Method for generating proximity correction model, proximity correction method and proximity correction system

Abstract

The invention provides a method executed by a computing system and used for generating a proximity correction model, a proximity correction method and a proximity correction system. The method for generating a proximity correction model includes: acquiring a plurality of captured images on a wafer after performing a first process using a first layout; generating an overlapping image obtained by overlapping the plurality of captured images; and performing machine learning on the proximity correction model using the image of the first layout and the overlay image.

Description

Technical Field

[0001] The present disclosure relates to a method for generating a proximity correction model, a method for performing a proximity correction process using the proximity correction model, and a proximity correction system. More specifically, the present disclosure relates to a method for generating a proximity correction model based on machine learning, applying an input layout image to the proximity correction model to predict an image of a pattern generated when the input layout is applied to a first process, and performing a proximity correction process using such a prediction. Background Art

[0002] As semiconductor processes become miniaturized, it becomes more difficult to accurately form a predetermined target pattern on a wafer. The target pattern is formed on a wafer by performing a photolithography process using a mask layout and performing an etching process using a photoresist pattern formed as a result of the photolithography process. In this process, the optical path deviation due to various process conditions called optical proximity effect (OPE) or process proximity effect (PPE) increases, which needs to be corrected.

[0003] Optical proximity correction (OPT) or process proximity correction (PPC) is performed to reduce the impact of OPE or PPE. OPC can be understood as a process of artificially adjusting a mask layout by comparing an expected photoresist pattern as a result of applying the mask layout to a photolithography process with a target photoresist pattern, while PPC can be understood as a process of artificially adjusting a photoresist pattern by comparing an expected etched pattern as a result of applying the photoresist pattern to an etching process with a target etched pattern.

[0004] A technology for generating an OPC model has been proposed, which uses machine learning to infer a photoresist pattern by receiving data about a mask layout. Incidentally, the learning data used in the machine learning process of the OPC model does not properly reflect the changes that occur in the process of measuring the photoresist pattern on the wafer. For example, due to the miniaturization of the above-mentioned semiconductor manufacturing process, the measurement data of the photoresist pattern is generated by repeatedly shooting microscope photography data on the FOV (field of view) on the integrated wafer. However, information about the changes that occur between each shooting is not properly reflected in the above-mentioned learning data. In other words, the learning data lacks information about the changes in the measurement results as the photoresist pattern, and as a result, an OPC model that is problematic in terms of robustness due to overfitting of the learning data alone is generated.

[0005] Similar to the above-mentioned OPC model, a technique for generating a PPC model that receives data about a photoresist pattern and uses machine learning to infer an etching pattern has been proposed. However, the learning data used in the machine learning process of the PPC model also does not properly reflect changes that occur in the process of measuring the etching pattern on the wafer. That is, the PPC model learned using learning data that does not reflect changes that exist in the process of measuring the etching pattern has problems in terms of robustness. Summary of the invention

[0006] The technical aspects that will be achieved through some embodiments of the present disclosure provide a method and system for generating a neighbor correction model that is robust to noise or changes that may be included in actual input data through a machine learning method by reflecting information about changes in measurement data included in the learning data during the machine learning process.

[0007] Another technical aspect to be achieved through some embodiments of the present disclosure provides an image-to-image deep learning method and a system thereof for reflecting information about changes in measurement data included in learning data during a machine learning process.

[0008] Another technical aspect to be achieved by some embodiments of the present disclosure provides a method and a processing system thereof for performing a proximity correction process such as OPC or PPC using a proximity correction model robust to noise or variations that may be included in data to be actually input.

[0009] However, aspects of the present disclosure are not limited to the aspects set forth herein. The above and other aspects of the present disclosure will become more apparent to those of ordinary skill in the art to which the present disclosure belongs by referring to the detailed description of the present disclosure given below.

[0010] According to one aspect of the present disclosure, a method for generating a proximity correction model performed by a computing system is provided. The method includes: acquiring a plurality of captured images of a wafer after performing a first process on the wafer using a first layout; generating an overlapped image obtained by overlapping the plurality of captured images; and performing machine learning on the proximity correction model based on the image of the first layout and the overlapped image.

