Classification AI learning support device, learning support method, and learning support program

Abstract

This facilitates the detection of defects caused by circuit layouts with a small sample size. [Solution] The learning support device includes an image acquisition unit that acquires a plurality of faulty circuit images 122 showing faulty circuit layouts that cause defects extracted from known circuit layouts and a plurality of normal circuit images 120 showing normal circuit layouts; a pseudo-image acquisition unit that inputs each of the plurality of faulty circuit images 122 to a generation AI (Artificial Intelligence) 130 and acquires a plurality of pseudo-images 142 generated by the generation AI 130 based on the plurality of faulty circuit images 122; a learning processing unit that inputs the plurality of faulty circuit images 122, the plurality of pseudo-images 142 and the plurality of normal circuit images 120 as training data to a classification AI 150; and an output unit that outputs the learned classification AI 150.

Description

Technical Field

[0004] , , , , , , , , , , ,

[0005] , , , , , , ,

[0003]

[0001] This disclosure relates to a learning support technology for classification AI.

Background Art

[0002] One of the causes of defects in semiconductor devices is defects caused by circuit layout. Defects caused by circuit layout include crystal defects that occur in a substrate having a specific circuit layout, and voids that occur in vias of a specific wiring layout. In order to detect such defects caused by circuit layout, circuit layout information of semiconductor devices may be used.

[0003] Regarding the failure analysis of semiconductor devices, for example, Japanese Patent Application Laid-Open No. 2007-335605 (Patent Document 1) discloses a failure analysis device for semiconductor devices. The failure analysis device includes an inspection information acquisition unit that acquires a failure observation image of a semiconductor device, a layout information acquisition unit that acquires layout information, and a failure analysis unit that performs failure analysis. The failure analysis unit uses wiring information in which the configuration of a plurality of wirings of a semiconductor device is described by a pattern data group of wiring patterns in each of a plurality of layers, extracts a wiring passing through an analysis region among the plurality of wirings as a candidate wiring for failure, and extracts the candidate wiring by performing equipotential tracing of the wiring pattern using the pattern data group (see [Summary]).

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] It is difficult to prepare a large number of samples of defects caused by circuit layout in advance. In particular, it is difficult to prepare samples of defects caused by circuit layout with a small sample size. Therefore, according to the technology disclosed in Patent Document 1, there is a possibility that defects caused by circuit layout with a small sample size cannot be detected.

[0006] This disclosure is made in view of the above-mentioned background, and can facilitate the detection of defects caused by circuit layouts with a small sample size. [Means for solving the problem]

[0007] Artificial Intelligence (AI) is used to generate pseudo-images based on multiple images of faulty circuits that show the circuit layout causing the failure. These pseudo-images are then used as training data for classification AI. [Effects of the Invention]

[0008] According to one embodiment, the technology of the present disclosure facilitates the detection of defects resulting from circuit layouts with a small sample size.

[0009] Other challenges and novel features will become apparent from the description and accompanying drawings in this specification. [Brief explanation of the drawing]

[0010] [Figure 1] This figure shows a first example of operation of a learning support device according to this embodiment. [Figure 2] This figure shows a second example of operation of a learning support device according to this embodiment. [Figure 3] This figure shows an example of the functional block configuration of a learning support device 300 according to this embodiment. [Figure 4] This figure shows an example of the hardware configuration of the learning support device 300 according to this embodiment. [Figure 5] This figure shows an example of a sample circuit image that may be causing a malfunction. [Figure 6] It is a diagram showing an example of the similarity between defective circuit images E1 to E7 shown in FIG. 5. [Figure 7] It is a diagram showing an example of experimental results 700 using the defective circuit images shown in FIG. 5. [Figure 8] It is a diagram showing an example of an application using classification AI 150. [Figure 9] It is a diagram showing an example of the flow of processing executed by the learning support device 300 according to the present embodiment.

Embodiments for Carrying Out the Invention

[0011] Hereinafter, embodiments of the technical idea according to the present disclosure will be described while referring to the drawings. In the following description, the same parts are denoted by the same reference numerals. Their names and functions are also the same. Therefore, detailed descriptions thereof will not be repeated. Also, each embodiment, each modification example, each software or program configuration, each hardware configuration, each function, and each process, etc. may be selectively combined as appropriate.

[0012] <A. Operation and Configuration of Learning Support Device for Classification AI> FIG. 1 is a diagram showing a first operation example of the learning support device according to the present embodiment. The learning support device 300 (see FIG. 3) according to the present embodiment supports the learning of classification AI 150 for detecting defects caused by the circuit layout of semiconductor devices. For this purpose, the learning support device 300 uses generation AI 130 to increase the number of teacher data input to classification AI 150. The teacher data here is an image of the circuit layout of semiconductor device 100.

[0013] In this disclosure, "circuit layout" refers to the wiring layout of a semiconductor device. Typically, semiconductor devices have a multilayer structure, and multiple layers included in this multilayer structure are used as wiring layers. Therefore, a circuit layout may include the wiring layouts of two or more layers. A circuit layout also includes via information for connecting wiring in different layers. Furthermore, in this disclosure, a circuit layout includes the wiring layout of the entire semiconductor device or a part of it. A circuit layout also includes images of the wiring layout extracted from the design data of the semiconductor device, and images of the wiring layout obtained by photographing the manufactured semiconductor device with a camera or the like. Furthermore, in this disclosure, "images of circuit layouts that may cause manufacturing defects" may be referred to as "defective circuit images." Furthermore, "images of normal circuit layouts" may be referred to as "normal circuit images." Defective circuit images and normal circuit images may be collectively referred to as "circuit images."

[0014] Generally, manufacturing defects in semiconductor devices occur rarely. Therefore, it is difficult to obtain a sufficient number of defective circuit images to use as training data for the classification AI 150. To address this, the learning support device 300 acquires defective circuit images 122 from design data of known semiconductor devices 100. Then, the learning support device 300 generates pseudo-images of the defective circuit images 122 using the generation AI 130. In other words, the learning support device 300 can generate a large number of defective circuit image samples from a small number of defective circuit images 122. In this way, the learning support device 300 can obtain the necessary sample defective circuit images for training the classification AI 150.

[0015] Next, a series of operations of the learning support device 300 will be described. First, the learning support device 300 acquires multiple normal circuit images 120 and multiple defective circuit images 122. Each of the multiple normal circuit images 120 is a part of a known semiconductor device 100 and is an image of a normal area 110 without defects. Each of the multiple defective circuit images 122 is a part of a known semiconductor device 100 and is an image of a defective area 112.

[0016] In a certain situation, the learning support device 300 may acquire an image of the circuit layout of the entire known semiconductor device 100 or design data, and the position information of the defective portion 112. In this case, the learning support device 300 may partition the image of the circuit layout and extract the image of the partition including the defective portion 112 as the defective circuit image 122. Similarly, the learning support device 300 may extract the image of the partition not including the defective portion 112 as the normal circuit image 120.

