Sample observation system and image processing method

Abstract

A sample observation system having a scanning electron microscope and a computer performs the following processes: (1) a plurality of images taken by the scanning electron microscope are acquired; (2) a learning defect image containing a defect site and a learning reference image not containing the defect site are acquired from the plurality of images; (3) a presumed processing parameter is calculated using the learning defect image and the learning reference image; (4) an inspection defect image containing a defect site is acquired; and (5) a pseudo reference image is presumed using the presumed processing parameter and the inspection defect image.

Description

Technical Field

[0001] This invention relates to a sample observation system and image processing method for observing circuit patterns and defects formed on a semiconductor wafer as a sample using a charged particle microscope or the like. Background Technology

[0002] In semiconductor wafer manufacturing, to ensure profitability, it is crucial to quickly initiate production processes and transition to a high-yield mass production system as early as possible. To this end, various inspection devices, instruments for observing defects on samples, and measuring devices are introduced into the production line.

[0003] The device for observing defects on a sample is a device that takes high-resolution images of the defect location on the wafer based on the defect location coordinates (coordinate information indicating the location of the defect on the sample (wafer)) output by the defect inspection device. The widely used defect observation device utilizes a scanning electron microscope (SEM) (hereinafter referred to as a re-examination SEM).

[0004] In semiconductor mass production lines, the automation of sample observation is desired. The review SEM is equipped with functions such as Automatic Defect Review (ADR) to automatically collect images of the defect locations within the sample, and Automatic Defect Classification (ADC) to automatically classify the collected defect images.

[0005] Furthermore, since the defect location coordinates output by the defect inspection device contain errors, the ADR has the following function: to re-detect the defect from an image taken with the defect location coordinates output by the defect inspection device as the center of the expanded field of view, to take a high-magnification image of the re-detected defect location, and to obtain an image for observation. As a defect detection method from SEM images, Japanese Patent Application Publication No. 2001-189358 (Patent Document 1) describes a method in which an image obtained by taking a picture of an area having the same circuit pattern as the defect location is used as a reference image, and an image obtained by taking a picture of the defect location (hereinafter referred to as the defect image) is compared with the reference image to detect the defect.

[0006] WO2019 / 216303 (Patent Document 2) describes a method for generating reference images based on database images and photographed images corresponding to design data, comparing the photographed images and reference images, and detecting defect candidates.

[0007] In addition, Non-Patent Document 1 discloses a method for learning the correspondence between input and output images using a neural network.

[0008] Existing technical documents

[0009] Patent documents

[0010] Patent Document 1: Japanese Patent Application Publication No. 2001-189358

[0011] Patent Document 2: WO2019 / 216303 Publication

[0012] Non-patent literature

[0013] Non-patent literature 1: Olaf Ronneberger, Philipp Fischer, Thomas Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation”, arXiv preprint arXiv:1505.04597 (2015) Summary of the Invention

[0014] The problem the invention aims to solve

[0015] The system for observing defects on a sample (hereinafter referred to as the sample observation system) of the present invention relates to a system for taking pictures of samples such as semiconductor wafers, obtaining images, and observing the images.

[0016] In sample observation systems, acquiring more images per unit time (high-throughput operation) is important. Patent Document 1 describes a method for defect detection by acquiring a reference image for each defect image. However, if the reference image can be inferred from the defect image, the acquisition of the reference image can be omitted, thus increasing the throughput of sample observation. Patent Document 2 describes a method for generating a reference image based on a database image corresponding to design data and captured images. However, design data is highly confidential information and cannot be retrieved from semiconductor production lines, especially mass production lines requiring high productivity, making it sometimes difficult to utilize. As mentioned above, when design data is unavailable, it is difficult to infer the reference image from the defect image, and none of the aforementioned known examples mention a method to solve this problem.

[0017] The purpose of this invention is to solve the problems of the prior art described above, enabling the estimation of reference images based on defect images without using design data, thereby improving the throughput of sample observation.

[0018] In addition, another objective of the present invention is to solve the problems of the prior art described above, so as to be able to infer the defect location from the defect image without using design data, thereby improving the throughput of sample observation.

[0019] Methods for solving problems

[0020] The sample observation system and method of the present invention, which includes a scanning electron microscope and a computer, or the computer included in the sample observation system, performs the following processing:

[0021] (1) Obtain multiple images captured by the scanning electron microscope;

[0022] (2) Obtain a learning defect image containing the defective part and a learning reference image not containing the defective part from the plurality of images;

[0023] (3) Using the learning defect image and the learning reference image, the estimated processing parameters are calculated;

[0024] (4) Obtain an inspection image containing the defective area; and

[0025] (5) Using the estimated processing parameters and the inspection defect image, an estimated pseudo-reference image is generated.

[0026] In the process described in (3):

[0027] (3A) Based on a predetermined evaluation value, the position of the learning defect image and the learning reference image is aligned to obtain the alignment amount;

[0028] (3B) Based on the above alignment, a local defect image for learning is cut out from the defect image for learning;

[0029] (3C) Based on the alignment amount, cut out a learning partial reference image from the learning reference image; and

[0030] (3D) The estimated processing parameters are calculated using the local defect image and the local reference image used for learning.

[0031] In another viewpoint, the sample observation system and method of the present invention, which includes a scanning electron microscope and a computer, or the computer included in the sample observation system, performs the following processing:

[0032] Multiple images were acquired using the scanning electron microscope;

[0033] Obtain a learning defect image containing the defect location from the plurality of images;

[0034] Using the defective images used for learning, estimated processing parameters are calculated;

[0035] Obtain defect images for inspection, including the defective areas; and

[0036] The estimated processing parameters and the inspection defect image are used to estimate the defect location in the inspection defect image.

[0037] In the calculation of the estimated processing parameters:

[0038] Based on a predetermined evaluation value, the positions of the learning defect image and the learning reference image are aligned to obtain the alignment amount.

[0039] Based on the alignment amount, a local defect image for learning is cut out from the defect image for learning;

[0040] Based on the alignment amount, a learning partial reference image is cut out from the learning reference image; and

[0041] The estimated processing parameters are calculated using the local defect image and the local reference image used for learning.

[0042] Invention Effects

[0043] According to the present invention, even when design data is unavailable, reference images can be estimated based on defect images during specimen observation. Furthermore, by estimating reference images, the acquisition of reference images can be omitted, thereby increasing the throughput of specimen observation.

[0044] Furthermore, according to the present invention, during sample observation, the defect location can be estimated based on the defect image. Moreover, by estimating the defect location based on the defect image, the acquisition of reference images can be omitted, thereby improving the throughput of sample observation. Attached Figure Description

[0045] Figure 1 This is a block diagram showing the general structure of the sample observation system of Example 1.

[0046] Figure 2 This is a flowchart illustrating the sample observation process of the sample observation system in Example 1.

[0047] Figure 3 This is a figure illustrating an example of the results of determining defect locations during sample observation using the sample observation system of Example 1.

[0048] Figure 4 This is a flowchart illustrating the learning sequence of the sample observation system in Example 1.

[0049] Figure 5 This is a flowchart of the processing of learning image pairs in the learning sequence of the sample observation system of Example 1.

[0050] Figure 6 This is a flowchart of the process for calculating the estimated processing parameters in the learning sequence of the sample observation system of Example 1.

