JP2021111365A - A recording medium in which a pattern edge detection method, a pattern edge detection device, and a program for causing a computer to execute pattern edge detection are recorded. - Google Patents

Abstract

The present invention provides a pattern edge detection method that can detect edges (outlines) of a pattern on an image without using pattern design data. [Solution] A pattern edge detection method generates a target image of a target pattern formed on a workpiece, generates a feature vector representing multiple feature amounts of each pixel of the target image, and uses machine learning to generate the feature vector. The model outputs a judgment result indicating whether a pixel with a feature vector is an edge pixel or a non-edge pixel, and a judgment result indicating an edge pixel is obtained. A virtual edge is generated by connecting multiple pixels with a feature vector with a line. [Selection diagram] Figure 5

Description

The present invention relates to a method and an apparatus for detecting an edge (contour line) of a pattern formed on a workpiece such as a wafer or a mask involved in semiconductor manufacturing from an image generated by a scanning electron microscope. The present invention also relates to a program for causing a computer to perform such pattern edge detection.
The present invention also relates to a method and an apparatus for creating an edge detection model by machine learning.

Conventionally, an edge (contour line) of a pattern formed on a workpiece such as a wafer is detected as follows. First, an image of the pattern on the workpiece is generated by a scanning electron microscope. Next, a CAD pattern is generated from the pattern design data (also referred to as CAD data), and the CAD pattern is superimposed on the pattern on the image. The CAD pattern is a virtual pattern created based on the design information (position, length, size, etc.) of the pattern included in the design data.

FIG. 25 is a schematic view showing a CAD pattern 505 superimposed on the pattern 501 on the image 500. As shown in FIG. 25, the computer generates a plurality of search lines 507 extending in the normal direction with respect to the edge of the CAD pattern 505, and creates a luminance profile of the image 500 along these search lines 507. In FIG. 25, only a part of the plurality of search lines 507 is drawn in order to simplify the drawing.

FIG. 26 is a diagram showing a luminance profile along the search line shown in FIG. 25. The vertical axis of FIG. 26 represents the luminance value, and the horizontal axis represents the position on the search line 507. The computer detects an edge point 510 where the luminance value on the luminance profile is equal to the threshold. The computer repeats the same operation to determine a plurality of edge points on the luminance profile along all search lines 507. The line connecting these plurality of edge points is determined to be the edge of the pattern 501 on the image 500.

However, in the example shown in FIGS. 27 to 29, the pattern edge on the image 500 may not be correctly determined (detected). That is, in the example shown in FIG. 27, a part of the edge of the pattern 501 is missing, and the edge of the pattern 501 does not exist on the search line 507 perpendicular to the CAD pattern 505. In the example shown in FIG. 28, the edge of the CAD pattern 505 is far away from the edge of the pattern 501 on the image, so that the edge of the pattern 501 does not exist on the search line 507. In the example shown in FIG. 29, the edge of the pattern 510, which does not exist in the CAD pattern 505, cannot be detected by the conventional method using the search line 507.

27-29 show examples of pattern defects, and it is important to detect the edges of such defective patterns. However, the actual pattern may deviate from the design data, and the conventional method using the design data may not be able to correctly detect the edge of the defective pattern.

On the other hand, the development of a technique for detecting the edge of a pattern formed on a workpiece such as a wafer by using a model created by machine learning is underway. In this technique, whether or not each pixel of an image in which a pattern appears is a pixel constituting a pattern edge is determined by an edge detection model (learned model).

The edge detection model is created by machine learning (for example, deep learning, decision tree learning, etc.) using the training data prepared in advance. The training data includes a pattern image generated by a scanning electron microscope and correct answer data of each pixel constituting the pattern image. The correct answer data is information that identifies each pixel as either a pixel that constitutes an edge of the pattern or a pixel that does not form an edge. By executing machine learning using such training data, the parameters (weighting coefficient, etc.) constituting the edge detection model are optimized.

However, the pattern edges used in the training data have fluctuations, and the boundary line between the edge and the non-edge region on the image is unclear. The edge detection model created using such training data may fail to detect the edge or erroneously detect the edge. Creating an accurate edge detection model requires a large amount of training data for machine learning, and as a result, machine learning takes a very long time.

Japanese Unexamined Patent Publication No. 2003-178314 Japanese Unexamined Patent Publication No. 2013-98267 Japanese Unexamined Patent Publication No. 2020-140518

The present invention provides a pattern edge detection method and a pattern edge detection device capable of detecting an edge (contour line) of a pattern on an image without using pattern design data.
The present invention also provides a method and an apparatus capable of creating an accurate edge detection model without spending a long time on machine learning.

In one aspect, a target image of a target pattern formed on a workpiece is generated, a feature vector representing a plurality of feature quantities of each pixel of the target image is generated, and the feature vector is constructed by machine learning. A feature that is input to the model, outputs a determination result indicating whether the pixel having the feature vector is an edge pixel or a non-edge pixel from the model, and obtains a determination result indicating an edge pixel. A pattern edge detection method is provided in which a plurality of pixels having a vector are connected by a line to generate a virtual edge.

In one aspect, the model is a decision tree.
In one aspect, the pattern edge detection method selects a plurality of training patterns from design data and generates a plurality of training images of a plurality of actual patterns created based on the plurality of training patterns. The edges of the plurality of actual patterns on the plurality of training images are detected, and the plurality of reference pixels constituting the plurality of training images are not formed with the plurality of first reference pixels constituting the edges. It is classified into a plurality of second reference pixels, a plurality of first feature vectors of the plurality of first reference pixels and a plurality of second feature vectors of the plurality of second reference pixels are generated, and the plurality of first features are generated. Further including a step of constructing the model by machine learning using the training data including the vector and the plurality of second feature vectors.

In one aspect, the plurality of training patterns are a plurality of patterns having an edge extending in at least a first direction, an edge extending in a second direction perpendicular to the first direction, a corner edge, and a terminal edge.
In one aspect, the plurality of real patterns are patterns formed on the workpiece.
In one aspect, in the step of selecting the plurality of training patterns from the design data, a design drawing including the plurality of patterns drawn based on the design data is displayed on the display screen, and the plurality of patterns included in the design drawings are displayed. This is a step of displaying the plurality of training patterns selected from the above patterns or the area in which the plurality of training patterns are located on the display screen in a visually emphasized manner.
In one aspect, the pattern edge detection method generates a plurality of luminance profiles of the target image along a plurality of search lines extending in the normal direction with respect to the virtual edge, and a plurality of luminance profiles are generated based on the plurality of luminance profiles. Further includes a step of generating an updated edge by determining the edge points of the above and connecting the plurality of edge points with a line.
In one aspect, the pattern edge detection method further comprises the step of generating a CAD pattern corresponding to the target pattern from the design data and measuring the distance from the edge of the CAD pattern to the updated edge.

In one aspect, an image generation device that generates a target image of a target pattern formed on a workpiece and a calculation system connected to the image generation device are provided, and the calculation system includes a plurality of pixels of each pixel of the target image. A feature vector representing the feature quantity of is generated, the feature vector is input to a model constructed by machine learning, and whether the pixel having the feature vector is an edge pixel or a non-edge pixel is determined. A pattern edge detection device is provided which outputs a determination result to be shown from the model and connects a plurality of pixels having a feature vector for which a determination result indicating edge pixels is obtained with a line to generate a virtual edge.

In one aspect, the model is a decision tree.
In one aspect, the arithmetic system selects a plurality of training patterns from design data, generates a plurality of training images of a plurality of actual patterns created based on the plurality of training patterns, and generates the plurality of training images. The plurality of reference pixels constituting the plurality of training images by detecting the edges of the plurality of actual patterns on the training image of the above, a plurality of first reference pixels constituting the edge, and a plurality of reference pixels not forming the edge. Classified into the second reference pixels, a plurality of first feature vectors of the plurality of first reference pixels and a plurality of second feature vectors of the plurality of second reference pixels are generated, and the plurality of first feature vectors and the plurality of first feature vectors are generated. The model is constructed by machine learning using the training data including the plurality of second feature vectors.

In one aspect, the plurality of training patterns are a plurality of patterns having an edge extending in at least a first direction, an edge extending in a second direction perpendicular to the first direction, a corner edge, and a terminal edge.
In one aspect, the plurality of real patterns are patterns formed on the workpiece.
In one aspect, the arithmetic system has a display screen, and the arithmetic system displays a design drawing including a plurality of patterns drawn based on the design data on the display screen, and the design drawing. The plurality of training patterns selected from the plurality of patterns included in the above, or an area in which the plurality of training patterns are located is configured to be displayed on the display screen in a visually emphasized manner. ..
In one aspect, the arithmetic system generates a plurality of luminance profiles of the target image along a plurality of search lines extending in the normal direction with respect to the virtual edge, and a plurality of edges based on the plurality of luminance profiles. It is configured to generate an updated edge by determining a point and connecting the plurality of edge points with a line.
In one aspect, the arithmetic system is configured to generate a CAD pattern corresponding to the target pattern from the design data and measure the distance from the edge of the CAD pattern to the updated edge.

