Defect inspection system, defect inspection method, computer

Abstract

The purpose of this invention is to efficiently obtain images for learning by a machine learning model when inspecting images of an object using a machine learner. The defect inspection system of this invention segments a layout image of the structure of a sample into small region images, groups these small region images, and captures images of the small regions within each group at a resolution higher than that of the layout image (see Figure 2).

Description

Technical Field

[0001] This invention relates to a defect inspection system. Prior Technology

[0002] Prior art in the technical field to which this invention pertains is Patent Document 1. Patent Document 1 describes a "teacher data generation apparatus that generates teacher data, characterized by comprising: an image receiving unit that receives a defect image of a predetermined size including a detection area of ​​a defect from an inspection device used to photograph an object for defect detection, and defect information indicating the range of the detection area in the defect image; a cropped image generation unit that, based on the defect information, crops out a region including the detection area from the defect image as a cropped image; a display control unit that displays at least a portion of the defect image on a display screen; a determination result receiving unit that receives a determination result of a defect category input by an operator for the defect image displayed on the display screen; and a teacher data generation unit that marks the cropped image with the determination result to generate teacher data" (Please 1). [Previous Technical Documents] [Patent Literature]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2023-122140 Summary of the Invention

[0004] [The problem that this invention aims to solve] Patent Document 1 generates teacher data by cropping an image from an inspection image, including the defect detection area, thereby generating learning data for a model used in inspection. Since the model used in inspection must learn from learning images acquired under the same camera conditions as during inspection, it is necessary to photograph the image before cropping under the same camera conditions as during inspection in order to create learning data through image cropping. Therefore, in order for the inspection model to learn for detecting minor defects, it is necessary to acquire learning images under high-resolution camera conditions, which requires a significant amount of camera time. Consequently, a considerable amount of preparation time is required until the start of inspection.

[0005] The present invention was developed in view of the aforementioned problems, and its purpose is to "efficiently obtain images for learning by a machine learning model when examining images of objects by a machine learner". [Methods used to solve problems]

[0006] The defect inspection system of the present invention divides the layout image of the structure of the display sample into small area images, groups the aforementioned small area images, and captures the aforementioned small area images in each group at a resolution higher than that of the aforementioned layout image. [Effects of the Invention]

[0007] According to the defect inspection system of the present invention, when inspecting an image of an object by means of a machine learner, it is possible to efficiently obtain an image for learning by the machine learning model. Other issues, configurations, and effects not described above will be clearly understood through the following description of embodiments. Simple Explanation of the Diagram

[0008] [Figure 1] A schematic diagram of the defect inspection system 1 in implementation form 1. [Figure 2] shows the flowchart of the processing flow of defect inspection system 1. [Figure 3] shows an example of the layout image obtained in step S1. [Figure 4] Flowchart of the grouping process using layout image 5 as the object in step S2. [Figure 5] shows an example of small region segmentation of layout image 5 in step S21. [Figure 6] shows an example of clustering processing using image features. [Figure 7] shows an example of the result of grouping the layout image. [Figure 8] shows an example of the setting of the second region 14 using the result of grouping processing. [Figure 9] shows an example of the camera coordinate setting based on the learning image of region 2 14. [Figure 10] shows an example of a learning image obtained under the second camera condition. [Figure 11] Flowchart of the process of learning the good product image estimation model in step S5. [Figure 12] Step S6 uses the flowchart of the inspection process of the learned good product image estimation model. [Figure 13] shows an example of the results of inspection using the good product image estimation model. [Figure 14A] An example of a GUI for a defect inspection system. [Figure 14B] An example of a GUI for a defect inspection system. [Figure 15] shows the pattern of obtaining the layout image in implementation mode 2. [Figure 16] shows the pattern of obtaining the layout image in implementation mode 3. [Figure 17] shows an example of camera coordinate settings that take into account differences in image quality within the camera's field of view. [Figure 18] shows an example of a GUI for setting up an inspection area using grouping results. [Figure 19] Flowchart of the learning process of the defect extraction model. [Figure 20] Flowchart of defect inspection using the defect extraction model. [Figure 21] shows an example of one method for selecting the second region. [Figure 22] Schematic diagram of the defect inspection system 1 in embodiment 8. Implementation

[0009] In this specification, the imaging device includes a charged particle beam device and an imaging device based on an optical microscope. Furthermore, the charged particle beam device includes Scanning Electron Microscope (SEM), Re-examination SEM, and Focused Ion Beam (FIB), etc. The following is an example of using an SEM as a charged particle beam device.

[0010] The following uses drawings to illustrate embodiments of the present invention. Furthermore, the embodiments described below are not limited to the invention covered by the claims, and not all of the elements and combinations described in the embodiments are necessarily necessary for the solution of the invention.

[0011] <Implementation Mode 1> Figure 1 is a schematic diagram of the defect inspection system 1 according to Embodiment 1 of the present invention. The defect inspection system 1 includes a charged particle beam device 2 and a computer 3. In Embodiment 1, the system comprising the charged particle beam device 2 and the computer 3 is described as the defect inspection system 1. Embodiment 1 will be described using the example of "using an image obtained by photographing a sample with an optical microscope, and setting an area for the charged particle beam device 2 to acquire a learning image".

[0012] <The Composition of a Defect Inspection System> The defect inspection system 1 includes a charged particle beam device 2 and a computer 3. The charged particle beam device 2 includes a camera device 100 and a controller 200. The camera device 100 has components such as a sample stage (platform) or drive circuitry within a housing (in other words, a column or sample chamber). The controller 200 is a control system that drives and controls the camera device 100, and can be installed via a computer or circuitry. The computer 3 is, in other words, a computer system. The defect inspection system 1 includes components necessary for generating signal waveforms or images based on the detection signals from the camera device 100 constituting the charged particle beam device 2. An example of the charged particle beam device 2 is a scanning electron microscope (SEM).