[0017] Next, the learning support device 300 inputs a plurality of normal circuit images 120 to the generation AI 130. The generation AI 130 outputs a plurality of pseudo images 140 based on the plurality of normal circuit images 120. As an example, assume that the first normal circuit image and the second normal circuit image are input to the generation AI 130. And assume that the generation AI 130 is set to generate 500 pseudo images per image. In this case, the generation AI 13 is output 500 pseudo images of the first normal circuit image and 500 pseudo images of the second normal circuit image. The learning support device 300 acquires a set of a plurality of pseudo images 140 from the generation AI 130. As an example, when the number of normal circuit images 120 input to the generation AI 130 is 100, the learning support device 300 acquires 100 sets of a plurality of pseudo images 140.

[0018] Similarly, the learning support device 300 inputs a plurality of defective circuit images 122 to the generation AI 130. The generation AI 130 outputs a plurality of pseudo images 142 based on the plurality of defective circuit images 122. The learning support device 300 acquires a set of a plurality of pseudo images 142 from the generation AI 130.

[0019] In a given scenario, the generating AI 130 may generate multiple pseudo-images 142 based on individual faulty circuit images 122. For example, suppose the generating AI 130 is given a first faulty circuit image and a second faulty circuit image as separate inputs. In this case, the generating AI 130 may generate a first set of pseudo-images based on the input first faulty circuit image, and further generate a second set of pseudo-images based on the input second faulty circuit image. The first set of pseudo-images may include features of the first faulty circuit image. The second set of pseudo-images may include features of the second faulty circuit image. That is, if the generating AI 130 generates multiple pseudo-images 142 based on individual faulty circuit images 122, the learning support device 300 may obtain a set of multiple pseudo-images 140 corresponding to each of the multiple faulty circuit images 122. Similarly, the generating AI 130 may generate multiple pseudo-images 140 based on individual normal circuit images 120.

[0020] In other scenarios, the generating AI 130 may generate multiple pseudo-images 142 based on multiple faulty circuit images 122. For example, suppose the generating AI 130 is input with 10 faulty circuit images, including the first to the tenth faulty circuit images. In this case, the generating AI 130 can generate a set of multiple pseudo-images based on the input 10 faulty circuit images. Each generated pseudo-image may contain one or more features of the 10 faulty circuit images. For example, one pseudo-image may contain only the features of the first faulty circuit image. As another example, another pseudo-image may contain features of the first, second, and seventh faulty circuit images. That is, if the generating AI 130 generates multiple pseudo-images 142 based on multiple faulty circuit images 122, the learning support device 300 may obtain a set of multiple pseudo-images 140 that contain at least one feature of the multiple faulty circuit images 122. Similarly, the generating AI 130 may generate multiple pseudo-images 140 based on multiple normal circuit images 120.

[0021] In one scenario, the generating AI 130 may generate pseudo-images such that the number of images in a set of multiple pseudo-images 140 is equal to the number of images in a set of multiple pseudo-images 142. In another scenario, the generating AI 130 may generate pseudo-images such that the total number of images in a set of multiple normal circuit images 120 and multiple pseudo-images 140 is equal to the total number of images in a set of multiple faulty circuit images 122 and multiple pseudo-images 142.

[0022] In some scenarios, the learning support device 300 may have a built-in generation AI 130. In other scenarios, the learning support device 300 may use an external generation AI 130. In either case, the learning support device 300 can input arbitrary prompts or parameters to the generation AI 130 and obtain a desired number of sets of pseudo-images 140 and sets of pseudo-images 142. The parameters input to the generation AI 130 may include multiple normal circuit images 120, multiple faulty circuit images 122, the number of pseudo-images generated for each image, and other arbitrary information.

[0023] In certain situations, the learning support device 300 may pre-train the generating AI 130 using multiple normal circuit images 120, multiple faulty circuit images 122, or other images. By pre-training the generating AI 130 with circuit images, it can generate pseudo-images of circuit images with high accuracy.

[0024] In the example shown in Figure 1, the learning support device 300 generates multiple pseudo-images 140 and multiple pseudo-images 142 using one generation AI 130, but this is just one example. In some situations, the learning support device 300 may use separate generation AIs for generating multiple pseudo-images 140 and multiple pseudo-images 142. For example, the learning support device 300 may use a first generation AI to generate multiple pseudo-images 140 from multiple normal circuit images 120, and a second generation AI to generate multiple pseudo-images 142 from multiple faulty circuit images 122.

[0025] Next, the learning support device 300 inputs a set of multiple normal circuit images 120, a set of multiple pseudo-images 140, a set of multiple faulty circuit images 122, and a set of multiple pseudo-images 142 as training data to the classification AI 150. In a given situation, the learning support device 300 may input a portion of the images included in the set of multiple normal circuit images 120, a set of multiple pseudo-images 140, a set of multiple faulty circuit images 122, and a set of multiple pseudo-images 142 as training data to the classification AI 150. The classification AI 150 learns to classify circuit layouts based on the images input as training data. That is, the classification AI 150 learns to classify the input circuit images into either normal circuit images or faulty circuit images.

[0026] Next, the learning support device 300 loads multiple test images 160 into the trained classification AI 150. Each of the test images 160 is associated with a correct classification result. The classification AI 150 returns the classification result 170 for each of the test images 160 to the learning support device 300. The learning support device 300 verifies the classification accuracy of the classification AI 150 by comparing the classification result 170 for each of the test images 160 with the correct classification result. If the accuracy rate of the classification AI 150 is above a predetermined threshold, the learning support device 300 outputs the classification AI 150. If the accuracy rate of the classification AI 150 is below a predetermined threshold, the learning support device 300 causes the classification AI 150 to perform additional training.

[0027] As explained with reference to Figure 1, the learning support device 300 utilizes the pseudo-images generated by the generation AI 130 as training data for the classification AI 150. This allows the learning support device 300 to prepare a sufficient number of faulty circuit samples for training the classification AI 150. As a result, the learning support device 300 can improve the classification accuracy of the circuit images by the classification AI 150.

[0028] Figure 2 shows a second example of operation of the learning support device according to this embodiment. Referring to Figure 2, the differences between the second and first examples of operation will be explained. In the second example of operation, the generation AI 130 does not generate a set of multiple pseudo-images 140 from multiple normal circuit images 120. That is, the learning support device 300 inputs multiple defective circuit images 122 to the generation AI 130, but does not input multiple normal circuit images 120 to the generation AI 130. The generation AI 130 outputs multiple pseudo-images 142 based on the multiple defective circuit images 122. The learning support device 300 obtains a set of multiple pseudo-images 142 from the generation AI 130.