In one aspect, a step of issuing a command to a scanning computer to generate a target image of a target pattern formed on a workpiece, and a step of generating a feature vector representing a plurality of feature quantities of each pixel of the target image. , The step of inputting the feature vector into the model constructed by machine learning, and the determination result indicating whether the pixel having the feature vector is an edge pixel or a non-edge pixel is output from the model. A computer-readable record in which a program for causing a computer to execute a step of connecting a plurality of pixels having a feature vector for which a judgment result indicating an edge pixel is obtained and a step of generating a virtual edge is recorded. The medium is provided.

In one aspect, the model is a decision tree.
In one aspect, the program includes a step of selecting a plurality of training patterns from design data and a plurality of actual patterns created based on the plurality of training patterns by issuing a command to the scanning electron microscope. An edge is formed by a step of generating a plurality of training images, a step of detecting edges of the plurality of actual patterns on the plurality of training images, and a plurality of reference pixels constituting the plurality of training images. A step of classifying into a plurality of first reference pixels and a plurality of second reference pixels that do not form an edge, a plurality of first feature vectors of the plurality of first reference pixels, and a plurality of the plurality of second reference pixels. Using the step of generating the second feature vector of the above and the training data including the plurality of first feature vectors and the plurality of second feature vectors, the computer is further made to execute the step of constructing the model by machine learning. It is configured as follows.

In one aspect, the plurality of training patterns are a plurality of patterns having an edge extending in at least a first direction, an edge extending in a second direction perpendicular to the first direction, a corner edge, and a terminal edge.
In one aspect, the plurality of real patterns are patterns formed on the workpiece.
In one aspect, the step of selecting the plurality of training patterns from the design data includes a step of displaying a design drawing including the plurality of patterns drawn based on the design data on the display screen and the step of displaying the design drawing in the design drawing. This is a step of displaying the plurality of training patterns selected from the plurality of patterns or the area in which the plurality of training patterns are located on the display screen in a visually emphasized manner.
In one aspect, the program includes a step of generating a plurality of luminance profiles of the target image along a plurality of search lines extending in the normal direction with respect to the virtual edge, and a plurality of luminance profiles based on the plurality of luminance profiles. The computer is configured to further perform a step of determining an edge point and a step of generating an updated edge by connecting the plurality of edge points with a line.
In one aspect, the program further performs on the computer a step of generating a CAD pattern corresponding to the target pattern from design data and a step of measuring the distance from the edge of the CAD pattern to the updated edge. It is configured to let you.

In one aspect, it is a method of creating an edge detection model for detecting the edge of a pattern on an image, in which a training image of a workpiece on which the pattern is formed is generated by a scanning electron microscope, and the training image is generated. The edge of the above pattern is detected, the feature vector of the pixels constituting the training image is calculated, and the target area in the training image is divided into an edge area, an edge vicinity area, and a non-edge area. A plurality of feature vectors of a plurality of first pixels in the edge region, a plurality of feature vectors of a plurality of second pixels in the edge vicinity region, and a plurality of feature vectors of a plurality of third pixels in the non-edge region. A method is provided in which training data including the above training data is created, and an edge detection model is created by machine learning using the training data.

In one aspect, when the number of the plurality of first pixels is A, the total number of the plurality of second pixels and the number of the plurality of third pixels is B, the number A is divided by the number B. The value (A / B) is a predetermined numerical value.
In one aspect, the value (A / B) obtained by dividing the number A by the number B is in the range 0.6-1.5.
In one aspect, the non-edge region is separated from the edge region by a predetermined number of pixels, and the edge neighborhood region is located between the edge region and the non-edge region.
In one aspect, the step of dividing the target region in the training image into an edge region, an edge neighborhood region, and a non-edge region divides the target region in the training image into an edge region, an exclusion region, and an edge neighborhood region. , And the step of dividing into a non-edge region, the exclusion region is adjacent to the edge region and is located between the edge region and the edge neighborhood region, and the training data is the exclusion region. Does not include the feature vector of the pixels inside.
In one aspect, the target region includes a first region that includes a first edge, a second region that includes a second edge that is perpendicular to the first edge, and a third region that includes a corner edge and a terminal edge.
In one aspect, the number of pixels in the first region, the number of pixels in the second region, and the number of pixels in the third region are in predetermined proportions.

In one aspect, a model generation device that creates an edge detection model for detecting the edges of a pattern on an image, a storage device that stores a program for creating the edge detection model, and the program. The model generation device includes a calculation device that executes a calculation according to an instruction included, and the model generation device acquires a training image of a workpiece on which a pattern is formed from a scanning electron microscope, and obtains an edge of the pattern on the training image. The feature vector of the pixel that is detected and constitutes the training image is calculated, the target area in the training image is divided into an edge area, an edge vicinity area, and a non-edge area, and a plurality of a plurality of in the edge area. Training data including a plurality of feature vectors of the first pixel, a plurality of feature vectors of a plurality of second pixels in the edge vicinity region, and a plurality of feature vectors of a plurality of third pixels in the non-edge region is created. , A model generator is provided that is configured to create an edge detection model by machine learning using the training data.

In one aspect, when the number of the plurality of first pixels is A, the total number of the plurality of second pixels and the number of the plurality of third pixels is B, the number A is divided by the number B. The value (A / B) is a predetermined numerical value.
In one aspect, the value (A / B) obtained by dividing the number A by the number B is in the range 0.6-1.5.
In one aspect, the non-edge region is separated from the edge region by a predetermined number of pixels, and the edge neighborhood region is located between the edge region and the non-edge region.
In one aspect, the model generator is configured to divide the target region in the training image into an edge region, an exclusion region, an edge neighborhood region, and a non-edge region. It is adjacent to the edge region and is located between the edge region and the edge neighborhood region, and the training data does not include the feature vector of the pixels in the exclusion region.
In one aspect, the target region includes a first region that includes a first edge, a second region that includes a second edge that is perpendicular to the first edge, and a third region that includes a corner edge and a terminal edge.
In one aspect, the number of pixels in the first region, the number of pixels in the second region, and the number of pixels in the third region are in predetermined proportions.

In one aspect, a step of acquiring a training image of a workpiece on which a pattern is formed from a scanning electron microscope, a step of detecting an edge of the pattern on the training image, and a pixel constituting the training image. A step of calculating the feature vector of the above, a step of dividing the target region in the training image into an edge region, an edge vicinity region, and a non-edge region, and a plurality of features of a plurality of first pixels in the edge region. A step of creating training data including a vector, a plurality of feature vectors of a plurality of second pixels in the edge vicinity region, and a plurality of feature vectors of a plurality of third pixels in the non-edge region, and the training data. A computer-readable recording medium is provided in which a program is recorded for causing a computer to perform a step of creating an edge detection model by using machine learning.

According to the present invention, edges are detected using a model created by machine learning instead of pattern design data. Specifically, a virtual edge is generated based on the determination result output from the model. This virtual edge is expected to have a shape very close to the edge of the pattern appearing in the image.

According to the present invention, training data including pixels in the edge region, pixels in the region near the edge, and pixels in the non-edge region is used for machine learning. In particular, since the training data includes pixels in the edge vicinity region, which is considered difficult to determine, the edge detection model created by machine learning accurately determines whether or not a given pixel is an edge. Can be done.
Further, according to the present invention, the pixels in the exclusion region are not used for machine learning. The pixels in this exclusion region may be edge pixels or non-edge pixels. That is, the pixels in the exclusion region are uncertain pixels. By excluding such an uncertain pixel feature vector from the training data, machine learning of the edge detection model can be completed at an early stage.