[0051] Figure 7 This diagram illustrates the image pair alignment and image cropping processes involved in calculating the estimated processing parameters of the sample observation system in Example 1.

[0052] Figure 8 This is a block diagram illustrating the structure of the neural network for estimating pseudo-reference images in the sample observation system of Example 1.

[0053] Figure 9 This is the GUI for setting the size of the learning image in the sample observation system of Example 1.

[0054] Figure 10 This is the GUI for setting the learning termination conditions in the sample observation system of Example 1.

[0055] Figure 11 It is a GUI in the sample observation system of Example 1 used to confirm the estimated error of each learning step.

[0056] Figure 12 It is a GUI in the sample observation system of Example 1 for displaying the image specified in the pseudo-reference image inferred from the defect image used for learning.

[0057] Figure 13 This is a timing diagram of the processing of defect images for observation obtained by a conventional sample observation system compared with the sample observation system of Example 1.

[0058] Figure 14 This is a timing diagram of the processing of defect images obtained through the sample observation system of Example 1.

[0059] Figure 15 This is a flowchart illustrating the learning sequence of the sample observation system in Example 2.

[0060] Figure 16 This is a block diagram showing the schematic structure of the computer in the sample observation system of Example 3.

[0061] Figure 17 This is a flowchart illustrating the sample observation process of the sample observation system in Example 3.

[0062] Figure 18 This is a flowchart of the estimated processing parameters in the calculation learning sequence of the sample observation system in Example 3.

[0063] Figure 19This is a block diagram illustrating the structure of the neural network for estimating defect locations in the sample observation system of Example 3.

[0064] Figure 20 This is a diagram illustrating the formula used in the case of estimating a pseudo-reference image using a local defect image in accordance with Examples 1 to 3.

[0065] Figure 21 This is a diagram illustrating the formulas used in the processing of Example 3. Detailed Implementation

[0066] The embodiments will be described with reference to the accompanying drawings. Furthermore, the embodiments described below do not limit the scope of the invention to which patent protection is sought, and not all of the elements and combinations described in the embodiments are necessarily necessary for the solution of the invention.

[0067] (Implementation Method 1)

[0068] <System Structure>

[0069] use Figure 1 This embodiment will describe the sample observation system. In this embodiment, a sample observation system equipped with a scanning electron microscope (SEM) will be used as the imaging device for photographing the sample. However, the imaging device involved in this embodiment can be any imaging device other than a SEM, such as an optical microscope or an imaging device using charged particles such as ions. Furthermore, the image of the object of observation will be described using an image obtained by photographing defects on a semiconductor wafer, but it can also be an image obtained by photographing other samples such as flat panel displays or biological samples.

[0070] Figure 1 The structure of the sample observation device 100 of this embodiment is shown. The sample observation device (also called the sample observation system) 100 includes an SEM 101 for taking pictures of the sample, a defect inspection device 102 for detecting defects in the sample, and a computer 103.

[0071] In addition to a stage 105 that carries a sample wafer 104 for observation and can move in the XY plane or XYZ space, an electron source 107 that generates an electron beam 106 that irradiates the sample wafer 104, and a detector 110 that detects secondary electrons 108 or reflected electrons 109 generated from the sample wafer 104 irradiated with the electron beam 106, the SEM 101 also includes an electron lens (not shown) that focuses the electron beam 106 onto the sample wafer 104, and a deflector (not shown) for scanning the electron beam 106 on the sample wafer 104.

[0072] The defect inspection device 102 is an apparatus that captures an optical image of the wafer surface and inspects for defects by comparing it with an image of a qualified part (e.g., an image of an adjacent chip). However, such an inspection device is limited by the illumination wavelength, and the resolution of the image is limited to about several hundred nanometers. Therefore, for defects on the order of tens of nanometers on the wafer, it can only detect their presence or absence and output the coordinates of the defects on the wafer.

[0073] Computer 103 includes a user interface (labeled User I / F) 111, a network interface (labeled Network I / F) 112, a control unit 113 for controlling SEM 101, a storage unit 114 for storing information, and a processing unit 115. The storage unit 114 may be, for example, a disk drive, a volatile or non-volatile semiconductor memory device, but may also be other storage devices. Furthermore, the storage unit may be composed of multiple of the aforementioned devices. The processing unit may be any one of a CPU, GPU, FPGA, or LSI, or a combination thereof. The control unit 113, described later, may also be a subsystem different from computer 103 (sometimes called a control subsystem). In this specification, computer 103 is described as an example of one computer, but multiple computers may be used. For example, GUI display processing may be performed by a display computer such as a tablet computer or smartphone, while other image processing may be performed by other computers. Furthermore, a computer may have multiple of the aforementioned components.

[0074] User interface 111 can be a touch panel, monitor, keyboard, mouse, etc., but it can also be any other device as long as it can accept operations from the operator (user) and display information. Network interface 112 is an interface used to communicate with external devices such as defect inspection equipment and SEM via a network.

[0075] The structure of this embodiment will be described with respect to the control unit 113, the processing unit 115, and the storage unit 114.

[0076] The control unit 113 includes a stage control unit 116, an electron beam control unit 117, and a detector control unit 118. The stage control unit 116 controls the movement and stopping of the stage 105. The electron beam control unit 117 controls a deflector (not shown) to irradiate the electron beam 106 within a predetermined field of view, thereby controlling the scanning area of ​​the electron beam 106 on the sample wafer 104. The detector control unit 118 samples the signal from the detector 110 synchronously with the scanning of the electron beam 106 driven by the deflector (not shown), adjusting gain, offset, etc., to generate a digital image. Furthermore, the control unit 113 can be implemented, for example, via a circuit, or via a CPU, GPU, FPGA, or LSI.

[0077] The storage unit 114 includes: an image storage area 119, which stores the digital image generated in the detector control unit 118 along with accompanying information; a process storage area 120, which stores process information including devices, manufacturing processes, image acquisition conditions, etc.; and an estimated processing parameter storage area 121, which stores parameters related to the estimated processing of the pseudo-reference image. Furthermore, these areas do not necessarily need to be separate. The internal data configuration of the storage unit 114 can be arbitrary, as long as it can store processes, parameters, and estimated processing parameters.

[0078] The processing unit 115 includes: an estimation processing parameter calculation unit 122, which calculates estimation processing parameters for estimating a pseudo-reference image based on a defect image; a pseudo-reference image estimation unit 123, which estimates a pseudo-reference image based on the estimation processing parameters; and a defect location determination unit 124, which determines the defect location in the defect image. Furthermore, if the processing unit 115 is a device such as a CPU or GPU that executes predetermined processing via a program, the program corresponding to the estimation processing parameter calculation unit 122, the pseudo-reference image estimation unit 123, and the defect location determination unit 124 (collectively referred to as an image processing program) is stored in the storage unit 114. The processing unit 115 then performs these processes by reading this program.

[0079] Next, the processing performed in this embodiment will be described. Furthermore, unless otherwise explicitly stated, the subsequent processing related to Embodiments 1 to 3 is described as processing performed by the computer 103. More specifically, the control of SEM 101 is a processing performed by the control unit 113, while other processing is a shared processing example performed by the processing unit 115. Additionally, if the processing unit 115 is a CPU or GPU, it reads and executes the program stored in the storage unit 114 (hereinafter sometimes referred to as an image processing program). However, in the case where the processing is described as being performed by the estimated processing parameter calculation unit 122, the pseudo-reference image estimation unit 123, or the defect location determination unit 124, since these units are included in the processing unit 115, it can also be considered that the processing is performed by the processing unit 115.