[0011] According to other aspects of the present disclosure, a proximity correction method for a manufacturing process performed by a computing system is provided. The proximity correction method includes: inputting image data of a first layout to a proximity correction model that has undergone machine learning; using output data of the proximity correction model to generate a predicted image of a pattern formed after performing a first process using the first layout; and performing a proximity correction process based on a difference between the predicted image and a target image. The proximity correction model may be machine-learned using an overlapped image obtained by overlapping a plurality of captured images of the wafer after performing the first process on the wafer.

[0012] According to another aspect of the present disclosure, a proximity correction system is provided. The proximity correction system includes: a communication interface configured to receive image data of a first layout; a storage device storing executable data of a computer program; a memory; and a processor configured to run a computer program loaded into the memory. The computer program includes instructions that, when executed by the processor, cause the proximity correction system to: input the received image data of the first layout into the proximity correction model; use the output data of the proximity correction model to generate a predicted image of a pattern formed after performing a first process using the first layout; and perform a proximity correction process based on the difference between the predicted image and the target image. The proximity correction model may be a model that has been machine-learned using an overlapping image obtained by overlapping a plurality of captured images on a wafer after performing the first process. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The above and other aspects and features of the present disclosure will become more apparent by describing in detail example embodiments of the present disclosure with reference to the accompanying drawings, in which:

[0014] Figure 1 is a configuration diagram of a proximity correction system according to at least one embodiment of the present disclosure;

[0015] Figure 2 For reference only Figure 1 A conceptual diagram of the results of the operation of the described proximity correction system;

[0016] Figure 3 is a diagram for explaining removal of information on changes in measurement data according to the related art in a process of calculating representative data;

[0017] Figure 4 is a diagram for explaining maintaining information about changes in learning data according to some embodiments of the present disclosure;

[0018] Figure 5is a diagram for explaining an image processing process performed when generating learning data in some embodiments of the present disclosure;

[0019] Figure 6 is a flow chart of a method for generating a proximity correction model according to at least one embodiment of the present disclosure;

[0020] Figure 7 It is used to explain the reference in detail Figure 6 A detailed flow chart of some operations of the method described for generating a neighbor correction model;

[0021] Figure 8 is a diagram for conceptually illustrating a machine learning process that may be performed in some embodiments of the present disclosure; and

[0022] Fig. 9 is a hardware configuration diagram of a computing system for proximity correction processing according to at least one embodiment of the present disclosure. DETAILED DESCRIPTION

[0023] Hereinafter, example embodiments of the present disclosure will be described with reference to the accompanying drawings. By referring to the following detailed description and the accompanying drawings of the example embodiments, the advantages and features of the present disclosure and the methods for implementing the present disclosure can be more easily understood. However, the present disclosure can be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure will be thorough and complete, and will fully convey the concept of the present disclosure to those skilled in the art, and the present disclosure will be defined by the attached claims and their equivalents. When describing the present disclosure, if it is determined that the detailed description of the relevant known configuration or function may obscure the main points of the present disclosure, the detailed description will be omitted.

[0024] Unless the context clearly specifies singularity, the singular expressions used in the following embodiments include plural concepts. Additionally, unless the context clearly specifies a plurality, plural expressions include singular concepts. In addition, terms such as first, second, A, B, (a), (b) used in the following embodiments are only used to distinguish one element from another element, and these terms do not limit the nature, order or sequence of the related elements.

[0025] The elements described with reference to terms such as computing instances, computing systems, units, modules, blocks, ~ devices, ~ devices, etc. used in the present disclosure and the functional blocks shown in the drawings can be implemented with processing circuit systems such as software, hardware, or a combination thereof. For example, the processing circuit system can more specifically include, but is not limited to, a central processing unit (CPU), an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a system on a chip (SoC), a programmable logic unit, a microprocessor, an application specific integrated circuit (ASIC), etc. For example, software can be machine code, firmware, embedded code, application software, and / or the like. For example, hardware can include circuits, electronic circuits, processors, computers, integrated circuits, integrated circuit cores, passive components, combinations thereof, and / or the like.

[0026] The following will refer to Figures 1 to 5 The configuration and operation of a proximity correction system according to at least one embodiment of the present disclosure are described. Figure 1 As shown, the proximity correction system may include a proximity correction model trainer 100. In addition, in some embodiments, the proximity correction system may also include a service server 200 and at least one of an electron microscope (SEM) image storage device 10 and a target layout image storage device 20, the service server 200 serving a proximity correction process such as optical proximity correction (OPC) and / or process proximity compensation (PPC), the SEM image storage device 10 storing SEM images as measurement results about an object such as a wafer, and the target layout image storage device 20 storing images of a target layout for forming a desired pattern on a wafer.