It is a schematic diagram which shows one Embodiment of the pattern edge detection apparatus. It is a schematic diagram which shows the target image. It is a schematic diagram which shows a virtual edge. It is a schematic diagram which shows one Embodiment of the model which consists of a decision tree. It is a figure which shows an example of the determination result at the time of inputting a certain feature vector into a plurality of decision trees shown in FIG. It is a figure which shows another example of the determination result when another feature vector is input to the plurality of decision trees shown in FIG. It is a figure which shows a plurality of search lines extending in the normal direction with respect to a virtual edge. It is a figure which shows an example of the luminance profile along one of the search lines shown in FIG. 7. It is a schematic diagram which shows an example of the training pattern used for making a training data. It is a schematic diagram which shows an example of the design drawing displayed on the display screen. It is a figure which shows the embodiment which displays the frame which shows the area where the selected training pattern is located on the display screen. It is a figure which shows the embodiment which displays the selected training pattern in a manner visually different from other patterns. It is a schematic diagram which shows the training data which includes the 1st feature vector of the 1st reference pixel which constitutes an edge, and the 2nd feature vector of the 2nd reference pixel which does not form an edge. This is a part of a flowchart showing the operation of the pattern edge detection device. This is the remaining part of the flowchart showing the operation of the pattern edge detection device. It is a figure explaining the embodiment which measures the distance from the edge of a CAD pattern to the updated edge. It is a schematic diagram which shows one Embodiment of the pattern edge detection apparatus. It is a figure which shows an example of the image of the workpiece in which a pattern was formed. It is a figure which shows the detected edge. It is a figure which superposed the detected edge shown in FIG. 19 on the image shown in FIG. It is a figure explaining one Embodiment which calculates a feature amount of a pixel. It is a figure explaining the operation which divides a target area in an image into an edge area, an edge neighborhood area, and a non-edge area. It is a figure explaining another embodiment which creates an edge detection model. It is a figure which shows an example of the target area including a plurality of areas set in an image. It is a schematic diagram which shows the CAD pattern superimposed on the pattern on the image. It is a figure which shows the luminance profile along the search line shown in FIG. It is a figure which shows an example of a defective pattern. It is a figure which shows another example of a defective pattern. It is a figure which shows still another example of a defective pattern.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a schematic view showing an embodiment of a pattern edge detection device. As shown in FIG. 1, the pattern edge detection device includes a scanning electron microscope 1 and an arithmetic system 3. The scanning electron microscope 1 is an example of an image generator that generates an image of a workpiece. Examples of workpieces include wafers or masks involved in semiconductor manufacturing. Wafers are used as examples of workpieces in the embodiments described below, but the present invention is not limited to the following embodiments. The pattern is a wiring pattern of an electronic device formed on the workpiece.

The scanning electron microscope 1 is connected to the arithmetic system 3, and the operation of the scanning electron microscope 1 is controlled by the arithmetic system 3. The calculation system 3 includes a storage device 6 in which a database 5 and a program are stored, a processing device 7 that executes a calculation according to an instruction included in the program, and a display screen 10 that displays an image and a GUI (graphical user interface). ing. The storage device 6 includes a main storage device such as a RAM and an auxiliary storage device such as a hard disk drive (HDD) and a solid state drive (SSD). Examples of the processing device 7 include a CPU (central processing unit) and a GPU (graphic processing unit). However, the specific configuration of the arithmetic system 3 is not limited to these examples.

The arithmetic system 3 further includes an input device 12 including a mouse 12a and a keyboard 12b. The user can operate the GUI appearing on the display screen 10 by using the mouse 12a and / or the keyboard 12b. The input device 12 including the mouse 12a and the keyboard 12b is an example, and the present invention is not limited to the input device 12 of the present embodiment.

The arithmetic system 3 includes at least one computer. For example, the arithmetic system 3 may be an edge server connected to the scanning electron microscope 1 by a communication line, or a cloud server connected to the scanning electron microscope 1 by a communication network such as the Internet or a local area network. May be good. The arithmetic system 3 may be a combination of a plurality of servers. For example, the arithmetic system 3 may be a combination of an edge server and a cloud server connected to each other by a communication network such as the Internet or a local area network, or a combination of a plurality of servers not connected by a communication network.

The scanning electron microscope 1 includes an electron gun 15 that emits an electron beam composed of primary electrons (charged particles), a focusing lens 16 that focuses the electron beam emitted from the electron gun 15, and an X deflector 17 that deflects the electron beam in the X direction. It has a Y deflector 18 that deflects an electron beam in the Y direction, and an objective lens 20 that focuses the electron beam on a wafer W which is an example of a workpiece.

The focusing lens 16 and the objective lens 20 are connected to the lens control device 21, and the operation of the focusing lens 16 and the objective lens 20 is controlled by the lens control device 21. The lens control device 21 is connected to the arithmetic system 3. The X deflector 17 and the Y deflector 18 are connected to the deflection control device 22, and the deflection operation of the X deflector 17 and the Y deflector 18 is controlled by the deflection control device 22. The deflection control device 22 is also connected to the arithmetic system 3 in the same manner. The secondary electron detector 25 and the backscattered electron detector 26 are connected to the image acquisition device 28. The image acquisition device 28 is configured to convert the output signals of the secondary electron detector 25 and the backscattered electron detector 26 into an image. The image acquisition device 28 is also connected to the arithmetic system 3 in the same manner.

The stage 31 arranged in the chamber 30 is connected to the stage control device 32, and the position of the stage 31 is controlled by the stage control device 32. The stage control device 32 is connected to the arithmetic system 3. The transfer device 34 for mounting the wafer W on the stage 31 in the chamber 30 is also connected to the arithmetic system 3.

The electron beam emitted from the electron gun 15 is focused by the focusing lens 16, then focused by the objective lens 20 while being deflected by the X deflector 17 and the Y deflector 18, and is irradiated on the surface of the wafer W. When the wafer W is irradiated with the primary electrons of the electron beam, the secondary electrons and backscattered electrons are emitted from the wafer W. Secondary electrons are detected by the secondary electron detector 25, and backscattered electrons are detected by the backscattered electron detector 26. The detected secondary electron signal and backscattered electron signal are input to the image acquisition device 28 and converted into an image. The image is transmitted to the arithmetic system 3.

The design data of the pattern formed on the wafer W is stored in advance in the storage device 6. The pattern on the wafer W is created based on the design data. The pattern design data includes pattern design information such as the coordinates of the vertices of the pattern, the position, shape and size of the pattern, and the number of layers to which the pattern belongs. A database 5 is constructed in the storage device 6. The pattern design data is stored in the database 5 in advance. The arithmetic system 3 can read the design data from the database 5 stored in the storage device 6. The design data is also called CAD data. CAD is an abbreviation for computer-aided design.

Next, a method of detecting the edge (contour line) of the pattern on the image will be described. First, the scanning electron microscope 1 generates a plurality of images of a plurality of patterns formed on the wafer W. The arithmetic system 3 acquires an object image, which is one of a plurality of images, from the scanning electron microscope 1. FIG. 2 is a schematic view showing a target image. In the target image 50, a target pattern 51 for detecting an edge appears. The target pattern 51 is a pattern formed on the wafer W.

The calculation system 3 generates a feature vector representing a plurality of feature quantities of each pixel of the target image 50. The feature vector is a multidimensional vector including a plurality of feature quantities of each pixel. The feature amount is a numerical value representing the feature of the pixel. In the present embodiment, the plurality of feature quantities of each pixel is the difference between the luminance value of the pixel and the luminance value of the other pixel. In one example, the luminance value is a discrete numerical value from 0 to 255 according to the gray scale. In this embodiment, the other pixels are adjacent pixels. In one embodiment, the other pixels may be non-adjacent pixels.

The arithmetic system 3 applies a differential filter to the target image 50 to generate a feature vector including a plurality of feature quantities. Specifically, the arithmetic system 3 calculates a plurality of differences between the luminance value of a certain pixel and the luminance values of a plurality of pixels existing around the pixel. These calculated differences constitute a plurality of feature quantities included in one feature vector.

For example, when the brightness value of the pixel P1 shown in FIG. 2 is 100 and the brightness values of a plurality of pixels existing around the pixel P1 are 200, 150, 100, 50, the calculated difference is -100, -50,0,50. Therefore, the feature vector of pixel P1 in this example is represented as (-100, -50, 0, 50). On the other hand, when the brightness value of the pixel P2 shown in FIG. 2 is 10, and the brightness values of the plurality of pixels existing around the pixel P2 are 20, 15, 10, and 10, the calculated difference is -10, It is -5,0,0. Therefore, the feature vector of pixel P2 in this example is represented as (-10, -5,0,0).

In the present embodiment, the number of feature quantities included in the feature vector is four, but the present invention is not limited to this embodiment. The feature vector may contain a number of features less than or greater than four.

The arithmetic system 3 inputs a plurality of feature quantities constituting a feature vector into a model constructed by machine learning, and outputs a determination result indicating edge pixels or non-edge pixels from the model. This model is a trained model created by machine learning using training data. The training data includes the feature vectors of each of the plurality of pixels and the correct answer data of these feature vectors. The correct answer data is information that identifies a pixel having a certain feature vector as either a pixel that constitutes an edge of a pattern or a pixel that does not form an edge. Each of the feature vectors of the plurality of pixels included in the training data is associated with the correct answer data.

The model created by machine learning using such training data can determine whether the unknown pixel is an edge pixel or a non-edge pixel from the feature vector of the unknown pixel. .. That is, when the feature vector of an unknown pixel is input to the model, the model outputs a determination result indicating an edge pixel or a non-edge pixel.