[0013] The camera device 100 outputs a detection signal obtained by the detector 112 based on the irradiation of the sample 109 on the platform 110 by the charged particle beam 103, and also outputs reflected or scattered light from the laser irradiated by the optical microscope 113 as a detection signal. The controller 200 processes the input or received detection signals, thereby generating and storing images and other signals as detection signals for the charged particle beam device 2. The controller 200 also outputs images and other signals. The computer 3 (computer system) inputs or receives images and other signals and processes them.

[0014] In the imaging device 100, the charged particle beam 103 drawn from the electron source 101 via the lead-out electrode 102 is concentrated by a focusing lens 104. The charged particle beam 103 is deflected by the scanning deflector 105, thereby scanning the surface of the sample 109 on the platform 110 in one or two dimensions. The charged particle beam 103 is decelerated by a negative voltage (deceleration voltage) applied to the electrodes built into the sample stage, i.e., the platform 110, and is focused and irradiated onto the sample 109 by the lensing action of the object lens 108. The platform 110 can also be a mechanism capable of moving in the vertical direction (perpendicular to the X and Y directions in the Z direction), and can also be a mechanism capable of rotating or tilting in any axial direction.

[0015] The details of the installation of the charged particle beam device 2 are not limited, and it can also be configured to irradiate multiple beams. Furthermore, although the diagram shows the charged particle beam device 2 as having one detector 112, it is not limited to this and can also be configured to have multiple detectors. For example, it could be configured to have an SE detector for detecting Secondary Electrons (SE) and a BSE detector for detecting Backscattered Electrons (BSE). Also, for example, it could be configured to have multiple detectors at multiple locations. That is, as a configuration for capturing images, it could also be configured to have multiple channels and multiple detection systems. Furthermore, when generating an image, multiple images can be processed by integration and other methods based on multiple signals repeatedly detected from the same area, thereby generating a single image.

[0016] Sample 109 is, for example, a semiconductor wafer. When sample 109 is irradiated with a beam of charged particles 103, electrons 111, such as secondary electrons (SE) and backscattered electrons (BSE), are emitted from within sample 109. These emitted electrons 111 are accelerated by a negative voltage (decelerating voltage) applied to sample 109 and captured by detector 112. The detection signal output from detector 112 is sent to controller 200. Controller 200 receives the detection signal via communication device 201.

[0017] The controller 200 includes a communication device 201, a processor 202, a memory 203 (such as RAM or non-volatile memory), a display device 204, and an input / output device 205. These components are interconnected via a bus architecture. The controller 200 controls the imaging performed by the camera device 100 according to a set imaging formula. The processor 202 generates an observation image (examined image) based on the detection signal obtained via the communication device 201, which is an image of the brightness of the amount of electrons captured by the detector 112. The processor 202 stores the generated observation image (examined image) and other data in the memory 203. The processor 202 sends a signal corresponding to the generated observation image (examined image) and other data to the computer 3 via the communication device 201. The computer 3 receives the signal via the communication device 310. In addition, the user U1 can also operate the display device 204 and the input / output device 205 to use the controller 200.

[0018] Communication device 201 is a device equipped with a communication interface with camera device 100 or computer 3. The communication interface may include, for example, a LAN, but is not limited to this. Input / output device 205 is an input or output device. The input or output device may be built-in or externally connected. Examples of input devices include a keyboard, mouse, or microphone. Examples of output devices include a monitor, printer, or speaker.

[0019] Computer 3 includes a communication device 310, a storage device 320, a processor 330, a memory 331 (such as RAM or non-volatile memory), a display device 332, and an input / output device 333. These components are interconnected via a bus architecture. Storage device 320 is a memory with a relatively large memory area. Storage device 320 stores a layout image DB 321, a learning image DB 322, an image to be inspected DB 323, learning completion parameters DB 324, and a program 325. Storage device 331 of computer 3 also stores necessary management information or a database (DB). Furthermore, computer 3 can be connected to external memory devices or servers, and necessary data and information can be stored on these external memory devices or servers, and their data and information can be read and written appropriately. Alternatively, the program 325 within storage device 320 can store the aforementioned programs. Furthermore, the memory 331 can be configured to appropriately store various data and information processed by the processor 330. The memory 331 may store, for example, program execution modules or image data, processing result information, etc. The storage 320, in addition to the database shown in the diagram, may also store processing result information, historical information, screen data, etc.

[0020] Communication device 310 is a device equipped with a communication interface with controller 200. External devices can also be connected to computer (computer system) 3 via a network such as a LAN. Input / output device 333 is an input device or an output device. Input or output devices can be built-in or externally connected. Examples of input devices are keyboards, mice, or microphones. Examples of output devices are monitors, printers, or speakers.

[0021] User U1 of defect inspection system 1 uses computer (computer system) 3 by operating display device 332 and input / output device 333. Thus, user U1 utilizes defect inspection system 1. The display screen on display device 332 displays the user interface of defect inspection system 1; in other words, it displays the screen associated with the graphical user interface (GUI). Alternatively, computer 3 can be configured as a client server system. In this case, user U1 operates a PC or similar device that acts as a client terminal, and the PC or similar device accesses computer (computer system) 3, which acts as a server, via communication.

[0022] Computer (computer system) 3 is the part that executes the characteristic processing of this embodiment. Computer (computer system) 3 is connected to controller 200. Alternatively, controller 200 may be configured as part of computer (computer system) 3, or computer 3 may be configured as part of controller 200. Controller 200 and computer (computer system) 3 may also be configured as an integrated computer system. Controller 200 may also execute at least a portion of the characteristic processing of this embodiment. The entire assembly including computer (computer system) 3 may also be configured as charged particle beam device 2.