[0080] <Sample Observation>

[0081] use Figure 2 The observation and treatment of the samples are explained.

[0082] First, the sample wafer 104, which will be the object of observation, is loaded onto the stage 105 (S201). Next, the process corresponding to the sample wafer is read from the process storage area (S202). The semiconductor pattern formed on the sample wafer 104 is manufactured through multiple manufacturing processes, and its appearance can vary significantly between processes. Furthermore, the characteristics of the sample, such as its ease of charging, can also differ. Therefore, the imaging conditions are generally adjusted and stored according to each device and manufacturing process. For the same reason, the estimation processing parameters for the pseudo-reference image are also managed for each process, thereby improving estimation accuracy.

[0083] Next, the defect coordinate information output by the defect inspection device 102 is received or read in (S203). Here, either all received or read defect coordinates can be used as the object of observation, or the result of sampling based on user-specified conditions can be used as the object of observation. Next, it is checked whether the estimated processing parameters corresponding to the process of the sample wafer 104 are stored in the estimated processing parameter storage area 121 (S204). If not stored... Figure 2 (If "none" is specified in the original text), the estimated processing parameters are calculated using the learning sequence described later, and the results are stored in the estimated processing parameter storage area 121 (S205). Furthermore, the defect coordinate information from the aforementioned defect inspection device can be received via the network interface 112, or it can be read from a removable storage medium such as a USB memory. Moreover, S205 can be implemented by the estimated processing parameter calculation unit 122, or it can be implemented entirely by the calculation unit 115 as described above.

[0084] Next, the estimated processing parameters of the pseudo-reference image are read from the estimated processing parameter storage area 121 (S206). Then, the SEM 101 is used to sequentially photograph the observed defects on the sample wafer 104, and a series of observations are performed from S207 to S213.

[0085] First, the stage 105 is moved by the control unit 113 so that the observation defect on the sample wafer 104 is included in the imaging field of the SEM 101 (S207). Next, an inspection defect image is obtained (S208). This image is obtained, for example, by irradiating a relatively wide area containing the observation defect on the sample wafer 104 with an electron beam 106 using the SEM 101, detecting the generated secondary electrons 108 and reflected electrons 109 using the detector 110, thereby taking an image of the relatively wide area containing the observation defect, and processing the detection signal obtained from the detector 110 by the detector control unit 118 to obtain a low-magnification image of the relatively wide area containing the observation defect.

[0086] Next, image preprocessing such as noise removal and brightness unevenness correction is applied to the defect image for inspection (S209). In the pseudo-reference image estimation unit 123, a pseudo-reference image is estimated based on the preprocessed defect image for inspection using estimation processing parameters read from the estimation processing parameter storage area 121 (S210). This pseudo-reference image is equivalent to an image in which the same circuit pattern as the defect image for inspection is observed, but without any defects. Furthermore, in the estimation of the pseudo-reference image, design data is not required; instead, the estimation processing parameters and the preprocessed defect image for inspection are used.

[0087] Next, in the defect location determination unit 124, the inspection defect image and the pseudo-reference image are compared, and the defect location is determined based on the inspection defect image (S211). The method for determining the defect location can be the method described in Patent Document 1, etc.

[0088] Next, the field of view is narrowed to take a high-magnification picture of the identified defect area to obtain a defect image for observation (hereinafter referred to as the defect image for observation) (S212), and the defect image for inspection, the pseudo-reference image, and the defect image for observation are stored in the image storage area 119 (S213).

[0089] At this point, Figure 2 The process ends. Furthermore, this series of processes can be repeatedly performed on all observed defects of the sample wafer 104, or it can be performed on a subset of observed defects based on other benchmarks. Additionally, S205 can be performed either by the estimated processing parameter calculation unit 122 or as a whole by the processing unit 115.

[0090] <<Example of the results of determining the location of defects>>

[0091] use Figure 3 This section describes an example of the result of identifying the defect location in S211. If the defect location is determined using the inspection defect image 301 and the pseudo-reference image 302, a specific defect location image 303 is obtained that can distinguish between the defect location and areas outside the defect location. In the specific defect location image 303, for example, the pixel value of the defect location is set to 1, and the pixel value of areas outside the defect location is set to 0.

[0092] <<Learning Sequence>>

[0093] use Figure 4 This describes the process of calculating the estimated processing parameters through the learning sequence of S205 and storing the results in the estimated processing parameter storage area 121.

[0094] First, obtain the size of the image acquired for learning (S401). Next, set the defect to be learned (S402). This can use all of the defect coordinates read in S203 as the learning object, or can use the result of sampling based on user-specified conditions as the learning object. Next, obtain a pair of an image including the defect set as the learning object (hereinafter, referred to as the learning defect image) and an image including a region designed (hereinafter, referred to as the learning reference image) in a manner to form the same circuit pattern as the periphery of the defect position serving as the learning object (S403).

[0095] <<<Details of obtaining the learning image pair in S403>>>

[0096] Here, use Figure 5 to explain S403. First, control the stage 105 to move so that a region designed (or assumed) in a manner to form the same circuit pattern as the periphery of the defect position serving as the learning object (hereinafter, referred to as the reference region) is included in the imaging field of view of the SEM 101 (S501). In addition, in the determination of the reference region, information with as good accuracy as design data is not required. In a semiconductor wafer, multiple chips designed in a manner to form the same circuit pattern are arranged on the wafer. Therefore, as the simplest method, consider a region offset by one chip from the defect coordinates as the reference region. However, the reference region can also be selected by other methods.

[0097] Next, irradiate the reference region on the specimen wafer 104 with the electron beam 106 for scanning, image the reference region by detecting the generated secondary electrons 108 and reflected electrons 109 with the detector 110, and process the detection signal from the detector 110 obtained by this imaging with the detector control unit 118 to obtain a learning reference image in a manner to be larger than the size obtained in S401 (S502).

[0098] Next, control the stage 105 to move so that a region including the defect serving as the learning object (hereinafter, referred to as the defect region) is included in the imaging field of view of the SEM 101 (S503). Next, irradiate the defect region on the specimen wafer 104 with the electron beam 106 for scanning, image the defect region by detecting the generated secondary electrons 108 and reflected electrons 109 with the detector 110, and process the detection signal from the detector 110 obtained by this imaging with the detector control unit 118 to obtain a learning defect image in a manner to be larger than the size obtained in S401 (S504).

[0099] After obtaining the learning defect image and the learning reference image, attach additional information in a paired manner of the learning defect image and the learning reference image and store them in the image storage area 119 (S505).

[0100] The above is the description of the use of Figure 5 . Additionally, the order of obtaining the learning reference image and the learning defect image can be either first. That is, the processes of S501 to S502 can also be executed after S503 to S504.

[0101] <<<Parallel processing of learning sequence and defect observation processing>>>

[0102] Return to Figure 4 's description. The following S404 and S405 described below are processes intended to perform the determination of the defect location and the processes required for defect observation in parallel with the learning sequence, so as to enable defect observation even during the time-consuming learning sequence.

