# US 12,762,685 B2 · Substrate inspection device, substrate inspection system, and substrate inspection method > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,762,685 B2 - Document ID: US12762685B2 - Status: Granted - Kind: B2 - Application number: 18/924,338 - Filed: 2024-10-23 - Published: 2025-02-06 - Granted: 2026-10-13 - Earliest claimed: 2019-06-06 - Assignees: Tokyo Electron Limited - Applicants: Tokyo Electron Limited - Inventors: Shuji Iwanaga; Tadashi Nishiyama - Web page: https://patentgrep.com/patent/US12762685B2 ## CPC - G06T 7/0004: {Industrial image inspection} - G06T 2207/20084: Artificial neural networks [ANN] - G06T 2207/30148: Semiconductor; IC; Wafer - G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL ## Abstract A substrate inspection apparatus for inspecting a substrate, includes: an acquisition part configured to acquire an estimated image of an inspection target substrate after a process by a substrate processing apparatus, based on an image estimation model created by a machine learning by using a captured image before the process by the substrate processing apparatus and a captured image after the process by the substrate processing apparatus for each of a plurality of substrates, and a captured image of the inspection target substrate before the process by the substrate processing apparatus; and a determination part configured to determine the presence or absence of a defect in the inspection target substrate, based on a captured image of the inspection target substrate and the estimated image of the inspection target substrate after the process by the substrate processing apparatus. ## Claims ### Claim 1 An information processing apparatus, comprising: a memory storing a captured image before a process and a captured image after the process for each of a plurality of substrates; and a controller configured to: select an image set for model creation, which is a captured image set including a combination of the captured image before the process and the captured image after the process for each of the plurality of substrates stored in the memory; and create an image estimation model by machine learning by using the captured image before the process and the captured image after the process for the image set for model creation, wherein the controller is further configured to select the image set for model creation based on correlation information between in-plane tendency of a pixel value in the captured image before the process and in-plane tendency of a pixel value in the captured image after the process. ### Claim 2 (depends on 1) The information processing apparatus of claim 1, wherein the image estimation model generates an estimated image after the process for an inspection target substrate. ### Claim 3 (depends on 2) The information processing apparatus of claim 2, wherein a presence or absence of a defect in the inspection target substrate is determined based on a captured image after the process for the inspection target substrate and the estimated image after the process for the inspection target substrate. ### Claim 4 (depends on 1) The information processing apparatus of claim 1, wherein the correlation information is a degree of abnormality determined by using a correlation distribution between the in-plane tendency of a pixel value in the captured image before the process and the in-plane tendency of a pixel value in the captured image after the process. ### Claim 5 (depends on 1) The information processing apparatus of claim 1, wherein the image estimation model is a generation network that converts an input arbitrary image to generate an estimated image in a conditional generative adversarial network, wherein in the conditional generative adversarial network, for an identification network that receives a captured image corresponding to the arbitrary image and obtained by capturing performed after the process, or the estimated image generated based on the arbitrary image, together with the arbitrary image, and identifies whether the image received together with the arbitrary image is the captured image corresponding to the arbitrary image or the estimated image, and an identification method is machine-learned so as to accurately perform the identification, and wherein for the generation network, an image conversion method is machine-learned so as to recognize the estimated image as the captured image corresponding to the arbitrary image in the identification network. ### Claim 6 (depends on 3) The information processing apparatus of claim 3, wherein the controller is further configured to: acquire a plurality of image sets as candidates for the image set for model creation; resolve a plane distribution of pixel values in each of captured images of a substrate included in the acquired image set into in-plane tendency components of a plurality of pixel values by using a Zernike polynomial and calculate a Zernike coefficient of each of the in-plane tendency components; calculate a Mahalanobis distance of each of the image sets for each term in the Zernike polynomial, based on a correlation distribution between a Zernike coefficient in a captured image of the substrate before the process and a Zernike coefficient in a captured image of the substrate after the process; and determine the correlation information based on the Mahalanobis distance for each of the image sets and extract the image set for model creation based on the correlation information among the image sets acquired by the acquiring the plurality of image sets as the candidates. ### Claim 7 (depends on 6) The information processing apparatus of claim 6, wherein the controller is further configured to determine the correlation information for each of the image sets by adding the Mahalanobis distance calculated for each term in the Zernike polynomial by the calculating the Mahalanobis distance and extract an image set in which the correlation information obtained by adding the Mahalanobis distance does not exceed a threshold value, among the image sets acquired by the acquiring the plurality of image sets as the candidates, as the image set for model creation. ### Claim 8 (depends on 6) The information processing apparatus of claim 6, wherein the correlation