obtained, the contour line data is corrected on the basis of the obtained displacement amount, and the dimensions of a desired portion of the pattern are measured on the basis of the corrected contour line data.

Description

Dimension measurement apparatus, dimension measurement method, and semiconductor device manufacturing system

[0001] The present invention relates to a dimension measurement apparatus, a dimension measurement method, and a semiconductor device manufacturing system.

[0002] In recent years, semiconductor device performance has been improved by introducing new materials and by making their structures more three-dimensional and complex. The fabrication of current advanced semiconductor devices requires nanometer-level precision for patterns consisting of lines and spaces. In process development, after processing using processing equipment such as an etching device, wafers are cut out and cross-sections of the sample are observed using a scanning electron microscope (SEM) to capture cross-sectional SEM images. Dimensions of the required locations are measured, and the processing results are then fed back into the process development process. In this process, if the pattern has a large aspect ratio, the pattern shape may deform between the time the wafer is cut out after etching and the time the cross-sectional SEM image is captured, or even during image capture, resulting in images in which adjacent patterns touch each other. Even when such images are obtained, it is necessary to obtain the dimensions of the shape before deformation immediately after etching, which is difficult to determine from cross-sectional SEM images.

[0003] To address this issue, Patent Document 1 (JP-A-2005-109523) describes a preprocessing method for approximating the contour of a single trench portion with an elliptical function, thereby correcting the deformed trench shape to a bilaterally symmetrical shape. In this process, the pattern pitch (= line width + space width) is assumed to be always maintained, and the left line width is subtracted from the pattern pitch to obtain the corrected trench width, which is then distributed bilaterally symmetrically about the trench centerline. Specifically, Patent Document 1 (JP-A-2005-109523) discloses the following as inventions of a contour analysis device, a processing dimension extraction system, a processing condition determination system, and a semiconductor device manufacturing system: "High-precision fitting is performed for complex shapes that may occur in semiconductor processing. The shape model is a curve drawn in one stroke from the start point to the end point on the periphery of a figure that is a combination of one or more ellipses and one or more line segments in an xy plane defined by mutually orthogonal x- and y-axes. The shape model represents a processing dimension identified based on a singular point in the shape model. The constraints on the shape model parameters are derived based on a processing dimension function with the shape model parameters as variables and a domain of the processing dimensions."

[0004] Patent No. 7345665

[0005] In the method of Patent Document 1, the slope of the left line pattern is corrected at the same time as the trench to be corrected, but because the corrected trench width is determined only from the left line width, there is a problem that the resulting trench shape does not match the shape of the right line. For this reason, Patent Document 1 is unable to reproduce the pattern before deformation, making it difficult to measure the dimensions of the pattern before deformation. Therefore, an object of the present invention is to provide a method that can measure the dimensions of a pattern before deformation for a cross-sectional image having a structure in which pattern deformation has occurred.

[0006] In order to solve the above-mentioned problems, one representative dimension measurement device of the present invention is a dimension measurement device that measures the dimension of a desired portion of an etched pattern using a cross-sectional image including the pattern, and when the pattern is in contact with another pattern, after the contacting portion of the pattern is cut, a segmentation image of the pattern with the contacting portion cut is obtained using a semantic segmentation model, contour data of the pattern is obtained using the segmentation image, the position of the pattern before contact is estimated using the contour data, the amount of displacement between the estimated position and the contour is determined, the contour data is corrected based on the determined amount of displacement, and the dimension of the desired portion of the pattern is measured based on the corrected contour data.

[0007] According to the present invention, it is possible to measure the dimensions of a pattern before deformation for a cross-sectional image having a structure in which pattern deformation has occurred. Problems, configurations, and effects other than those described above will become apparent from the following description of the embodiments.

[0008] FIG. 1 is a diagram illustrating an example of the system configuration of a semiconductor device manufacturing system. FIG. 2 is a diagram illustrating an example of the internal configuration of a server. FIG. 3 is a diagram illustrating an example of a semantic segmentation model. FIG. 4 is a diagram illustrating an example of a cross-sectional image included in stored data. FIG. 5 is a diagram illustrating an example of a segmentation image, which is annotation data included in a training dataset. FIG. 6 is a table showing the correspondence between label names, label numbers, and colors in annotation data used to train a semantic segmentation model. FIG. 7 is a diagram illustrating an example of an image input to a semantic segmentation model. FIG. 8 is a diagram illustrating a segmentation image for an input image. FIG. 9 is a diagram illustrating an enlarged portion of a segmentation image illustrating a start point A and an end point B of a contact point. FIG. 10 is a diagram illustrating an enlarged portion of a segmentation image after cutting the contact point by a line segment AB. FIG. 11 is a diagram illustrating a segmentation image after cutting the contact point of FIG. 8. FIG. 12 is a diagram illustrating a contour line determined from the segmentation image after cutting. FIG. 13 is a diagram showing a contour image after contour deformation correction. FIG. 14 is a diagram explaining a method for correcting the shape of a deformed contour (hereinafter also referred to as a "deformed pattern"). FIG. 15 is a diagram showing the definition of six types of feature points extracted from a divided contour. FIG. 16 is a diagram showing a foreground image obtained by cutting out a foreground portion from an input image. FIG. 17 is a diagram showing a foreground image after correction in which a pattern deformation correction has been performed on the foreground image cut out from the input image. FIG. 18 is a diagram showing a background image obtained by cutting out a background portion from an input image. FIG. 19 is a diagram showing an artificial background image created from the background image cut out from the input image. FIG. 20 is a diagram showing an example of a generated corrected image. FIG. 21 is a diagram showing an example of dimension measurement performed using six types of feature points extracted from corrected contour data, and the measurement values ​​printed and output on the corrected image. FIG. 22 is a flowchart showing an example of a processing procedure in a dimension measurement system. FIG. 23 is a diagram showing an example of a GUI screen displayed during model learning. FIG. 24 is a diagram showing an example of a GUI screen displayed during the inference and measurement step.

[0009] In this disclosure, an image recognition model is used to measure the dimensions of a semiconductor device from a cross-sectional image of the semiconductor device having a repetitive pattern. Here, "semiconductor device" refers not only to finished products but also to semiconductor devices undergoing processing, whether in wafer form or individually separated chip form. Specifically, the image recognition model is a semantic segmentation model that segments a cross-sectional image. If the inference results of this image recognition model indicate contact areas in the segmentation image due to deformation caused by a process other than the processing process, the contact areas are cut, contour line data is obtained from the cut segmentation image, the reference positions of each line before deformation are estimated from the contour line data, the displacement of the contour line data measured from the reference positions is calculated, and the contour line data is corrected, thereby enabling measurement of the pre-deformation dimensions. Note that the image recognition model used may also be a different type of segmentation model, such as an instance segmentation model or a panoptic segmentation model, as long as the necessary dataset is available. When using these models, even if masks are in contact with each other, the contacting masks are distinguished as separate masks, eliminating the need for a cutting process for the contact areas. Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings.