[0103] Compare the learning defect image obtained in S403 with the learning reference image, and determine the defect location (S404) in the same way as in S211. Then, narrow the field of view and take a high-magnification photograph of the determined defect location to obtain an observation defect image, store the observation defect image in the image storage area 119, and exclude it from the defects to be observed in S207 to S213 (S405). The above processes of S403 to S405 are repeatedly executed for all or part of the learning target defects of the sample wafer 104.

[0104] The above is the parallel processing of the learning sequence and defect observation processing.

[0105] Finally, obtain the learning end condition for estimating the processing parameters (S406), and in the estimation processing parameter calculation unit 122, calculate the estimation processing parameters for estimating the pseudo-reference image (S407: detailed content will be described later).

[0106] The above is Figure 4 's description. Here, in addition to the advantages of the above parallel processing, S404 and S405 do not require obtaining the observation defect image in S212, so they have the advantage of being able to efficiently perform sample observation. In addition, S404 to S405 can also be omitted. In the case of omission, for the learning target defects, they are not excluded from the defects to be observed in S207 to S213.

[0107] <<<Detailed content of the estimation processing parameter calculation in S407>>>

[0108] Use Figure 6 to explain the estimation processing parameter calculation process of step S407. In the calculation of the estimation processing parameters, design data is not required, and the learning defect image and the learning reference image are used.

[0109] First, all the learning defect images and learning reference images obtained in S403 are subjected to image preprocessing such as noise removal and brightness unevenness correction, similar to S209 (S601). Next, in order to absorb the errors in positioning caused by table movement and electron beam irradiation, the positions of the learning defect images and learning reference images are aligned based on predetermined evaluation values, and the alignment amounts AX and AY between the images are calculated (S602). As predetermined evaluation values, the positions are aligned according to the position where the evaluation value is the maximum or minimum, using normalized cross-correlation coefficients or mean square errors, etc. In addition, if the image resolution (the number of pixels in each image of the same field of view) is different, the image resolution is made consistent by linear interpolation or the like before position alignment.

[0110] Next, based on the alignment, a local defect image for learning is cut out from the defect image for learning, and a local reference image for learning is cut out from the reference image for learning (S603).

[0111] Here, use Figure 7 The processing of S601 and S602 is explained. Figure 7 The diagram shows a learning defect image 701 and a learning reference image 702. In S601, the two images are aligned according to a predetermined evaluation value to obtain an image alignment result 703. However, in the image alignment result 703, the defective portion contained only in the learning defect image 701 is not shown. In S602, based on alignment amounts AX and AY, a learning partial defect image 704 is cut from the learning defect image 701, and a learning partial reference image 705 is cut from the learning reference image 702, representing a shared area between the learning defect image and the learning reference image.

[0112] Return to using Figure 6 The following steps describe the process: First, the learning estimation parameters are initialized (S604). At this time, previously learned estimation parameters can also be read from the estimation parameter storage area 121 and used as initial values ​​for the learning estimation parameters. Next, based on the learning estimation parameters, a pseudo-defect image is estimated from the local defect image used for learning (S605), and the estimation error between the local reference image used for learning and the pseudo-defect image is calculated (S606). The learning estimation parameters are then updated to minimize the estimation error (S607). Next, it is checked whether the learning termination condition obtained in S406 is met (S608). If the learning termination condition is met, the learning estimation parameters are stored as estimation parameters in the estimation parameter storage area 121 (S609). If the learning termination condition is not met, the process returns to step S605.

[0113] above, Figure 6 The process is complete.

[0114] In addition, the following conditions could be considered as conditions for the conclusion of learning:

[0115] (Learning termination condition 1) Compare the estimated error with the pre-set estimated error threshold TH. The estimated error is less than the estimated error threshold TH.

[0116] (Learning termination condition 2) The user accepts the action to terminate the learning process.

[0117] (Learning termination condition 3) Repeat the S605 to S608 processes of MR a predetermined number of times.

[0118] Applications of Neural Networks

[0119] Furthermore, as a method for estimating a pseudo-reference image based on a learned local defect image in S605, the neural network described in Non-Patent Document 1 can be used. Moreover, this neural network is also used in S210 when estimating a pseudo-reference image based on an inspection defect image. Specifically, using... Figure 8 A U-shaped neural network, such as the one shown, called U-net, is sufficient. Here, Y represents the input image. F11(Y), F12(Y), F13(Y), F21(Y), F22(Y), F23(Y), F24(Y), F31(Y), and F32(Y) represent intermediate data. F(Y) represents the estimation result from the pseudo-reference image.

[0120] pass Figure 20 Equations 1 to 10 are used to calculate intermediate data and the final result. In Equations 1 to 10, "*" represents convolution operation, DS represents the operation of applying a 2×2 maximum filter to the input image and reducing it by half in the spatial (XY) direction, US represents the operation of upsampling the input image to twice its size in the spatial direction, and CC represents the operation of combining the two input images in the channel direction.

[0121] Here, the meanings of the variables used in Equations 1 to 10 are as follows:

[0122] W1 is c1 filters of size c0×f1×f1.

[0123] c0 is the number of channels in the input image.

[0124] f1 is the size of the spatial filter.

[0125] By convolving the input image with a filter of size c0×f1×f1 times, a c1-dimensional feature map is obtained.

[0126] The meanings of the variables used in Equations 1 to 10 are again recorded as follows:

[0127] B1 is a c1-dimensional vector (corresponding to the bias components of c1 filters).

[0128] W2 is a filter of size c1×f2×f2.

[0129] B2 is a c2-dimensional vector.

[0130] W3 is a filter with dimensions c2×f3×f3.

[0131] B3 is a vector with dimensions c3.

[0132] W4 is a filter with dimensions of c3×f4×f4.

[0133] B4 is a c2-dimensional vector.

[0134] W5 is a filter of size (c²×2)×f5×f5.

[0135] B5 is a c2-dimensional vector.

[0136] W6 is a filter with dimensions of c2×f6×f6.

[0137] B6 is a vector with c1 dimensions.

[0138] W7 is a filter of size (c1×2)×f7×f7.

[0139] B7 is a c4-dimensional vector.

[0140] W8 is a filter with dimensions of c4×f8×f8.

[0141] B8 is a vector with c5 dimensions.

[0142] Here, c0 and c5 are values ​​determined by the number of channels in the local defect image and the local reference image used for learning. Additionally, f1–f8 and c1–c4 are hyperparameters determined by the user before the learning sequence; for example, they can be set to f1–f8 = 3, c1 = 8, c2 = 16, c3 = 32, and c4 = 64. The parameters calculated by the estimated processing parameters calculation process (S405) are W1–W8 and B1–B8.

[0143] Furthermore, other structures can also be used as the neural network structure shown above. Figure 8 The text describes the construction of a U-net with a maximum depth of 3, but the depth can be changed, and networks with a maximum depth of 4 or higher can also be used.

[0144] The estimation error calculation process (S607) is a process that evaluates the difference (error) between the estimation result F(Y) and the partially referenced images used for learning, and updates the parameters in such a way that the estimation error obtained in this process decreases. As a method for quantifying the difference (error) between images, mean square error or similar methods can be used.

[0145] In the estimation parameter update process (S608), the general error backpropagation method can be used in the learning of the neural network. Furthermore, when calculating the estimation error, all pairs of the acquired learning defect image and learning reference image can be used, but a mini-batch approach can also be employed. That is, multiple image pairs can be randomly extracted from the pairs of learning defect images and learning reference images and the parameters can be updated repeatedly. Moreover, patch images can be randomly cut from an image pair and used as the input image Y of the neural network. This allows for efficient learning.

