Learning model generation method, information processing method, computer program, and information processing device

Abstract

The invention provides a learning model generation method, an information processing method, a computer program, and an information processing device which are expected to remove noise from an image obtained by shooting a substrate with good precision. In the method for generating a learning model according to the present embodiment, an information processing device performs: a process for acquiring a plurality of images of a target substrate captured in time series; a learning data generation unit that generates learning data in which one of two images selected from the plurality of acquired images is used as an input and the other image is used as an output and the two images are associated with each other; and generating, by machine learning using the learning data, a learning model for receiving, as an input, an image obtained by capturing an image of the target substrate, and outputting an image from which the noise of the image is removed. It is preferable that the plurality of images include images of different conditions, and the two images associated with each other as input and output in the learning data are images of the same condition.

Description

Technical Field

[0001] The present disclosure relates to a method for generating a learning model, an information processing method, a computer program, and an information processing apparatus. Background Art

[0002] In Patent Document 1, a pattern inspection / measurement apparatus is proposed that uses the position of an edge extracted from image data obtained by photographing an object pattern with edge extraction parameters to inspect or measure the object pattern. The pattern inspection / measurement apparatus generates edge extraction parameters using a reference pattern representing a shape that is a reference for inspection or measurement and the image data.

[0003] Patent Document 1: Japanese Patent Application Laid-Open No. 2014-81220 Summary of the Invention

[0004] The present disclosure provides a method for generating a learning model, an information processing method, a computer program, and an information processing apparatus that can be expected to remove noise from an image obtained by photographing a substrate with good accuracy.

[0005] In a method for generating a learning model according to one embodiment, an information processing apparatus performs the following processing: acquiring a plurality of images of an object substrate taken in time series; generating learning data in which one of two images selected from the acquired plurality of images is used as an input and the other image is used as an output and is made to correspond; and generating a learning model by machine learning using the above learning data, wherein the learning model receives an image obtained by photographing an object substrate as an input and outputs an image from which the noise of the image has been removed.

[0006] According to the present disclosure, it is possible to expect to remove noise from an image obtained by photographing a substrate with good accuracy. Brief Description of the Drawings

[0007] Figure 1 is a schematic diagram for explaining the outline of the information processing system of the present embodiment.

[0008] Figure 2 is a block diagram showing an example of a configuration of the information processing apparatus of the present embodiment.

[0009] Figure 3 is a schematic diagram showing an example of a configuration of the learning model of the present embodiment.

[0010] Figure 4 is a flowchart showing an example of steps of photographing processing for data collection performed by the substrate inspection apparatus of the present embodiment.

[0011] Figure 5It is a flowchart showing an example of a prescribed procedure for data collection and learning model generation performed by the information processing apparatus according to the present embodiment.

[0012] Figure 6 It is a schematic diagram for explaining an example of the process of position offset correction performed by the learning data generation unit.

[0013] Figure 7 It is a schematic diagram for explaining an example of the generation process of an image pair.

[0014] Figure 8 It is a flowchart showing an example of the steps of the length measurement and inspection processes performed by the information processing apparatus according to the present embodiment.

[0015] Figure 9 It is a schematic diagram showing an example of noise removal performed by the information processing system according to the present embodiment. Detailed Embodiments

[0016] Hereinafter, specific examples of the information processing system according to the embodiment of the present disclosure will be described with reference to the drawings. In addition, the present disclosure is not limited to these examples, and is represented by the claims, and is intended to include all modifications within the meaning and scope equivalent to the claims.

[0017] <System Outline>

[0018] Figure 1 It is a schematic diagram for explaining the outline of the information processing system according to the present embodiment. The information processing system according to the present embodiment is configured to include an information processing apparatus 1 and a substrate inspection apparatus 3. The substrate inspection apparatus 3 according to the present embodiment has a photographing function such as a SEM (Scanning Electron Microscope) or a TEM (Transmission Electron Microscope) that photographs a wafer (substrate) to be inspected. The substrate inspection apparatus 3 is, for example, an apparatus that photographs a wafer and acquires a photographed image for inspecting the wafer processed by an apparatus such as a substrate processing apparatus that performs processes such as etching on a semiconductor wafer.

[0019] The information processing apparatus 1 is an apparatus that performs processes related to control and monitoring of the operation of the substrate inspection apparatus 3. The information processing apparatus 1 according to the present embodiment performs a process of generating a learning model 5 by collecting SEM images of a wafer photographed by the substrate inspection apparatus 3 and performing machine learning using the collected multiple SEM images. The learning model 5 is a learning model that receives a SEM image photographed by the substrate inspection apparatus 3 as input and outputs a SEM image with noise removed from the input image.

[0020] In addition, the information processing device 1 uses the generated learning model 5 to perform processes such as length measurement and defect inspection related to the formations on the wafer. That is, the information processing device 1 acquires the SEM image of the wafer captured by the substrate inspection device 3 and inputs the acquired SEM image into the learning model 5. The information processing device 1 acquires the noise-removed SEM image output by the learning model 5 and performs processes such as length measurement and inspection based on the acquired SEM image. Thus, the information processing device 1 can perform processes such as length measurement and inspection using the noise-removed SEM image of the wafer, and therefore, it is possible to expect an improvement in the accuracy of processes such as length measurement and inspection compared to the case of using an SEM image without noise removal. In addition, the processes such as length measurement and defect inspection of the wafer using the learning model 5 may be performed not by the information processing device 1 but by the substrate inspection device 3.

[0021] Figure 2 FIG. is a block diagram showing a structural example of the information processing device 1 according to the present embodiment. The information processing device 1 according to the present embodiment can be realized, for example, by installing a prescribed application program or the like in a general information processing device such as a personal computer or a server computer. The information processing device 1 is configured to include a processing unit 11, a storage unit 12, a communication unit 13, a display unit 14, an operation unit 15, and the like. In addition, in the present embodiment, it is described that the processing is performed by one information processing device 1, but the processing of the information processing device 1 may be performed dispersedly by a plurality of devices.