[0010] [Example] In the dimension measurement system of this example, the learning data set of the semantic segmentation model, which is an image recognition model, consists of cross-sectional SEM (Scanning Electron Microscope) images, which are image data of an object to be measured, and segmentation images, which are annotation data.

[0011] In the pre-learning step, a semantic segmentation model is trained. The input data, that is, cross-sectional SEM images and segmentation images, are provided as training data to the semantic segmentation model, and the model is made to learn the region shapes.

[0012] In the inference step, a segmentation image is obtained from the cross-sectional SEM image to be measured using the trained semantic segmentation model.

[0013] In the deformation correction step, if there are any contact points due to deformation caused by processes other than the machining process, the contact points are cut off, the contour line is found, and the deformation of the contour line is corrected. (Details will be described later.) The contour line data is then converted into a format that makes dimensional measurement easy and saved.

[0014] Furthermore, if image correction is also required, the input cross-sectional SEM image is divided into two images: a foreground image consisting of the target structure and a background image consisting of the rest of the target structure, and the foreground image is corrected using the displacement amount of the contour line data to generate a corrected (corrected) image.

[0015] The measurement step is a step of measuring the dimensions of a predetermined measurement point using the data obtained in the inference step. Specifically, the desired dimensional values ​​are calculated by reading data related to the measurement point specified by the user and converting it into dimensions.

[0016] <Configuration Example of Dimension Measurement System> First, the configuration of a dimension measurement system will be described with reference to FIG. 1 . FIG. 1 is a diagram showing an example of the configuration of a semiconductor device manufacturing system 10. The semiconductor device manufacturing system 10 includes a processing device 111 and a dimension measurement system 110. The dimension measurement system 110 includes an evaluation device 100, a server 101, a database 102, and one or more input / output devices 103, all of which are connected via a network. In this disclosure, a case will be described in which the processing device (semiconductor manufacturing device) 111 and the dimension measurement system 110 are connected via a network. When connected to a network, the processing device 111 transmits data indicating processing conditions to the evaluation device 100 via the network. Note that the present disclosure can also be applied when there is no network between the processing device 111 and the dimension measurement system 110. Processing condition data may be exchanged between the processing device 111 and the evaluation device 100 via hardware, for example.

[0017] The input / output device 103 is, for example, a terminal equipped with a display and keyboard, or a PC or tablet equipped with a built-in storage medium, and as shown in the figure, is used by a system operator such as a measurement engineer who uses the evaluation device 100, a process engineer who uses the processing device 111, or a programmer who manages the server 101 or database 102. In the following description, when the term "input / output device 103" is used, it means a collective term for input / output device 103-1, input / output device 103-2, and input / output device 103-3, and is intended to describe characteristics common to all input / output devices. The number of input / output devices 103 is not limited to three.

[0018] The processing device 111 is a device for processing semiconductors or semiconductor devices including semiconductors. The processing performed by the processing device 111 is, for example, pattern processing, and includes processing performed by at least one of a wet etching device, a dry etching device (plasma etching device), an electron beam processing device, and a laser processing device. In the following description, the processing device 111 is an etching device.

[0019] The evaluation apparatus 100 is an apparatus for evaluating semiconductors or semiconductor devices containing semiconductors. Specifically, the evaluation apparatus 100 is an apparatus that outputs a cross-sectional image as an evaluation result of a sample obtained from a wafer processed by the processing apparatus 111, and includes, for example, an SEM, a TEM (Transmission Electron Microscope), a processing dimension measurement apparatus using an optical monitor (e.g., a cross-sectional SEM apparatus), or an FIB apparatus. Because the evaluation apparatus 100 acquires a cross-sectional image including a pattern etched by the processing apparatus 111, the evaluation apparatus 100 can also be referred to as a charged particle beam apparatus. The sample obtained from the wafer may be a sample (coupon) obtained by cutting the wafer and cutting a portion thereof, or the entire wafer. Alternatively, a lamella production apparatus may be disposed midway during transport of the wafer from the processing apparatus 111 to the evaluation apparatus 100, and the apparatus may extract a portion of the semiconductor or semiconductor device as a fragment, and the extracted sample may be used as the measurement target sample.

[0020] The server 101 is connected to at least the processing device 111 and the evaluation device 100. The server 101 is a platform on which an application (described in detail below) for measuring the dimensions of a desired portion of a pattern using a cross-sectional image including an etched pattern is implemented. The server 101 can be configured with a normal computer, and an OS, a framework, a library, a programming language, a model, and the like required for deep learning are installed. It is desirable for the server 101 to be equipped with a high-performance GPU (Graphics Processing Unit) to perform model learning in a short period of time. Model learning and inference are performed by logging in to the server 101 from the input / output device 103. While the case where the server 101 is configured by hardware and software such as an OS has been described, the configuration of the server 101 is not limited to this. The server 101 may function as a platform, which is an environment in which an application is executed, and may be configured, for example, as a cloud environment.

[0021] The database 102 is an external storage device that stores the processing conditions of the processing device 111 for each sample, the imaging conditions of the evaluation device 100, cross-sectional images, annotation data, model parameters representing the trained model, processed data of contour line data, measurement results, etc.

[0022] A process engineer uses the annotation tool 127 via the input / output device 103 to create annotation data for images obtained by the evaluation device 100 and stored in the database 102, stores the data in the database 102, and also performs model learning on the server 101 in response to a command from the input / output device 103. After the model has been learned, when an image to be measured obtained by the evaluation device 100 is provided, inference is performed using the learned model, and the inference results, that is, a segmentation image and processed data of contour data, are stored in the database 102. When measuring, when the desired measurement location is input, the necessary data is read from the database 102 and the dimensions are calculated. If necessary, a programmer uses the input / output device 103 to switch models, etc.

[0023] The above tasks do not necessarily have to be shared among a measurement engineer, a process engineer, and a programmer, and may be performed by a single system operator.

[0024] <Example of Internal Configuration of Server 101 with Dimension Measurement Function> Next, the internal configuration of the server 101 will be described with reference to FIG. 2 . FIG. 2 is a diagram showing an example of the internal configuration of the server 101. The server 101 can also be referred to as a dimension measurement device that measures the dimensions of a semiconductor device having a repetitive pattern from a cross-sectional image of the semiconductor device. In the server 101, dimensions of a desired portion of a pattern are measured using a cross-sectional image including an etched pattern. More specifically, in the server 101, when a pattern is in contact with another pattern, the contacting portion of the pattern is cut off, and then a segmentation image of the pattern where the contacting portion is cut off is obtained using a semantic segmentation model. Contour data of the pattern is acquired using the segmentation image. The position of the pattern before the contact is estimated using the contour data. The amount of displacement between the estimated position and the contour is calculated. The contour data is corrected based on the calculated amount of displacement. The dimension of the desired portion of the pattern is measured based on the corrected contour data. Specifically, the server 101 includes a processor 116 and memory (117, 118, 126), and the memory stores a dimension measurement program (application) that causes the processor 116 to function (execute) as a learning unit 122 that implements a semantic segmentation model, an inference unit 123 that outputs a segmentation image for an input image, a deformation correction processing unit 124 that cuts contact points in the segmentation image to obtain a contour line, corrects deformation of the contour line and deformation of the image, and obtains feature points that satisfy predetermined criteria from the contour line, and a dimension measurement unit 125 that measures dimensions of predetermined measurement points using data obtained by the deformation correction processing unit 124. Each component will be described below.