The arithmetic system 3 selects a plurality of pixels having a feature vector for which a determination result indicating an edge pixel is obtained, and connects the selected plurality of pixels with a line to generate a virtual edge. FIG. 3 is a schematic diagram showing a virtual edge. The arithmetic system 3 forms a virtual edge 55 by connecting a plurality of pixels PX having a feature vector for which a determination result indicating an edge pixel is obtained with a line. The virtual edge 55 is expected to have a shape close to the edge of the target pattern 51 (see FIG. 2) on the wafer W.

In this embodiment, a decision tree is used as the model. A decision tree is a model (trained model) constructed according to a random forest algorithm, which is an example of a machine learning algorithm.

FIG. 4 is a schematic diagram showing an embodiment of a model composed of a decision tree. As shown in FIG. 4, the model 60 includes a plurality of decision trees 60A, 60B, 60C. The feature vector of each pixel is input to each of these decision trees 60A, 60B, and 60C. The plurality of decision trees 60A, 60B, and 60C determine whether the pixel having the feature vector is an edge pixel or a non-edge pixel according to the algorithm of each decision tree. In the example shown in FIG. 4, the model 60 is composed of three decision trees 60A, 60B, and 60C, but the number of decision trees is not particularly limited. In one embodiment, the model 60 may include only one decision tree.

FIG. 5 is a diagram showing an example of a determination result when a feature vector (-100, -50, 0, 50) is input to a plurality of decision trees 60A, 60B, 60C shown in FIG. The feature vectors (-100, -50, 0, 50) are input to the three decision trees 60A, 60B, 60C, respectively. The first decision tree 60A and the second decision tree 60B determine that the pixel having the feature vector (-100, -50, 0, 50) is an edge pixel, and the third decision tree 60C is the feature. A pixel having a vector (-100, -50, 0, 50) is determined to be a non-edge pixel.

FIG. 6 is a diagram showing an example of a determination result when a feature vector (-10, -5, 0, 0) is input to a plurality of decision trees 60A, 60B, 60C shown in FIG. The feature vectors (-10, -5, 0, 0) are input to the three decision trees 60A, 60B, and 60C, respectively. All decision trees 60A, 60B, 60C determine that pixels with feature vectors (-10, -5,0,0) are non-edge pixels.

There are as many determination results as there are decision trees 60A, 60B, 60C, and the determination results may differ depending on the decision trees 60A, 60B, 60C. The calculation system 3 employs the larger number of the determination results indicating the edge pixels or the determination results indicating the non-edge pixels. In the example shown in FIG. 5, two of the three decision trees 60A, 60B, and 60C output a determination result indicating edge pixels, and the other one outputs a determination result indicating non-edge pixels. In this case, the arithmetic system 3 adopts the determination result having the larger number, and determines that the pixel having the input feature vector (-100, -50, 0, 50) is the edge pixel. In the example shown in FIG. 6, all the decision trees 60A, 60B, and 60C output determination results indicating non-edge pixels. In this case, the arithmetic system 3 determines that the pixel having the input feature vector (-10, -5,0,0) is a non-edge pixel.

Decision trees have the advantage that machine learning can be completed faster than other models such as neural networks. For example, machine learning for constructing a plurality of decision trees using training data is completed in about 1 to 5 minutes. Therefore, by adopting the model 60 provided with the decision tree, the time from the start of machine learning to the generation of the virtual edge 55 can be shortened.

In general, even if the pattern is created from the same design data, the edge shape of the pattern will be slightly different from wafer to wafer. Models created using images of patterns on one wafer may fail to detect edges of patterns on other wafers. According to this embodiment, the actual pattern used for creating the training data and the target pattern 51 for which the virtual edge 55 should be generated are formed on the same wafer (workpiece) W. That is, the machine learning of the model 60 in the learning phase and the generation of the virtual edge 55 in the edge detection phase are executed using the image of the same wafer (workpiece) W. Therefore, the arithmetic system 3 can generate the virtual edge 55 of the target pattern 51 with high accuracy by using the model 60 constructed by machine learning using the training data.

In the present embodiment, a plurality of decision trees are used as the model 60 constructed by machine learning, but the present invention is not limited to the present embodiment. In one embodiment, the model 60 constructed by machine learning may be a model consisting of a support vector machine or a neural network. When the model 60 is a neural network, the feature vector is input to the input layer of the neural network, and the determination result is output from the output layer of the neural network. Deep learning is suitable for machine learning of neural networks.

The virtual edge 55 shown in FIG. 3 is expected to have a shape extremely close to the edge of the target pattern 51 shown in FIG. In one embodiment, the arithmetic system 3 may further perform a step of detecting the edge of the target pattern 51 based on the virtual edge 55. The edge detection of the target pattern 51 is performed in the same manner as the conventional edge detection method described with reference to FIGS. 25 and 26. However, the virtual edge 55 is used instead of the CAD pattern. Specifically, as shown in FIG. 7, the arithmetic system 3 generates a plurality of luminance profiles of the target image 50 along a plurality of search lines 65 extending in the normal direction with respect to the virtual edge 55, and a plurality of luminance profiles are generated. A plurality of edge point EPs are determined based on the brightness profile, and the plurality of edge point EPs are connected by a line to generate an updated edge 67.

FIG. 8 is a diagram showing an example of a luminance profile along one of the search lines 65 shown in FIG. 7. The arithmetic system 3 determines the edge point EP whose luminance value on the luminance profile is equal to the threshold value. As shown in FIG. 7, the arithmetic system 3 determines a plurality of edge point EPs on a plurality of luminance profiles along the plurality of search lines 65, and connects these edge point EPs with a line to update the edge. 67 is generated and the updated edge 67 is drawn on the target image 50. The updated edge 67 is expected to have a shape very close to the actual edge of the target pattern 51 (see FIG. 2).

Next, the training data used for machine learning for constructing the model 60 will be described. As described above, the training data is created from images of a plurality of real patterns on the wafer W on which the target pattern 51 on which the virtual edge 55 should be generated is formed. The arithmetic system 3 selects a plurality of training patterns from the design data. The design data (also referred to as CAD data) is the design data of the pattern formed on the wafer W.

In order to improve the edge determination accuracy of the model 60, it is desirable that the training data is created from images of patterns having various edge shapes. From this point of view, the training patterns used to create the training data are, as shown in FIG. 9, an edge E1 extending in the first direction and an edge E2 extending in the second direction perpendicular to the first direction. , Includes a plurality of patterns PT1, PT2, PT3 having a corner edge E3 and a terminal edge E4. The arithmetic system 3 extracts (selects) a plurality of training patterns PT1, PT2, and PT3 having edges E1 to E4 having various shapes as described above from the design data.

The calculation system 3 is configured to display a design drawing drawn based on the design data on the display screen 10 (see FIG. 1). FIG. 10 is a schematic view showing an example of the design drawing 75 displayed on the display screen 10. The design drawing 75 includes various patterns drawn based on the design data. The user can visually check the design drawing 75 on the display screen 10 and select the patterns PT1, PT2, and PT3 having edges extending in multiple directions as shown in FIG. More specifically, the user operates the input device 12 provided with the mouse 12a and the keyboard 12b shown in FIG. 1, and as shown in FIG. 10, from the plurality of patterns on the design drawing 75, for a plurality of trainings. Patterns PT1, PT2 and PT3 can be selected.

The user can operate the input device 12 to delete or change a part of the training patterns PT1, PT2, PT3, or change another pattern on the design drawing 75 into the training patterns PT1, PT2, PT3. It is also possible to add.

The arithmetic system 3 displays a plurality of selected training patterns PT1, PT2, PT3, or an area in which these patterns PT1, PT2, PT3 are located in a visually emphasized manner. For example, as shown in FIG. 11, a frame 80 indicating an area in which the selected training patterns PT1, PT2, and PT3 are located may be displayed on the display screen 10, or as shown in FIG. The selected training patterns PT1, PT2, PT3 themselves may be displayed in a manner visually different from other patterns. In the example shown in FIG. 12, the selected training patterns PT1, PT2, PT are displayed with thicker lines than the other patterns, but in the other example, the selected training patterns PT1, PT2, PT3 are displayed. May be displayed in a color different from other patterns. The user can visually confirm the plurality of training patterns PT1, PT2, and PT3 on the display screen 10.

The scanning electron microscope 1 generates a plurality of training images of a plurality of actual patterns created based on a plurality of selected training patterns. At this time, the scanning electron microscope 1 may generate an image of the target pattern 51 on which the virtual edge 55 should be generated. The arithmetic system 3 acquires a plurality of training images from the scanning electron microscope 1 and stores them in the storage device 6.