[0027] In at least one embodiment, the proximity correction model trainer 100 may be composed of one or more computing devices. For example, the proximity correction model trainer 100 may be composed of one or more cloud computing instances, and / or the proximity correction model trainer 100 may be composed of computing instances included in (or including) one or more virtual machines and / or one or more containers.

[0028] In addition, the proximity correction model trainer 100 can be configured to include both a physical server (e.g., in the form of an on-premise instance) and a cloud computing instance. For example, considering the need to process data with high security requirements, a module that analyzes or stores data at least temporarily can be implemented on an on-premise physical server located on a local network that is blocked from accessing the Internet by a firewall, and other modules can be configured using a cloud computing instance.

[0029] In the present disclosure, the term "proximity correction" refers to a process for correcting the occurrence of optical path deviation due to various process conditions in the manufacturing process called optical proximity effect (OPE) and / or process proximity effect (PPE), and OPC, inverse lithography technology (ILT) and PPC can be considered as some examples of proximity correction.

[0030] In the following, OPC and PPC will be described as example proximity correction processes to facilitate understanding of the present disclosure. However, the proximity correction of the present disclosure should not be limited to proximity correction performed in a semiconductor manufacturing process. It should be understood that the proximity correction of the present disclosure generally also includes processes for performing corrections for optical proximity effects and / or process proximity effects in a manufacturing process in fields other than semiconductors (such as display panels).

[0031] The proximity correction model trainer 100 is configured to generate a proximity correction model using machine learning. For example, the proximity correction model can be understood as a model that receives layout data and is configured to infer (corresponding to the layout after executing a specific process) the data of the formed pattern as its output.

[0032] In some embodiments, the proximity correction model may be an image-to-image type model. For example, the proximity correction model may be a model that maps an input image to an output image. The data input to the proximity correction model may be an image of a layout, and the data to be output may be an image of a pattern formed corresponding to the layout. The proximity correction model may be an artificial neural network (ANT)-based model that is configured to learn (input layout image and pattern image) using an image of an actual pattern formed as a result of executing a specific process for an input layout, and when a new layout image is input, a pattern image formed as a result of the operation of a specific process is output.

[0033] In some embodiments, the neighbor correction model can be a model based on a convolutional neural network (CNN).For example, the neighbor correction model can be a model based on a deep convolutional generative adversarial network (DCGAN).

[0034] In some embodiments, the neighbor correction model can be a model based on a generative adversarial network (GAN). For example, the neighbor correction model can be configured to include a generator that is trained to generate a fake pattern image so that it is difficult for the discriminator to distinguish between a true pattern image and a fake pattern image generated by the generator. Figure 8 The architecture of the neighbor correction model is described in more detail.

[0035] In at least some embodiments, the service server 200 is configured to use the proximity correction model to provide a service for performing an OPC process or a PPC process to the user terminal 30. For example, the service server 200 may be part of a semiconductor design automation (EDA) server and / or may be a server that works in conjunction with a semiconductor design automation service server.

[0036] For example, the service server 200 may use the difference between the output image of the proximity correction model and the after-development inspection (ADI) target image to transmit data of one or more user interfaces for adjusting the layout of the photoresist pattern to the user terminal 30. At this time, the layout input to the proximity correction model may be the layout of the mask, and the proximity correction model may be a model for inferring the photoresist pattern developed on the wafer after performing a photoresist process using the layout of the mask.

[0037] In addition, for example, the service server 200 may provide one or more user interfaces for adjusting the layout of the photoresist pattern based on the difference between the output image of the proximity correction model and the post-cleaning inspection (ACI) target image to the user terminal 30. At this time, the layout (e.g., the input of the proximity correction model) may be and / or include the layout of the photoresist pattern, and the proximity correction model may be a model configured to infer the etching pattern formed on the wafer (e.g., after an etching process using the layout of the photoresist pattern is performed). Therefore, the user interface may receive a user input for indicating a layout to be input into the proximity correction model, which is configured to predict the etching pattern from the new photoresist pattern, thereby allowing the new photoresist pattern to be confirmed without producing the wafer, thereby reducing the cost for developing the new photoresist pattern. In addition, in at least one embodiment, using the proximity correction model, the user terminal 30 may be further configured to suspend the production of the wafer when the difference between the predicted etching pattern and the target pattern is greater than the manufacturing tolerance. Additionally, in at least some embodiments (as discussed in further detail below), a trained neighborhood correction model can be used to infer corrections for a photoresist pattern, and the inferred corrections can then be applied to new photoresists and the corrected etch patterns formed thereby.