[0025] The server 101 contains an interface 115, a processor 116, a non-volatile memory (ROM) 117, a volatile memory (RAM) 118, a storage 126, and the like. The input / output device 103 outputs user-specified data 132, including designation of a cross-sectional image to be measured and designation of a measurement location, to the server 101 via the interface 115, or accepts dimension measurement results 130 output from the server 101. The stored data 131 input to the server 101 includes image data such as a cross-sectional SEM image, annotation data (including segmentation images and label information attached to the segmentation images), model parameters representing a trained model, processed data (segmentation images), and measurement results. The stored data 131 also includes information defining the measurement location, processing conditions for the semiconductor device corresponding to the measurement image, and imaging conditions for the cross-sectional SEM image. The stored data 131 is stored in the database 102 and retrieved when needed. Although not shown, the input / output device 103 is provided with input / output devices such as a mouse for operating the GUI screen and a keyboard for inputting various setting values.

[0026] The storage 126 stores an annotation tool 127 and dimension measurement software 128 having the dimension measurement function of this embodiment. The annotation tool 127 may be installed in the server 101 or in any or all of the input / output devices 103. When the annotation tool 127 is installed in the server 101, the tool is operated from the input / output device 103. The dimension measurement software 128 is loaded into the RAM 118 as needed. The processor 116 executes the dimension measurement software 128, causing the dimension measurement system 110 to realize the dimension measurement function.

[0027] Furthermore, the dimension measurement software 128 of this embodiment mainly includes a learning unit 122, an inference unit 123, a deformation correction processing unit 124, and a dimension measurement unit 125, which are incorporated into the dimension measurement software 128 in the form of software modules. Fig. 2 shows a state in which the learning unit 122, the inference unit 123, the deformation correction processing unit 124, and the dimension measurement unit 125 are deployed in the RAM 118. Note that the term "memory" includes the ROM 117 and the RAM 118 in which the dimension measurement software 128 is deployed, as well as the storage 126 in which the dimension measurement software 128 is stored.

[0028] <Example of Semantic Segmentation Model> Next, a semantic segmentation model will be described with reference to FIG. 3 . FIG. 3 is a diagram showing an example of the semantic segmentation model. The semantic segmentation model is used in the learning unit 122 and the inference unit 123 of the server 101. Note that in this embodiment, an example using a neural network 30 will be described, but this is not limiting, and machine learning models such as decision trees can also be used. In the neural network 30, pixel information input to the input layer is propagated to the intermediate layer and the output layer in order and calculated, and a label number of the area to which each pixel belongs (the label number will be described later) is output from the output layer. The intermediate layer is made up of multiple repetitions of convolutional layers, pooling layers, etc. The specific layer structure varies depending on the model employed. During learning, the parameters of the intermediate layer are adjusted so that the error between the output label of each pixel and the correct annotation data is minimized.

[0029] During training of the semantic segmentation model, parameters of the intermediate layer are adjusted so as to minimize the error between the label of each output pixel and the correct annotation data. The semantic segmentation model shown in Fig. 3 is implemented in the training unit 122 and the inference unit 123 of Fig. 2, and the main body of the dimension measurement software 128 itself is stored in the storage 126 of Fig. 2. During training and inference, the semantic segmentation model of Fig. 3 is deployed in the RAM 118 and executed by the processor 116.

[0030] <Cross-Sectional Image> Next, with reference to FIGS. 4 to 8 , a cross-sectional image to be processed by the dimension measurement method will be described. FIG. 4 is a diagram illustrating an example of a cross-sectional image included in the saved data 131. Specifically, FIG. 4 illustrates an example of a cross-sectional SEM image included in the training data set. In this example, the image size is 1280 pixels horizontally and 960 pixels vertically. The sample has a trench structure in which unit patterns of lines and spaces are repeated, with silicon dioxide as the mask 302 and silicon as the substrate 301. The image includes three regions: a background 300, the substrate 301, and the mask 302. Note that, although the present disclosure describes a case in which an SEM image (hereinafter also referred to as a "cross-sectional SEM image") is used as the cross-sectional image, the present disclosure is not limited thereto. The cross-sectional image may also be a TEM (Transmission Electron Microscope) image. Hereinafter, each structure consisting of the mask 302 and the underlying columnar portion of the substrate 301 will be referred to as a "line pattern," and each background 300 of the etched portion of the substrate 301 will be referred to as a "trench." In this disclosure, the target structure for dimension measurement will be a line pattern, but this disclosure is not limited to this case. This disclosure can also be applied to structures other than line patterns. With respect to a line pattern, the direction in which the substrate 301 and the mask 302 overlap will be referred to as the up-down direction or vertical direction, and the direction perpendicular to the up-down direction will be referred to as the left-right direction or horizontal direction. For example, it may be expressed as if the line pattern extends in the up-down direction (vertical direction), and the bottom of the trench is located in the left-right direction (horizontal direction) of the line pattern. In the following description, in relation to a trench structure, a case will be described in which the direction in which the columnar portions of the trench structure extend (depth direction) will be referred to as the up-down direction (vertical direction), and the direction in which the columnar portions are repeated (direction perpendicular to the depth direction) will be referred to as the left-right direction (horizontal direction), but the method of setting the directions is not limited to this case.

[0031] FIG. 5 shows an example of a segmentation image, which is annotation data included in a training dataset. The segmentation image is created either manually by a user using a dedicated annotation tool for a cross-sectional SEM image, which is input data to the model, or by using a trained semantic segmentation model. The segmentation image is an image in which a label number for each region is assigned to each pixel. In the present disclosure, three regions, namely, a background, a mask, and a substrate, are extracted from the cross-sectional SEM image, and the segmentation image is divided into three regions: a background region 310, a substrate region 311, and a mask region 312. The background region 310, the substrate region 311, and the mask region 312 in the segmentation image correspond to the background 300, the substrate 301, and the mask 302, respectively, of the cross-sectional SEM image.

[0032] FIG. 6 is a table showing the correspondence between label names, label numbers, and colors in annotation data used to train a semantic segmentation model. The table shown in FIG. 6 is stored in the database 102. The label names of the three regions are "background," "substrate," and "mask," the label numbers are "0," "1," and "2," and the colors are "black," "gray," and "white." The label names, label numbers, and colors assigned to each label are arbitrary, but must be fixed within a single dataset. Each pixel included in the segmentation image is assigned one of the three label numbers included in the correspondence.

[0033] Fig. 7 is a diagram showing an example of an image input to the semantic segmentation model. Fig. 8 is a diagram showing a segmentation image for the input image. After training the model using the training data sets such as those shown in Figs. 4 and 5, when an image to be measured (cross-sectional SEM image 303) (Fig. 7) is input to the trained model, a segmentation image 304 (Fig. 8) corresponding to the image is obtained as an inference result.