[0146] <gui>

[0147] Next, the GUI displayed on the user interface 111 will be described.

[0148] Figure 9 This GUI 900 represents the GUI used to set the size of the learning image in S401. The GUI 900 displays the defect inspection image size 901 and the field of view 902 included in the process read in S202. Furthermore, an input section 903 for setting the learning image size is displayed on the GUI 900, and after setting the learning image size via the GUI 900, the processing after step S402 is executed by pressing the "OK" button 904.

[0149] Here, the method for determining the size of the image used for learning will be explained. Figure 8 In the neural network shown, at depth d, the input image is reduced to a size of 1 / (2^d) (XY direction). Therefore, when using a neural network with a maximum depth D, the input image is required to be at least (2^D) × (2^D). However, in S602, the sizes of the learning local defect image and the learning local reference image obtained by aligning the learning defect image and the learning reference image are less than the sizes of the learning defect image and the learning reference image, respectively. Specifically, the image size is reduced by an amount corresponding to the alignment amounts AX and AY. In order to ensure that the size of the learning local defect image and the learning local reference image is at least (2^D) × (2^D), it is necessary to obtain the learning defect image and the learning reference image with a size of (2^D + AX) × (2^D + AY), but the alignment amounts AX and AY cannot be known before obtaining the learning defect image and the learning reference image. Since the alignment values ​​AX and AY depend on the stage movement error and the electron beam irradiation positioning error, the maximum values ​​MX and MY of the alignment values ​​can be calculated based on these errors. The size (2^D+MX)×(2^D+MY) can then be set as the learning image size. For example, with a maximum neural network depth of 7, MX = 50, and MY = 30, 178×158 can be set as the learning image size.

[0150] Alternatively, instead of obtaining the size of the learning image in S401, the learning defect image and the learning reference image can be obtained in S502 and S504 at the same size as the inspection defect image. In S602, the positions of the learning defect image and the learning reference image are aligned. After cutting out the learning local defect image and the learning local reference image in S603, it is checked that at least one of the learning local defect image and the learning local reference image is at or above a predetermined size. If the size is smaller than the predetermined size, the image is not used for processing in S605 to S608. When using a neural network with a maximum depth D, the predetermined size is set to (2^D) × (2^D).

[0151] Figure 10 The GUI 1000 for setting the learning termination condition in S406 is shown. The GUI 1000 displays an input section 1001 for setting the number of times S605 to S608 processes are repeated (MR), an input section 1002 for setting the estimated error threshold TH, and an input section 1003 for setting whether to accept the user's learning termination operation. After setting the learning termination condition through the GUI 1000, pressing the "Learning Start" button 1004 executes the calculation of the estimated processing parameters in S407. Pressing the "Cancel" button 1006 during the calculation of the estimated processing parameters interrupts the calculation. Additionally, pressing the "Progress Confirmation" button 1005 during the calculation of the estimated processing parameters switches to... Figure 11 The GUI 1100 shows the progress of the estimated processing parameter update process during the learning process. In the GUI 1100, the number of repetitions of the estimated parameter update and the shift of the estimated error are displayed in graph 1101. When the "Learning End" button 1103 of the GUI 1100 is pressed, it is considered that the user has accepted the end of the learning process, and the update of the estimated processing parameters during learning ends; that is, in S608, it is determined that the learning end condition has been met. When the "Estimated Image Confirmation" button 1102 of the GUI 1100 is pressed, it switches to... Figure 12 The GUI1200 is as shown. On the GUI1200, pressing the image ID selection button 1201 specifies the number of the image to be displayed. The channel selection unit 1202 selects the type of image to be displayed, such as secondary electron image (SE) or reflected electron image (BSE). Using the learning estimation parameters, pseudo-defect image estimation processing is performed for the specified image ID, displaying the learning local defect image 1203, the estimated pseudo-defect image 1204, and the learning local reference image 1205 for the specified image ID. Pressing the "OK" button 1206 on the GUI1200 switches to... Figure 11 The original GUI1100 display is shown as shown.

[0152] According to this embodiment, a learning defect image and a learning reference image are obtained, and the learning defect image and the learning reference image are used to calculate the estimated processing parameters. In sample observation, the pseudo reference image is estimated based on the inspection defect image, thereby eliminating the need to obtain the reference image and improving the throughput of sample observation.

[0153] <Comparison of Defect Observation Sequences>

[0154] Figure 13 The sequence of observing defects (1) and (2) in the specimen observation system is shown. The horizontal axis represents time, and the vertical axis represents the defects in the specimen.

[0155] First, the sequence of steps 1301 related to the observation of the defect (1) of the object under observation includes the following steps:

[0156] Move the worktable so that the field of view of the SEM101 includes a reference area (S) corresponding to the defect (1) of the observed object.

[0157] A reference image (RI) for learning purposes is obtained by photographing the reference area using SEM101.

[0158] Move the worktable so that the area (S) containing the defect (1) of the observed object is included in the field of view of the SEM101.

[0159] A relatively wide area containing the defect (1) of the observed object was photographed using an SEM101 to obtain a learning defect image (DI).

[0160] The defect location (D) in the learning defect image is determined using the learning defect image and the learning reference image.

[0161] An observation defect image (HI) is obtained by taking pictures of a relatively narrow area containing the identified defect location using an SEM101.

[0162] The same sequence is also included in step 1302, which is related to the observation of the next observation object defect (2). Here, the table movement (S) of step 1302 must be performed after the acquisition (HI) of the defect image for observation of the observation object defect (1) is completed. This is because, in step 1301, the observation object defect (1) remains within the shooting field of view of SEM 101 until the acquisition (HI) of the defect image for observation of the observation object defect (1) is completed.

[0163] Figure 14 This is the sequence related to the processing of S207 to S213 in this embodiment. The relationship between the horizontal and vertical axes is... Figure 13 The situation is the same. Figure 14 In the sequence, as Figure 13 There is no time available for estimation based on a pseudo-reference image (P), and on the other hand, there is no reference image acquisition (RI). In addition, there is one less table movement (S) in each process.

[0164] First, the sequence of steps 1401 related to the observation of the defect (1) of the object under observation includes the following steps:

[0165] Move the worktable so that the area (S) containing the defect (1) of the observed object is included in the field of view of the SEM101.

[0166] A defect image (DI) for inspection is obtained by taking pictures of a relatively wide area containing the defect (1) of the object under observation using a SEM101.

[0167] Based on the estimated processing parameters, a pseudo-reference image (P) is estimated from the defect image used for inspection.

[0168] Using the inspection defect image and the pseudo-reference image, determine the defect location (D) in the inspection defect image.

[0169] An observation defect image (HI) is obtained by taking pictures of a relatively narrow area containing the identified defect location using an SEM101.

[0170] Next, in step 1402 related to the observation of the object defect (2), the same processing is performed on the object defect (2).

[0171] As mentioned above, in Figure 14 In the sequence, by inferring the pseudo-reference image (P) based on the defect image, it is not necessary to Figure 13 The sequence described herein includes the first stage movement (S) and the acquisition of the reference image (RI). This reduces the number of stage movements to half and eliminates the need for image acquisition, thereby increasing throughput when sequentially observing multiple defects on the sample wafer 104 using the sample observation system.