[0022] The processing unit 11 is configured to use an arithmetic processing device such as a CPU (Central Processing Unit), an MPU (Micro-Processing Unit), a GPU (Graphics Processing Unit), or a quantum processor, a ROM (Read Only Memory), a RAM (Random Access Memory), and the like. The processing unit 11 performs various processes such as a process of collecting SEM images of the wafer captured by the substrate inspection device 3 to generate the learning model 5 and a process of removing the noise of the SEM image using the generated learning model 5 by reading and executing the program 12a stored in the storage unit 12.

[0023] The storage unit 12 is configured to use a large-capacity storage device such as a hard disk, for example. The storage unit 12 stores various programs executed by the processing unit 11 and various data required for the processing of the processing unit 11. In the present embodiment, the storage unit 12 stores the program 12a executed by the processing unit 11. In addition, a learning data storage unit 12b for storing learning data used in machine learning for generating the learning model 5 and a model information storage unit 12 for storing information related to the generated learning model 5 are provided in the storage unit 12.

[0024] In the present embodiment, the program (computer program, program product) 12a is provided in a manner recorded on a recording medium 99 such as a memory card or an optical disc. The information processing device 1 reads the program 12a from the recording medium 99 and stores it in the storage unit 12. However, the program 12a can also be written into the storage unit 12, for example, during the manufacturing stage of the information processing device 1. In addition, for example, the information processing device 1 can obtain the program 12a distributed by a remote server device or the like through communication. For example, a writing device can read the program 12a recorded on the recording medium 99 and write it into the storage unit 12 of the information processing device 1. The program 12a can be provided either in a distribution manner via a network or in a manner recorded on the recording medium 99.

[0025] The learning data storage unit 12b stores learning data generated based on the SEM images of wafers acquired and collected from the substrate inspection device 3. As will be described in detail later, the learning model 5 of the present embodiment has a structure that receives an SEM image as input and outputs an SEM image from which noise has been removed from the SEM image. The learning data used in the machine learning of the learning model 5 is data that establishes a correspondence between two SEM images, namely, the SEM image corresponding to the input of the learning model 5 and the SEM image corresponding to the output of the learning model 5. In the present embodiment, the information processing device 1 performs multiple shootings of one inspection object based on the substrate inspection device 3. The information processing device 1 selects two SEM images from the multiple SEM images obtained by shooting, generates data that establishes a correspondence with one of the two SEM images as input and the other as output as learning data, and stores it in the learning data storage unit 12b.

[0026] In addition, in the present embodiment, the information processing apparatus 1 collects SEM images captured according to various objects and various conditions, generates a plurality of learning data based on the plurality of SEM images captured according to various conditions, and stores the generated learning data in the learning data storage unit 12b for accumulation. In a wafer that is an inspection object of the substrate inspection apparatus 3, for example, irregularities corresponding to circuit elements or wirings constituting a semiconductor circuit are formed on the surface. In the present embodiment, it is preferable that the SEM images collected by the information processing apparatus 1 for learning data include images having various patterns of the shapes of the irregularities formed on the wafer.

[0027] In addition, in the case of using, as learning data, captured TEM images of a wafer or the like, the images collected by the information processing apparatus 1 may include, for example, images of the internal shape or cross-sectional shape of the wafer. In this case, it is preferable that the information processing apparatus 1 includes images having various patterns of boundary information of a plurality of films constituting the wafer formation. In the present embodiment, it is preferable that the information processing apparatus 1 collects images having various patterns of the structure of the wafer formation to generate learning data.

[0028] In addition, an SEM image is an image obtained by scanning an electron beam over the surface of a wafer, and the SEM image obtained by the imaging of the substrate inspection apparatus 3 may include an image obtained by accumulating (averaging) the results of multiple imaging (scanning of the electron beam). In the present embodiment, it is preferable that the SME images collected by the information processing apparatus 1 for learning data include images having various accumulation numbers. In addition, it is preferable that the collected SEM images include images captured at various scanning speeds, or images captured at various resolutions. It is possible to expect that the information processing apparatus 1 of the present embodiment generates the following learning model 5: by collecting SEM images captured according to various conditions to generate learning data, noise removal from the SEM images is performed with good accuracy.

[0029] The model information storage unit 12c stores information related to the learning model 5 generated by the information processing apparatus 1 through machine learning. The learning model 5 of the present embodiment may adopt, for example, a learning model having a structure such as DNN (Deep Neural Network), CNN (Convolutional Neural Network), FCN (Fully Convolution Network), or U-Net. Among the information related to the learning model 5 stored in the model information storage unit 12c, for example, structure information indicating what structure the learning model is, and information such as the values of the parameters inside the learning model may be included.

[0030] The communication unit 13 is connected to the substrate inspection device 3 via a cable such as a communication line or a signal line, and data is transmitted and received between the communication unit 13 and the substrate inspection device 3 via this cable. In the present embodiment, the communication unit 13 receives data of the SEM image of the wafer transmitted from the substrate inspection device 3, and provides the received data to the processing unit 11.

[0031] The display unit 14 is configured using a liquid crystal display or the like, and displays various images, characters, etc. based on the processing of the processing unit 11. In the present embodiment, the display unit 14, for example, displays the SEM image of the wafer obtained from the substrate inspection device 3, and displays the results of length measurement and inspection based on the SEM image, etc.

[0032] The operation unit 15 accepts the operation of the user, and notifies the accepted operation to the processing unit 11. For example, the operation unit 15 accepts the operation of the user through an input device such as a mechanical button or a touch panel provided on the surface of the display unit 14. In addition, for example, the operation unit 15 can be an input device such as a mouse and a keyboard, and these input devices can also be configured to be detachable from the information processing device 1.