[0038] Reference Figure 2 The operation result of the proximity correction system according to the present embodiment is described. The proximity correction system applies the image 40 of the mask layout specified from the user terminal 30 in the target layout image storage device 20 to the OPC model, thereby providing the user terminal 30 with data on the expected photoresist pattern 41 to be obtained when the photolithography process is performed using the mask layout. The user of the user terminal 30 can check the expected photoresist pattern 41 and can apply an input for adjusting the mask layout to the user terminal 30 accordingly.

[0039] In addition, the proximity correction system applies the image 41 of the photoresist pattern designated from the user terminal 30 to the PPC model, thereby providing the user terminal 30 with data on an expected etching pattern 42 to be obtained when an etching process is performed using the photoresist pattern.

[0040] The user of the user terminal 30 may check the intended etch pattern 42 and accordingly apply input for adjusting the photoresist pattern to the user terminal 30. Then, a pattern may be generated on the wafer based on the adjusted photoresist pattern.

[0041] The proximity correction model trainer 100 may receive data from the SEM image storage device 10 and the target layout image storage device 20 for machine learning of the proximity correction model, and may use the received data to configure the learning data. In at least some embodiments, such a learning data configuration task may be performed outside the proximity correction model trainer 100. In these cases, the proximity correction model trainer 100 is provided with a learning data set, and may use the provided learning data set to perform the learning task of the proximity correction model.

[0042] The image pair (layout image and pattern image) may be configured to configure the above-mentioned learning data. For example, the identifier of the layout image may be matched with the pattern image data stored in the SEM image storage device 10 to configure the image pair. In addition, in some embodiments, each of the layout image and the pattern image includes an execution identifier of a process that causes pattern formation, and the image pair (layout image and pattern image) may be configured based on such an execution identifier.

[0043] The proximity correction model trainer 100 may perform machine learning of the proximity correction model using the learning data. Machine learning may be performed until the proximity correction model reaches a target performance. For example, the proximity correction model trainer 100 may repeat machine learning on the proximity correction model until an error in a critical dimension (CD) of a pattern image inferred by the proximity correction model becomes less than a reference value.

[0044] In at least some embodiments, the neighbor correction model trainer 100 can use the functionality provided by machine learning operations (MLOps) to perform machine learning of the neighbor correction model. For example, the neighbor correction model trainer 100 can be a computing system configured to include one or more computing devices that perform computations related to machine learning.

[0045] In at least some embodiments, as the process is miniaturized, it may be necessary to perform multiple electron microscope shots (multiple shots) rather than just one electron microscope shot to accurately measure the pattern formed as a result of the process execution, and as a result, multiple shot images may be generated as measurement results for the same pattern. At this time, the pattern image of the learning data may include multiple shot images.

[0046] An average image of the plurality of captured images may be generated as a representative image of the plurality of captured images. In some embodiments, an image including an average value of the outline of a pattern extracted from each of the plurality of captured images may be generated as a representative image. In the machine learning process, the representative image may be used as a correct image of the layout image.

[0047] There may be variations between multiple captured images, and a representative image in the form of an average image may not effectively reflect such variations. Figure 3 This is explained. Figure 3 An example image 51 obtained by overlapping raw data of the original contour of a pattern extracted from each of a plurality of captured images is shown. The original contour may be a linear representation of the contour of the pattern included in the captured image with a predetermined (or otherwise determined) thickness. For example, the thickness may be understood as the minimum width value of the contour of the pattern to be expressed linearly. For example, the thickness may be 1 or 2. That is, the linear representation of the original contour may have a width of one pixel or two pixels.

[0048] The line end region 50 a included in the SEM image 50 on the measurement target region (FIV) of the pattern is a region where there may be a variation between a plurality of captured images. Figure 3 The example image 51 also shows that the variation of the original outline of each captured image significantly expands in the line end region 51 a .