[0172] (Implementation Method 2)

[0173] In Example 1, the following method was described: A SEM was used to photograph the defect area and a reference area to obtain a learning defect image and a learning reference image. The learning defect image and the learning reference image were used to calculate estimated processing parameters. During sample observation, a pseudo-reference image was estimated based on the inspection defect image, thereby improving the throughput of sample observation. When calculating the estimated processing parameters, the more pairs of learning defect images and learning reference images there are, the more efficient the learning can be. In this example, the following method is described: Even when there are no defects in the learning object or the number of defects in the learning object is small, a learning defect image is generated by assigning pseudo-defects to the learning reference image. The estimated processing parameters are calculated using the pairs of the learning reference image and the generated learning defect image.

[0174] The structure of the sample observation system in this embodiment is the same as that described in Embodiment 1. Figure 1 The structures shown are basically the same. The difference lies in the processing flow of the learning sequence; the sample observation flow outside the learning sequence is the same as that described in Example 1. Figure 2 The processing flow shown is the same. Furthermore, the GUI of the sample observation system in this embodiment is the same as that described in Embodiment 1. Figures 9-12 The interface shown is the same as that in Example 1. Hereinafter, only the parts that differ from Example 1 will be described.

[0175] <Learning Sequence>

[0176] use Figure 15 The learning sequence of S205 is explained.

[0177] First, the area to be studied is set (S1501). This can be set to one or more areas on the sample wafer specified by the user, or one or more areas on the sample wafer 104 can be set randomly. However, the defect coordinates output by the defect inspection device are not included in the area to be studied. Next, the stage 105 is moved so that the set area to be studied is included in the field of view of the SEM 101 (S1502). The area to be studied on the sample wafer 104 is irradiated with an electron beam 106 and scanned. The area to be studied is photographed by secondary electrons 108 and reflected electrons 109 generated by the detector 110. The detector control unit 118 processes the detection signal from the detector 110 obtained from the photograph and obtains a reference image for learning (S1503).

[0178] Next, a pseudo-defect is assigned to the learning reference image to generate a learning defect image, thereby obtaining the learning defect image (S1504). Additional information is then attached to the learning defect image and the learning reference image in pairs and stored in the image storage area 119 (S1505). The pseudo-defect area PR can be randomly set in terms of its center position and size (width and height) within the plane of the learning reference image.

[0179] As pseudo-defects, compensation can also be applied to the density of the region PR. Alternatively, the region PR can be set as an edge containing the circuit pattern, and the circuit pattern can be deformed based on the edge strength. The above simulates tiny defects, but it is also possible to generate huge defects that cover the entire surface of the image. The types of pseudo-defects are not limited to these; any type of defect can be modeled and generated.

[0180] The processes S1502 to S1505 described above are repeatedly performed on the entire area of ​​the sample wafer 104 that is used as the learning target. Next, the learning termination condition (S1506: equivalent to...) is obtained based on the estimated processing parameters. Figure 4 In S406), in the estimation processing parameter calculation unit 122, estimation processing parameters for estimating the pseudo-reference image are calculated (S1507: equivalent to...). Figure 4 (S407).

[0181] Furthermore, in the above description, the method of using the image containing spurious defects generated in S1504 as the learning defect image for calculating the estimated processing parameters has been described. However, in addition to the image containing spurious defects generated in S1504, the learning defect image can also be used, similarly to S403 described in Example 1, as an image containing all defect coordinates read in S203 or defect coordinates sampled based on user-specified conditions. That is, in S1507, the estimated processing parameters can also be calculated using a pair of images obtained by photographing the defect-containing area on the sample wafer 104 (first defect image) and a reference image corresponding to the first defect image (first reference image), and a pair of second reference images and an image containing spurious defects generated from the second reference image (second defect image).

[0182] According to this embodiment, even when there are no defects in the learning object or few defects in the learning object, a learning defect image can be generated by assigning a pseudo-defect to the learning reference image, and the estimated processing parameters can be calculated by using the pair between the learning reference image and the generated learning defect image. In sample observation, the pseudo-reference image is estimated based on the inspection defect image, thereby eliminating the need to obtain reference images and improving the throughput of sample observation.

[0183] (Implementation Method 3)

[0184] In Examples 1 and 2, the following method was described: using estimation processing parameters obtained by learning the correspondence between a defect image and a reference image, during sample observation, a pseudo-reference image was estimated based on the defect image for inspection, and the defect image for inspection was compared with the pseudo-reference image to determine the defect location in the defect image for inspection. In this example, a method for calculating estimation processing parameters for estimating the defect location in the defect image for inspection, and for estimating the defect location in the defect image for inspection based on the estimation processing parameters, is described.

[0185] <System Structure>

[0186] Regarding the structure of the sample observation system 100 in this embodiment, the SEM 101 and the defect inspection device 102 are the same as those described in Embodiment 1. Figure 1 The structures shown are the same, except for the structure of computer 103. Hereinafter, only the parts that differ from Embodiment 1 will be described.

[0187] use Figure 16 The computer 103 of the sample observation system of this embodiment will be described. The storage unit 114 includes: an image storage area 119, which stores digital images generated in the detector control unit 118 along with accompanying information; a process storage area 120, which stores process information including devices, manufacturing processes, image acquisition conditions, etc.; and an estimation processing parameter storage area 1601, which stores parameters related to the estimation processing of defect locations in the defect image. The processing unit 115 includes: an estimation processing parameter calculation unit 1602, which calculates estimation processing parameters for estimating defect locations in the defect image; and a defect location estimation unit 1603, which estimates defect locations based on the estimation processing parameters.

[0188] <Sample Observation Method>

[0189] use Figure 17 The observation methods for the samples are explained. S1701~S1703 and Figure 2 S201 to S203 are the same.

[0190] In S1704, it is confirmed whether the estimated processing parameters corresponding to the process of the processed sample wafer 104 are stored in the estimated processing parameter storage area 1601. If not stored (in... Figure 17 (If "none" is specified in the original text), in the estimation processing parameter calculation unit 1602, the estimation processing parameters are calculated and stored using the learning sequence described later (S1705). Next, the estimation processing parameters are read from the estimation processing parameter storage area 1601 (S1706).

[0191] Next, SEM101 was used to sequentially photograph the defects on the sample wafer 104 for observation. Steps S1707 to S1709 are... Figure 2 Steps S207 to S209 are the same, so their description is omitted. Next, in the defect location estimation unit 1603, the defect location in the inspection defect image that has undergone image preprocessing is estimated using estimation processing parameters (S1710). Furthermore, in the estimation of the defect location, design data is not required, and the inspection defect image that has undergone estimation processing parameters and image preprocessing is used. Next, the estimated defect location is photographed at high magnification with the field of view narrowed to obtain an observation defect image (S1711), and the inspection defect image and the observation defect image are stored in the image storage area 119 (S1712). The above processing of S1707 to S1712 is repeatedly performed on all the defects to be observed on the sample wafer 104.

[0192] <<Learning Sequence>>

[0193] Regarding the learning sequence of S1705, apart from S407, it has the same characteristics as described in Example 1. Figure 4 The processing flow shown is the same sequence.