[0033] In addition, the storage unit 12 can be an external storage device connected to the information processing device 1. In addition, the information processing device 1 can be a multi-computer composed of multiple computers, or can be a virtual machine virtually constructed by software. In addition, the information processing device 1 is not limited to the above structure, and for example, it may not include the display unit 14 and the operation unit 15, etc.

[0034] In addition, in the information processing device 1 of the present embodiment, the processing unit 11 reads out the program 12a stored in the storage unit 12 and executes it, whereby the information acquisition unit 11a, the learning data generation unit 11b, the model generation unit 11c, the noise removal unit 11d, and the length measurement / inspection processing unit 11e, etc. are implemented by the processing unit 11 as software functional units.

[0035] The information acquisition unit 11a communicates with the substrate inspection device 3 using the communication unit 13, thereby performing the process of acquiring the SEM image of the inspection target wafer photographed by the substrate inspection device 3. In addition, the information acquisition unit 11a acquires information related to the shooting conditions when shooting the SEM image, such as the cumulative number, scanning speed, and resolution, etc. from the substrate inspection device 3 together with the SEM image. The information acquisition unit 11a stores the SEM image obtained from the substrate inspection device 3 and the information related to the shooting conditions in the storage unit 12 in a corresponding manner.

[0036] In addition, in the present embodiment, when acquiring SEM images for learning data used for machine learning to generate the learning model 5, the information acquisition unit 11a acquires a plurality of SEM images obtained by photographing the same inspection object a plurality of times in time series from the substrate inspection device 3. The information acquisition unit 11a stores these plurality of SEM images in the storage unit 12 in correspondence with conditions related to the photographing conditions. In addition, it is assumed that the photographing conditions are the same when photographing the same inspection object a plurality of times. When photographing an inspection object different from this inspection object, the photographing conditions can be changed. Preferably, the SEM images collected by the information acquisition unit 11a for generating learning data include images photographed under various photographing conditions.

[0037] In addition, in the present embodiment, when acquiring an SEM image for measuring the length or inspecting a wafer as an inspection object, the information acquisition unit 11a only needs to acquire one SEM image obtained by photographing the inspection object at least once. However, in this case, the information acquisition unit 11a may also acquire a plurality of SEM images obtained by photographing a plurality of times. In addition, in this case, the information acquisition unit 11a may or may not acquire information related to the photographing conditions together with the SEM image.

[0038] The learning data generation unit 11b performs the following processing: Based on the plurality of SEM images acquired by the information acquisition unit 11a, learning data (teacher data) for performing machine learning for generating the learning model 5 is generated. As described above, in the present embodiment, the information acquisition unit 11a acquires a plurality of SEM images obtained by photographing the same inspection object a plurality of times in time series. The learning data generation unit 11b appropriately extracts two SEM images from the plurality of SEM images. The learning data generation unit 11b generates data in which one of the two extracted SEM images is set as an input image (source image) to the learning model 5 and the other SEM image is set as an output image (target image, correct value) of the learning model 5. The learning data generation unit 11b stores the generated data as learning data in the learning data storage unit 12b.

[0039] In addition, the learning data generation unit 11b may perform processing for correcting the shift of the photographing position of the inspection object on the plurality of SEM images obtained by photographing the same inspection object a plurality of times in time series. For example, the learning data generation unit 11b can determine the position where the inspection object is reflected in the plurality of SEM images, cut out an image region of a specified size where the inspection object is reflected from each SEM image, and use the cut-out image as the corrected SEM image. The learning data generation unit 11b generates learning data based on the corrected image.

[0040] The model generation unit 11c generates the learning model 5 by performing machine learning processing using the learning data stored in the learning data storage unit 12b. In the present embodiment, the learning model 5 is a learning model that receives the input of an SEM image and generates and outputs an SEM image with noise removed from the input SEM image. In the present embodiment, the machine learning performed by the model generation unit 11c to generate the learning model 5 is machine learning that employs the Noise2Noise method. In the machine learning of Noise2Noise, by having the learning model learn the conversion from an image containing noise to an image containing noise, a learning model that converts an image containing noise into an image with noise removed can be generated. The machine learning of Noise2Noise is a prior art, so a detailed description is omitted in the present embodiment.

[0041] In the present embodiment, the SEM image acquired from the substrate inspection apparatus 3 is an image containing noise, and all of the multiple SEM images obtained by performing multiple shootings in time series are images containing noise. As described above, the learning data generation unit 11b generates learning data by extracting two images from these multiple SEM images and establishing a correspondence, so both of the two SEM images included in the learning data are images containing noise. The model generation unit 11c can determine the internal parameters of the learning model 5 and generate the learning model 5 by performing so-called supervised machine learning with one SEM image included in the learning data as the input image to the learning model 5 and the other SEM image as the correct value of the output image of the learning model 5. The model generation unit 11c stores information related to the generated learning model 5 in the model information storage unit 12c.

[0042] The noise removal unit 11d performs the following processing: using the learning model 5 generated in advance by machine learning, it removes the noise from the SEM image acquired from the substrate inspection apparatus 3. The noise removal unit 11d reproduces the learning model 5 that has completed machine learning based on the information stored in the model information storage unit 12c. The noise removal unit 11d removes the noise of the SEM image by inputting the SEM image of the wafer to be inspected, which is acquired by the information acquisition unit 11a from the substrate inspection apparatus 3, into the learning model 5 and obtaining the SEM image output by the learning model 5.