[0049] In the prior art, in the machine learning of the neighbor correction model, an image of an average profile 52a obtained by calculating the average value of the pattern profile extracted from each of a plurality of captured images has been used as a representative image 52 of the pattern measured in the field of view (FOV). However, the representative image 52 does not include information that the variation of the line end region 50a is greater than that of other regions. Therefore, there is room for the neighbor correction model (trained using learning data including the representative image 52) to become overfitted to the learning data and thus not compensate for (or take into account) the profile inconsistency.

[0050] On the other hand, the proximity correction model trainer 100 of the present embodiment may perform machine learning of the proximity correction model using an overlapped image obtained by overlapping a plurality of captured images to reflect the above-mentioned changes in the learning of the proximity correction model.

[0051] In some embodiments, the neighbor correction model trainer 100 may perform a process of preparing a learning data set, and if so, the neighbor correction model trainer 100 may generate an overlap image. In some other embodiments, the neighbor correction system of the present disclosure may further include a data preparer (not shown) configured to provide a learning data set used by the neighbor correction model trainer 100. The learning data may be configured in the form of (layout image, overlap image, and representative image), and the data preparation part may generate an overlap image.

[0052] Reference Figure 4 Description Overlapping images. Figure 4 It also shows that Figure 3 The same SEM image 50 and example image 51 are shown in the example image 51. The line end region 50a and the line region 51b are shown in the example image 51. As mentioned above, in general, the variation in the line end region 50a may be greater than the variation in the line region 51b, and the variation in the line end region 50a may be greater than that in the line region 51b. Figure 4 The original contour of each of the plurality of captured images represented in the example image 51 also shows large variations.

[0053] The overlapping image may be obtained by overlapping a plurality of captured images. In the process of generating the overlapping image, the original outline of the pattern may be extracted from each of the plurality of captured images, a predetermined image processing may be performed on the image expressing the original outline, and the resulting images subjected to the predetermined image processing may be overlapped.

[0054] Figure 4 An example overlapping image 53 is shown.

[0055] exist Figure 4 , it can be seen that the width of the line end region 53a is expressed as wider than the width of the line region 53b, and this means that the variation of the line end region 53a is greater than the variation of the line region 53b. In addition to the line end region of the pattern, the corner region of the pattern or the convex region of the pattern can also be expressed as a region with lower measurement reliability than the line region.

[0056] In some embodiments, predetermined image processing may be performed so that the line end region 53a of the overlapped image 53 has a first sharpness indicating a first measurement confidence, and the line end region 53b of the overlapped image 53 has a second sharpness indicating a second measurement confidence. At this time, the higher the measurement confidence, the higher the sharpness of the overlapped image 53 may be.

[0057] In the process of overlapping a plurality of captured images, a task of summing the pixel values ​​of each captured image with each other may be performed. With this in mind, each pixel value may be suppressed to a predetermined (or otherwise determined) ratio by a predetermined image processing. For example, the predetermined image processing may include a task of uniformly adjusting each pixel value to a predetermined ratio.

[0058] For example, each pixel of the overlapping image 53 may have a value between 0 (black) and 255 (white). At this time, it should be understood that the background of the overlapping image 53 is black and the pixel value of the maximum sharpness is 255. When the predetermined ratio is 20%, the pixel at the position corresponding to the outline of the pattern and having the maximum sharpness pixel value of 255 may have a pixel value adjusted from 255 to 50 by the predetermined image processing. Even if the pixel value at the position corresponding to the outline of each captured image is suppressed by the above-mentioned image processing, the overlapping image 53 is generated as a result of overlapping a plurality of captured images. Therefore, the pixel value at the position corresponding to the outline of the overlapping image 53 will have a higher (brighter) value according to the change of the position and the ratio of the suppressed pixel value.

[0059] In some embodiments, the predetermined ratio may have a fixed value. In some other embodiments, the predetermined ratio may be a value determined corresponding to the number of shots. At this time, as the number of shots increases, the predetermined ratio may be determined to be lower. For example, in at least one embodiment, if the number of shots is 10, the predetermined ratio may be determined to be 10% (1 / 10), and if the number of shots is 50, the predetermined ratio may be determined to be 2% (1 / 50).