[0023] Regarding computer 3 (computer system), when configured as a client-server system, its operation is as follows: User U1 accesses the server, i.e., computer (computer system) 3, from a client terminal such as a PC via the network. The server provides a screen with a graphical user interface (GUI) to the client terminal. The server sends GUI screen data (e.g., a webpage) to the client terminal. The client terminal displays the GUI screen on its monitor based on the received screen data. User U1 views the GUI screen and inputs instructions or settings. The client terminal sends input information to the server. The server performs processing corresponding to the received input information. For example, the server performs image evaluation and device status estimation processing, remembers the processing results, and sends GUI screen data (or simply update information) to the client terminal to display the processing results. The client terminal updates the GUI screen display based on the received screen data. User U1 can view the GUI screen and confirm the processing results, such as the device status estimation results.

[0024] <Defect Inspection System Processing Flow> Figure 2 is a flowchart illustrating the processing flow of defect detection system 1. The steps in Figure 2 are explained below.

[0025] Step S1: The imaging device 100 acquires an image of the first region 4 (described later) on the sample 109 based on the first imaging conditions, and saves it together with accompanying information such as imaging coordinates in the layout image DB321 as a layout image 5 (described later) representing the structure of the first region 4. In this embodiment, the example described is "based on the first imaging conditions including a setting value of imaging magnification, the first region 4 on the sample 109 is irradiated with laser by the optical microscope 113, and an Optical Microscope (OM) image is generated based on the detection signal obtained from the reflected light or scattered light, and set as the layout image 5".

[0026] Step S2: The processor 330 divides the layout image 5 obtained in step S1 into small region images 6 (described later), calculates image feature quantities for each of the small region images 6 and performs clustering processing, thereby outputting a pattern diagram 12, which represents the result of dividing the small regions with the same circuit pattern in the first region 4 into the same cluster 13.

[0027] Step S3: The processor 330 selects at least one representative small region from each of the clusters 13 in the pattern diagram 12 as the second region 14 (described later).

[0028] Step S4: The processor 330, based on the second imaging conditions, acquires an image of the second region 14 on the sample 109 using the imaging device 100 as a learning image and saves it in the learning image DB322. In this embodiment, the example described is "based on the second imaging conditions, including settings such as magnification or pixel resolution, a charged particle beam 103 is irradiated onto the second region 14 on the sample 109, thereby generating an SEM image and saving it in the learning image DB322".

[0029] Step S5: The processor 330 uses the image stored in the learning image DB322 to learn an inspection model for inspecting the defective parts of the sample, and saves the internal parameters of the learned inspection model in the learned parameters DB324.

[0030] Step S6: The camera device 100 irradiates the area of ​​the sample 109 to be inspected with a charged particle beam 103, thereby acquiring an image to be inspected and saving it to the inspected image DB323. The processor 330 reads the internal parameters of the inspection model used for inspection from the learned parameters DB324, reads the inspected image of the sample 109 as the inspection object from the inspected image DB323, and inputs it into the inspection model, thereby outputting a good product image corresponding to the input image. The processor 330 compares the input inspected image with the output good product image to extract the defective areas.

[0031] <Example of Defect Inspection> Using Figures 3 to 13, assuming a defect inspection system is used to inspect defects on a semiconductor wafer, illustrate a specific example up to the point of defect inspection and processing.

[0032] Figure 3 shows an example of a layout image obtained in step S1. In the design of semiconductor devices, structures that repeat the same circuit patterns are sometimes used. While it is necessary to systematically learn the circuit patterns within a specified area when examining a model, when the same circuit pattern is contained within the area, the same circuit pattern is repeated during the acquisition of the learning image, resulting in lengthy and time-consuming imaging, thus increasing the time required to acquire the learning image. Therefore, in this embodiment, an optical microscope with an imaging area wider than that of a SEM is used to acquire layout image 5 and grasp the layout of the circuit pattern. Layout image 5 is acquired to quickly grasp the layout of the circuit pattern formed within the first region 4. The layout image 5 is obtained by "irradiating a laser with an optical microscope 113 on a first imaging condition, including a setting value of magnification, for a first region 4 designated as the inspection area on the sample 109 (e.g., a die formed on a semiconductor wafer), and generating an Optical Microscope (OM) image based on the detection signal obtained from the reflected or scattered light." The semiconductor wafer is a structure formed by the stacking of circuit patterns. In the OM image, information about both the circuit patterns formed on the upper and lower layers is superimposed on the detection signal. Since the SEM image reflects the lower layer circuit pattern through imaging conditions, using the OM image as the layout image allows for consideration of variations in the learning image that includes the lower layer reflected in the SEM.

[0033] Figure 4 is a flowchart of the grouping process in step S2, where the layout image 5 is the object. The grouping process is the process of "using the layout image 5 to identify the regions in the first region 4 that have the same circuit pattern".

[0034] Step S21: The processor 330 divides the layout image 5 into small regions and generates small region images 6.

[0035] Step S22: The processor 330 calculates image feature quantities for each of the small region images 6. In this embodiment, the small region image 6 is input into a trained deep learning model (e.g., VGG-16 or ResNet-50), and the feature maps output from the intermediate layers are calculated as image feature quantities. A deep learning model trained on hierarchical classification can obtain general feature representations to appropriately classify images. Training can also be autoencoder processing (processing the output in a manner similar to the input) or segmentation processing (segmenting each target segmentation region contained in the input image). The intermediate layer feature representations of the trained deep learning model are used as feature quantities representing the construction of the circuit patterns contained in the small region image 6.

[0036] Step S23: The processor 330 uses the image feature quantity calculated in step S22 to cluster the small region image 6.

[0037] Figures 5-7 illustrate an example of clustering treatment.

[0038] Figure 5 illustrates an example of small-region segmentation of the layout image 5 in step S21. In this embodiment, the size of the small region is set to be the same as the field of view size of the second camera condition, and the layout image 5 is segmented in a non-repeating, mesh-like manner. When the size of the small region is the same as the field of view size of the second camera condition, computational efficiency can be well performed in subsequent processing. However, the size of the small region may also be different from the field of view size of the second camera condition, and the region may be repeated during segmentation.