[0029] Next, the learning support device 300 inputs a set of multiple normal circuit images 120, multiple defective circuit images 122, and multiple pseudo-images 142 as training data to the classification AI 150. The learning support device 300 can easily obtain normal circuit images 120 from design data of known semiconductor devices 100 or other semiconductor device design data. Therefore, the learning support device 300 may obtain the same number of multiple normal circuit images 120 as the total number of sets of multiple defective circuit images 122 and multiple pseudo-images 142.

[0030] Subsequently, similar to the first example of operation, the learning support device 300 loads multiple test images 160 into the trained classification AI 150 and tests the classification result 270 of the learning support device 300. If the accuracy rate of the classification AI 150 is above a predetermined threshold, the learning support device 300 outputs the classification AI 150. If the accuracy rate of the classification AI 150 is below a predetermined threshold, the learning support device 300 causes the classification AI 150 to perform additional training.

[0031] In the first and second operational examples, the training data input to the classification AI 150 is different. Therefore, the classification results 170 and 270 may also be different. In the first operational example, the learning support device 300 can easily prepare more training data than in the second operational example by using the generation AI 130. Also, in the second operational example, the learning support device 300 can increase the proportion of actual circuit images in the training data.

[0032] In certain situations, the learning support device 300 may be configured to select and execute an operation shown in either the first or second operation example based on settings input from the user. In this case, the user can choose which of the first or second operation example to use, depending on the number of semiconductor device design data available, etc.

[0033] Figure 3 shows an example of the functional block configuration of a learning support device 300 according to this embodiment. Each functional block shown in Figure 3 is a configuration for realizing the functions of the learning support device 300 described in this embodiment, and can be realized by a program, hardware, or a combination thereof. In some cases, each functional block shown in Figure 3 may be realized by the execution of a program on the hardware shown in Figure 4. In other cases, a part of each functional block shown in Figure 3 may be realized as hardware. In this case, the learning support device 300 includes hardware corresponding to one or more functional blocks shown in Figure 3, in addition to the hardware shown in Figure 4.

[0034] The learning support device 300 comprises a pre-processing unit 301, an image acquisition unit 302, a similarity calculation unit 303, a pseudo-image acquisition unit 304, a post-processing unit 305, a learning processing unit 306, an evaluation unit 307, and an output unit 308.

[0035] The preprocessing unit 301 performs arbitrary processing on the data before it is input to the generation AI 130. The preprocessing unit 301 outputs the processed data to the image acquisition unit 302. The data here includes the circuit layout of the entire semiconductor device (hereinafter referred to as the "overall circuit image"), multiple normal circuit images, and multiple defective circuit images.

[0036] In a given scenario, the preprocessor 301 can extract normal circuit images and faulty circuit images from the overall circuit image. More specifically, the preprocessor 301 receives input of the overall circuit image and location information of faulty areas on the overall circuit image. The preprocessor 301 can extract multiple normal circuit images and multiple faulty circuit images from the overall circuit image on a section-by-section basis. The preprocessor 301 can extract sections containing faulty areas as faulty circuit images and extract the remaining sections as normal circuit images.

[0037] Furthermore, in other aspects, the preprocessing unit 301 may set coefficients or weights for each section. These coefficients are parameters input to the generation AI 130 and are used as adjustment parameters to adjust the magnitude of the change in the pseudo-image relative to the original image. For example, the preprocessing unit 301 may set a first coefficient for the defective section and a second coefficient different from the first coefficient for the normal section. The preprocessing unit 301 may also associate the coefficients with each extracted normal circuit image and each defective circuit image as tags or metadata. The learning support device 300 can adjust the magnitude of the change in the pseudo-image relative to the original image in the defective circuit image and the magnitude of the change in the pseudo-image relative to the original image in the normal circuit image individually by including these coefficients in the parameters of the generation AI 130. For example, the learning support device 300 may set the magnitude of the change in the pseudo-image relative to the original image in the defective circuit image to be smaller than the magnitude of the change in the pseudo-image relative to the original image in the normal circuit image.

[0038] Furthermore, in other aspects, the preprocessing unit 301 may generate a faulty circuit image by rotating each of the multiple faulty circuit images by 90 degrees. Typically, the circuit layout of a semiconductor device includes a layout in which the wiring of multiple layers is orthogonal. For example, suppose a certain circuit layout includes a first wiring layer and a second wiring layer. In this case, the first wiring layer may mainly include wiring in a first direction, and the second wiring layer may mainly include wiring in a second direction perpendicular to the first direction. However, the wiring direction of each wiring layer may change depending on the design. For example, suppose the first wiring layer mainly includes wiring in a second direction, and the second wiring layer mainly includes wiring in a first direction. Even in this case, the circuit layout is still valid. It is desirable that the classification AI 150 can detect faults in any of the above circuit layouts. Therefore, the preprocessing unit 301 generates an image by rotating each of the multiple faulty circuit images by 90 degrees and includes the generated image in the faulty circuit image. By rotating the circuit image by 90 degrees, the wiring of the two wiring layers becomes an image in which they appear to have been swapped. The classification AI150 can classify normal and faulty circuits regardless of the orientation of the two wiring layers by learning from the original faulty circuit image and an image of the original faulty circuit image rotated 90 degrees.

[0039] The image acquisition unit 302 acquires data from the preprocessing unit 301 for input to the generation AI 130. This data includes at least multiple normal circuit images and multiple defective circuit images. The data may also include adjustment parameters for adjusting the magnitude of the change in the pseudo-image relative to the original image. The adjustment parameters are associated with each circuit image as coefficients or weights. Alternatively, if preprocessing of the circuit images is not required, the image acquisition unit 302 may acquire multiple normal circuit images and multiple defective circuit images from the user's terminal or the like as data for input to the generation AI 130. The image acquisition unit 302 outputs the acquired data to the similarity calculation unit 303.

[0040] The similarity calculation unit 303 calculates the similarity of each of the multiple faulty circuit images. This similarity can be defined by the difference in the feature quantities of each faulty circuit image. Based on the feature quantities, the similarity calculation unit 303 classifies each of the multiple faulty circuit images into either a minor faulty circuit image or a major faulty circuit image. A major faulty circuit image is a faulty circuit image that is similar to other faulty circuit images. A minor faulty circuit image is a faulty circuit image that is not similar to other faulty circuit images. In other words, a major faulty circuit image is a faulty circuit image that contains features that are detected more frequently than minor faulty circuit images.

[0041] In some scenarios, the similarity calculation unit 303 may output classification information to the image acquisition unit 302. In this case, the image acquisition unit 302 outputs multiple normal circuit images and multiple classified defective circuit images to the pseudo-image acquisition unit 304. In addition, the image acquisition unit 302 may output adjustment parameters (coefficients or weights for each circuit image) to the pseudo-image acquisition unit 304. In other scenarios, the similarity calculation unit 303 may output multiple normal circuit images and multiple classified defective circuit images to the pseudo-image acquisition unit 304. In this case, the similarity calculation unit 303 may further output adjustment parameters (coefficients or weights for each circuit image) to the pseudo-image acquisition unit 304. An example of the operation of the similarity calculation unit 303 using sample data will be described later with reference to Figure 6.