[0060] Because the pixel value of the position corresponding to the contour is suppressed compared to the pixel value of the maximum sharpness, the pixel value of the overlapped image is different when the contours of the plurality of captured images match and do not match. Therefore, the sharpness can be expressed differently. For example, in each case, the overlapped image 53 expresses the area in which the change between the respective captured images is high as a lower sharpness, and expresses the area in which the change is low as a higher sharpness.

[0061] That is, when a first region of the overlapping image 53 has a first change, a second region that does not overlap the first region has a second change, and the first change is greater than the second change, the sharpness of the first region will be lower than the sharpness of the second region. Again, on the overlapping image 53, the region with a smaller change will represent the pattern more clearly than the region with a larger change. The overlapping image 53 not only expresses information about the outline of the pattern in the form of the sharpness of the outline, but also expresses information about the change that may be different for each region of the pattern. Therefore, in terms of learning data, the overlapping image 53 is consistent with Figure 3 The representative image 52 includes more diverse types of information than the data.

[0062] In some other embodiments, the predetermined image processing may include performing dithering on the original outline. Dithering refers to an image processing technique used in computer graphics to create an illusion of color depth in an image on a system with a limited palette. In a dithered image, colors that are not available in the palette are approximated by diffusing colored pixels from within the available palette. In this case, the overlapped image may be understood as an image obtained by overlapping the dithered original outline of each of the multiple captured images. Reference will be made to Figure 5 Describe the process of generating overlapping images.

[0063] First, a first original contour 54a may be extracted from a first captured image included in a plurality of captured images as measurement results for the same FOV, and when a change occurs in the measurement for the FOV, a second original contour 54b different from the first original contour 54a may be extracted from a second captured image. Figure 5 , the width of the first original profile 54a is W1, whereas the width of the second original profile 54b is W2 which is different from W1.

[0064] Then, dithering can be performed on each of the first original contour 54a and the second original contour 54b to generate dithered contour band images 55a and 55b as a result. The original contour of the line shape can be modified into a smooth band shape by dithering. Next, an overlapped image 56 in which the contour band images 55a and 55b overlap can be generated.

[0065] although Figure 5 The overlap of two captured images is shown, but the example is not limited thereto, so it should be understood that the overlap image 56 is generated by performing the overlap of a plurality of captured images that may be greater than two captured images. Of course, in some embodiments, an abnormal captured image that fails to pass a predetermined standard among the plurality of captured images may be excluded from the overlap target.

[0066] Figure 5The overlapped image 56 includes a line region 56a with a small change and a line end region 56b with a large change. Figure 4 The example image 51 (which is generated as a result of direct overlap of the original contour) contains change information of a wide contour band area. In addition, due to the characteristics of the jitter, change information is not missed in the contour band area. According to this feature, the overlap image 56 is information including change information about the measurement pattern, and can be used as good learning data for reflecting changes in the machine learning process of the proximity correction model.

[0067] In some embodiments, the dithering of each captured image may be performed at the same level so that the overlapped image 56 may be used as learning data.

[0068] Additionally, in some other embodiments, the degree of jitter for each captured image can be adaptively determined based on the number of shots on the FOV. For example, the degree of jitter can be determined to be lower as the number of shots on the FOV increases. As the number of shots increases, the contour band area will cover a wide area on the overlapping image through changes between each shot. Taking this into account, by applying a lower degree of jitter as the number of shots increases, the contour band area of ​​each captured image on the overlapping image can be overlapped to an appropriate degree. In addition, the change of each position represented by the overlapping image can be made to have an appropriate range of values ​​accordingly.

[0069] Additionally, in some other embodiments, the quality index of the captured image can be used to dynamically determine the degree of jitter for each captured image. For example, since the degree of jitter is determined to be high when the quality index is low, the reliability of the original contour extracted from the low-quality captured image can be reduced and the band area of ​​the original contour can be widened. The quality index can be a value determined using metadata of the SEM image. For example, when the sharpness of the SEM image is high, the quality index can be determined to have a higher value.

[0070] The neighbor correction model trainer 100 can use the above-mentioned overlapping images to perform machine learning on the neighbor correction model. In the machine learning process, the aforementioned representative data is used as the correct image, and the overlapping image is used as reference data for determining the weight reflected in the loss data for error back propagation. Figure 8 The operation of the proximity correction model trainer 100 to perform machine learning using overlapping images is described in detail.