Next, the arithmetic system 3 detects edges of a plurality of real patterns on the plurality of training images. This edge detection is performed according to the conventional edge detection method described with reference to FIGS. 25 and 26. That is, the arithmetic system 3 generates a plurality of CAD patterns corresponding to the plurality of training patterns from the design data. The arithmetic system 3 may apply corner round processing to each CAD pattern to form rounded corner edges. Next, the arithmetic system 3 superimposes these CAD patterns on the plurality of patterns on the training image. The arithmetic system 3 generates a plurality of search lines extending in the normal direction with respect to the edge of the CAD pattern, and creates a plurality of luminance profiles of the image along the search lines. The arithmetic system 3 determines an edge point at which the luminance value on one luminance profile is equal to the threshold value. Further, the arithmetic system 3 repeats the same operation to determine a plurality of edge points on the luminance profile along all the search lines. The arithmetic system 3 connects these plurality of edge points with a line, and sets the line connecting the edge points as the edge of the actual pattern on the training image. In this way, the edge of the actual pattern on the training image is detected (determined).

As a result of such edge detection of the actual pattern on the training image, the arithmetic system 3 can label each reference pixel constituting the training image with an edge pixel label or a non-edge pixel label. That is, the arithmetic system 3 classifies the plurality of reference pixels constituting the plurality of training images into a first reference pixel forming an edge and a second reference pixel not forming an edge.

The arithmetic system 3 generates a plurality of first feature vectors of the plurality of first reference pixels and a plurality of second feature vectors of the plurality of second reference pixels. As described above, each feature vector is a multidimensional vector including a plurality of feature quantities of each reference pixel. The arithmetic system 3 creates training data including a plurality of first feature vectors, a plurality of second feature vectors, and correct answer data of these feature vectors. The correct answer data is information that identifies that a pixel having a certain feature vector is either a pixel that constitutes an edge of a pattern or a pixel that does not form an edge. Each of the first feature vector and the second feature vector included in the training data is associated with the correct answer data.

FIG. 13 is a schematic diagram showing training data 70 including the first feature vector of the first reference pixel forming the edge and the second feature vector of the second reference pixel not forming the edge. The first feature vector is labeled with correct data representing edge pixels, and the second feature vector is labeled with correct data representing non-edge pixels. The arithmetic system 3 builds the model 60 by machine learning using the training data 70.

In this embodiment, the model 60 is composed of a plurality of decision trees 60A, 60B, 60C. The arithmetic system 3 creates a plurality of data groups 70A, 70B, 70C each including a plurality of first feature vectors and a plurality of second vectors randomly extracted from the training data 70, and these data groups 70A, 70B, A plurality of decision trees 60A, 60B, 60C are constructed using 70C. More specifically, the arithmetic system 3 uses the data group 70A to determine the model parameters of the decision tree 60A. In the same manner, the arithmetic system 3 uses the data group 70B to determine the model parameters of the decision tree 60B and the data group 70C to determine the model parameters of the decision tree 60C.

The arithmetic system 3 verifies the model 60 including the plurality of decision trees 60A, 60B, 60C having the model parameters determined as described above by using the training data 70. Specifically, the arithmetic system 3 inputs the first feature vector included in the training data 70 to the model 60, outputs the determination result from the model 60, and collates whether or not the determination result indicates an edge pixel. .. Similarly, the arithmetic system 3 inputs the second feature vector included in the training data 70 to the model 60, outputs the determination result from the model 60, and collates whether or not the determination result indicates non-edge pixels. The arithmetic system 3 repeatedly executes such verification to acquire a plurality of determination results, and calculates the determination accuracy at which the plurality of determination results match the correct answer data.

If the determination accuracy is equal to or higher than the set value, the arithmetic system 3 executes the generation of the virtual edge 55 using the model 60 as described above. If the determination accuracy is smaller than the set value, the arithmetic system 3 re-executes the training data creation and the machine learning of the model. In one embodiment, the arithmetic system 3 does not use the model 60 when the determination accuracy is smaller than the set value, and the edge of the target pattern 51 according to the conventional edge detection method described with reference to FIGS. 25 and 26. May be detected.

14 and 15 are flowcharts showing the operation of the pattern edge detection device described so far.
In step 1, the arithmetic system 3 selects (extracts) a plurality of training patterns from the design data. As shown in FIG. 9, a plurality of training patterns selected have an edge E1 extending in the first direction, an edge E2 extending in the second direction perpendicular to the first direction, a corner edge E3, and a terminal edge E4. Includes the pattern of.
In step 2, the scanning electron microscope 1 generates a plurality of training images of a plurality of real patterns, each of which is created based on a plurality of selected training patterns. At this time, the scanning electron microscope 1 may generate an image of the target pattern 51 on which a virtual edge should be generated.

In step 3, the arithmetic system 3 detects edges of a plurality of real patterns on the plurality of training images. This edge detection is performed according to the conventional edge detection method described with reference to FIGS. 25 and 26.
In step 4, the arithmetic system 3 classifies the plurality of reference pixels constituting the plurality of training images into a first reference pixel forming an edge and a second reference pixel not forming an edge.
In step 5, the arithmetic system 3 generates a plurality of first feature vectors of the plurality of first reference pixels and a plurality of second feature vectors of the plurality of second reference pixels. Specifically, the arithmetic system 3 generates a feature vector representing a plurality of feature quantities of each reference pixel.
In step 6, the arithmetic system 3 creates training data 70 including a plurality of first feature vectors, a plurality of second feature vectors, and correct answer data of these feature vectors (see FIG. 13).

In step 7, the arithmetic system 3 executes machine learning using the training data 70 to build the model 60. More specifically, the arithmetic system 3 adjusts the model parameters so that when a certain feature vector is input to the model 60, the model 60 outputs a correct determination result.
In step 8, the arithmetic system 3 verifies the determination accuracy of the model 60 using the training data 70. Specifically, the arithmetic system 3 inputs a plurality of feature vectors included in the training data 70 into the model 60 one by one, and outputs a plurality of determination results from the model 60. The calculation system 3 calculates the determination accuracy, which is the ratio of these plurality of determination results matching the correct answer data.

In step 9, the arithmetic system 3 compares the determination accuracy with the set value. If the determination accuracy is smaller than the set value, the operation flow returns to step 6. In one embodiment, when the determination accuracy is smaller than the set value, the operation flow does not return to step 6, and the arithmetic system 3 performs the target pattern according to the conventional edge detection method described with reference to FIGS. 25 and 26. The edge of 51 may be detected.

If the determination accuracy is equal to or higher than the set value in step 9 described above, the arithmetic system 3 generates a virtual edge using the model 60 in step 10 as shown in FIG. Specifically, the arithmetic system 3 generates a feature vector representing a plurality of feature quantities of each pixel of the target image 50, inputs the feature vector into the model 60, and determines a determination result indicating edge pixels or non-edge pixels. Is output from the model 60. The arithmetic system 3 generates a virtual edge by connecting a plurality of pixels having a feature vector for which a determination result indicating an edge pixel is obtained with a line.

In step 11, the arithmetic system 3 uses the virtual edge as a reference edge to perform edge detection according to a conventional edge detection method and generate an updated edge. Specifically, as shown in FIGS. 7 and 8, the arithmetic system 3 generates a plurality of search lines 65 extending in the normal direction with respect to the virtual edge 55, and the target image 50 along these search lines 65. A plurality of luminance profiles are generated, a plurality of edge point EPs are determined based on the plurality of luminance profiles, and the plurality of edge point EPs are connected by a line to generate an updated edge 67.

In step 12, the arithmetic system 3 checks how far the updated edge 67 generated in step 11 is from the edge of the CAD pattern. Specifically, as shown in FIG. 16, the arithmetic system 3 generates a CAD pattern 75 corresponding to the target pattern 51 from the design data, superimposes the CAD pattern 75 on the target pattern 51 on the target image 50, and CAD. The distance from the edge of the pattern 75 to the updated edge 67 of the target pattern 51 is measured at a plurality of measurement points. The plurality of measurement points are arranged on the edges of the CAD pattern. According to this step 12, it is possible to know how much the updated edge 67 deviates from the design data (or how close it is to the design data).

The arithmetic system 3 including at least one computer operates according to the instructions contained in the program electrically stored in the storage device 6. That is, the arithmetic system 3 issues a command to the scanning electron microscope 1 to generate a target image 50 of the target pattern 51 formed on the workpiece, and a feature representing a plurality of feature quantities of each pixel of the target image 50. A step of generating a vector, a step of inputting the feature vector into the model 60 constructed by machine learning, and indicating whether the pixel having the feature vector is an edge pixel or a non-edge pixel. A step of outputting the determination result from the model 60 and a step of connecting a plurality of pixels having a feature vector for which the determination result indicating the edge pixels is obtained with a line to generate a virtual edge 55 are executed.

The program for causing the arithmetic system 3 to execute these steps is recorded on a computer-readable recording medium which is a non-temporary tangible object, and is provided to the arithmetic system 3 via the recording medium. Alternatively, the program may be input to the arithmetic system 3 via a communication network such as the Internet or a local area network.