[0042] The pseudo-image acquisition unit 304 inputs the acquired data to the generation AI 130 and acquires the pseudo-images generated by the generation AI 130. The data acquired by the pseudo-image acquisition unit 304 includes multiple normal circuit images and multiple classified defective circuit images. The data may also include adjustment parameters (coefficients or weights for each circuit image). The pseudo-image acquisition unit 304 outputs multiple normal circuit images, multiple pseudo-images containing at least some of the features of the multiple normal circuit images, multiple defective circuit images, and multiple pseudo-images containing at least some of the features of the multiple defective circuit images to the post-processing unit 305 or the learning processing unit 306. From this point onward, in order to distinguish them from the generated pseudo-images, the normal circuit images and defective circuit images input to the generation AI 130 may be referred to as "original circuit images."

[0043] The number of faulty circuit images is likely to be less than the number of normal circuit images. Therefore, the pseudo-image acquisition unit 304 can adjust the parameter for the number of each circuit image generated so that the total number of normal circuit images and their pseudo-images is equal to the total number of faulty circuit images and their pseudo-images.

[0044] Furthermore, the pseudo-image acquisition unit 304 can adjust the parameter for the number of images generated for each circuit image so that the total number of minor faulty circuit images and their pseudo-images is equal to the total number of major faulty circuit images and their pseudo-images. Normally, the number of minor faulty circuit images is less than the number of major faulty circuit images. Therefore, if the number of pseudo-images for each of the multiple faulty circuit images generated is equal, the total number of major faulty circuit images and their pseudo-images is likely to be significantly larger than the total number of minor faulty circuit images and their pseudo-images. In this case, the classification AI 150 is greatly influenced by the major faulty circuit images and their pseudo-images during training. As a result, even if the classification AI 150 can detect major faults with high accuracy, it may not be able to detect minor faults. Therefore, the pseudo-image acquisition unit 304 increases the number of pseudo-images generated for minor faulty circuit images to make the total number of minor faulty circuit images and their pseudo-images equal to the total number of major faulty circuit images and their pseudo-images. As a result, the classification AI 150 becomes more likely to detect minor faults. Furthermore, the pseudo-image acquisition unit 304 can input adjustment parameters (coefficients or weights for each circuit image) to the generation AI 130.

[0045] The post-processing unit 305 can perform arbitrary processing on the pseudo-image output by the generating AI 130. The post-processing unit 305 also outputs the processed pseudo-image to the learning processing unit 306. Furthermore, the post-processing unit 305 can also perform arbitrary processing on the original circuit image. In this case, the post-processing unit 305 outputs the processed original circuit image along with the processed pseudo-image to the learning processing unit 306.

[0046] As an example, the post-processing unit 305 may gray out the pseudo-image output by the generation AI 130. The pseudo-image generated by the generation AI 130 may have different colors from the original circuit image. Therefore, the color difference between the pseudo-image and the original circuit image may affect the learning of the classification AI 150. By graying out the raw pseudo-image, the post-processing unit 305 can suppress the influence of the color difference between the pseudo-image and the original circuit image on the classification AI 150. In addition, the post-processing unit 305 may gray out not only the raw pseudo-image but also the original circuit image.

[0047] The learning processing unit 306 inputs the original circuit image and pseudo-image acquired from the pseudo-image acquisition unit 304 or the post-processing unit 305 as training data to the classification AI 150, allowing the classification AI 150 to learn. After the classification AI 150 has finished learning, the learning processing unit 306 outputs a learning completion notification to the evaluation unit 307.

[0048] The evaluation unit 307 evaluates the classification AI 150 based on the acquisition of a learning completion notification. As an example, the evaluation unit 307 inputs multiple circuit images (corresponding to multiple test images 160) that are not used as training data to the classification AI 150 and has the classification AI 150 classify these circuit images. Each of the multiple circuit images used to evaluate the classification AI 150 is linked to answer information. For each of the multiple circuit images, the evaluation unit 307 compares the classification result of the classification AI 150 with the answer information and calculates the accuracy rate of the classification result of the classification AI 150. If the accuracy rate is above a predetermined threshold, the evaluation unit 307 outputs an output permission notification for the classification AI 150 to the output unit 308. If the accuracy rate is below a predetermined threshold, the evaluation unit 307 outputs a retraining command for the classification AI 150 to the learning processing unit 306.

[0049] The output unit 308 outputs the trained classification AI 150 based on receiving an output permission notification for the classification AI 150. In one scenario, the output unit 308 may transmit the classification AI 150 itself to the user's terminal based on receiving a request from the user's terminal. In another scenario, the output unit 308 may provide the classification AI 150 as a service. In this case, the classification AI 150 classifies the circuit image received by the learning support device 300 from the user's terminal and outputs the classification result to the output unit 308. The output unit 308 transmits the classification result to the user's terminal. Furthermore, in yet another scenario, the classification AI 150 may be incorporated into an application and used, as shown in Figure 8. In this case, the learning support device 300 may have the functionality of the application. Alternatively, another device having the functionality of the application may use the classification AI 150.

[0050] Figure 4 shows an example of the hardware configuration of a learning support device 300 according to this embodiment. The learning support device 300 may not include some of the configurations shown in Figure 4. Furthermore, the learning support device 300 may include configurations not shown in Figure 4. In addition, the learning support device 300 may include two or more of each configuration shown in Figure 4. Each functional block shown in Figure 3 can be realized by executing a program on the hardware shown in Figure 4.

[0051] The learning support device 300 includes a processor 401, memory 402, storage 403, external device IF 404, input IF 405, output IF 406, and communication IF 407. In some cases, the learning support device 300 may have two or more of each component, or it may not have some of the components.

[0052] The processor 401 can execute programs to realize various functions of the learning support device 300. The processor 401 is composed of, for example, at least one integrated circuit. According to one embodiment, the integrated circuit may include at least one CPU (Central Processing Unit), at least one GPU (Graphics Processing Unit), at least one FPGA (Field Programmable Gate Array), at least one ASIC (Application Specific Integrated Circuit), at least one AI chip, or a combination thereof.

[0053] Memory 402 functions as a workspace for processor 401. Memory 402 stores programs executed by processor 401 and data referenced by processor 401. In some cases, memory 402 may be implemented as DRAM (Dynamic Random Access Memory) or SRAM (Static Random Access Memory), etc.

[0054] Storage 403 is non-volatile memory that stores programs executed by processor 401 and data referenced by processor 401. Processor 401 executes programs read from storage 403 into memory 402 and references data read from storage 403 into memory 402. In some cases, storage 403 can be implemented by an HDD (Hard Disk Drive), SSD (Solid State Drive), EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically Erasable Programmable Read Only Memory), or flash memory, etc.