[0039] Figure 6 illustrates an example of clustering using image features. In feature space 7, image features 8 of each small region image 6 are plotted along the axis of image features calculated from the small region image 6. In feature space 7, the distances between image features 8 of small region images 6 with similar circuit pattern layouts become closer. Therefore, in feature space 7, clustering is applied based on the distances between image features 8, thereby dividing each layout of the circuit pattern reflecting the small region image 6 into clusters 9. For example, the g-means algorithm can be used for clustering. Since the g-means algorithm uses statistical tests to dynamically determine the appropriate number of clusters, it prevents small region images 6 with different circuit pattern layouts from being mixed in clusters 9. Clustering can also be achieved using the k-means algorithm or DBSCAN (Density-Based Spatial Clustering of Applications with Noise), etc.

[0040] Figure 7 is an example of the result of grouping the layout image. Pattern 12 is the same size as the layout image 5, and it is a diagram in which "each of the small area images 6 is assigned a cluster index 13 corresponding to its respective cluster 9 and assigned to the corresponding area of ​​the layout image 5". By referring to this pattern 12, the area in the first region 4 where the same circuit pattern is formed can be understood.

[0041] Figure 8 is a diagram illustrating one example of the setting of the second region 14 after grouping processing. In step S3, based on the pattern diagram 12 corresponding to the first region 4, the small region image 6 closest to the centroid is selected as the second region 14 from each of the clusters 9. This is because the small region image 6 closest to the centroid in cluster 9 is the image of the average image feature quantity in its cluster, and can be said to represent the pattern of the cluster. However, it is also possible to select not only the small region image 6 closest to the centroid but also multiple small region images 6 from each of the clusters 9 as the second region 14. For example, if the small region image 6 farthest from the centroid is also selected as the second region, the variation among the clusters 9 can be increased.

[0042] Figure 9 illustrates an example of setting the camera coordinates based on the learning image of the second region 14. The center of the camera field of view 16 in the second camera condition, i.e., the camera coordinates 15, is determined and set as the camera coordinates of the learning image by encompassing the second region 14. Since the second region 14 represents the layout of each circuit pattern, by capturing the second region 14 in a comprehensive manner, a learning image reflecting all variations in the layout of the circuit pattern can be obtained without capturing the entire first region 4.

[0043] Figure 10 shows an example of a learning image obtained under the second camera condition. The second camera condition for obtaining the learning image is the same as the camera condition used during inspection. In this embodiment, the second camera condition is the SEM camera condition, and it is a camera condition that allows obtaining an image with a higher resolution than the OM image obtained under the first camera condition. Learning image 17 is an example of an image captured by SEM under the second camera condition based on the camera coordinates set in the second region 14. By capturing images under the higher resolution second camera condition, detailed shapes of circuit patterns that cannot be captured at the resolution of the first camera condition can be captured. Since it is the same camera condition as during inspection, it can be used for learning the inspection model.

[0044] Figure 11 is a flowchart of the process of learning the good product image estimation model in step S5. In this embodiment, the example is described using the "good product image estimation model as the inspection model". The good product image estimation model is based on a machine learning model, such as an autoencoder, U-Net, or GAN (Generative Adversarial Networks). Furthermore, the good product image estimation model is a model that can be learned using only good product images and does not require collecting defective images, which are considered costly to collect. Each step is implemented by the processor 330.

[0045] S502: First, create a learning data set. The learning data set is created by reading the learning images stored in the learning image DB322.

[0046] S503: Through repeated learning cycles, the following learning procedures are repeatedly implemented.

[0047] S504: Divide the images contained in the learning dataset into an arbitrary number of images and set each as a learning mini-batch, thereby dividing the learning dataset into learning mini-batches.

[0048] S505: For each learning mini-batch, skip S506~S510.

[0049] S506: Obtain a small batch of learning data.

[0050] S507~S510: For each segmented learning mini-batch, perform data augmentation (S507), learning loss calculation (S508~S509), and internal parameter update (S510).

[0051] S507: In data augmentation, based on the reference values ​​of the random number table, brightness conversion, contrast conversion, and deformation addition are applied to each of the images contained in the learning mini-batch with random intensity within a constant range, thereby creating the input image for the good image estimation model.

[0052] S508~S509: To enable the good image estimation model to learn the estimation process from defective images to good images, noise is added to the input image as a defective feature, creating a defective image. This defective image is then input into the good image estimation model to obtain the estimated good image as output. At this point, the average of the squared errors of the brightness of each pixel between the estimated good image and the image before adding noise (i.e., the input image) is calculated as the learning loss. The learning loss can also include image quality-related metrics such as SSIM (Structural Similarity) or regularization terms that penalize the internal parameters to prevent overfitting of the training dataset.

[0053] S510: Based on the calculated learning loss, use optimization techniques such as gradient descent to search for the values ​​of the learner's internal parameters such as the learning loss decreases, and update the internal parameters of the good image estimation model according to the search results.

[0054] S511~S512: After repeating the processing from data expansion (S507) to internal parameter update (S510) an equal number of times as the number of learning mini-batches (S511), the learning dataset is split into learning mini-batches again, and the internal parameters are updated for each learning mini-batch (S512=>S505). By repeating this learning loop an arbitrary number of times, the internal parameters of the good image estimation model are determined.