[0043] The length measurement / inspection processing unit 11e performs processing such as length measurement and inspection on the wafer to be inspected imaged in the SEM image based on the SEM image from which noise has been removed by the noise removal unit 11d. For example, the length measurement / inspection processing unit 11e performs image processing such as edge detection on the SEM image, thereby identifying the uneven pattern such as the surface or cross-section of the wafer imaged in the SEM image. The length measurement / inspection processing unit 11e can perform length measurement, for example, by measuring the distance between two specified corner portions included in the uneven pattern. In addition, the length measurement / inspection processing unit 11e can inspect the wafer to be inspected for defects, for example, by investigating whether the uneven pattern matches a specified pattern or whether the length measurement result is within a specified range. The length measurement / inspection processing unit 11e displays the length measurement result or inspection result, etc. on the display unit 14.

[0044] In addition, these processes performed by the length measurement / inspection processing unit 11e are just examples and are not limited thereto. The length measurement / inspection processing unit 11e can perform length measurement by any method and can also perform inspection by any method. The information processing device 1 may also be configured to perform only one of length measurement and inspection based on the SEM image and not perform the other. In addition, the information processing device 1 may perform the process of removing noise from the SEM image, and other devices may perform processing such as length measurement and inspection based on the SEM image from which noise has been removed.

[0045] <Generation of learning model>

[0046] Figure 3 FIG. is a schematic diagram showing a structural example of the learning model 5 of the present embodiment. The learning model 5 of the present embodiment forms a structure such as DNN or CNN, accepts a SEM image of a specified size as input, and outputs a SEM image of a specified size. The SEM image output by the learning model 5 is a SEM image from which noise has been removed from the input SEM image. By using the learning model 5, the information processing device 1 can accurately perform processing such as length measurement and inspection on the wafer processed by the substrate inspection device 3 based on the SEM image captured by the substrate inspection device 3. Before performing such processing as length measurement and inspection, the information processing device 1 performs a process of collecting SEM images captured by the substrate inspection device 3 and generating the learning model 5 through machine learning using the collected SEM images.

[0047] Figure 4It is a flowchart showing an example of the steps of the imaging process for data collection performed by the substrate inspection apparatus 3 of the present embodiment. The substrate inspection apparatus 3 of the present embodiment performs an imaging process of a wafer for data collection under the control from the information processing apparatus 1, for example. The substrate inspection apparatus 3 positions the part to be imaged on the wafer to be imaged (step S1). The substrate inspection apparatus 3 performs SEM imaging by scanning an electron beam over the target part of the wafer at the determined imaging position (step S2). The substrate inspection apparatus 3 determines whether a specified number of SEM images, specified by the information processing apparatus 1, for example, have been taken (step S3). If the specified number of images has not been taken (S3: No), the substrate inspection apparatus 3 returns the process to step S2 and repeats the imaging for the same part of the wafer.

[0048] If the specified number of images has been taken (S3: Yes), the substrate inspection apparatus 3 sends the specified number of captured SEM images to the information processing apparatus 1 (step S4). At this time, the substrate inspection apparatus 3 may send information such as the imaging conditions together with the multiple SEM images to the information processing apparatus 1. The substrate inspection apparatus 3 determines whether the imaging of all the parts of the wafer to be inspected has been completed (step S5). If the imaging of all the required parts has not been completed (S5: No), the substrate inspection apparatus 3 returns the process to step S1, positions other parts of the wafer, and performs SEM imaging. If the imaging of all the required parts has been completed (S5: Yes), the substrate inspection apparatus 3 ends the process.

[0049] In addition, the number of times of imaging a single part (the specified number determined in step S3) may not be constant. For example, different numbers of images may be taken for each imaging part. In this case, the number of images taken at each part may be determined by the substrate inspection apparatus 3 or by the user. The substrate inspection apparatus 3 may, for example, randomly determine the number of images, or may evaluate the amount of noise or sharpness, etc. of the first captured image and determine the number of images based on the evaluation result, or may determine the number of images by other methods. Further, the substrate inspection apparatus 3 may, for example, accept input of the number of images from the user for each imaging part, or may determine the number of images based on the settings of the number of images previously made by the user.

[0050] Figure 5It is a flowchart showing an example of a prescribed procedure for data collection and learning model generation performed by the information processing apparatus 1 of the present embodiment. The information acquisition unit 11a of the processing unit 11 of the information processing apparatus 1 of the present embodiment communicates with the substrate inspection apparatus 3 using the communication unit 13, thereby acquiring a plurality of SEM images obtained by the substrate inspection apparatus 3 photographing the same part of the wafer multiple times (step S21).

[0051] Next, the learning data generation unit 11b of the processing unit 11 performs a position offset correction process for correcting the offset of the position of the photographed object part between the images on the plurality of SEM images acquired in step S21 (steps S22 to S25).

[0052] Figure 6 It is a schematic diagram for explaining an example of the position offset correction process performed by the learning data generation unit 11b. In Figure 6 , a plurality of SEM images (first image to third image) obtained by photographing the same part of the wafer are arranged vertically. In addition, in Figure 6 , the results of the processing for each SEM image are arranged horizontally from left to right in time series in the order of pre-correction image → calculating the offset amount → determining the shear size → post-correction image. Figure 6 Each image illustrated in is an SEM image obtained by photographing a part of the wafer to be inspected, for example, a part where four circular formations are arranged vertically and horizontally.

[0053] The substrate inspection apparatus 3 converges and irradiates an electron beam on the wafer to be inspected, and scans the electron beam on the wafer, thereby obtaining an SEM image. Therefore, even when the wafer is fixedly held and the same part is continuously photographed, due to the error in the scanning position of the electron beam, the position of the inspection object reflected in the photographed image may shift. By pre-correcting the position offset of a plurality of SEM images and generating learning data based on the corrected SEM images, the information processing apparatus 1 of the present embodiment can expect to improve the noise removal accuracy of the learning model 5 generated by machine learning using this learning data.