[0034] The tilt of the entire image is then corrected. Typically, measurement engineers capture images so that the substrate surface is horizontal, but manual capture can result in some tilt. In this embodiment, dimensions are measured assuming that the substrate surface is horizontal, so the image is corrected so that the substrate surface is horizontal. The required rotation angle is obtained by calculating the coordinate values ​​of the vertices in the mask area 312 of the segmentation image (in FIG. 8 , the vertices are the uppermost points of each of the 14 line patterns (when the direction from the substrate area 311 to the mask area 312 is considered to be upward)), and then using a regression equation to find the slope of the line passing through these vertices. The image is rotated around the center of the image as the origin so that the corrected substrate surface is horizontal. The same rotation is performed on the original cross-sectional SEM image 303. To ensure that the rotated image is the same size as the original image, any excess portions are deleted and any missing portions are interpolated by extrapolation.

[0035] <Method of Cutting Contacting Points> Next, a method of cutting contacting points among dimension measurement methods will be described with reference to FIGS. 9 to 11 . As post-processing of the segmentation image 304, contacting points of line patterns are detected and cut. FIG. 9 is an enlarged view of a portion of the segmentation image illustrating the start point A and end point B of the contacting points. FIG. 9 is a view illustrating how to determine the cutting points. FIG. 9 shows the vicinity of mask regions 312 of two contacting line patterns cut out from FIG. 8. FIG. 9 shows how the masks of adjacent line patterns are in contact with each other. First, the trench side is traced upward from the bottom, and if there is no dead end, it is determined that there is a contacting point, and the dead end point is set as point A 313, which is the start point of the contacting point. To determine whether a dead end has been reached, for example, the process is carried out by selecting pixels at the boundary where the label changes (here, the boundary between the substrate area 311 or the mask area 312 and the background area 310) and tracing upward (from the substrate area 311 to the mask area 312), and the point A313 above which there is no boundary part of the label is set as the starting point of the contact point of the mask.

[0036] Next, the pixel in the background region 310 closest to point A is searched upward, and this point is set as the end point B 314 of the contact point of the line pattern. FIG. 10 shows an enlarged view of a portion of the segmentation image after the contact point has been cut with a line segment AB. FIG. 10 is a diagram explaining the cutting method. A line segment AB is drawn connecting the start point A 313 and end point B 314 of the contact point, and the label of the pixel where the line segment intersects is changed from the "mask" label to the "background" label. This cuts off the mask region 312 in the segmentation image.

[0037] Fig. 11 is a diagram showing a segmentation image after cutting the contact points in Fig. 8. A segmentation image 304a is an image obtained by performing cutting processing on the contact points of the line patterns in the segmentation image 304 in Fig. 8.

[0038] <Deformation Correction Method Using Contour Lines> Next, a deformation correction method using contour lines will be described with reference to Figs. 12 to 14. Next, the deformation correction processing unit 124 obtains contour lines from the corrected segmentation image 304a. In addition, a portion in the segmentation image 304a where two adjacent pixels have different labels is detected as a contour line, and the contour line is classified by using the combination of the labels of the two pixels.

[0039] FIG. 12 is a diagram showing contours determined from the segmentation image 304a after cutting. FIG. 12 shows a contour map 323 obtained from the segmentation image 304a of FIG. 11. The classified contours are a 0th contour 320 which is the boundary between the background region 310 and the substrate region 311, a 1st contour 321 which is the boundary between the background region 310 and the mask region 312, and a 2nd contour 322 which is the boundary between the substrate region 311 and the mask region 312. In FIG. 12, the 0th contour 320 is displayed in black, the 1st contour 321 in dark gray, and the 2nd contour 322 in light gray. By classifying the contours, pattern division of the contour data is performed automatically.

[0040] Next, the deformation correction processing unit 124 performs deformation correction on the contour lines shown in Fig. 12. Fig. 13 is a diagram showing a contour map 324 after the deformation correction on the contour lines. When the contact points of the line patterns are divided, the line patterns are deformed from their original shapes, and therefore the contour lines are also displayed in a deformed state. By performing deformation correction on the contour lines, it is possible to display contour lines that can be estimated to be the contour lines of the original, undistorted shape of the line patterns.

[0041] FIG. 14 is a diagram illustrating a method for correcting the shape of a deformed contour line (hereinafter also referred to as a "deformed pattern"). FIG. 14 shows the contour line for one line pattern. For simplicity of explanation, FIG. 14 omits the second contour line 322, and shows the 0th contour line 320 and the first contour line 321 without distinction. The x-axis corresponds to the substrate surface, and the y-axis corresponds to the direction perpendicular to the substrate surface. The positive y-axis direction is the direction from the mask toward the substrate. The line pattern is a trench pattern or a line-and-space pattern. When the pattern is a trench pattern, the contour line data is corrected by calculating the amount of displacement from a reference position 333 (details will be described later) for each coordinate of the contour line data in the depth direction of the trench, and subtracting the calculated amount of displacement from the coordinate of the contour line data in the direction perpendicular to the depth direction of the trench. First, a curve corresponding to the center line 330 in the x-direction is calculated from the x-coordinate of the contour line of the line pattern to be corrected (hereinafter also referred to as a "target line"). The center line 330 is a line connecting the midpoints of the x-coordinates of two points on the 0th contour line 320 and the 1st contour line 321 that share a common y-coordinate value. Next, the midpoint of the x-coordinates of the bottoms 332 of the trenches on both sides of the line pattern in the x-axis direction is set as the original reference position 333 of the target line. Next, a correction region 334 is set from the shallower of the left and right trench bottoms (the trench bottom with the smaller y-coordinate value) to the vertex of the target line (the point with the smallest y-coordinate). Next, for each y in the correction region 334, a displacement d(y) 335 of the center line 330 from the reference position 333 is calculated. Next, the x-coordinate of the contour data, which indicates the coordinates of the contour of the target line, is shifted by -d(y) (moved in a direction parallel to the x-axis direction) to obtain the corrected contour, thereby correcting the deformed pattern. The displacement d(y) of each deformed pattern can also be used to quantitatively evaluate the deformation. Furthermore, the method of setting the x-axis and y-axis is not limited to the above. For example, the positive direction of the y-axis may be the direction from the substrate surface toward the mask.

[0042] <Dimension Measurement Method> Next, a method for measuring dimensions using the contour data of the corrected contour line will be described with reference to FIGS. 15 to 21 . In this disclosure, an example is described in which the contour data of the classified contour line is converted in advance into a data format that allows dimensions to be easily calculated. However, a method in which the contour data is saved as is and the data is read during measurement and then processed to determine the points required for dimension measurement may also be used. The endpoints (feature points) typically required for dimension measurement are the top and bottom ends of the contour line, such as the vertex of a mask or the bottom of a trench. Including the left and right, four points are required: top, bottom, left, and right. Furthermore, if the contour line is viewed as a single open curve, two endpoints (hereinafter referred to as the start point and end point) can be defined. Therefore, the required feature points are a total of six points: the top, bottom, left, and right endpoints, as well as the start and end points.