[0055] Figure 12 is a flowchart of the inspection process using the learned good product image estimation model in step S6. The camera device 100 scans the area of ​​the sample 109 to be inspected using a charged particle beam 103 and saves the acquired image to be inspected in the image DB323. The processor 330 reads the image of the object to be inspected from the image DB323. The processor 330 reads the internal parameters of the good product image estimation model from the learned parameters DB324. The processor 330 inputs the image to be inspected into the good product image estimation model with the internal parameters read in, and outputs the estimated good product image. When an image to be inspected containing defects is input into the good product image estimation model, an image of the defective part is output, transformed like a good product. The processor 330 compares the image to be inspected and the estimated good product image to capture the defective part, thereby creating a difference image where each pixel represents the difference in pixel values ​​between the image to be inspected and the estimated good product image. Processor 330, based on the image features of the differential image, excludes false positives that indicate areas that, although showing differences, are not actually defects. Processor 330 outputs information related to the captured defect areas as the inspection result.

[0056] Figure 13 shows an example of the inspection result using a good product image estimation model. In the inspection, an image 18 to be inspected and an estimated good product image 19 are used. The estimated good product image 19 is obtained by inputting the image 18 to the good product image estimation model with learned parameters and outputting it. In the comparison processing unit, the difference values ​​of each pixel between the image 18 to be inspected and the estimated good product image 19 are calculated, thereby outputting a difference image 20. Since the difference regions generated in the comparison processing unit are regions where the image to be inspected is transformed into different shapes, these are regions with a high probability of defects. Next, false alarm detection processing is performed on the difference image 20 to exclude regions that, although showing differences, are not actually defects, and an inspection result 21 representing the defect location is output. In the false alarm detection process, connected regions in the difference image are considered as potential defects. For each potential defect, the mean, variance, maximum value, roundness, and area of ​​the difference values ​​are calculated as feature quantities. It is then determined whether each feature quantity is outside the normal range, thereby identifying whether it is a defect or a false alarm. Judgment criteria include, for example, area below a threshold, maximum difference value below a threshold, and average difference value below a threshold.

[0057] Figures 14A and 14B show examples of the GUI of a defect inspection system. Figure 14A shows the first half of the screen, and Figure 14B shows the second half. As shown in Figure 14A, the GUI of this embodiment includes an input field 1400 for learning parameters of a good product image estimation model and a display field 1410 for learning images. The input field 1400 for learning parameters includes an input field 1401 for the learning area and an input field 1406 for learning parameters. The input field 1401 for the learning area includes a display field 1405 containing information on the arrangement and coordinates of the wafers manufactured on the sample, an input field 1402 for the wafer number of the learning area, an input field 1403 for the wafer coordinates of the upper left vertex of the rectangle used to specify the learning area within the specified wafer, and an input field 1404 for the wafer coordinates of the lower right vertex of the rectangle. The input field 1406 for learning parameters is used to set the parameters for learning the good product image estimation model. It includes an input field 1407 for the number of learning iterations and an input field 1408 for determining the batch size of the small batch. After entering the parameters, when the learning start button 1409 is pressed, the processing equivalent to steps S1 to S5 is executed.

[0058] The display bar 1410 for the learning image includes a display bar 1411 for the acquired layout image, a pattern diagram 1412 output as the result of grouping the layout image, and a display bar 1413 for accompanying information such as the acquired learning image, coordinates, and cluster name. The display bar 1411 for the layout image displays the area in the layout image where the learning image was acquired, using rectangles or similar shapes.

[0059] As shown in Figure 14B, the GUI of this embodiment has an input field 1420 for inspection parameters and a display field 1430 for inspection results. The input field 1420 for inspection parameters includes an input field 1421 for the inspection area and an input field 1426 for specifying the learning completion parameters of the model used for inspection. The input field 1421 for the inspection area includes a display field 1425 for the arrangement and coordinate information of the wafers manufactured on the sample, an input field 1422 for specifying the wafer number of the wafers to be inspected using numbers or "ALL" to represent all wafers, an input field 1423 for specifying the wafer coordinates of the upper left vertex of the rectangle used to specify the inspection area within the specified wafer, and an input field 1424 for specifying the wafer coordinates of the lower right vertex of the rectangle. The input field 1426 for learning completion parameters allows inputting the path to a data file containing the learning completion parameters of the model used for inspection, thereby specifying the learning completion parameters used for inspection. After inputting the parameters, when the check start button 1427 is pressed, the process equivalent to step S6 is executed. During and after the execution, information related to the check results is displayed in the check result display bar 1430.

[0060] The inspection results display panel 1430 includes an image display panel 1431 related to the inspection and a defect distribution display panel 1432 on the sample. The image display panel 1431 displays the inspected image, the output of the good product image estimation model (i.e., the estimated good product image), and an image showing the captured defect locations, along with accompanying information such as the wafer number or coordinate information. The defect distribution display panel 1432 displays the number and density of defects detected in each area of ​​the sample (e.g., each wafer manufactured on a semiconductor wafer), and a wafer map 1433 showing the defect density using color bars.

[0061] <Implementation Mode 1: Summary> According to the defect inspection system 1 of embodiment 1, the lengthy shooting of learning images with the same circuit pattern layout can be reduced, and an imaging method and defect inspection system that "can efficiently obtain learning images of the model for inspection" can be provided.

[0062] <Implementation Mode 2> Figure 15 shows the layout image obtained in Embodiment 2 of the present invention. The basic structure of Embodiment 2 is the same as / shared with Embodiment 1. Hereinafter, we will mainly describe the components in Embodiment 2 that are different from those in Embodiment 1. In Embodiment 2, as the first imaging condition for obtaining the layout image, a SEM with a lower magnification than the second imaging condition is used instead of an optical microscope.