[0054] The learning data generation unit 11b, for example, uses the first SEM image as a basic image, and calculates the offset amount of the position of the inspection object reflected in the image for this basic image and each subsequent image. At this time, the learning data generation unit 11b, for example, moves the second image in the up, down, left, and right directions relative to the reference image, and searches for a position where the inspection objects reflected in the two images match. In Figure 6In the example shown, by moving the second image a few pixel amounts in the upper left direction, the inspection objects respectively reflected in the reference image and the second image are made to coincide. The learning data generation unit 11b calculates the amount (number of pixels) by which the second image has moved from its original position during this movement as the offset amount relative to the reference image. In Figure 6 In the example shown, the offset amount of the second image is a few pixels to the right and a few pixels downward. In this figure, this offset amount is represented by the areas blackened on the right side and the lower side of the second image.

[0055] The learning data generation unit 11b similarly calculates the offset amount relative to the reference image for the images after the third image as well. In Figure 6 In the example shown, the offset amount of the third image is a few pixels upward and a few pixels to the left. The learning data generation unit 11b calculates the offset amounts relative to the reference image from the second image to the last image for each of the four directions of up, down, left, and right (step S22). Then, the learning data generation unit 11b calculates the maximum value of the offset amounts for each direction based on the multiple offset amounts calculated for each of the four directions of up, down, left, and right (step S23).

[0056] Next, the learning data generation unit 11b determines the size of the image regions to be cut from the multiple images based on the maximum values of the offset amounts related to the four directions of up, down, left, and right that have been calculated. The learning data generation unit 11b uses the size obtained by reducing the size of the original image by the maximum values of the offset amounts calculated for each of the four directions of up, down, left, and right as the cutting size (step S24). In Figure 6 In the example shown, the cutting sizes for each image are overlapped and displayed with a square frame.

[0057] Next, the learning data generation unit 11b cuts an image region of the determined cutting size from the first image (reference image). In addition, for the images after the second image, the learning data generation unit 11b cuts an image region of the determined cutting size from the image after alignment with the reference image ( Figure 6 the image of "calculating the offset amount"). Thus, the learning data generation unit 11b cuts image regions of the determined cutting size from the multiple images respectively (step S25), and uses the multiple cut images as the corrected images. In addition, Figure 6 The offset correction method shown is an example, and is not limited thereto. The information processing apparatus 1 may also correct the offset by any method.

[0058] Next, the learning data generation unit 11b generates multiple sets of image pairs based on the multiple SEM images with the position offset corrected (step S26). Figure 7It is a schematic diagram for explaining an example of the generation process of an image pair. In Figure 7 In the example shown, N SEM images are arranged horizontally in order from the first to the Nth, and further, the same SEM images are arranged in two rows, top and bottom. However, in Figure 7 the top row, the Nth image is omitted and represented by a dashed box, and in Figure 7 the bottom row, the first image is omitted and represented by a dashed box. In addition, Figure 7 each of the images shown is an SEM image obtained by photographing a part of the wafer to be inspected, for example, a part where a plurality of linearly arranged formations extending in the up-down direction are arranged in the left-right direction.

[0059] The learning data generation unit 11b sets the first image and the second image as a pair, sets the second image and the third image as a pair, …, sets the (N - 1)th image and the Nth image as a pair, thereby generating (N - 1) groups of image pairs. That is, the learning data generation unit 11b generates (N - 1) groups of image pairs by setting two images, one being the image at a certain time point and the other being the image at the next time point, from the N SEM images obtained by photographing the same part in time series. The learning data generation unit 11b stores, as learning data, the data in which one of the two images of the generated image pair is set as the input image to the learning model 5 and the other is set as the output image of the learning model 5 in the learning data storage unit 12b (step S27).

[0060] The learning data generation unit 11b determines whether the collection of learning data is completed based on whether there is sufficient accumulation of learning data for generating the learning model 5 (step S28). If the collection of learning data is not completed (S28: No), the learning data generation unit 11b returns the process to step S21, obtains a plurality of SEM images captured by the substrate inspection device 3 for different parts of the same wafer or different wafers, and repeats the above process. In addition, in the present embodiment, the number of SEM images captured by the substrate inspection device 3 for the same part may be different for each captured part. Therefore, the number of SEM images obtained by the learning data generation unit 11b in step S21 may also be different for each cycle of steps S21 to S28.

[0061] When the collection of learning data is completed (S28: Yes), the model generation unit 11c of the processing unit 11 reads out a plurality of pieces of learning data stored in the learning data storage unit 12b (step S29). The model generation unit 11c uses the learning data read out in step S29, takes one of the pair of two SEM images included in the learning data as the input image to the learning model 5, and takes the other as the output image (correct value) of the learning model 5, and performs so-called supervised machine learning (step S30), thereby generating the learning model 5. The model generation unit 11c stores information related to the learning model 5 generated by the machine learning in step S30 in the model information storage unit 12c (step S31), and ends the processing.

[0062] <Utilization of Learning Model>

[0063] Figure 8 It is a flowchart showing an example of the steps of the length measurement and inspection processing performed by the information processing apparatus 1 of the present embodiment. In the information processing system of the present embodiment, for example, various processes for a wafer are performed in a substrate processing apparatus, and the wafer is photographed by the substrate inspection apparatus 3 during or after these processes. The information acquisition unit 11a of the processing unit 11 of the information processing apparatus 1 communicates with the substrate inspection apparatus 3 using the communication unit 13, thereby acquiring the SEM image of the wafer photographed by the substrate inspection apparatus 3 (step S41).