The pattern edges used in the training data have fluctuations, and the boundary between the edges and non-edge areas on the image is unclear. A model created using such training data (hereinafter referred to as an edge detection model) may fail to detect an edge or erroneously detect an edge. Creating an accurate model requires a large amount of training data for machine learning, and as a result, machine learning takes a very long time.

Therefore, the embodiment described below provides a method and an apparatus capable of creating an accurate edge detection model without spending a long time on machine learning. FIG. 17 is a schematic view showing another embodiment of the pattern edge detection device. Since the configuration and operation of the present embodiment, which are not particularly described, are the same as those of the embodiments described with reference to FIGS. 1 to 16, the duplicated description will be omitted.

The arithmetic system 3 includes a model generation device 80 that generates an edge detection model for detecting the edges of the pattern formed on the workpiece W. The image acquisition device 28 is connected to the model generation device 80.

The model generator 80 is composed of at least one computer. The model generation device 80 includes a storage device 80a in which the program is stored, and a processing device 80b that executes an operation according to an instruction included in the program. The storage device 80a includes a main storage device such as a RAM and an auxiliary storage device such as a hard disk drive (HDD) and a solid state drive (SSD). Examples of the processing device 80b include a CPU (central processing unit) and a GPU (graphic processing unit). However, the specific configuration of the model generator 80 is not limited to these examples. The storage device 80a may be integrated with the storage device 6, and the processing device 80b may be integrated with the processing device 7.

The model generation device 80 is configured to create an edge detection model for detecting the edge of the pattern on the image sent from the image acquisition device 28 by machine learning. The creation of the edge detection model will be described below.

First, the workpiece W on which the pattern is formed is prepared. The scanning electron microscope 1 generates a training image of the workpiece W, and the model generator 80 acquires the training image from the scanning electron microscope 1. FIG. 18 is a diagram showing an example of a training image of the workpiece W on which the pattern is formed. In the example shown in FIG. 18, a plurality of patterns appear in the training image. These patterns are training patterns used for training data. The model generator 80 detects the edges of the pattern on the training image. Known image processing techniques such as the Sobel filter and the Canny method are used to detect edges. Alternatively, the edge detection may be performed according to the conventional edge detection method described with reference to FIGS. 25 and 26. In one embodiment, the user may manually correct the detected edges. Further, in one embodiment, the edges may be drawn by the user.

FIG. 19 is a diagram showing the detected edges. As shown in FIG. 19, the detected edges are represented by lines. The model generator 80 superimposes the detected edge on the training image of the workpiece W. FIG. 20 is a diagram in which the detected edges shown in FIG. 19 are superimposed on the training image shown in FIG.

Next, the model generation device 80 generates a feature vector representing a plurality of feature quantities of each pixel constituting the training image shown in FIG. The feature vector is a multidimensional vector including a plurality of feature quantities of each pixel. The feature amount is a numerical value representing the feature of the pixel. In the present embodiment, the plurality of feature quantities of each pixel is the difference between the luminance value of the pixel and the luminance value of the other pixel. In one example, the luminance value is a discrete numerical value from 0 to 255 according to the gray scale. In this embodiment, the other pixels are adjacent pixels. In one embodiment, the other pixels may be non-adjacent pixels.

An embodiment for calculating the feature amount of a pixel will be described with reference to FIG. As shown in FIG. 21, the model generator 80 applies a differential filter to the training image and calculates the feature amount of each pixel. More specifically, the model generator 80 differentiates the brightness values of the pixels constituting the training image along a plurality of directions, and is a feature quantity consisting of the difference in the brightness values between the two pixels arranged in each direction. Is calculated for each pixel.

In the example shown in FIG. 21, the model generator 80 differentiates the luminance values of the pixels along the directions of 0 degrees, 45 degrees, 90 degrees, and 135 degrees. That is, the model generator 80 is arranged in the 0 degree direction, the difference in the brightness values of the pixels arranged in the 0 degree direction, the difference in the brightness values of the pixels arranged in the 45 degree direction, the difference in the brightness values of the pixels arranged in the 90 degree direction, and the difference in the 135 degree direction. Calculate the difference in the brightness values of the pixels. Therefore, for each pixel, a feature quantity consisting of four numerical values can be obtained. For example, the feature amount of the pixel represented by the reference numeral P1 in FIG. 21 is represented by a feature vector composed of 200, 50, 0, and -50. However, the angle of differentiation, the number of angles, and the number of features per pixel are not limited to this embodiment.

Next, as shown in FIG. 22, the model generation device 80 divides the target region 100 in the training image into an edge region R1, an edge neighborhood region R2, and a non-edge region R3. The target region 100 shown in FIG. 22 is a part of the training image of the workpiece W shown in FIG. More specifically, the target region 100 shown in FIG. 22 is a region including the edge of the pattern on the training image shown in FIG. The edge region R1 is an region including the pixels constituting the detected edge shown in FIG. The width of the edge region R1 is constant. For example, the width of the edge region R1 may be a width corresponding to one pixel, or may be a width corresponding to a plurality of predetermined numbers (for example, three) of pixels.

The non-edge region R3 is separated from the edge region R1 by a predetermined number of pixels. The edge neighborhood region R2 is located between the edge region R1 and the non-edge region R3. That is, the edge neighborhood region R2 is adjacent to the edge region R1 and the non-edge region R3, and extends along the edge region R1 and the non-edge region R3. The width of the edge vicinity region R2 is constant. In one embodiment, the width of the edge vicinity region R2 is wider than the width of the edge region R1. Normally, the edge region R1 is smaller than the edge neighborhood region R2, and the edge neighborhood region R2 is smaller than the non-edge region R3.

The model generator 80 includes a plurality of feature vectors of a plurality of pixels in the edge region R1, a plurality of feature vectors of a plurality of pixels in the edge neighborhood region R2, and a plurality of feature vectors of a plurality of pixels in the non-edge region R3. Create training data including, and create an edge detection model by machine learning using the training data. Examples of edge detection models include decision trees and neural networks. Examples of machine learning include decision trees and learning deep learning.

The training data includes correct answer data (or correct answer label) for each pixel. This correct answer data is information that identifies each pixel as either a pixel that constitutes an edge of a pattern or a pixel that does not form an edge. The pixels in the edge region R1 are pixels that form an edge, and the pixels in the edge neighborhood region R2 and the non-edge region R3 are pixels that do not form an edge. Machine learning uses edge detection model parameters (such as weighting factors) to correctly determine whether a pixel with a feature vector input to an edge detection model is an edge pixel or a non-edge pixel. Optimize. The edge detection model created by machine learning in this way can determine whether the pixel is an edge pixel or a non-edge pixel based on the pixel feature vector.

According to this embodiment, training data that inevitably includes pixels in the edge region R1, pixels in the edge neighborhood region R2, and pixels in the non-edge region R3 is used for machine learning. In particular, since the training data includes the pixels in the edge neighborhood region R2, which is considered difficult to determine, the edge detection model created by machine learning accurately determines whether or not the given pixel is an edge. be able to.

If the number of pixels in the non-edge region R3 included in the training data is too large compared to the number of pixels in the edge region R1 included in the training data, the edge detection model created using such training data The algorithm is biased towards non-edge pixel detection. As a result, the edge detection model cannot correctly determine that the input pixel is an edge pixel. Therefore, in order to improve the edge detection accuracy of the edge detection model, the plurality of pixels used for machine learning of the edge detection model include edge pixels (that is, pixels in the edge region R1) and non-edge pixels (that is, pixels). It is preferable that the pixels in the edge vicinity region R2 and the non-edge region R3 are evenly included.

From this point of view, when the number of pixels in the edge region R1 is A, the total number of pixels in the edge neighborhood region R2 and the number of pixels in the non-edge region R3 is B, the number A is the number B. The value (A / B) obtained by dividing is a predetermined value. The value (A / B) obtained by dividing the number A by the number B is in the range of 0.6 to 1.5. In order to improve the edge detection accuracy of the edge detection model, in one embodiment, the number A of pixels in the edge region R1 included in the training data is different from the number of pixels in the edge vicinity region R2 included in the training data. It is the same as the total B with the number of pixels in the edge region R3.

The pixel feature vector in the edge neighborhood region R2 has a value between the pixel feature vector in the edge region R1 and the pixel feature vector in the non-edge region R3. Therefore, it is difficult to accurately determine whether the pixels in the edge vicinity region R2 are edge pixels or non-edge pixels. From another point of view, it is possible to generate an edge detection model with high edge detection accuracy by using training data including many feature vectors of pixels in the edge vicinity region R2. Therefore, in one embodiment, the number of pixels in the edge neighborhood region R2 included in the training data is larger than the number of pixels in the non-edge region R3 included in the training data.