[0063] Figure 15 shows an example of a layout image obtained using a SEM under low-magnification imaging conditions. While reducing the magnification of an SEM makes it difficult to resolve the shape of fine circuit patterns, it allows for a wide imaging area at any given time. Since the purpose of obtaining a layout image is to grasp the approximate structure within a specified area rather than for inspection, a low-magnification SEM image can be used. Therefore, in this embodiment, a layout image is obtained using a SEM with a lower magnification than the second imaging condition, serving as the first imaging condition. The layout image 22, obtained at low magnification, is a blurred image of the circuit pattern compared to image 23 obtained under the same imaging conditions as inspection (i.e., the second imaging condition), but the approximate structure can still be grasped. Even with a low-magnification SEM, the resolution is higher than that of an optical microscope, and the shape of the circuit pattern is clearly visible. Therefore, compared to using an image obtained with an optical microscope as the layout image, the area for obtaining the learning image can be determined based on detailed structural information.

[0064] According to the defect inspection system 1 of embodiment 2, in addition to the effects of embodiment 1, it can further increase the effect of reducing the number of times the learning image is captured.

[0065] <Implementation Mode 3> Figure 16 shows the state of obtaining a layout image in Embodiment 3 of the present invention. The basic structure of Embodiment 3 is the same as / shared with Embodiment 1. Hereinafter, we will mainly describe the components in Embodiment 3 that are different from those in Embodiment 1. In Embodiment 3, as the first imaging condition for obtaining the layout image, an SEM with a lower pixel resolution than the second imaging condition is used instead of an optical microscope.

[0066] Figure 16 shows an example of a layout image obtained using a SEM under low pixel resolution imaging conditions. SEM, by reducing the resolution of each pixel (pixel resolution) without changing the camera's field of view, can broaden the imaging area per unit time. The layout image 24, obtained under the same camera conditions as the second method but with reduced pixel resolution, is a blurred image of the circuit pattern, but its general structure is still discernible. Even with SEM under reduced pixel resolution imaging conditions, because the resolution is higher than that of an optical microscope and the shape of the circuit pattern is clearly shown, the area of ​​the learning image can be determined based on detailed structural information compared to using an image obtained with an optical microscope as the layout image. Furthermore, because the total number of pixels in the obtained layout image is reduced by lowering the pixel resolution, the computational load in subsequent processing is reduced, thus shortening the processing time.

[0067] According to the defect inspection system 1 of embodiment 2, in addition to the effects of embodiment 1, it can shorten the group processing time and reduce the number of times the learning images are photographed.

[0068] <Implementation Mode 4> Using Figure 17, Embodiment 4 of the present invention will be explained. The basic structure of Embodiment 4 is the same as / shared with Embodiment 1. Hereinafter, we will mainly explain the components in Embodiment 4 that are different from those in Embodiment 1. In Embodiment 4, the setting of camera coordinates based on the learning image of the second region in step S4 is implemented considering the difference in image quality within the field of view of the second camera condition.

[0069] Figure 17 illustrates an example of camera coordinate settings that take into account differences in image quality within the field of view. In imaging systems such as optical microscopes or SEMs, aberrations occur because the actual imaging position sometimes deviates from the ideal imaging position due to differences in the incident positions of light rays or electron beams to the lens. For example, in electron beams passing near the center of the lens and those passing away from the center, the different imaging positions and curved imaging planes result in images with different levels of blur near the center and at the periphery. Since inspection models process learned image features, if the relationship between the circuit pattern and its position within the field of view differs between the learning and inspection periods, it can lead to incorrect detection of good parts or undetected defects. As a countermeasure, in this embodiment, when setting the camera coordinates based on the second region, the camera coordinates are set to capture the second region while repeating the camera field of view, so that when photographing the circuit patterns contained in each of the second regions, learning images are obtained near the center and periphery of the camera field of view.

[0070] According to the defect inspection system 1 of embodiment 4, in addition to the effects of embodiment 1, it can reduce the false detection of good products or the failure to detect defects due to "the change in the projection mode caused by the positional relationship of the circuit pattern in the camera field of view", and can perform defect inspection with higher accuracy.

[0071] <Implementation Mode 5> Using Figure 18, we will explain Embodiment 5. The basic structure of Embodiment 5 is the same as / shared with Embodiment 1. The following mainly explains the components in Embodiment 5 that are different from those in Embodiment 1. In Embodiment 5, the defect inspection in step S6 is performed on the area determined based on the grouping result, i.e., the pattern diagram.

[0072] The circuit patterns formed on a semiconductor wafer include those intended for inspection and those not requiring inspection. For example, some circuit patterns that do not require inspection may be virtual patterns formed to ensure uniform wiring density within the wafer without affecting wafer operation. Other times, circuit patterns that significantly impact manufacturing yield require focused inspection. Therefore, the importance of inspection varies for each circuit pattern. Since inspecting less important circuit patterns increases inspection time, the inspection area must be set according to the user's inspection objectives. Therefore, in this embodiment, a pattern diagram visualizing the circuit pattern layout is provided, allowing selection of the circuit patterns and areas to be inspected.

[0073] Figure 18 is an example of a GUI for setting an inspection area using grouping results. The inspection area input field 1800 includes a layout image display field 1801, a learning image display field 1801, and an area selection input field 1803. The area selection input field 1803 includes a checkbox for selecting either manual setting 1804 or pattern specification 1805 as the inspection area. In the case of manual setting, the area on the pattern image to be inspected is selected by clicking or using the cursor 1806, such as with a mouse. When the entire area of ​​a specified pattern is selected as the inspection object, the cluster name of the pattern to be inspected is entered in the input field.

[0074] According to the defect inspection system 1 of embodiment 5, in addition to the effects of embodiment 1, the inspection area is reduced to the area with circuit patterns that the user of the defect inspection system wants to focus on, thereby enabling high-speed defect inspection and efficient defect analysis.

[0075] <Implementation Mode 6> Using Figures 19 and 20, Embodiment 6 of the present invention will be described. The basic structure of Embodiment 6 is the same as / shared with Embodiment 1. Hereinafter, we will mainly describe the components in Embodiment 6 that are different from those in Embodiment 1. In Embodiment 6, instead of using a good product image estimation model, a defect extraction model that learns the features of the defective part is used as the inspection model.