[0064] The noise removal unit 11d of the processing unit 11 reads out the information stored in the model information storage unit 12c to construct the learning model 5, and inputs the SEM image acquired in step S41 into the learning model 5 (step S42). The noise removal unit 11d acquires the SEM image output by the learning model 5 (the SEM image with noise removed) based on the input of the SEM image in step S42 (step S43).

[0065] Figure 9 It is a schematic diagram showing an example of noise removal performed by the information processing system of the present embodiment. In Figure 9Among them, nine SEM images are arranged in a 3×3 matrix. The three SEM images in the upper row are the SEM images obtained by the information processing device 1 from the substrate inspection device 3. The SEM image taken once shown on the left side of the upper row is the SEM image obtained by scanning the electron beam once by the substrate inspection device 3. The SEM image averaged over four shots shown in the center of the upper row is the SEM image obtained by calculating the average value of the four SEM images obtained by scanning the electron beam four times by the substrate inspection device. The SEM image averaged over sixteen shots shown on the right side of the upper row is the SEM image obtained by calculating the average value of the sixteen SEM images obtained by scanning the electron beam sixteen times by the substrate inspection device. In addition, Figure 9 The three SEM images in the middle row are the SEM images obtained by performing image processing using a noise removal filter on the three SEM images in the upper row respectively. In addition, the noise removal filter uses a filter of the BM3D (Block Matching and 3D collaborative filtering) method, but the noise removal based on BM3D is a prior art, so the detailed description is omitted. Figure 9 The three SEM images in the lower row are the SEM images obtained by performing noise removal on the three SEM images in the upper row using the learning model 5 of the present embodiment respectively.

[0066] In the case of performing noise removal based on the filter Figure 9 on the SEM image taken once shown, the noise is not completely removed, and an unclear SEM image of the object to be inspected is obtained. In contrast, in the case of performing noise removal using the learning model 5 of the present embodiment based on the SEM image taken once, the noise is greatly removed, and a clear SEM image of the object to be inspected is obtained. In the noise removal using the learning model 5, even in the case of the SEM image taken once, the same degree of noise removal as in the case of the SEM image averaged over sixteen shots is achieved. In the noise removal based on the filter, in the case of the SEM image averaged over sixteen shots, although the same degree of noise removal as in the case of using the learning model 5 can be performed, sixteen shots require a lot of time.

[0067] After removing the noise from the SEM image in steps S42 and S43, the length measurement / inspection processing unit 11e of the processing unit 11 performs a length measurement process on the formation on the wafer reflected in the SEM image (step S44). In the length measurement process, the length measurement / inspection processing unit 11e, for example, performs edge detection on the SEM image to grasp the shape of the unevenness of the formation, and calculates the distance between specified parts (such as corner parts, etc.) of the unevenness. The length measurement / inspection processing unit 11e calculates the actual distance based on the calculated distance on the SEM image and the magnification of the shot. In the case of the SEM image shown in Figure 9 , the length measurement / inspection processing unit 11e can, for example, measure distances such as the diameter or circumference for a circular ring or cylindrical structure.

[0068] The length measurement / inspection processing unit 11e determines whether there is an abnormality in the formation on the wafer based on a comparison between the result of the length measurement in step S44 and a pre-determined threshold value (step S45). In the case where an abnormality is determined (S45: Yes), the length measurement / inspection processing unit 11e, for example, displays a warning message on the display unit 14 to notify the abnormality (step S46), and ends the process. In the case where no abnormality is determined (S45: No), the length measurement / inspection processing unit 11e ends the process without notifying the abnormality.

[0069] <Summary>

[0070] In the information processing system of the present embodiment having the above structure, the information processing device 1 acquires a plurality of SEM images related to the wafer to be inspected, which are obtained by the substrate inspection device 3 by shooting multiple times in time series. The information processing device 1 generates learning data in which one SEM image selected from the plurality of acquired SEM images is used as an input and the other SEM image is used as an output and is made to correspond. The information processing device 1 generates a learning model 5 that takes the SEM image obtained by shooting the wafer to be inspected as an input and outputs the SEM image with the noise removed from the received image by performing machine learning processing using the generated learning data. Thus, the information processing system can use the generated learning model 5 to remove the noise from the SEM image, and it is possible to expect an improvement in the accuracy of length measurement or inspection of the wafer using the SEM image.

[0071] In addition, in the present embodiment, the SEM image is processed as the image acquired from the substrate inspection device 3, but it is not limited thereto, and TEM images or various other images can be adopted.

[0072] In addition, in the information processing system of the present embodiment, the plurality of SEM images acquired by the information processing device 1 from the substrate inspection device 3 may include images under different conditions. Among the images with different conditions, for example, there may be images with different patterns of the uneven shapes formed on the substrate to be inspected, images with different cumulative counts, images with different scanning speeds, or images with different resolutions. However, in the learning data, the two SEM images that are input-output corresponding are SEM images under the same conditions. By performing machine learning using these multiple SEM images with different conditions, the learning model 5 can learn images under various conditions, and it is expected to improve the accuracy of noise removal of the learning model 5. In addition, the above conditions are just examples, and are not limited thereto. The plurality of images may include SEM images under various conditions other than these. For example, when the images processed by the information processing system are color images, for conditions such as the number of colors, resolution, or grayscale related to various colors such as RGB or CMY, there may be images with different conditions.

[0073] In addition, in the information processing system of the present embodiment, based on the plurality of SEM images captured by the substrate inspection device 3 in time series, the information processing device 1 generates learning data in which one image included in the plurality of SEM images is used as an input and the image after the one image is used as an output and they are corresponding. For example, for two consecutive images in time series, the information processing device 1 can generate learning data in which the previous image is used as an input and the next image is used as an output and they are corresponding. The information processing device 1 can generate a plurality of learning data by extracting multiple sets of two consecutive images from the plurality of SEM images. Thus, by making the plurality of SEM images corresponding in time series order, the information processing device 1 can easily generate learning data.