The arithmetic system 3 detects the edge on the target image of the workpiece W as follows by using the edge detection model created by machine learning. The scanning electron microscope 1 generates a target image of the workpiece W, and the arithmetic system 3 receives the target image of the workpiece W from the scanning electron microscope 1 and calculates a feature vector of pixels constituting the target image of the workpiece W. , The feature vector is input to the edge detection model, and the judgment result indicating whether the pixel having the feature vector is an edge pixel or a non-edge pixel is output from the edge detection model, and the judgment indicating the edge pixel is performed. An edge is generated by connecting a plurality of pixels having the obtained feature vector with a line.

The model generator 80 including at least one computer operates according to the instructions contained in the program electrically stored in the storage device 80a. That is, the model generator 80 acquires a training image of the workpiece W on which the pattern is formed from the scanning electron microscope 1, detects the edge of the pattern on the training image, and has pixels that constitute the training image. The feature vector is calculated, and the target region in the training image is divided into an edge region R1, an edge neighborhood region R2, and a non-edge region R3, and the feature vectors of a plurality of pixels in the edge region R1 and the edge neighborhood region R2 are included. The training data including the feature vectors of the plurality of pixels and the feature vectors of the plurality of pixels in the non-edge region R3 is created, and the step of creating an edge detection model by machine learning is executed using the training data.

The program for causing the model generation device 80 to execute these steps is recorded on a computer-readable recording medium which is a non-temporary tangible object, and is provided to the model generation device 80 via the recording medium. Alternatively, the program may be input to the model generator 80 via a communication network such as the Internet or a local area network.

Next, another embodiment for creating an edge detection model will be described with reference to FIG. Since the steps of the present embodiment, which are not particularly described, are the same as those of the above-described embodiments described with reference to FIGS. 17 to 22, the duplicated description will be omitted.

As shown in FIG. 23, in the present embodiment, the model generator 80 divides the target region 100 in the training image into an edge region R1, an edge neighborhood region R2, a non-edge region R3, and an exclusion region R4. It is configured in. The exclusion region R4 is adjacent to the edge region R1 and is located between the edge region R1 and the edge neighborhood region R2. The width of the exclusion region R4 shown in FIG. 23 is constant. In one embodiment, the width of the exclusion region R4 is narrower than the width of the edge neighborhood region R2. Further, in one embodiment, the width of the exclusion region R4 is the same as the width of the edge region R1 or narrower than the width of the edge region R1.

The training data does not include the feature vector of the pixels in the exclusion region R4. That is, the training data includes feature vectors of a plurality of pixels in the edge region R1, feature vectors of a plurality of pixels in the edge neighborhood region R2, and feature vectors of a plurality of pixels in the non-edge region R3, but excludes regions. It does not include the feature vector of the pixel in R4. Therefore, the pixels in the exclusion region R4 are not used for machine learning.

The exclusion region R4 is adjacent to the edge region R1, and the feature vector of each pixel in the exclusion region R4 is almost the same as the feature vector of each pixel in the edge region R1. Therefore, the pixels in the exclusion region R4 may be edge pixels or non-edge pixels. That is, the pixels in the exclusion region R4 are uncertain pixels. When such uncertain pixels are included in the training data, it is necessary to continue machine learning until the edge detection model satisfies the desired correct answer rate. As a result, machine learning takes a long time to complete. According to the present embodiment, since the feature vector of the pixel in the exclusion region R4 is excluded from the training data, the machine learning of the edge detection model can be completed at an early stage.

In order to further improve the edge detection accuracy of the edge detection model, in one embodiment, the target region 100 includes a plurality of regions including various pattern edges in the training image. This is because the sharpness of the edges of the pattern on the training image can change depending on the direction in which the edges extend.

FIG. 24 is a diagram showing an example of a target region 100 including a plurality of regions set in the training image. As shown in FIG. 24, the target region 100 is a first region T1 including the first edge E1 of the pattern in the image of the workpiece W and a second region including the second edge E2 perpendicular to the first edge E1. Includes T2 and a third region T3 that includes the corner edge E3 and end edge E4 of the pattern. The training data includes feature vectors of pixels in a plurality of regions T1, T2, T3 including edges E1, E2, E3, E4 extending in different directions. Machine learning using such training data can improve the detection accuracy of edges extending in various directions.

In order to further improve the detection accuracy of the edge extending in the other direction, in one embodiment, the number of pixels in the first region T1, the number of pixels in the second region T2, and the number of pixels in the third region T3 are set in advance. It is in a fixed ratio. When the number of pixels in the first region T1 is represented by S1, the number of pixels in the second region T2 is represented by S2, and the number of pixels in the third region T3 is represented by S3, the relationship between S1 and S2 and S3 is expressed by the following equation. Will be done.
S1 = m × S2 = n × S3
However, m is 0.9 to 1.1 and n is 0.01 to 0.1.

The above-described embodiment is described for the purpose of enabling a person having ordinary knowledge in the technical field to which the present invention belongs to carry out the present invention. Various modifications of the above embodiment can be naturally made by those skilled in the art, and the technical idea of the present invention can be applied to other embodiments. Therefore, the present invention is not limited to the described embodiments, but is construed in the broadest range according to the technical idea defined by the claims.

1 Scanning electron microscope 3 Arithmetic system 5 Database 6 Storage device 7 Processing device 10 Display screen 12 Input device 15 Electronic gun 16 Focusing lens 17 X deflector 18 Y deflector 20 Objective lens 21 Lens control device 22 Deflection control device 25 Secondary electron Detector 26 Reflective electron detector 28 Image acquisition device 30 Chamber 31 Stage 32 Stage control device 34 Conveyor device 50 Target image 51 Target pattern 55 Virtual edge 60 Model 60A, 60B, 60C Decision tree 65 Search line 67 Updated edge 70 Training Data 70A, 70B, 70C Data group 75 Design drawing 80 Model generator 100 Target area W Wafer (workpiece)

Claims (39)