[0076] Figure 19 is a flowchart of the learning process for the defect extraction model. The defect extraction model is based on a machine learning model, such as a CNN (Convolutional Neural Network) or U-Net. First, defect teacher labels representing the locations of defects in the learning image are created (S1902). The defect teacher label is, for example, an image where the defective part is set to a pixel value of 1 and everything else is set to a pixel value of 0. The creation of the defect teacher labels can be done manually, such as through visual inspection, or based on the results of inspection methods such as image processing. The learning dataset, which matches the learning image with the created defect teacher labels, is used to learn the defect extraction model (S1903~S1912). The learning is performed by minimizing the error between the defect certainty of each pixel output by the defect extraction model and the defect teacher label. Other processing is the same as in Embodiment 1.

[0077] Figure 20 is a flowchart of defect inspection using a defect extraction model. When the learned defect extraction model is used for inspection, the inspection result representing the defect confidence level of each pixel in the inspected image can be obtained by inputting the image being inspected. Other processing is the same as in Embodiment 1.

[0078] Since the good product image estimation model in implementation form 1 only learns from good product images, it cannot learn the image features of defects. However, since the defect extraction model can learn the image features of defects, it can inspect defects with higher sensitivity.

[0079] According to the defect inspection system 1 of embodiment 6, in addition to the effects of embodiment 1, the sensitivity of defect inspection can be improved by learning the image features of the defect.

[0080] <Implementation Mode 7> Using Figure 21, we will explain regarding Embodiment 7. The basic structure of Embodiment 7 is the same as / shared with that of Embodiment 1. The following mainly explains the components in Embodiment 7 that differ from those in Embodiment 1. In Embodiment 7, in step S3, as the second region, not only are those closest to the centroid of each cluster selected, but also those at a distance of more than a threshold from the centroid are selected.

[0081] Figure 21 illustrates one example of a method for selecting the second region. Clustering is used to form clusters for each small region image that projects the same circuit pattern layout. However, even within the same cluster, the layout of the circuit pattern projected within the image will not be completely identical if the image features differ. For example, there may be differences in the position of the circuit pattern projected within the image, or in noise and minor distortions caused by the imaging device. In cases where these differences are significant, even within the same cluster, there is a possibility that small region images may actually have different layouts. In this case, if only the small region image closest to the centroid of the cluster is selected, there is a possibility of learning omissions in the circuit pattern layout. To prevent this, in this embodiment, the second region is selected not only from those closest to the centroid of each cluster, but also from those at a distance greater than a threshold. The threshold is set, for example, using the average and standard deviation of the distance from the centroid of the cluster. By setting a threshold, since a deviation value boundary 26 centered on the cluster center can be set, and the small area image 27 outside the deviation value boundary can be determined to have a different tendency from the small area image that occupies a large proportion within the cluster, it is selected as the second region.

[0082] According to the defect inspection system 1 of embodiment 7, in addition to the effects of embodiment 1, it can prevent omissions in the learning of circuit patterns and stabilize the inspection performance of the inspection model.

[0083] <Implementation Mode 8> Using Figure 22, we will explain Embodiment 8. The basic structure of Embodiment 6 is the same as / shared with that of Embodiment 1. The following mainly explains the structural parts in Embodiment 8 that are different from those in Embodiment 1. In Embodiment 8, a structure of "interconnecting a computer and multiple devices via a network" is used.

[0084] Figure 22 is a schematic diagram of the defect inspection system 1 in Embodiment 8. The computer 3, multiple charged particle beam devices 2, multiple optical inspection devices 1901, multiple re-inspection SEMs 1902, and multiple manufacturing devices 1903 are interconnected via a network 1900. The computer 3 collects data output from each device via a communication device 310 and stores it in a storage device 320 as needed. Each device connected to the computer 3 can receive action instructions or data through its communication device. In a manufacturing plant, multiple units of these devices may operate simultaneously.

[0085] According to the defect inspection system 1 of embodiment 8, in addition to the effects of embodiment 1, by connecting multiple devices and computers, it is possible to improve the performance of the inspection model by accumulating various learning images or to perform more detailed defect analysis by coordinating the inspection results of multiple manufacturing processes.

[0086] <Regarding variations of the present invention> This invention is not limited to the embodiments described above and may include various modifications. For example, the embodiments described above are embodiments explained in detail for the purpose of easily understanding and illustrating this invention, and are not necessarily limited to having all the components described. Furthermore, it is possible to replace a part of the structure of a certain embodiment with the structure of another embodiment, and it is also possible to add the structure of another embodiment to the structure of a certain embodiment.

[0087] Furthermore, the aforementioned components, functions, processing units, and processing methods can also be partially or entirely implemented in hardware, for example, through integrated circuit design. Additionally, the aforementioned components and functions can also be implemented in software by having a processor interpret and execute programs that implement each function. The information such as programs, tables, and files that implement each function can be stored in memory or recording devices such as hard drives, SSDs (Solid State Drives), or recording media such as IC cards, SD cards, and DVDs.

[0088] Furthermore, the term "control line" or "information line" indicates those deemed necessary in the description, but not all control lines or information lines may be indicated on the product. In fact, it can be assumed that almost all components are interconnected.

[0089] In the above implementation, the model for extracting image features in the intermediate layer (feature extraction model) can be, for example, the following, or any of them can be used: (a) a model that classifies the input image into clusters based on its features; (b) a model that segments the input image into partial images based on its features; or a model that extracts the features of the input image and uses them to regenerate the input image.