[0055] The external device IF404 can be connected to any external device such as a printer, scanner, and external HDD. In some cases, the external device IF404 can be implemented using a USB (Universal Serial Bus) terminal or the like.

[0056] The input interface 405 can be connected to any input device such as a keyboard, mouse, touchpad, or gamepad. In some cases, the input interface 405 can be implemented using a USB terminal, PS / 2 terminal, or Bluetooth® module.

[0057] Output IF406 can be connected to any output device such as a cathode ray tube display, liquid crystal display, or organic EL display. In some cases, output IF406 can be implemented using a USB terminal, D-sub terminal, DVI (Digital Visual Interface) terminal, HDMI® (High-Definition Multimedia Interface) terminal, or DisplayPort terminal.

[0058] The communication IF407 is connected to other devices via a wired or wireless network. In some cases, the communication IF407 can be implemented using a wired LAN (Local Area Network) port and a Wi-Fi® (Wireless Fidelity) module, etc. In other cases, the communication IF407 can send and receive data using communication protocols such as TCP / IP (Transmission Control Protocol / Internet Protocol) and UDP (User Datagram Protocol).

[0059] In some cases, the learning support device 300 is composed of one device or a combination of multiple devices. The devices comprising the learning support device 300 may include personal computers, workstations, server devices, tablets, smartphones, SoCs (System-on-a-chip), and SoMs (System-on-Module). Furthermore, the devices comprising the learning support device 300 may include any peripheral devices such as switches, routers, displays, keyboards, and mice. In addition, the learning support device 300 may include virtual machines and instances built on a cloud environment. In some cases, the learning support device 300 may be connected to input / output devices such as displays and keyboards and used as a standalone device. In other cases, the learning support device 300 may provide various functions as a service or web application via a network. In this case, users can use the functions of the learning support device 300 via a browser or client software installed on their terminal. Furthermore, if the learning support device 300 is composed of two or more devices, it can also be called a learning support system.

[0060] As explained with reference to Figures 1 to 4, the learning support device 300 includes an image acquisition unit 302 that acquires multiple faulty circuit images 122 showing faulty circuit layouts extracted from known circuit layouts and multiple normal circuit images 120 showing normal circuit layouts. The learning support device 300 also includes a pseudo-image acquisition unit 304 that inputs each of the multiple faulty circuit images 122 to a generation AI 130 and acquires multiple pseudo-images 142 generated by the generation AI 130 based on the multiple faulty circuit images 122. The learning support device 300 also includes a learning processing unit 306 that inputs the multiple faulty circuit images 122, the multiple pseudo-images 142, and the multiple normal circuit images 120 as training data to a classification AI 150. Furthermore, the learning support device 300 includes an output unit 308 that outputs a trained classification AI 150. By having these configurations, the learning support device 300 can increase the number of faulty circuit image samples 122 that serve as training data for the classification AI 150.

[0061] In a given scenario, the pseudo-image acquisition unit 304 is configured to generate multiple pseudo-images 142 such that the sum of the multiple defective circuit images 122 and the multiple pseudo-images 142 equals the sum of the extracted multiple normal circuit images 120. This allows the learning support device 300 to adjust the number of defective circuit images, which are fewer than the number of normal circuit images, to an appropriate number.

[0062] In a given scenario, the pseudo-image acquisition unit 304 inputs each of the multiple normal circuit images 120 to the generation AI 130, and the generation AI 130 acquires multiple pseudo-images 140 generated based on the multiple normal circuit images 120. Inputting the multiple defective circuit images 122, the multiple pseudo-images 140, and the multiple normal circuit images 120 as training data to the classification AI 150 includes further including the multiple pseudo-images 140 generated based on the multiple normal circuit images 120 in the training data. This allows the learning support device 300 to adjust the number of normal circuit images as needed.

[0063] In a given scenario, each of the multiple faulty circuit images 122 and each of the multiple normal circuit images 120 contains more than a predetermined threshold of wiring. This allows the learning support device 300 to use regions containing more than a predetermined number of wiring as training data.

[0064] In a certain phase, the learning support device 300 further includes a preprocessing unit 301 for preprocessing a known circuit layout. The preprocessing unit 301 divides the known circuit layout into sections, sets a first coefficient in the sections containing defective areas, and sets a second coefficient different from the first coefficient in the sections containing normal areas. The preprocessing unit 301 is also configured to extract multiple defective circuit images 122 from the defective area sections and multiple normal circuit images 120 from the normal area sections, and to output the extracted multiple defective circuit images 122, multiple normal circuit images 120, the first coefficient, and the second coefficient to the image acquisition unit 302. The pseudo-image acquisition unit 304 inputs the first coefficient and the second coefficient as parameters to the generation AI 130. By including these coefficients in the parameters of the generation AI 130, the learning support device 300 can individually adjust the magnitude of the change in the pseudo-image relative to the original image in the defective circuit images and the magnitude of the change in the pseudo-image relative to the original image in the normal circuit images.

[0065] In a given scenario, the first and second coefficients are adjustment parameters used to adjust the magnitude of the change in the pseudo-image generated by the generating AI 130 relative to the original image. The larger the adjustment parameter, the greater the difference between the pseudo-image and the original image. The preprocessing unit 301 sets the value of the first coefficient to be smaller than the value of the second coefficient. As a result, the learning support device 300 can obtain a pseudo-image for the faulty circuit image 122 that is more similar to the original image than the normal circuit image 120. Furthermore, the learning support device 300 can obtain a pseudo-image for the normal circuit image 120 that is more varied and differs from the original image than the faulty circuit image 122.

[0066] Setting the value of the first coefficient to be smaller than the value of the second coefficient in a given situation includes setting the values ​​of the first and second coefficients so that they are in a predetermined ratio. This allows the learning support device 300 to adjust the amount of change from the original circuit image for the pseudo-images corresponding to the normal circuit image 120 and the defective circuit image 122 using the ratio.

[0067] In a certain situation, the preprocessing unit 301 is configured to generate a plurality of rotated images obtained by rotating each of the plurality of defective circuit images 122 by 90 degrees, include the plurality of rotated images in the plurality of defective circuit images 122, and output the plurality of defective circuit images 122 to the image acquisition unit 302. The classification AI 150 can classify normal circuits and defective circuits regardless of the orientations of the two wiring layers by learning the original defective circuit images and the images obtained by rotating the original defective circuit images by 90 degrees.

[0068] In a certain situation, the learning support device 300 further includes a postprocessing unit 305 that grays out the plurality of pseudo-images 142 before the plurality of pseudo-images 142 are input to the classification AI 150. By graying out the raw pseudo-images, the postprocessing unit 305 can suppress the influence of the color difference between the pseudo-images and the original circuit images on the classification AI 150.