[0043] FIG. 15 is a diagram illustrating the definitions of six types of feature points extracted from a segmented contour. Specifically, FIG. 15 illustrates the positions of the six feature points for the first contour 321 in a mask portion included in a segmentation image. The vertical (up-down) axis is the y-axis, and the horizontal (left-right) axis is the x-axis. The six feature points are point T336 (the vertex of the mask, with the smallest y-coordinate), point B337 (the bottom of the mask, with the largest y-coordinate), point L338 (the leftmost point of the mask, with the smallest x-coordinate), point R339 (the rightmost point of the mask, with the largest x-coordinate), point S340 (the starting point S of the first contour 321 (the point with the smaller x-coordinate of the two end points)), and point E341 (the ending point of the first contour 321 (the point with the larger x-coordinate of the two end points)). Depending on the shape of the first contour line 321, more than one of these six points may coincide. In the example of FIG. 15 , point B337 coincides with point S340. The deformation correction processing unit 124 stores the determined feature points that satisfy the predetermined criteria in the database 102. During the inference step, the inference unit 123 determines the coordinates of these six feature points for each divided contour line and stores them in a file in the database 102.

[0044] In addition to the illustrated first contour line 321, feature points are also extracted from the zeroth contour line 320 and the second contour line 322. Although the case where six feature points are extracted from the first contour line 321 has been described, the present disclosure is not limited to this. Seven or more feature points may be defined and extracted.

[0045] The measurement location is specified by two points on the contour line selected by the user. During measurement, the measurement direction is fixed, for example, from left to right or top to bottom, and the measurement start point and measurement end point are specified by the symbol "contour line number_feature point symbol." The measurement direction is fixed so that the relative positions of the feature point symbols at the measurement start point and measurement end point can be used to determine whether the measurement target is inside or outside the contour line. That is, if the measurement start point is point L and the measurement end point is point R, the measurement direction is correct, and the measurement target is the distance inside the contour line. Conversely, if the measurement start point is point R and the measurement end point is point L, the measurement target is determined to be outside the contour line, i.e., the distance between two adjacent contour lines. Furthermore, if the measurement start point and measurement end point have the same symbol, the measurement target is determined to be the distance between adjacent contour lines with the same number.

[0046] <Correction of Deformed Pattern in Cross-Sectional SEM Image> If necessary, cross-sectional SEM image correction can be performed in conjunction with contour line correction as follows. The cross-sectional image is divided into a foreground image (foreground image 345) containing the pattern and a background image (background image 347) other than the foreground using a segmentation image, and a foreground image (corrected foreground image 346) corrected using the displacement is created. In addition, a second background image (artificial background image 348) is created in place of the background image, and the corrected foreground image and the second background image are combined to create a corrected cross-sectional image (corrected image 349). This correction utilizes a characteristic of cross-sectional SEM images, namely, that the image consists of two parts: the foreground, which is the workpiece, and the background, which contains no object, and therefore correction of the foreground does not result in inconsistency with the background, which contains no object.

[0047] Fig. 16 is a diagram showing foreground image 345 obtained by cutting out the foreground portion from the input image. Fig. 17 is a diagram showing foreground image 346 obtained by performing pattern deformation correction on foreground image 345 cut out from the input image. Fig. 18 is a diagram showing background image 347 obtained by cutting out the background portion from the input image. Fig. 19 is a diagram showing artificial background image 348 created from background image 347 cut out from the input image.

[0048] First, the deformation correction processing unit 124 separates the input cross-sectional SEM image (cross-sectional SEM image 303 in FIG. 7) into a foreground image 345 (FIG. 16) consisting of the mask and the substrate and a background image 347 (FIG. 18) consisting of the rest, using a contour map 323, which is a segmentation image in which the contact points in FIG. 12 have been cut out.

[0049] For the foreground image 345, the same correction as that applied to the contour lines is applied to each line pattern to correct deformation of the line patterns, and a corrected foreground image 346 (FIG. 17) is obtained.

[0050] An artificial background image (second background image) 348 (FIG. 19) is created from the background image 347. The artificial background image 348 is, for example, an image having a single luminance distribution or an arbitrary luminance distribution.

[0051] Finally, the corrected foreground image 346 is overlaid on the artificial background image 348 to generate a corrected image.

[0052] 20 is a diagram showing an example of a generated corrected image 349. The deformation of the line pattern has been corrected in the corrected image 349. To distinguish the corrected image 349 from the original input image (the cross-sectional SEM image 303 in FIG. 7), it is desirable to clearly indicate that the corrected image 349 is an artificial image, such as by labeling it as 'corrected image'.

[0053] The necessity of the artificial background image 348 and how to create it will be explained below. Correcting the foreground image 345 requires a new image with background pixels. The brightness of the background changes gradually in the y direction (it is dark at the top of the screen and slightly brighter in the trench at the bottom), but is highly position-dependent in the x direction due to the influence of scattered electrons from the edges of the processed pattern (it is slightly bright near the line pattern and darker away). For this reason, it is difficult to accurately calculate the brightness of the pixels in the new background image.

[0054] Therefore, an artificial background image 348, resembling wallpaper, is created for the entire image, including pixels unaffected by the correction of the foreground image 345. The artificial background image 348 may be a uniform image with the same brightness, but if an image closer to the background of the actual image is desirable, it is created based on the background image 347. Various methods for creating the artificial background image are conceivable, and one example is described below. First, for all pixels included in the background image 347, the mean brightness μ(y) and variance σ(y) of the pixels in the background image 347 in the y direction are calculated. Next, random brightness values ​​according to a normal distribution with mean μ(y) and variance σ(y) are assigned to all pixels at x coordinates corresponding to each y value. As a result, the brightness distribution in the y direction of the artificial background image 348 approximately matches the mean brightness distribution μ(y) of the original image (cross-sectional SEM image 303 in FIG. 7 ).