[0074] In addition, the method of generating learning data based on the plurality of SEM images is not limited to the above method. For example, the information processing device 1 can also randomly extract two images from the plurality of SEM images with the same conditions and appropriately establish correspondence with the input and output to generate learning data. It is also possible to use the SEM image after the time series as the input of the learning model 5 and the SEM image before the time series as the output (correct value) to generate learning data.

[0075] In addition, in the information processing system of the present embodiment, for a plurality of SEM images obtained from the substrate inspection apparatus 3, the information processing apparatus 1 performs a process of correcting the positions of the inspection objects reflected in each SEM image, and generates learning data based on the corrected images. For example, the information processing apparatus 1 selects one reference image from the plurality of SEM images, and calculates the offset amount of each image other than the reference image with respect to the position of the reference image. The information processing apparatus 1 can determine the shearing size based on the maximum value of the calculated offset amount, shear an image of this size from each image, and use the sheared image as the corrected image. By correcting the position offset in advance to generate learning data, it is possible to expect an improvement in the noise removal accuracy of the learning model 5 generated by machine learning using this learning data.

[0076] In addition, in the information processing system of the present embodiment, the information processing apparatus 1 acquires an SEM image obtained by the substrate inspection apparatus 3 photographing a wafer of an inspection object, inputs the acquired SEM image into the learned learning model 5, and acquires the SEM image output by the learning model 5. Thus, the information processing apparatus 1 can acquire an SEM image from which noise has been removed by the learning model 5. Based on the acquired SEM image, the information processing apparatus 1 performs length measurement or inspection of the inspection object reflected in the SEM image. Thus, the information processing apparatus 1 can perform length measurement and inspection and other processes based on the SEM image from which noise has been removed, and an improvement in the accuracy of length measurement and inspection and the like can be expected.

[0077] Furthermore, in the present embodiment, both the process of generating the learning model 5 and the process of using the generated learning model are performed by the information processing apparatus 1, but it is not limited thereto. The two processes may be separately performed by different apparatuses, or may be appropriately shared by three or more apparatuses. In addition, the following process may be performed: the substrate inspection apparatus 3 uses the generated learning model 5 to remove noise from the SEM image. In addition, as the substrate processed by the substrate processing apparatus and inspected by the substrate inspection apparatus 3, a semiconductor wafer has been exemplified, but the substrate is not limited to a semiconductor wafer, and may be various substrates such as a glass substrate. In addition, the photographing based on the substrate inspection apparatus 3 is not limited to SEM or TEM, and for example, photographing based on a CCD (Charge Coupled Device) or the like may also be performed. In addition, Figure 6 , Figure 7 and Figure 9 the SEM images are merely examples, and the images processed by the information processing apparatus 1 are not limited to the exemplified images, and may be arbitrary images.

[0078] <Examples of Variations Related to Generation of Learning Data>

[0079] In the above-described embodiment, learning data is generated by combining two SEM images including noise through multiple SEM images obtained by photographing the same part multiple times. In photographing such as EUV (Extreme Ultraviolet Lithography) exposure or TEM, it is sometimes difficult to photograph the same part multiple times. The information processing apparatus 1 according to a modified example generates learning data based on a plurality of noisy images (images including noise) obtained by photographing different parts.

[0080] The information processing apparatus 1 according to the modified example acquires, for example, a noisy image A obtained by photographing part A of a substrate and a noisy image B obtained by photographing part B from among a plurality of noisy images obtained by photographing. The information processing apparatus 1 performs image processing using an appropriate noise removal filter on the noisy images A and B respectively, and obtains clean images (images from which noise has been removed) A and B obtained by removing the noise of the noisy images A and B. The information processing apparatus 1 performs a deconvolution process using a group of one noisy image B and the clean image B, and extracts a noise component B included in the noisy image B. The information processing apparatus 1 can obtain a noisy image A' by performing a process of convolving the extracted noise component B on the other clean image A.

[0081] The original noisy image A and the noisy image A' obtained through the above-described process are images obtained by photographing the same part A of the substrate, but are images including different noise. The information processing apparatus 1 can use the image obtained by combining these two noisy images A and A' as learning data for machine learning of Noise2Noise. In addition, in the case of performing machine learning other than Noise2Noise, for example, the image obtained by combining the clean image A and the noisy image A' can also be used as learning data.

[0082] In addition, in the above example, noise is removed from the noisy image A to generate the clean image A, and the noise component B extracted from the noisy image B is overlapped on the clean image A, but is not limited thereto. For example, an image with less noise that can be processed as a clean image may be acquired at the stage of photographing, and the noise component B may be overlapped on the image as the above-described clean image A to generate a noisy image A'. In addition, the two images A and B may be images obtained by photographing different substrates instead of the same substrate. For example, for the clean image A (or a clean image A with less noise at the stage of photographing) obtained by removing noise from the noisy image A obtained by photographing the first substrate, the noise component B extracted from the noisy image B obtained by photographing the second substrate may be overlapped to generate a noisy image A'.

[0083] The information processing apparatus 1 according to a modified example of the above structure includes: a noise removal unit that removes noise from a noisy image to generate a clean image; a noise component extraction unit that extracts a noise component from the noisy image and the clean image by deconvolution; and a noise overlapping unit that convolves and overlaps the extracted noise component with another clean image. The information processing apparatus 1 acquires two noisy images obtained by photographing different parts, obtains two clean images through the noise removal unit, extracts a noise component based on one of the noisy image and the clean image, and overlaps the noise component with the other clean image to obtain another noisy image. The information processing apparatus 1 can combine the obtained noisy image with the original noisy image as learning data and perform machine learning of Noise2Noise. Alternatively, the information processing apparatus 1 may also combine the obtained noisy image with the clean image and use it as learning data.