[0194] use Figure 18 Explanation and Figure 4 The method for calculating the estimated processing parameters corresponding to S407. In calculating the estimated processing parameters, design data is not required; learning defect images and learning reference images are used.

[0195] Steps S1801 to S1803 and Figure 6 Steps S601 to S603 are the same, so their description is omitted. Next, the local defect image used for learning and the local reference image used for learning are compared to determine the defect location in the local defect image used for learning, and a specific defect location image representing the defect location is obtained (S1804: equivalent to...). Figure 2 (S211). Alternatively, in S1804, instead of using the learning defect image and the learning reference image to determine the defect location, the learning defect image can be displayed in the GUI, the user specifies the defect location, and the image that can distinguish the specified defect location from areas other than the specified defect location is used as the specific defect location image. Next, the estimation processing parameters during learning are initialized (S1805: equivalent to...). Figure 6 (S604). At this time, previously learned estimated processing parameters can also be read from the estimated processing parameter storage area 1601 and used as the initial values ​​of the estimated processing parameters during learning.

[0196] Next, in the defect location estimation unit 1603, based on the estimation processing parameters learned, the defect location is estimated using the local defect image learned (S1806), the estimated defect location image is obtained, the estimation error is calculated using the estimated defect location image and the specific defect location image (S1807), and the estimation processing parameters learned are updated to reduce the estimation error (S1808: equivalent to...). Figure 6 (S607). Next, confirm whether the learning termination condition obtained in S406 is met (S1809: equivalent to...). Figure 6 In step S608, if the learning termination condition is met, the estimated processing parameters during learning are stored as estimated processing parameters in the estimated processing parameter storage area 1601 (S1810). The learning termination condition is as described in Example 1. If the learning termination condition is not met, the process returns to step S1806. Figure 18 The process is complete.

[0197] As a method for estimating defect locations based on local defect images learned in S1806, the following is used: Figure 19 The neural network shown can be used, where the input image Y is the local defect image used for learning, and F(Y) is the image of the inferred defect location. Figure 21 Equation 12 is used to calculate the estimated defect area image F(Y). In Equation 12, F(Y)(x, y, c) represents the pixel value of the estimated defect area image F(Y) with coordinates x in the X direction, y in the Y direction, and channel c. The same applies to F15(Y)(x, y, c). This is illustrated in Example 1. Figure 20 Equations 1 to 8 are used to calculate F11(Y), F12(Y), F21(Y), F22(Y), F23(Y), F24(Y), F31(Y), and F32(Y). Additionally, in S1807, when calculating the estimated error of the defect location, [the following is used]... Figure 21 Equation 13 converts the image DD (a 1-channel binary image) of a specific defect location into a 2-channel image MSK, and calculates the difference (error) between the image MSK and the inferred defect location image F(Y). As a method to quantify the difference (error) between images, mean square error can be used, similar to S607.

[0198] Here, W9 consists of two filters of size (c1×2)×f9×f9, and B9 is a 2D vector. f9 is a hyperparameter determined by the user before learning the sequence; for example, f9 = 3. The parameters calculated through the estimation processing (S1801~S1810) are W1~W6, W9, B1~B6, and B9.

[0199] According to this embodiment, a learning defect image is used to calculate an estimation processing parameter for estimating the defect location in the defect image. During sample observation, the defect location in the inspection defect image is estimated based on the estimation processing parameter, thereby eliminating the need to obtain reference images and improving the throughput of sample observation.

[0200] Summary

[0201] The following content was described in Examples 1 to 3 above. Furthermore, the numbers and letters enclosed in parentheses in the following description do not indicate the execution order of the processes.

[0202] <<Viewpoint 1>>

[0203] A sample observation system comprising a scanning electron microscope and a computer, wherein the computer performs the following processing:

[0204] (1) Obtain multiple images captured by the scanning electron microscope;

[0205] (2) Obtain a learning defect image containing the defective part and a learning reference image not containing the defective part from the plurality of images;

[0206] (3) Using the learning defect image and the learning reference image, the estimated processing parameters are calculated;

[0207] (4) Obtain an inspection image containing the defective area; and

[0208] (5) Using the estimated processing parameters and the inspection defect image, an estimated pseudo-reference image is generated.

[0209] In the process described in (3):

[0210] (3A) Based on a predetermined evaluation value, the position of the learning defect image and the learning reference image is aligned to obtain the alignment amount;

[0211] (3B) Based on the above alignment, a local defect image for learning is cut out from the defect image for learning;

[0212] (3C) Based on the alignment amount, cut out a learning partial reference image from the learning reference image; and

[0213] (3D) The estimated processing parameters are calculated using the local defect image and the local reference image used for learning.

[0214] <<Viewpoint 2>>

[0215] In the sample observation system described in viewpoint 1,

[0216] The computer performs the following processing:

[0217] (6) Compare the pseudo-reference image with the inspection defect image to determine the defect location of the inspection defect image.

[0218] <<Viewpoint 3>>

[0219] In the sample observation system described in viewpoint 1,

[0220] The estimated processing parameters are the parameters of the neural network.

[0221] In the neural network, the minimum size of the image input to the input layer is the first size.

[0222] As part of the process described in (1), the computer acquires the plurality of images, each of which is a first size or larger.

[0223] <<Viewpoint 4>>

[0224] In the sample observation system described in viewpoint 1,

[0225] The estimated processing parameters are the parameters of the neural network.

[0226] In the neural network, the minimum size of the image input to the input layer is the first size.

[0227] The scanning electron microscope sets the image size based on the maximum value of the first size and the alignment amount.

[0228] Take multiple images based on the set dimensions.

[0229] The maximum value of the alignment amount is determined based on the stage movement error and the electron beam irradiation positioning error of the scanning electron microscope.

[0230] <<Viewpoint 5>>

[0231] In the sample observation system described in viewpoint 1,

[0232] The computer obtains the termination condition for the calculation and processing of the estimated processing parameters.

[0233] As part of the process described in (3), the computer performs the following process:

[0234] (3F) When the termination condition is detected, the update of the estimated processing parameters is terminated.

[0235] <<Viewpoint 6>>

[0236] In the sample observation system described in viewpoint 5,

[0237] The computer performs the following processing in parallel with the calculation of the estimated processing parameters:

[0238] (7) Compare the learning defect image with the learning reference image to determine the defect location of the learning defect image.

[0239] <<Viewpoint 7>>

[0240] In the sample observation system described in viewpoint 1,

[0241] The computer omits the acquisition of reference images corresponding to the defect images used for inspection.

[0242] <<Viewpoint 8>>

[0243] A sample observation system comprising a scanning electron microscope and a computer, wherein,

[0244] The computer performs the following processing:

[0245] Multiple images were acquired using the scanning electron microscope;

[0246] Obtain a learning defect image containing the defect location from the plurality of images;

[0247] Using the defective images used for learning, estimated processing parameters are calculated;

[0248] Obtain defect images for inspection, including the defective areas; and

[0249] The estimated processing parameters and the inspection defect image are used to estimate the defect location in the inspection defect image.

[0250] In the calculation of the estimated processing parameters:

[0251] Based on a predetermined evaluation value, the positions of the learning defect image and the learning reference image are aligned to obtain the alignment amount.

[0252] Based on the alignment amount, a local defect image for learning is cut out from the defect image for learning;

[0253] Based on the alignment amount, a learning partial reference image is cut out from the learning reference image; and

[0254] The estimated processing parameters are calculated using the local defect image and the local reference image used for learning.