[0055] 21 shows an example of dimension measurement performed using six feature points extracted from corrected contour data, and the measurement values ​​printed and output on a corrected image. A dimension print image 350 is an image in which dimensions are printed by measuring each line pattern included in the corrected image 349 using feature points. The numerical values ​​are actual dimensions, and although units are omitted, they are in nanometers (nm). In this example, three dimension measurement locations are specified: mask height ([1_T, 2_T]), trench depth ([0_T, 0_B]), and minimum line width ([0_R, 0_L]). The mask height ([1_T, 2_T]) indicates the length between point T (the point with the smallest y coordinate) on the first contour line 321, which is the boundary between the background region 310 and the mask region 312, and point T on the second contour line 322, which is the boundary between the substrate region 311 and the mask region 312. In Figure 21, the measured values ​​are 37.7 nm, 39.0 nm, 38.4 nm, 38.7 nm, 39.7 nm, 39.7 nm, 39.0 nm, 39.4 nm, 38.7 nm, 38.4 nm, 38.0 nm, 38.0 nm, 38.7 nm, and 38.7 nm from the left side of the paper. It also shows the trench depth ([0_T, 0_B]) and the length between point T on the 0th contour line 320, which is the boundary between the background region 310 and the substrate region 311, and point B on the 0th contour line 320 (the point with the largest y coordinate), and in Figure 21, from the left side of the paper, the measured values ​​are 225.6 nm, 197.4 nm, 223.9 nm, 201.1 nm, 218.6 nm, 197.1 nm, 228.5 nm, 188.5 nm, 228.5 nm, 189.5 nm, 222.2 nm, 192.8 nm, 216.3 nm, 203.1 nm, and 220.6 nm. Furthermore, the minimum line width ([0_R, 0_L]) indicates the length between point R (the point with the maximum x-coordinate) on the 0th contour line 320 and point L (the point with the minimum x-coordinate) on the 0th contour line 320, and in Figure 21, from the left side of the paper, the measured values ​​are 8.6 nm, 7.0 nm, 7.3 nm, 7.3 nm, 8.6 nm, 7.6 nm, 8.3 nm, 7.6 nm, 7.9 nm, 7.0 nm, 7.6 nm, 7.0 nm, 7.6 nm, and 7.6 nm.

[0056] In this way, by using this technique, even if a cross-sectional SEM image (cross-sectional SEM image 303 in FIG. 7 ) in which the line pattern has been deformed is obtained, it is possible to reproduce an image equivalent to the line pattern before deformation, and it is also possible to measure dimensions equivalent to the dimensions of the pattern before deformation. Furthermore, it is also possible to generate an image in which the deformation of the line pattern has been corrected. The corrected image 349 in which the deformation has been corrected can also be used as an input image for other dimension measurement tools or measurement systems. Furthermore, because the corrected line pattern roughly reproduces the line pattern before deformation, it is possible to display dimensions superimposed on the corrected line pattern, making it easier for the user to recognize the line pattern and the measurement points for the dimensions.

[0057] <Flow from Learning to Measurement> Next, the processing flow from learning to measurement will be described with reference to Fig. 22. Fig. 22 is a flowchart showing an example of a processing procedure in the dimension measurement system 110. Fig. 22 shows a flowchart from the start of learning to measuring the dimensions of an input image. Fig. 22 is a flowchart executed by the server 101 from the start of learning to measuring the dimensions of an input image. The learning step corresponds to steps S100 and S101. The inference step corresponds to steps S102 to S108 and steps S112 and S114, and the measurement step corresponds to steps S109 to S111 and S115.

[0058] First, stored data 131 including processing conditions, imaging conditions, and a training data set is prepared in the server 101. The training data set is a set of cross-sectional SEM images, which are input data, and annotation images (segmentation images in this disclosure) that are color-coded for each region. These training data sets are specified from the input / output device 103 (step S100).

[0059] Next, the processor 116 transfers the training data set from the database 102 to the model training unit 122, and causes the training unit 122 to train the model (step S101). The trained model parameters are stored in the database 102.

[0060] Next, in the inference step, an image for measuring dimensions is input from the input / output device 103 (step S102). The magnification of the input image is read from the database 102. The processor 116 passes the model and trained model parameters from the database 102 to the inference unit 123, as well as the input image, which is then inferred by the inference unit 123 (step S103). As a result of the inference, a segmentation image, which is an image that has been subjected to semantic segmentation, is obtained (step S104).

[0061] Next, the inference unit 123 performs rotation correction of the tilt of the entire segmentation image so that the surface of the substrate becomes horizontal (step S105). Specifically, after determining the contour line, the inference unit 123 performs processing to rotate the segmentation image and correct the tilt so that the surface of the cross-sectional structure in the segmentation image (for example, the straight line connecting the uppermost points of a line pattern) becomes horizontal with respect to the image.

[0062] Next, the deformation correction processing unit 124 cuts off contact points of the line patterns in the segmentation image (step S106). Next, the classified contour lines are extracted from the segmentation image after cutting, and the line pattern deformation is corrected (step S107). At this time, the deformation correction processing unit 124 also extracts feature points from the corrected line patterns. After the correction, the inference unit 123 converts the format of the contour line data (step S108). The converted contour line data (hereinafter also referred to as "converted data") and data indicating the feature points are stored in the database 102.

[0063] Next, in the measurement step, when the user inputs the required measurement locations and whether image correction is required (step S109), the processor 116 reads the required converted data from the database 102 (step S110) and passes it to the dimension measurement unit 125. The dimension measurement unit 125 obtains the dimensions by reading the data required for measuring (calculating the dimensions) at the specified measurement locations from the database 102. The dimension measurement unit 125 calculates the dimensions based on the specified measurement locations and segmentation images, etc. The specified measurement locations are specified by the user, for example, as described below. Since the units of the obtained dimensions are pixels, the imaging magnification stored in the database 102 is called up and converted into actual dimensions (step S111).

[0064] The processor 116 checks whether image correction is necessary (step S112), and if correction is not necessary, the processor 116 outputs the measured dimension results to the input / output device 103 and stores the numerical values ​​in the database 102 (step S113).

[0065] If image correction is required, the deformation correction processing unit 124 corrects the image (step S114). The dimension measurement unit 125 prints the measured dimension results on the corrected image and outputs them to the input / output device 103, and also stores the corrected image and the dimension values ​​in the database 102 (step S115).

[0066] The processor 116 determines whether the necessary measurements have been completed (step S116). If they have not been completed and an image to be measured needs to be added, the process returns to step S102 and the process is repeated. If the measurements have been completed, the process ends (step S117).

[0067] <GUI displayed in dimension measurement system during learning step> Next, the GUI in the dimension measurement system 110 will be described with reference to Figs. 23 and 24. Fig. 23 is a diagram showing an example of a GUI screen displayed during model learning. The GUI screen is displayed, for example, on the input / output device 103. The GUI screen shown in Fig. 23 is configured so that a learning screen used during learning and a measurement screen used when executing inference and measurement can be switched using tabs, and this screen is displayed when the learning tab 400 displayed as "train" is selected. A group of buttons for instructing module execution and model learning is arranged on the upper row, and a terminal display window (terminal window) 406 is arranged on the lower row.

[0068] First, in specifying the training data (step S100 in FIG. 22), the user selects the input folder specification button 410 to specify the folder storing the training data, and specifies the folder from the dialog that appears. The specified folder name is displayed in the input folder name display cell 413. Next, the user presses the output folder specification button 411 to specify the folder in which to store the trained model after training, and specifies the folder from the dialog that appears. The specified folder name is displayed in the output folder name display cell 414. To cancel the specified folder name, the user selects the clear button 412. Then, the user operates the input folder specification button 410 or the output folder specification button 411 again.

[0069] To start model learning (step S101 in FIG. 22), the user selects a learning start button 402. A status display cell 405 indicating the status of the learning process is displayed next to the learning start button 202. When "Finished" is displayed in the status display cell 405, the learning step of step S101 is complete. Since learning takes time, the progress of the learning is output to a terminal display window 406 as needed, and can be checked by the user.