[0069] <B. Classification of Defects and Adjustment of the Number of Generated Pseudo-Images> Next, referring to FIGS. 5 to 7, the learning process of the classification AI 150 using sample circuit images and the experimental results will be described.

[0070] Figure 5 shows an example of a sample circuit image that causes a defect. Figure 5 shows seven defective circuit images E1, E2, E3, E4, E5, E6, and E7. Each defective circuit image E1 to E7 includes multiple orthogonal wires and vias 500. Each defective circuit image E1 to E7 is an image of a circuit where a manufacturing defect occurred in the area where the central via 500 is located. Defective circuit images E1 to E7 correspond to the multiple defective circuit images 122 in Figure 1. Defective circuit images E1 to E7 and multiple pseudo-images generated from defective circuit images E1 to E7 are used as training data for classification AI 150. In some cases, generation AI 130 may generate a set of multiple pseudo-images based on individual defective circuit images. In this case, the set of multiple pseudo-images output by generation AI 130 corresponds to the individual defective circuit image (defective circuit image E1, etc.) input to generation AI 130. In other scenarios, the generating AI 130 may generate a set of multiple pseudo-images based on multiple faulty circuit images. In this case, the set of multiple pseudo-images output by the generating AI 130 may include one or more features of the multiple faulty circuit images (faulty circuit images E1 to E7) input to the generating AI 130.

[0071] Figure 6 shows an example of the similarity between the defective circuit images E1 to E7 shown in Figure 5. Table 600 shows the calculation results of the similarity between the defective circuit images E1 to E7. As an example, the similarity is calculated using the SSIM (Structural Similarity) formula 650. The similarity may also be calculated using other evaluation methods.

[0072] Similarity is calculated for each pair of faulty circuit images, and a higher similarity value indicates that the two faulty circuit images are more similar. For example, cell 602 shows that the similarity between faulty circuit image E1 and faulty circuit image E2 is "0.638". Cell 604 shows that the similarity between faulty circuit image E1 and faulty circuit image E4 is "0.408". The value in cell 602 is greater than the value in cell 604. That is, faulty circuit image E1 is more similar to faulty circuit image E2 than to faulty circuit image E4. Also, cells with a value of "1.000" indicate the similarity of identical faulty circuit images.

[0073] Referring again to Table 600, we can see that the values ​​in each cell in rows 610, 620, and 630 are all less than "0.500", except for the similarity of identical faulty circuit images, and are all smaller than the values ​​in other rows. The values ​​in each cell in rows 610, 620, and 630 indicate the similarity between faulty circuit images E3, E4, and E7 and other faulty circuits. From this, we can see that faulty circuit images E3, E4, and E7 are not similar to any of the other faulty circuit images and are minor faulty circuit images. Therefore, the learning support device 300 classifies faulty circuit images E3, E4, and E7 as minor faulty circuit images. Faulty circuit images E1, E2, E5, and E6 contain cells with a similarity of "0.500" or higher, and are therefore similar to other faulty circuit images. Therefore, the learning support device 300 classifies faulty circuit images E1, E2, E5, and E6 as major faulty circuit images.

[0074] The learning support device 300 adjusts the number of pseudo-images generated for each faulty circuit image so that the sum of faulty circuit images E1, E2, E5, E6 and their pseudo-images is equal to the sum of faulty circuit images E3, E4, E7 and their pseudo-images.

[0075] As explained with reference to Figures 5 and 6, the multiple faulty circuit images 122 include minor faulty circuit images and major faulty circuit images. The pseudo-image acquisition unit 304 is configured to generate more pseudo-images 142 containing the features of minor faulty circuit images than pseudo-images 142 containing the features of major faulty circuit images. In a given situation, the pseudo-image acquisition unit 304 is configured to generate multiple pseudo-images 142 such that the sum of the pseudo-images 142 containing the features of major faulty circuit images is equal to the sum of the pseudo-images 142 containing the features of minor faulty circuit images. The learning support device 300 adjusts the number of pseudo-images generated for minor and major faulty circuit images to make it easier for the trained classification AI 150 to detect minor faulty circuits.

[0076] Figure 7 shows an example of experimental results 700 using the defective circuit image shown in Figure 5. Experimental results 700 include the judgment results 710 of the classification AI 150 trained using only the original circuit image as training data, and the judgment results 720 of the classification AI 150 trained using both the original circuit image and a simulated image as training data. The vertical axis of experimental results 700 shows the detection rate. The horizontal axis of experimental results 700 shows the type of defective circuit image.

[0077] In the experiment, the classification threshold of AI150 was set so that the rate of misidentifying a normal circuit as a faulty circuit was 0.2 (20%). The test data used in the experiment consisted of images of faulty circuits E1 to E7, with the faulty location (via 500) shifted from the center.

[0078] Referring to the judgment result 710, classification AI 150 completely failed to detect the minor faulty circuit image E4. This is likely because classification AI 150 was heavily influenced by the major faulty circuit images during training.

[0079] On the other hand, referring to the determination result 720, the classification AI 150 can detect the defective circuit image E4, which is a minor defective circuit image, to some extent. This is presumably because by increasing the number of pseudo-images of the defective circuit image E4 included in the training data, the classification AI 150 has become able to detect circuit images similar to the defective circuit image E4.

[0080] As described with reference to FIGS. 5 to 7, by adjusting the number of generated pseudo-images of the minor defective circuit image and the major defective circuit image, the classification AI 150 after learning may be more likely to detect minor defective circuits.

[0081] <C. Utilization of Classification Results> FIG. 8 is a diagram showing an example of an application using the classification AI 150. The classification AI 150 may be configured to output the probability that the input circuit image is a defective circuit image. In this case, the classification AI 150 can output the probability of being a defective circuit image for each section of the circuit image of the input semiconductor device 800. The application can output a heat map 820 of the semiconductor device 800 using the probability. The heat map 820 can display a portion 830 where a defect is likely to occur in a dark color and a portion 840 where a defect is less likely to occur in a light color. Alternatively, the heat map 820 may be expressed like a contour line. By referring to the heat map 820, the designer of the semiconductor device can grasp the locations on the semiconductor device that are likely to cause defects.

[0082] In some situations, the application may output a heat map 820 of the entire semiconductor device 800. In other situations, the application may output a heat map 820 of a partial region 810 of the semiconductor device 800. Alternatively, the application may receive a circuit image of the entire semiconductor device 800 and information specifying the partial region 810 to be heat-mapped. In this case, the classification AI 150 may output a probability of being a defective circuit image for each section of the partial region 810. The application may use the probability to output a heat map 820 of the partial region 810.

[0083] In some situations, the learning support device 300 may have the function of the application. In this case, the learning support device 300 may include the application as a heat map creation unit (not shown). In this case, the output unit 308 outputs the learned classification AI 150 to an area where the heat map creation unit can refer to it. The area where the heat map creation unit can refer to includes within the learning support device 300, within the storage of other devices, on a cloud environment, or any other arbitrary area.