Generates a target image of the target pattern formed on the workpiece and
A feature vector representing a plurality of feature quantities of each pixel of the target image is generated.
The feature vector is input to the model constructed by machine learning,
A determination result indicating whether the pixel having the feature vector is an edge pixel or a non-edge pixel is output from the model.
A pattern edge detection method in which a virtual edge is generated by connecting a plurality of pixels having a feature vector for which a judgment result indicating an edge pixel is obtained with a line. The pattern edge detection method according to claim 1, wherein the model is a decision tree. Select multiple training patterns from the design data
A plurality of training images of a plurality of actual patterns created based on the plurality of training patterns are generated, and a plurality of training images are generated.
The edges of the plurality of real patterns on the plurality of training images are detected, and the edges are detected.
The plurality of reference pixels constituting the plurality of training images are classified into a plurality of first reference pixels forming an edge and a plurality of second reference pixels not forming an edge.
A plurality of first feature vectors of the plurality of first reference pixels and a plurality of second feature vectors of the plurality of second reference pixels are generated.
The pattern edge detection method according to claim 1 or 2, further comprising a step of constructing the model by machine learning using the training data including the plurality of first feature vectors and the plurality of second feature vectors. The plurality of training patterns are a plurality of patterns having an edge extending in at least a first direction, an edge extending in a second direction perpendicular to the first direction, a corner edge, and a terminal edge. The pattern edge detection method described. The pattern edge detection method according to claim 3 or 4, wherein the plurality of actual patterns are patterns formed on the workpiece. The process of selecting the plurality of training patterns from the design data is
A design drawing including a plurality of patterns drawn based on the design data is displayed on the display screen, and the design drawing is displayed.
This is a step of displaying the plurality of training patterns selected from the plurality of patterns included in the design drawing or the area where the plurality of training patterns are located on the display screen in a visually emphasized manner. , The pattern edge detection method according to any one of claims 3 to 5. A plurality of luminance profiles of the target image along a plurality of search lines extending in the normal direction with respect to the virtual edge are generated.
A plurality of edge points are determined based on the plurality of brightness profiles, and a plurality of edge points are determined.
The pattern edge detection method according to any one of claims 1 to 6, further comprising a step of generating an updated edge by connecting the plurality of edge points with a line. A CAD pattern corresponding to the target pattern is generated from the design data, and the CAD pattern is generated.
The pattern edge detection method according to claim 7, further comprising the step of measuring the distance from the edge of the CAD pattern to the updated edge. An image generator that generates the target image of the target pattern formed on the workpiece,
A computing system connected to the image generator is provided.
The arithmetic system
A feature vector representing a plurality of feature quantities of each pixel of the target image is generated.
The feature vector is input to the model constructed by machine learning,
A determination result indicating whether the pixel having the feature vector is an edge pixel or a non-edge pixel is output from the model.
A pattern edge detection device that generates a virtual edge by connecting a plurality of pixels having a feature vector for which a judgment result indicating an edge pixel is obtained with a line. The pattern edge detection device according to claim 9, wherein the model is a decision tree. The arithmetic system
Select multiple training patterns from the design data
A plurality of training images of a plurality of actual patterns created based on the plurality of training patterns are generated, and a plurality of training images are generated.
The edges of the plurality of real patterns on the plurality of training images are detected, and the edges are detected.
The plurality of reference pixels constituting the plurality of training images are classified into a plurality of first reference pixels forming an edge and a plurality of second reference pixels not forming an edge.
A plurality of first feature vectors of the plurality of first reference pixels and a plurality of second feature vectors of the plurality of second reference pixels are generated.
The pattern edge detection device according to claim 9 or 10, wherein the model is constructed by machine learning using the training data including the plurality of first feature vectors and the plurality of second feature vectors. .. 11. The plurality of training patterns are a plurality of patterns having an edge extending in at least a first direction, an edge extending in a second direction perpendicular to the first direction, a corner edge, and a terminal edge. The pattern edge detector described. The pattern edge detection device according to claim 11, wherein the plurality of actual patterns are patterns formed on the workpiece. The arithmetic system has a display screen and
The arithmetic system
A design drawing including a plurality of patterns drawn based on the design data is displayed on the display screen.
The plurality of training patterns selected from the plurality of patterns included in the design drawing, or the area in which the plurality of training patterns are located is configured to be displayed on the display screen in a visually emphasized manner. The pattern edge detection device according to any one of claims 11 to 13. The arithmetic system
A plurality of luminance profiles of the target image along a plurality of search lines extending in the normal direction with respect to the virtual edge are generated.
A plurality of edge points are determined based on the plurality of brightness profiles, and a plurality of edge points are determined.
The pattern edge detection device according to any one of claims 9 to 14, which is configured to generate an updated edge by connecting the plurality of edge points with a line. The arithmetic system
A CAD pattern corresponding to the target pattern is generated from the design data, and the CAD pattern is generated.
The pattern edge detection device according to claim 15, which is configured to measure the distance from the edge of the CAD pattern to the updated edge. The step of issuing a command to the scanning electron microscope to generate the target image of the target pattern formed on the workpiece,
A step of generating a feature vector representing a plurality of feature quantities of each pixel of the target image, and
The step of inputting the feature vector into the model constructed by machine learning,
A step of outputting a determination result indicating whether the pixel having the feature vector is an edge pixel or a non-edge pixel from the model, and
A computer-readable recording medium in which a program for causing a computer to execute a step of connecting a plurality of pixels having a feature vector for which a judgment result indicating an edge pixel is obtained with a line to generate a virtual edge is recorded. The computer-readable recording medium of claim 17, wherein the model is a decision tree. The program
Steps to select multiple training patterns from design data,
A step of issuing a command to the scanning electron microscope to generate a plurality of training images of a plurality of actual patterns created based on the plurality of training patterns, and a step of generating a plurality of training images.
A step of detecting the edges of the plurality of real patterns on the plurality of training images, and
A step of classifying the plurality of reference pixels constituting the plurality of training images into a plurality of first reference pixels forming an edge and a plurality of second reference pixels not forming an edge.
A step of generating a plurality of first feature vectors of the plurality of first reference pixels and a plurality of second feature vectors of the plurality of second reference pixels.
17 or claim 17, wherein the computer is further configured to perform a step of constructing the model by machine learning using the training data including the plurality of first feature vectors and the plurality of second feature vectors. 18. The computer-readable recording medium according to 18. 19. The plurality of training patterns are a plurality of patterns having an edge extending in at least a first direction, an edge extending in a second direction perpendicular to the first direction, a corner edge, and a terminal edge. The computer-readable recording medium described. The computer-readable recording medium according to claim 19 or 20, wherein the plurality of actual patterns are patterns formed on the workpiece. The step of selecting the plurality of training patterns from the design data is
A step of displaying a design drawing including a plurality of patterns drawn based on the design data on the display screen, and
It is a step of displaying the plurality of training patterns selected from the plurality of patterns included in the design drawing, or the area where the plurality of training patterns are located on the display screen in a visually emphasized manner. , A computer-readable recording medium according to any one of claims 19 to 21. The program
A step of generating a plurality of luminance profiles of the target image along a plurality of search lines extending in the normal direction with respect to the virtual edge, and a step of generating a plurality of luminance profiles.
A step of determining a plurality of edge points based on the plurality of brightness profiles, and
The computer readable according to any one of claims 17 to 22, wherein the computer is configured to further perform a step of generating an updated edge by connecting the plurality of edge points with a line. Recording medium. The program
A step of generating a CAD pattern corresponding to the target pattern from the design data, and
23. The computer-readable recording medium of claim 23, wherein the computer is further configured to perform a step of measuring the distance from the edge of the CAD pattern to the updated edge. A method of creating an edge detection model for detecting the edges of a pattern on an image.
A training image of the workpiece on which the pattern is formed is generated with a scanning electron microscope.
Detecting the edge of the pattern on the training image,
The feature vector of the pixels constituting the training image is calculated, and the feature vector is calculated.
The target area in the training image is divided into an edge area, an edge neighborhood area, and a non-edge area.
A plurality of feature vectors of a plurality of first pixels in the edge region, a plurality of feature vectors of a plurality of second pixels in the edge vicinity region, and a plurality of feature vectors of a plurality of third pixels in the non-edge region. Create training data including
A method of creating an edge detection model by machine learning using the training data. When the number of the plurality of first pixels is A, the total number of the plurality of second pixels and the number of the plurality of third pixels is B, the value obtained by dividing the number A by the number B (A). / B) is the method according to claim 25, which is a predetermined numerical value. The method according to claim 26, wherein the value (A / B) obtained by dividing the number A by the number B is in the range of 0.6 to 1.5. The non-edge region is separated from the edge region by a predetermined number of pixels.
The method according to any one of claims 25 to 27, wherein the edge neighborhood region is located between the edge region and the non-edge region. The step of dividing the target region in the training image into an edge region, an edge neighborhood region, and a non-edge region divides the target region in the training image into an edge region, an exclusion region, an edge neighborhood region, and a non-edge region. It is a process of dividing into areas,
The exclusion region is adjacent to the edge region and is located between the edge region and the edge vicinity region.
The method according to any one of claims 25 to 28, wherein the training data does not include a feature vector of pixels in the exclusion region. 25 to 25. The target region includes a first region including a first edge, a second region including a second edge perpendicular to the first edge, and a third region including a corner edge and a terminal edge. 29. The method according to any one of paragraphs. The method according to claim 30, wherein the number of pixels in the first region, the number of pixels in the second region, and the number of pixels in the third region are in a predetermined ratio. A model generator that creates an edge detection model for detecting the edges of a pattern on an image.
A storage device that stores a program for creating the edge detection model, and
It is provided with an arithmetic unit that executes arithmetic according to the instructions included in the program.
The model generator
A training image of the workpiece on which the pattern is formed is obtained from a scanning electron microscope.
Detecting the edge of the pattern on the training image,
The feature vector of the pixels constituting the training image is calculated, and the feature vector is calculated.
The target area in the training image is divided into an edge area, an edge neighborhood area, and a non-edge area.
A plurality of feature vectors of a plurality of first pixels in the edge region, a plurality of feature vectors of a plurality of second pixels in the edge vicinity region, and a plurality of feature vectors of a plurality of third pixels in the non-edge region. Create training data including
A model generator configured to create an edge detection model by machine learning using the training data. When the number of the plurality of first pixels is A, the total number of the plurality of second pixels and the number of the plurality of third pixels is B, the value obtained by dividing the number A by the number B (A). / B) is the model generation device according to claim 32, which is a predetermined numerical value. The model generator according to claim 33, wherein the value (A / B) obtained by dividing the number A by the number B is in the range of 0.6 to 1.5. The non-edge region is separated from the edge region by a predetermined number of pixels.
The model generation device according to any one of claims 32 to 34, wherein the edge neighborhood region is located between the edge region and the non-edge region. The model generator is configured to divide the target region in the training image into an edge region, an exclusion region, an edge neighborhood region, and a non-edge region.
The exclusion region is adjacent to the edge region and is located between the edge region and the edge vicinity region.
The model generator according to any one of claims 32 to 35, wherein the training data does not include a feature vector of pixels in the exclusion region. 32. The model generator according to any one of 36. The model generation device according to claim 37, wherein the number of pixels in the first region, the number of pixels in the second region, and the number of pixels in the third region are in a predetermined ratio. Steps to obtain a training image of the patterned workpiece from a scanning electron microscope, and
The step of detecting the edge of the pattern on the training image, and
The step of calculating the feature vector of the pixels constituting the training image, and
A step of dividing the target area in the training image into an edge area, an edge neighborhood area, and a non-edge area.
A plurality of feature vectors of a plurality of first pixels in the edge region, a plurality of feature vectors of a plurality of second pixels in the edge vicinity region, and a plurality of feature vectors of a plurality of third pixels in the non-edge region. Steps to create training data, including
A computer-readable recording medium on which a program for causing a computer to execute a step of creating an edge detection model by machine learning using the training data is recorded.

Images