[0070] <GUI displayed on the measurement system during the inference / measurement step> Figure 24 is a diagram showing an example of a GUI screen displayed on the input / output device 103 when the inference / measurement step is performed. The GUI screen shown in Figure 24 can be switched between and displayed with the GUI screen shown in Figure 23 using tabs, and when the measurement tab 401 labeled "measure" is selected, the screen shown in Figure 24 is displayed. The screen includes an inference window 420, a contour window 421, a measurement point setting window 422, a measurement result window 423, a measurement value window 424, and an image correction window 440.

[0071] In step S102 of inputting an image for measurement, the user selects the folder button 425 in the inference window 420 and selects the folder containing the image to be measured from the dialog box that appears. The name of the selected folder is displayed in the folder cell 428 on the right. The clear button 426 is used to cancel the folder selection. In step S103 of model-based estimation (step S103 of FIG. 22), the user selects the inference start button 427 to start inference. Once the inference is complete, the status is displayed in the status display cell 429 on the right. During the inference, the following steps are automatically performed: output of a segmentation image, which is the inference result (step S104 of FIG. 22), image tilt correction (step S105 of FIG. 22), cutting of contact points (step S106 of FIG. 22), contour extraction and pattern deformation correction (step S107 of FIG. 22), and data conversion (step S108 of FIG. 22).

[0072] Upon completion of the inference, a divided contour map 430 for one measurement image is displayed in different colors in the contour window 421. The user can confirm the correspondence between each contour number and the color coding.

[0073] In the measurement point input (step S109 in FIG. 22 ), the user inputs the name, measurement start point, and measurement end point of the measurement point into the measurement point cell 431 in the measurement point setting window 422. The naming method is arbitrary. The measurement start point and measurement end point are specified using the symbols "contour number_feature point symbol." Note that FIG. 22 shows an example in which the user inputs the measurement start point and measurement end point into the measurement point cell 431 using a keyboard. However, since the options are determined by the number of contour line types and the number of feature point symbol types, they may also be implemented so that selection is made from a pull-down menu. Alternatively, the user may click the corresponding point on the contour map 430 with the mouse to detect its coordinate values, and the symbol names of the measurement start point and measurement end point may be automatically entered into the measurement point cell 431.

[0074] Next, the user specifies whether or not to correct deformation in the image in the image correction window 440 (step S112 in FIG. 22). If correction is to be performed, the user selects the execute button 441, and if correction is not to be performed, the user selects the not execute button 442.

[0075] Next, the user starts measurement by pressing the measurement start button 432 in the measurement value window 424. In step S110, data required for measurement is read for all images contained in the folder cell 428, dimensions are calculated, and after dimension scale conversion in step S111, the status is displayed in the status display cell 435 on the right.

[0076] In the dimension output / save process (step S113 in FIG. 22 ), the measurement results are saved in the database 102. To display the results, the user selects the desired measurement image by pressing the select button 433. Upon selection, the image name is displayed in the sample name display cell 438 in the measurement result window 423. If the execute button 441 is selected, a corrected image with measurement values ​​437, in which measurement values ​​are added to the measurement image, is displayed. If the no execute button 442 is selected, the corrected image with measurement values ​​437 is not displayed. Furthermore, the processing conditions for this image are retrieved from the database 102, and the processing recipe is displayed in the recipe display cell 439. At the same time, statistical values ​​such as the number of detected particles, the average dimension value, and the standard deviation are displayed in the measurement value display cell 436. To display the results for a different image, the user selects the select button 433 and reselects the measurement image. To clear the displayed results, the user selects the clear button 434.

[0077] It should be noted that the present invention is not limited to the above-described embodiments, and includes various modifications and equivalent configurations within the spirit and scope of the appended claims. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to configurations including all of the described configurations. Furthermore, part of the configuration of the embodiments may be replaced with other configurations. Furthermore, other configurations may be added to the configuration of the embodiments. Furthermore, other configurations may be added to, deleted from, or substituted for part of the configuration of the embodiments.

[0078] Furthermore, the above-described configurations, functions, processing units, processing means, etc. may be partly or entirely implemented in hardware, for example, by designing them as integrated circuits, or may be implemented in software by a processor interpreting and executing a program that implements each function. Information such as the program, table, and file that implements each function can be stored in a storage device such as a memory, a hard disk, or an SSD (Solid State Drive), or in a recording medium such as an IC (Integrated Circuit) card, an SD card, or a DVD (Digital Versatile Disc).

[0079] In addition, the control lines and information lines shown are those that are considered necessary for explanation, and do not necessarily represent all the control lines and information lines that are necessary for implementation. In reality, it can be assumed that almost all components are interconnected.

[0080] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the gist of the present invention.

[0081] The following are embodiments that may be included in the present invention, but the present invention is not limited to these. (Aspect 1) A dimension measurement device that measures dimensions of a desired portion of an etched pattern using a cross-sectional image including the pattern, wherein when the pattern is in contact with another pattern, the contacting portion of the patterns is cut, and then a segmentation image of the pattern with the contacting portion cut is obtained using a semantic segmentation model, contour data of the pattern is obtained using the segmentation image, the position of the pattern before the contact is estimated using the contour data, the amount of displacement between the estimated position and the contour is calculated, the contour data is corrected based on the calculated amount of displacement, and the dimension of the desired portion of the pattern is measured based on the corrected contour data. (Aspect 2) The dimension measurement device described in Aspect 1, wherein the cross-sectional image is divided into an image of a foreground portion including the pattern and an image of a background portion other than the foreground portion using the segmentation image, and an image of the foreground portion corrected using the amount of displacement is created. (Aspect 3) The dimension measurement device according to Aspect 1 or Aspect 2, wherein a second background image is created in place of the background image, and the corrected cross-sectional image is created by combining the corrected foreground image and the second background image. (Aspect 4) The dimension measurement device according to any one of Aspects 1 to 4, wherein the pattern is a trench pattern or a line and space pattern, and when the pattern is the trench pattern, an amount of displacement from the position is calculated for each coordinate of the contour data in the depth direction of the trench, and the contour data is corrected by subtracting the calculated amount of displacement from the coordinate of the contour data in a direction perpendicular to the depth direction of the trench. (Aspect 5) The dimension measurement device according to any one of Aspects 1 to 4, wherein the second background image is an image having a single luminance distribution or an arbitrary luminance distribution.(Aspect 6) The dimension measuring device according to any one of Aspects 1 to 5, wherein the cross-sectional image is a cross-sectional SEM (Scanning Electron Microscope) image or a TEM (Transmission Electron Microscope) image. (Aspect 7) A dimension measurement method for measuring dimensions of a desired portion of an etched pattern using a cross-sectional image including the pattern, wherein when the pattern is in contact with another pattern, the dimension measurement method comprises the steps of: after the contacting portion of the pattern is cut, using a semantic segmentation model to obtain a segmentation image of the pattern with the contacting portion cut; acquiring contour data of the pattern using the segmentation image; estimating the position of the pattern before contact using the contour data; determining the amount of displacement between the estimated position and the contour; correcting the contour data based on the determined amount of displacement; and measuring the dimensions of the desired portion of the pattern based on the corrected contour data. (Aspect 8) A semiconductor device manufacturing system having a platform on which an application for measuring dimensions of a desired portion of an etched pattern using a cross-sectional image including the pattern is implemented, wherein, when the pattern is in contact with another pattern, the application performs the following steps: after the contact point of the pattern is cut, a segmentation image of the pattern with the contact point cut is obtained using a semantic segmentation model; contour data of the pattern is obtained using the segmentation image; the position of the pattern before contact is estimated using the contour data; the amount of displacement between the estimated position and the contour; correcting the contour data based on the amount of displacement; and measuring the dimensions of the desired portion of the pattern based on the corrected contour data.(Aspect 9) The semiconductor device manufacturing system according to Aspect 8, wherein a semiconductor manufacturing apparatus and a charged particle beam apparatus that acquires a cross-sectional image including a pattern etched by the semiconductor manufacturing apparatus are connected to the platform. (Aspect 10) The semiconductor device manufacturing system according to Aspect 8 or Aspect 9, wherein the semiconductor manufacturing apparatus is a plasma etching apparatus, and the charged particle beam apparatus is a cross-sectional SEM (Scanning Electron Microscope) apparatus. (Aspect 11) The semiconductor device manufacturing system according to any one of Aspects 8 to 10, wherein the platform is a server.