[0084] As described with reference to FIG. 8, the learned classification AI 150 may be configured to output a probability that the input circuit image is a defective circuit image. The learning support device 300 further includes a heat map creation unit that outputs a heat map of the entire or a partial region of the semiconductor device to be inspected. The heat map creation unit inputs a plurality of input circuit images constituting the entire or a partial region of the semiconductor device to be inspected to the learned classification AI 150, and based on the probability of each of the plurality of input circuit images output by the learned classification AI 150, may create a heat map 820 of the entire or a partial region of the semiconductor device to be inspected. The person in charge of designing the semiconductor device can grasp the locations on the semiconductor device that are likely to cause defects by referring to the heat map 820.

[0085] <D. Processes Executed by the Learning Support Device> Figure 9 shows an example of the processing flow performed by the learning support device 300 according to this embodiment. In one phase, the processor 401 may read a program for performing the processing shown in Figure 9 from the storage 403 into the memory 402 and execute the program. In another phase, part or all of the processing can also be realized as a combination of circuit elements configured to perform the processing. Furthermore, in yet another phase, the following steps may be executed in a different order.

[0086] In step S905, the learning support device 300 acquires multiple normal circuit images. In step S910, the learning support device 300 acquires multiple defective circuit images. In some cases, the learning support device 300 may acquire design data of a semiconductor device or a circuit image of the entire semiconductor device. In this case, the learning support device 300 can extract multiple normal circuit images and multiple defective circuit images from the design data or circuit image of the semiconductor device.

[0087] In step S915, the learning support device 300 generates pseudo-images based on multiple normal circuit images using the generation AI 130. For example, the generation AI 130 may receive individual normal circuit images as input and generate multiple pseudo-images corresponding to individual normal circuit images. As another example, the generation AI 130 may receive multiple normal circuit images as input and generate multiple pseudo-images from multiple normal circuit images.

[0088] In step S920, the learning support device 300 creates a rotated image of each of the multiple faulty circuit images. Each rotated image is an image obtained by rotating the original faulty circuit image by 90 degrees. In subsequent processing, the learning support device 300 includes the created rotated images in the multiple faulty circuit images and performs each process. In some situations, the learning support device 300 does not need to perform the processing in this step. Alternatively, the learning support device 300 may decide whether or not to perform the processing in this step based on whether or not there is an input from the user instructing the creation of a rotated image.

[0089] In step S925, the learning support device 300 calculates the similarity between each of the multiple defective circuit images. In step S930, the learning support device 300 determines the number of pseudo-images to generate that include the features of each of the multiple defective circuit images, according to the similarity. More specifically, the learning support device 300 may determine that two circuits are similar if the similarity between them is greater than or equal to a predetermined threshold. The learning support device 300 determines that a defective circuit image that is not similar to any of the other defective circuit images is a minor defective circuit image. The learning support device 300 determines that a defective circuit image that is similar to any of the other defective circuit images is a major defective circuit image. The learning support device 300 determines the number of pseudo-images to generate for each defective circuit image such that the sum of minor defective circuit images and their pseudo-images is equal to the sum of major defective circuit images and their pseudo-images. In a given situation, the learning support device 300 may determine a defective circuit image as a minor defective circuit image if the number of other defective circuit images similar to a given defective circuit image is less than or equal to a predetermined number. For example, the learning support device 300 may determine a particular faulty circuit image to be a minor faulty circuit image if the number of images similar to that faulty circuit image is two or less.

[0090] In step S935, the learning support device 300 generates pseudo-images based on multiple faulty circuit images using the generation AI 130. For example, the generation AI 130 may accept individual faulty circuit images as input and generate multiple pseudo-images corresponding to individual faulty circuit images. As another example, the generation AI 130 may accept multiple faulty circuit images as input and generate multiple pseudo-images from multiple faulty circuit images. The generation AI 130 may generate pseudo-images for each circuit image such that the sum of multiple normal circuit images and their pseudo-images is equal to the sum of multiple faulty circuit images and their pseudo-images.

[0091] In step S940, the learning support device 300 converts each generated pseudo-image into a grayscale image. In some cases, if the multiple normal circuit images and multiple faulty circuit images are not grayscale images, the learning support device 300 may also convert these circuit images into grayscale images. In other cases, the learning support device 300 may convert each generated pseudo-image into an image of a uniform color other than grayscale. Furthermore, in other cases, the learning support device 300 does not have to perform the processing in this step. Alternatively, the learning support device 300 may decide whether or not to perform the processing in this step based on whether or not there is an input from the user instructing the creation of a grayscale image.

[0092] In step S945, the learning support device 300 causes the classification AI 150 to learn using the original circuit image and pseudo-images as training data. The training data includes multiple normal circuit images, multiple faulty circuit images, and multiple pseudo-images generated based on the multiple faulty circuit images. Furthermore, the training data may also include multiple pseudo-images generated based on the multiple normal circuit images.

[0093] In step S950, the learning support device 300 tests the trained classification AI 150. The learning support device 300 has the classification AI classify multiple circuit images associated with the answers. The learning support device 300 compares the answers and classification results for each of the multiple circuit images to inspect the classification performance of the classification AI 150. If the accuracy rate of the classification AI 150 is above a predetermined threshold, the learning support device 300 outputs the classification AI 150. If the accuracy rate of the classification AI 150 is below a predetermined threshold, the learning support device 300 has the classification AI 150 perform additional training.

[0094] In step S955, the learning support device 300 outputs the trained classification AI 150. In one scenario, the learning support device 300 may transmit the classification AI 150 itself to the user's terminal based on receiving a request from the user's terminal. In another scenario, the learning support device 300 may provide the classification AI 150 as a service.

[0095] As explained with reference to Figure 9, the learning support device 300 can be realized by having a computer execute a learning support program for the classification AI 150. The learning support program causes the computer to acquire multiple faulty circuit images 122 showing faulty circuit layouts extracted from known circuit layouts, and multiple normal circuit images 120 showing normal circuit layouts. The learning support program inputs each of the multiple faulty circuit images 122 into the generation AI 130, and causes the computer to acquire multiple pseudo-images 142 generated by the generation AI 130 based on the multiple faulty circuit images 122. The learning support device 300 causes the computer to input the multiple faulty circuit images 122, the multiple pseudo-images 142, and the multiple normal circuit images 120 as training data into the classification AI 150, and to output the trained classification AI 150.