information is the Mahalanobis distance calculated by the controller for each term in the Zernike polynomial, and wherein the controller is configured to extract an image set having no term in the Zernike polynomial in which the correlation information exceeds the threshold value, among the image sets, as the image set for model creation. ### Claim 9 (depends on 1) A substrate processing system, comprising: the information processing apparatus of claim 1; and a substrate processing apparatus configured to process a substrate. ### Claim 10 An information processing method, the method comprising: selecting an image set for model creation, which is a captured image set including a combination of a captured image before a process and a captured image after the process for each of the plurality of substrates, and creating an image estimation model by machine learning by using the captured image before the process and the captured image after the process for the image set for model creation, wherein the selecting the image set for model creation includes selecting the image set for model creation based on correlation information between in-plane tendency of a pixel value in the captured image before the process and in-plane tendency of a pixel value in the captured image after the process. ### Claim 11 (depends on 10) The information processing method of claim 10, wherein the image estimation model generates an estimated image after the process for an inspection target substrate. ### Claim 12 (depends on 11) The information processing method of claim 11, wherein a presence or absence of a defect in the inspection target substrate is determined based on a captured image after the process for the inspection target substrate and the estimated image after the process for the inspection target substrate. ### Claim 13 (depends on 10) The information processing method of claim 10, wherein the correlation information is a degree of abnormality determined by using a correlation distribution between the in-plane tendency of a pixel value in the captured image before the process and the in-plane tendency of a pixel value in the captured image after the process. ### Claim 14 (depends on 10) The information processing method of claim 10, wherein the image estimation model is a generation network that converts an input arbitrary image to generate an estimated image in a conditional generative adversarial network, wherein in the conditional generative adversarial network, for an identification network that receives a captured image corresponding to the arbitrary image and obtained by capturing performed after the process, or the estimated image generated based on the arbitrary image, together with the arbitrary image, and identifies whether the image received together with the arbitrary image is the captured image corresponding to the arbitrary image or the estimated image, and an identification method is machine-learned so as to accurately perform the identification, and wherein for the generation network, an image conversion method is machine-learned so as to recognize the estimated image as the captured image corresponding to the arbitrary image in the identification network. ### Claim 15 (depends on 12) The information processing method of claim 12, wherein the selecting the image set for model creation further includes: acquiring a plurality of image sets as candidates for the image set for model creation; resolving a plane distribution of pixel values in each of captured images of a substrate included in the acquired image set into in-plane tendency components of a plurality of pixel values by using a Zernike polynomial and calculate a Zernike coefficient of each of the in-plane tendency components; calculating a Mahalanobis distance of each of the image sets for each term in the Zernike polynomial, based on a correlation distribution between a Zernike coefficient in a captured image of the substrate before the process and a Zernike coefficient in a captured image of the substrate after the process; and determining the correlation information based on the Mahalanobis distance for each of the image sets and extract the image set for model creation based on the correlation information among the image sets acquired by the acquiring the plurality of image sets as the candidates. ### Claim 16 (depends on 15) The information processing method of claim 15, wherein the selecting the image set for model creation further includes determining the correlation information for each of the image sets by adding the Mahalanobis distance calculated for each term in the Zernike polynomial by the calculating the Mahalanobis distance and extract an image set in which the correlation information obtained by adding the Mahalanobis distance does not exceed a threshold value, among the image sets acquired by the acquiring the plurality of image sets as the candidates, as the image set for model creation. ### Claim 17 (depends on 15) The information processing method of claim 15, wherein the correlation information is the Mahalanobis distance for each term in the Zernike polynomial, and wherein the selecting the image set for model creation further includes extracting an image set having no term in the Zernike polynomial in which the correlation information exceeds the threshold value, among the image sets, as the image set for model creation. ### Claim 18 A non-transitory computer-readable recording medium storing a program that causes, by a computer, a substrate processing apparatus to perform a process comprising: selecting an image set for model creation, which is a captured image set including a combination of a captured image before a process and a captured image after the process for each of the plurality of substrates, and creating an image estimation model by machine learning by using the captured image before the process and the captured image after the process for the image set for model creation, wherein the selecting the image set for model creation includes selecting the image set for model creation based on correlation information between in-plane tendency of a pixel value in the captured image before the process and in-plane tendency of a pixel value in the captured image after the process.