[0082] 10 Semiconductor device manufacturing system, 30 Neural network, 100 Evaluation device, 101 Server, 102 Database, 103 Input / output device (terminal or PC), 110 Dimension measurement system, 111 Processing device, 115 Interface, 116 Processor, 117 ROM, 118 RAM, 122: Learning unit, 123 Inference unit, 124 Deformation correction processing unit, 125 Dimension measurement unit, 126 Storage, 127 Annotation tool, 128 Dimension measurement software, 130 Dimension measurement results, 131 Saved data, 132 Designated data, 300 Background, 301 Substrate, 302 Mask, 303 Cross-sectional SEM image, 304 Segmentation image, 304a Segmentation image, 310 Background region, 311 Substrate region, 312 Mask region, 313 Start point of contact location, 314 End point of contact point, 320 0th contour line, 321 1st contour line, 322 2nd contour line, 323 Contour line map, 324 Contour line map after deformation correction, 330 Center line of contour line relative to target line, 332 Bottom of trenches on both sides of target line, 333 Reference position of target line, 334 Area where inclination should be corrected (correction area), 335 Displacement amount from original center position, 336 T point, 337 B point, 338 L point, 339 R point, 340 S point, 341 E point, 345 Foreground image, 346 Foreground image after correction, 347 Background image, 348 Artificial background image, 349 Corrected image, 350 Dimension print image, 400 Learning tab, 401 Measurement tab, 402 Start learning button, 405 Status display cell, 406 Terminal display window, 410 Input folder specification button, 411 Output folder specification button, 412 Clear button, 413 Input folder name display cell, 414 Output folder name display cell, 420 Inference window, 421 Contour window, 422 Measurement point setting window, 423 Measurement result window, 424 Measurement value window, 425 Folder button, 426 Clear button, 427 Inference start button, 428 Folder cell, 429 Status display cell, 430 Contour diagram, 431 Measurement point cell, 432 Measurement start button, 433 Select button, 434 Clear button, 435 Status display cell, 436 Measurement value display cell, 437 Corrected image with measurement value, 438 Sample name display cell,439 Recipe display cell, 440 Image correction window, 441 Execute button, 442 Non-execute button,

Claims

1. A dimension measurement device that measures the dimensions of a desired portion of an etched pattern using a cross-sectional image including the pattern, wherein when the pattern is in contact with another pattern, the contacting portion of the pattern is cut off, and then a segmentation image of the pattern with the contacting portion cut off is obtained using a semantic segmentation model, contour line data of the pattern is obtained using the segmentation image, the position of the pattern before contact is estimated using the contour line data, the amount of displacement between the estimated position and the contour line is determined, the contour line data is corrected based on the determined amount of displacement, and the dimension of the desired portion of the pattern is measured based on the corrected contour line data.

2. A dimension measuring device according to claim 1, wherein the cross-sectional image is divided into an image of a foreground portion including the pattern and an image of a background portion other than the foreground portion using the segmentation image, and an image of the foreground portion corrected using the amount of displacement is created.

3. A dimension measuring device according to claim 2, wherein a second background image is created in place of the background image, and the corrected cross-sectional image is created by combining the corrected foreground image and the second background image.

4. A dimension measuring device according to claim 1, wherein the pattern is a trench pattern or a line and space pattern, and when the pattern is a trench pattern, a displacement amount from the position is calculated for each coordinate of the contour line data in the depth direction of the trench, and the contour line data is corrected by subtracting the calculated displacement amount from the coordinate of the contour line data in a direction perpendicular to the depth direction of the trench.

5. A dimension measuring device according to claim 3, wherein the image of the second background portion is an image having a single luminance distribution or an arbitrary luminance distribution.

6. A dimension measuring device according to claim 1, wherein the cross-sectional image is a cross-sectional SEM (Scanning Electron Microscope) image or a TEM (Transmission Electron Microscope) image.

7. A dimension measurement method for measuring dimensions of a desired portion of an etched pattern using a cross-sectional image including the pattern, comprising, when the pattern is in contact with another pattern, the steps of: after the contacting portion of the pattern is cut, obtaining a segmentation image of the pattern with the contacting portion cut using a semantic segmentation model; acquiring contour data of the pattern using the segmentation image; estimating the position of the pattern before contact using the contour data; determining the amount of displacement between the estimated position and the contour; correcting the contour data based on the determined amount of displacement; and measuring the dimensions of the desired portion of the pattern based on the corrected contour data.

8. A semiconductor device manufacturing system having a platform on which an application is implemented for measuring dimensions of a desired portion of an etched pattern using a cross-sectional image including the pattern, wherein, when the pattern is in contact with another pattern, the application performs the following steps: after the contacting portion of the pattern is cut, a segmentation image of the pattern with the contacting portion cut is obtained using a semantic segmentation model; acquiring contour data of the pattern using the segmentation image; estimating the position of the pattern before contact using the contour data; calculating the amount of displacement between the estimated position and the contour; correcting the contour data based on the calculated amount of displacement; and measuring the dimensions of the desired portion of the pattern based on the corrected contour data.

9. A semiconductor device manufacturing system according to claim 8, characterized in that a semiconductor manufacturing apparatus and a charged particle beam apparatus for acquiring a cross-sectional image including a pattern etched by the semiconductor manufacturing apparatus are connected to the platform.

10. A semiconductor device manufacturing system according to claim 9, wherein the semiconductor manufacturing equipment is a plasma etching equipment, and the charged particle beam equipment is a cross-sectional SEM (Scanning Electron Microscope) equipment.

11. A semiconductor device manufacturing system according to claim 8, wherein the platform is a server.

Images