[0090] 100: Camera device 101: Electronic Source 102: Lead-out electrode 103: Charged Particle Beam 104: Condensing Lens 105: Scan deflector 108: Object Lens 109: Sample 110: Platform 111: Electronics 112: Detector 113: Optical Microscope 200: Controller 201,310:Communication devices 202,330: Processor 203,331: Memory 204,332: Display device 205,333: Input / output devices 320: Storage 321: Layout Image DB 322: Learning Image DB 323: The image being inspected (DB) 324: Learning completed for parameter DB 325: Program

Claims

1. A defect inspection system that uses an observation image of a sample to inspect for defects contained in the sample, the defect inspection system being characterized by comprising: a camera device for acquiring the observation image; and a computer for inspecting the defects using the observation image, wherein the camera device acquires a layout image representing the structure of a first region of the sample under first imaging conditions; the computer generates small region images by dividing the layout image into smaller regions; the computer calculates image feature quantities of the small region images; the computer groups the small region images based on the image feature quantities; the computer selects small region images from each group obtained by the grouping; the computer sets the small region images selected from the groups as a second region; the camera device acquires an image of the second region under second imaging conditions having a resolution equal to or higher than the first imaging conditions; and the computer learns an inspection model for inspecting the defects using the image of the second region. The aforementioned computer uses the aforementioned inspection model to inspect the aforementioned defects.

2. As in request item 1, a defect inspection system, wherein, The aforementioned computer is equipped with a feature extraction model, which extracts the feature quantities of the aforementioned observed image. The aforementioned computer inputs the aforementioned small region image into the aforementioned feature extraction model, thereby obtaining the feature quantities that can be obtained from the intermediate layer of the aforementioned feature extraction model as the image feature quantities of the aforementioned small region image.

3. As in request item 1, a defect inspection system, wherein, The aforementioned computer selects the small region image that is closest to the centroid of each of the aforementioned groups from the feature space represented by the aforementioned image features.

4. As in request item 1, a defect inspection system, wherein, The aforementioned camera device is used to capture images of all planar areas within the area designated as the aforementioned second region.

5. As in request item 1, a defect inspection system, wherein, The aforementioned inspection model is configured as a good product image estimation model that "presumes to display a good product image that is presumed to have the appearance of the aforementioned input image when there is no aforementioned defect in the input image". The aforementioned computer compares the aforementioned good product image obtained by inputting the aforementioned input image into the aforementioned good product image estimation model with the aforementioned input image, thereby extracting the aforementioned defects contained in the aforementioned input image.

6. As in request item 1, a defect inspection system, wherein, The aforementioned computer provides a user interface that displays at least one of the aforementioned observed image, the aforementioned layout image, the aforementioned small area image, the aforementioned grouping results, and the results of checking the aforementioned defects.

7. As in request item 1, a defect inspection system, wherein, The aforementioned imaging device is composed of an optical microscope and a charged particle beam device. The aforementioned first imaging condition refers to the imaging condition of the aforementioned optical microscope, and the aforementioned second imaging condition refers to the imaging condition of the aforementioned charged particle beam device.

8. As in request item 1, a defect inspection system, wherein, The aforementioned camera device is composed of a charged particle beam device. The aforementioned first camera condition and the aforementioned second camera condition are camera conditions of the aforementioned charged particle beam device. The aforementioned first camera condition is a magnification of the aforementioned second camera condition or lower.

9. The defect inspection system as described in Request 1, wherein, The aforementioned camera device is composed of a charged particle beam device. The aforementioned first camera condition and the aforementioned second camera condition are the camera conditions of the aforementioned charged particle beam device. The aforementioned first camera condition refers to the image resolution below the aforementioned second camera condition.

10. The defect inspection system as described in Request 1, wherein, The aforementioned computer sets the aforementioned second region by "multiplely acquiring the region containing the center of the image of the aforementioned small region selected from the aforementioned group while staggering the camera range".

11. As in request item 1, a defect inspection system, wherein, The aforementioned computer displays a pattern diagram showing the positions of the aforementioned groups and the aforementioned small area images obtained through the aforementioned grouping. The aforementioned computer sets the aforementioned second area according to the results of the aforementioned small area images specified by the user as the objects of inspection on the aforementioned pattern diagram.

12. As in request item 1, a defect inspection system, wherein, The aforementioned inspection model is a defect extraction model configured to "predict the location of the aforementioned defect in the input image". The aforementioned computer inputs the aforementioned input image into the aforementioned defect extraction model, thereby extracting the aforementioned defect contained in the aforementioned input image.

13. As in request item 1, a defect inspection system, wherein, The aforementioned computer selects, from the aforementioned group, the aforementioned small region image that is closest to the centroid of the aforementioned group and the aforementioned small region image that has a distance of more than a threshold from the aforementioned centroid in the feature space represented by the aforementioned image features.

14. A defect inspection method, which uses an observation image of a sample to inspect defects contained in the sample, the defect inspection method being characterized by comprising: a step of acquiring a layout image representing the structure of a first region of the sample under a first imaging condition; a step of creating a small region image that divides the layout image into small regions; a step of calculating image feature quantities of the small region images; a step of grouping the small region images based on the image feature quantities; a step of selecting the small region images from each group obtained by the grouping; a step of setting the small region images selected from the groups as a second region; a step of acquiring an image of the second region under a second imaging condition having a resolution equal to or higher than the first imaging condition; a step of learning an inspection model for inspecting the defects using the image of the second region; and a step of inspecting the defects using the inspection model.

15. A computer that uses observed images of a sample to inspect for defects contained in the sample, the computer being characterized by comprising: a processor configured to perform the following steps: acquiring a layout image of the structure of a first region of the sample captured by a first camera condition; generating small region images by dividing the layout image into small regions; calculating image feature quantities of the small region images; grouping the small region images based on the image feature quantities; selecting the small region images from each group obtained by the grouping; setting the small region images selected from the groups as a second region; acquiring an image of the second region captured by a second camera condition having a resolution equal to or higher than that of the first camera condition; learning an inspection model for inspecting the defects using the image of the second region; and inspecting the defects using the inspection model.

Images