[0084] In addition, the method for generating learning data by the information processing apparatus 1 according to the modified example can be applied to SEM images, TEM images, or various images including other noises.

[0085] The embodiments disclosed herein are illustrative in all aspects and should be considered non-limiting. The scope of the present disclosure is not represented by the above description but by the claims, and is intended to include all modifications within the meaning and scope equivalent to the claims.

[0086] The matters described in each embodiment can be combined with each other. In addition, the independent claims and dependent claims described in the claims can be combined with each other in all combinations regardless of the citation form. And, although the form of describing a claim that cites two or more other claims (multiple claim form) is used in the claims, it is not limited thereto. It may also be described in a form of describing a multiple claim (multiple claim citing multiple claims) that cites at least one multiple claim.

[0087] Description of Reference Numerals

[0088] 1... Information processing apparatus (computer); 3... Substrate inspection apparatus; 5... Learning model; 11... Processing unit; 11a... Information acquisition unit; 11b... Learning data generation unit; 11c... Model generation unit; 11d... Noise removal unit; 11e... Length measurement / inspection processing unit; 12... Storage unit; 12a... Program (computer program); 12b... Learning data storage unit; 12c... Model information storage unit; 13... Communication unit; 14... Display unit; 15... Operation unit.

Claims

1. A method for generating a learning model, An information processing apparatus performs the following processing: Obtain a plurality of images of an object substrate taken in time series; Generate learning data in which one of two images selected from the obtained plurality of images is used as an input and the other image is used as an output and they are corresponding to each other; and A learning model is generated by machine learning using the above learning data, where The learning model receives an image obtained by photographing the object substrate as an input and outputs an image with the noise removed from the image.

2. The method for generating a learning model according to claim 1, wherein The plurality of images include images under different conditions, The two images corresponding to each other as an input and an output in the learning data are images under the same conditions.

3. The method for generating a learning model according to claim 2, wherein The conditions include a pattern of the structure of the formation formed on the object substrate.

4. The method for generating a learning model according to claim 2, wherein Each obtained image is an image obtained by accumulating the results of multiple shootings, The conditions include the number of accumulations of each image.

5. The method for generating a learning model according to claim 2, wherein The obtained image is an image obtained by photographing the object substrate with a scanning electron microscope, The conditions include the scanning speed of the scanning electron microscope.

6. The method for generating a learning model according to claim 2, wherein The conditions include the resolution of the image obtained by photographing.

7. The method for generating a learning model according to claim 1, wherein Generate learning data in which one of the plurality of images is used as an input and the image after the one image in the shooting order in time series is used as an output and they are corresponding to each other.

8. The method for generating a learning model according to claim 7, wherein Extract multiple sets of two consecutive images in time series from the plurality of images and generate a plurality of the learning data.

9. The method for generating a learning model according to claim 1, wherein For the obtained plurality of images, correct the positions of the object substrate reflected in each image, Generate the learning data based on the corrected plurality of images.

10. The method for generating a learning model according to claim 9, wherein Select a reference image from the plurality of images, For each image other than the reference image, calculate the offset amount relative to the position of the reference image, Determine the shear size based on the maximum value of the calculated offset amount, Shear an image of the size from each image, Use the sheared image as the corrected image.

11. The method for generating a learning model according to claim 9, wherein The learning data is data in which two images each containing noise and corrected to make the positions of the object substrate reflected therein consistent are corresponding to an input and an output.

12. An information processing method, An information processing apparatus performs the following processing: Obtain an image obtained by photographing an object substrate; Input the acquired image into the learning model to obtain the denoised image output by the above learning model, where, The above learning model is a model obtained by machine learning in such a way that it takes an image obtained by photographing a target substrate as input and outputs an image with the noise of the image removed; and Output the acquired image, The above learning model is generated by machine learning using learning data, where the learning data is data established by taking one of two images selected from multiple images of a target substrate photographed in time series as input and the other image as output.

13. The information processing method according to claim 12, wherein Based on the image with noise removed, perform length measurement or inspection related to the above target substrate.

14. A computer program that causes a computer to execute the following processing: Acquire multiple images of a target substrate photographed in time series; Generate learning data established by taking one of two images selected from the acquired multiple images as input and the other image as output; and A learning model is generated by machine learning using the above learning data, where The above learning model takes an image obtained by photographing a target substrate as input and outputs an image with the noise of the image removed.

15. A computer program that causes a computer to execute the following processing: Acquire an image obtained by photographing a target substrate; Input the acquired image into the learning model to obtain the denoised image output by the above learning model, where The above learning model is a model obtained by machine learning in such a way that it takes an image obtained by photographing a target substrate as input and outputs an image with the noise of the image removed; and Output the acquired image, The above learning model is generated by machine learning using learning data, where the learning data is data established by taking one of two images selected from multiple images of a target substrate photographed in time series as input and the other image as output.

16. An information processing apparatus Comprising a processing unit, The above processing unit performs the following processing: Acquire multiple images of a target substrate photographed in time series; Generate learning data established by taking one of two images selected from the acquired multiple images as input and the other image as output; and A learning model is generated by machine learning using the above learning data, where The above learning model takes an image obtained by photographing a target substrate as input and outputs an image with the noise of the image removed.

17. An information processing apparatus, wherein Comprising a processing unit, The above processing unit acquires an image obtained by photographing a target substrate, inputs the acquired image to a learning model to obtain an image with noise removed output by the above learning model, and outputs the acquired image, where the learning model is a model obtained by machine learning in such a way that it takes an image obtained by photographing a target substrate as input and outputs an image with the noise of the image removed, The above learning model is generated by machine learning using learning data, where the learning data is data established by taking one of two images selected from multiple images of a target substrate photographed in time series as input and the other image as output.

Images