[0255] Furthermore, it has been explained that the processing described above can also be implemented by an image processing program executed by the processing unit. The image processing program can be distributed via a computer-readable storage medium or from a distribution server computer. Here, the distribution server computer has a storage unit, a processing unit, and a network interface 112. Specific examples of each unit can also be the same as in the case of computer 103. The image processing program can also be stored in the storage unit of the distribution server computer with such a structure, and upon receiving a distribution request from computer 103, the processing unit reads the image processing program and sends it to computer 103 via the network interface 112.

[0256] Symbol Explanation

[0257] 100: Sample observation system

[0258] 101: SEM

[0259] 102: Defect Inspection Device

[0260] 103: Computer.< / gui>

Claims

1. A sample observation system comprising a scanning electron microscope and a computer, characterized in that, The computer performs the following processing: (1) Obtain multiple images captured by the scanning electron microscope; (2) Obtain a learning defect image containing the defective part and a learning reference image not containing the defective part from the plurality of images; (3) Using the learning defect image and the learning reference image, calculate the estimated processing parameters; (4) Obtain an inspection image containing the defective area; and (5) Using the estimated processing parameters and the inspection defect image, an estimated pseudo-reference image is derived. In the process described in (3): (3A) Based on a predetermined evaluation value, the position of the learning defect image and the learning reference image is aligned to obtain the alignment amount; (3B) Based on the alignment amount, cut out a local defect image for learning from the defect image for learning; (3C) Based on the alignment amount, cut out a learning partial reference image from the learning reference image; and (3D) The estimated processing parameters are calculated using the local defect image and the local reference image used for learning.

2. The sample observation system according to claim 1, characterized in that, The computer performs the following processing: (6) Compare the pseudo-reference image with the inspection defect image to determine the defect location of the inspection defect image.

3. The sample observation system according to claim 1, characterized in that, The estimated processing parameters are the parameters of the neural network. In the neural network, the minimum size of the image input to the input layer is the first size. As part of the process described in (1), the computer performs the following process: Each image is one of the plurality of images that is larger than or equal to the first size.

4. The sample observation system according to claim 1, characterized in that, The estimated processing parameters are the parameters of the neural network. In the neural network, the minimum size of the image input to the input layer is the first size. The scanning electron microscope sets the image size based on the maximum value of the first size and the alignment amount. Take multiple images based on the set dimensions. The maximum value of the alignment amount is determined based on the stage movement error and the electron beam irradiation positioning error of the scanning electron microscope.

5. The sample observation system according to claim 1, characterized in that, The computer obtains the termination condition for the calculation and processing of the estimated processing parameters. As part of the process described in (3), the computer performs the following process: When the termination condition is detected, the update of the estimated processing parameters is terminated.

6. The sample observation system according to claim 5, characterized in that, In parallel with the calculation of the estimated processing parameters, the computer performs the following processing: (7) The defect location of the learning defect image is determined by comparing the learning defect image with the learning reference image.

7. The sample observation system according to claim 1, characterized in that, The computer omits the acquisition of reference images corresponding to the defect images used for inspection.

8. The sample observation system according to claim 1, characterized in that, The learning reference image is an image of an area designed to form the same circuit pattern as the circuit pattern captured in the learning defect image. Through the processing described in (3A), the coordinate difference between the learning reference image and the learning defect image when the position is aligned is obtained as the alignment amount.

9. The sample observation system according to claim 8, characterized in that, The learning reference image is an image of the area of ​​the circuit pattern formed in a chip offset by one chip from the chip with the circuit pattern captured in the learning defect image, among a plurality of identical chips arranged in a wafer as photographed by the scanning electron microscope.

10. A sample observation system comprising a scanning electron microscope and a computer, characterized in that, The computer performs the following processing: Multiple images were acquired using the scanning electron microscope; Obtain a learning defect image containing the defective area and a learning reference image not containing the defective area from the plurality of images; The learning defect image and the learning reference image are used to calculate the estimated processing parameters; Obtain defect images for inspection, including the defective areas; as well as The estimated processing parameters and the inspection defect image are used to estimate the defect location in the inspection defect image. In the calculation of the estimated processing parameters: Based on a predetermined evaluation value, the positions of the learning defect image and the learning reference image are aligned to obtain the alignment amount. Based on the alignment amount, a local defect image for learning is cut out from the defect image for learning; Based on the alignment amount, a learning partial reference image is cut out from the learning reference image; and The estimated processing parameters are calculated using the local defect image and the local reference image used for learning.

11. A computer-executed image processing method, characterized in that, The image processing method performs the following steps: (M1) Obtain multiple images taken with a scanning electron microscope; (M2) Obtain a learning defect image containing the defect location and a learning reference image not containing the defect location from the plurality of images; (M3) Using the learning defect image and the learning reference image, the estimated processing parameters are calculated; (M4) Obtain an inspection image containing the defect replacement area; as well as (M5) Using the estimated processing parameters and the inspection defect image, a pseudo-reference image is estimated. As mentioned above (M3), it includes: (M3A) Based on a predetermined evaluation value, the position of the learning defect image and the learning reference image are aligned to obtain the alignment amount; (M3B) Based on the alignment amount, a local defect image for learning is cut out from the defect image for learning; (M3C) Based on the alignment amount, cuts out a learning partial reference image from the learning reference image; and (M3D) uses the learning local defect image and the learning local reference image to calculate the estimated processing parameters.

12. The image processing method according to claim 11, characterized in that, (M6) The pseudo-reference image is compared with the inspection defect image to determine the defect location of the inspection defect image.

13. The image processing method according to claim 11, characterized in that, The estimated processing parameters are the weight values ​​of the neural network. In the neural network, the minimum size of the image input to the input layer is the first size. As stated in (M1): Each of the plurality of images is the first size or larger.

14. The image processing method according to claim 11, characterized in that, The estimated processing parameters are the weight values ​​of the neural network. In the neural network, the minimum size of the image input to the input layer is the first size. The scanning electron microscope sets the image size based on the maximum value of the first size and the alignment amount. Take multiple images based on the set dimensions. The maximum value of the alignment amount is determined based on the stage movement error and the electron beam irradiation positioning error of the scanning electron microscope.

15. The image processing method according to claim 11, characterized in that, The termination condition for obtaining the estimated processing parameters is... As stated in (M3): When the termination condition is detected, the update of the estimated processing parameters is terminated.

16. The image processing method according to claim 15, characterized in that, The calculation of the estimated processing parameters is performed in parallel: (M7) the defect location of the learning defect image is determined by comparing the learning defect image with the learning reference image.

17. The image processing method according to claim 11, characterized in that, The acquisition of the reference image corresponding to the defect image used for inspection is omitted.

18. The image processing method according to claim 11, characterized in that, The learning reference image is an image of an area designed to form the same circuit pattern as the circuit pattern captured in the learning defect image. Through the processing described in (M3A), the coordinate difference between the learning reference image and the learning defect image when position alignment has been performed is obtained as the alignment amount.

19. The image processing method according to claim 18, characterized in that, The learning reference image is an image of the area of ​​the circuit pattern formed in a chip offset by one chip from the chip with the circuit pattern captured in the learning defect image, among a plurality of identical chips arranged in a wafer as photographed by the scanning electron microscope.

Images