[0096] <E.まとめ> As described above, the learning support device 300 according to this embodiment utilizes pseudo-images generated by the generation AI 130 as training data for the classification AI 150. This allows the learning support device 300 to prepare a sufficient number of faulty circuit samples for training the classification AI 150. As a result, the learning support device 300 can improve the classification accuracy of the circuit images of the classification AI 150. Furthermore, the learning support device 300 determines the number of pseudo-images to generate for each faulty circuit image such that the total number of minor faulty circuit images and their pseudo-images is equal to the total number of major faulty circuit images and their pseudo-images. This allows the learning support device 300 to increase the proportion of minor faulty circuit images in the training data. As a result, the classification AI 150 becomes more adept at detecting minor faults.

[0097] Although the present invention has been specifically described above based on embodiments, it goes without saying that the present invention is not limited to the above embodiments and can be modified in various ways without departing from its essence. [Explanation of Symbols]

[0098] 100,800 Semiconductor devices, 110 Normal areas, 112 Defective areas, 120 Normal circuit image, 122,E1,E2,E3,E4,E5,E6,E7 Defective circuit image, 130 Generating AI, 140,142 Pseudo-images, 150 Classification AI, 160 Test images, 170,270 Classification results, 300 Learning support device, 301 Pre-processing unit, 302 Image acquisition unit, 303 Similarity calculation unit, 304 Pseudo-image acquisition unit, 305 Post-processing unit, 306 Learning processing unit, 307 Evaluation unit, 308 Output unit, 401 Processor, 402 Memory, 403 Storage, 404 External device IF, 405 Input IF, 406 Output IF, 407 Communication IF, 500 Via, 602,604 Cells, 610, 620, 630 Rows, 650 Formulas, 700 Experimental Results, 710, 720 Judgment Results, 810 Partial Area, 820 Heatmap, 830 Areas with a High Probability of Failure, 840 Areas with a Low Probability of Failure.

Claims

1. An image acquisition unit that acquires multiple faulty circuit images showing the faulty circuit layout extracted from known circuit layouts, and multiple normal circuit images showing a normal circuit layout. A pseudo-image acquisition unit inputs each of the plurality of faulty circuit images to a generation AI (Artificial Intelligence), and acquires a plurality of pseudo-images generated by the generation AI based on the plurality of faulty circuit images. A learning processing unit that inputs the aforementioned multiple defective circuit images, the aforementioned multiple pseudo-images, and the aforementioned multiple normal circuit images as training data to a classification AI, A learning support device comprising an output unit that outputs the learned classification AI.

2. The aforementioned plurality of faulty circuit images include minor faulty circuit images and major faulty circuit images. The learning support device according to claim 1, wherein the pseudo-image acquisition unit is configured to generate more pseudo-images containing the characteristics of the minor faulty circuit image than the number of pseudo-images containing the characteristics of the major faulty circuit image.

3. The learning support device according to claim 2, wherein the pseudo-image acquisition unit is configured to generate the plurality of pseudo-images such that the sum of the plurality of pseudo-images including the characteristics of the major faulty circuit image is equal to the sum of the plurality of pseudo-images including the characteristics of the minor faulty circuit image.

4. The learning support device according to claim 1, wherein the pseudo-image acquisition unit is configured to generate the plurality of pseudo-images such that the sum of the plurality of defective circuit images and the plurality of pseudo-images is equal to the sum of the extracted plurality of normal circuit images.

5. The pseudo-image acquisition unit inputs each of the plurality of normal circuit images to the generation AI, and acquires a plurality of pseudo-images generated by the generation AI based on the plurality of normal circuit images. The learning support device according to claim 1, wherein inputting the plurality of defective circuit images, the plurality of pseudo-images, and the plurality of normal circuit images as training data to the classification AI further includes including the plurality of pseudo-images corresponding to each of the plurality of normal circuit images in the training data.

6. The learning support device according to claim 1, wherein each of the plurality of faulty circuit images and each of the plurality of normal circuit images includes wiring exceeding a predetermined threshold.

7. The system further includes a preprocessing unit for performing preprocessing on the known circuit layout, The aforementioned pre-processing unit, The known circuit layout is divided into sections, A first coefficient is set for the defective area section. A second coefficient different from the first coefficient is set in the section of the normal area. Multiple defective circuit images are extracted from the defective area section. Multiple normal circuit images are extracted from the normal area section. The extracted plurality of faulty circuit images, the plurality of normal circuit images, the first coefficient, and the second coefficient are configured to be output to the image acquisition unit. The learning support device according to claim 1, wherein the pseudo-image acquisition unit inputs the first coefficient and the second coefficient as parameters to the generating AI.

8. The first coefficient and the second coefficient are adjustment parameters for adjusting the magnitude of the change in the pseudo-image generated by the generation AI relative to the original image. The larger the adjustment parameter, the more the pseudo-image differs from the original image. The learning support device according to claim 7, wherein the preprocessing unit sets the value of the first coefficient to be smaller than the value of the second coefficient.

9. The learning support device according to claim 8, wherein setting the value of the first coefficient to be smaller than the value of the second coefficient includes setting the values ​​of the first coefficient and the second coefficient so that the values ​​of the first coefficient and the second coefficient are in a predetermined ratio.

10. The aforementioned pre-processing unit, Multiple rotated images are generated by rotating each of the above multiple defective circuit images by 90 degrees. The learning support device according to claim 7, configured to include the plurality of rotation images in the plurality of faulty circuit images and output the plurality of faulty circuit images to the image acquisition unit.

11. The learning support device according to claim 1, further comprising a post-processing unit that grays out the plurality of pseudo-images before the plurality of pseudo-images are input to the classification AI.

12. The trained classification AI is configured to output the probability that the input circuit image is a defective circuit image. The system further includes a heat map creation unit that outputs a heat map of all or part of the semiconductor device being inspected. The heat map creation unit, Multiple input circuit images constituting all or part of the semiconductor device to be inspected are input to the trained classification AI. The learning support device according to claim 1, which creates a heat map of all or part of the semiconductor device to be inspected based on the probability of each of the plurality of input circuit images output by the trained classification AI.

13. A method for supporting the learning of classification AI, Obtain multiple faulty circuit images showing the faulty circuit layout extracted from known circuit layouts, and multiple normal circuit images showing the normal circuit layout, Each of the aforementioned multiple faulty circuit images is input to the generation AI, and the generation AI obtains multiple pseudo-images generated based on the aforementioned multiple faulty circuit images. The plurality of defective circuit images, the plurality of pseudo-images, and the plurality of normal circuit images are input to the classification AI as training data. A learning support method, which includes outputting the pre-trained classification AI.

14. A learning support program for classification AI executed by a computer, Obtain multiple faulty circuit images showing the faulty circuit layout extracted from known circuit layouts, and multiple normal circuit images showing the normal circuit layout, Each of the aforementioned multiple faulty circuit images is input to the generation AI, and the generation AI obtains multiple pseudo-images generated based on the aforementioned multiple faulty circuit images. The plurality of defective circuit images, the plurality of pseudo-images, and the plurality of normal circuit images are input to the classification AI as training data. A learning support program that causes the computer to output the pre-trained classification AI.

Images