[0071] The configuration and operation of the proximity correction system according to the present embodiment have been described above. In the above disclosure, the proximity correction system according to the present embodiment can perform machine learning on a proximity correction model that infers a pattern formed as a result of performing a specific process using an input layout, and can use the proximity correction model that has undergone machine learning to provide a service for performing proximity correction processes such as OPC and PPC.

[0072] However, in some other embodiments, the proximity correction model may infer the final pattern formed as a result of performing two or more processes performed successively. Figure 2 Describe an example. Considering that the photoresist pattern 41 is an intermediate pattern for forming an etching pattern 42 on a wafer in a semiconductor manufacturing process, in some embodiments, the proximity correction system can provide a function that facilitates adjustment so as to form the target etching pattern 42. To this end, the proximity correction model trainer 100 receives an image of a mask layout, and can perform machine learning on an integrated model that infers an etching pattern formed by sequentially performing a photolithography process and an etching process using a mask layout. The mask layout image and the etching pattern image can be used as learning data for machine learning of the integrated model.

[0073] Already referenced above Figures 1 to 5 The proximity correction system according to the present embodiment is described. The method of operation of the proximity correction system of the present disclosure can be understood in more detail with reference to the other embodiments described below. In addition, even if not specified separately, the technical ideas that can be understood by the above-mentioned embodiments of the proximity correction system can be reflected in other embodiments described later.

[0074] In the following, reference will be made to Figures 6 to 8 A method for generating a proximity correction model according to at least one embodiment of the present disclosure is described. The method for generating a proximity correction model according to the present embodiment may be performed by one or more computing systems. In addition, in the method for generating a proximity correction model according to the present embodiment, some operations may be performed by a first computing device, while the remaining operations may be performed by a second computing device. For example, some operations of the method for generating a proximity correction model may be performed by a resident physical server, while the remaining operations may be performed by a cloud computing instance. In the following, when the execution entity of each operation is omitted, the execution entity may be understood as the computing system mentioned above.

[0075] Reference Figure 6The method for generating a proximity correction model according to the present embodiment is schematically described. The method for generating a proximity correction model according to the present embodiment can be understood as a method for generating a proximity correction model applied in, for example, a manufacturing process. Hereinafter, for ease of understanding, a method for generating a proximity correction model targeting a semiconductor manufacturing process will be described.

[0076] In operation S100, a plurality of captured images SEM of a target FOV on a wafer are acquired for machine learning of a proximity correction model. As mentioned above, the proximity correction model can be understood as a machine learning-based model configured to infer a pattern formed on a wafer after performing a first process (e.g., an etching process) using a first layout.

[0077] In operation S200, learning data is prepared using a plurality of captured images. The learning data may include a plurality of captured images and data of a first layout of a specific FOV for a specific wafer. For example, when the proximity correction model is an OPC model, the learning data may include an image of a mask layout and a plurality of captured images of a specific FOV of a photoresist pattern formed on a wafer as a result of performing a photolithography process using the mask layout. In addition, for example, when the proximity correction model is a PPC model, the learning data may include an image of a layout of a photoresist pattern and a plurality of captured images of a specific FOV of an etching pattern formed on a wafer as a result of performing an etching process using the photoresist pattern. In addition, the learning data may also include an overlapped image obtained by overlapping each of the plurality of captured images.

[0078] In operation S300, machine learning is performed on the neighbor correction model using the learning data. In the machine learning process, a representative image of a plurality of captured images included in the learning data is used as a correct image of the layout image, and an overlapping image included in the learning data can be used to determine a weight of a loss value based on a comparison result with the correct image.

[0079] In operation S400, a proximity correction process using a proximity correction model may be performed. Figure 1 Describes the operation of the service server and user terminals.

[0080] The following will refer to Figure 7 The learning data preparation process of operation S200 is described in detail.

[0081] In operation S210, a plurality of shot images of a specific FOV for measuring a target wafer are acquired. Next, the current shot image is set as the first shot image (S220). Additionally, the current overlapping image will be initialized to a predetermined background pixel value. Thereafter, operations S230 to S280 may be repeatedly performed for each of the plurality of shots.

[0082] In operation S230, the original contour of the current captured image is extracted. In operation S240, deformation parameters of the original contour image of the current captured image are determined. The deformation parameters may be parameters for performing a predetermined image processing on the original contour image. The deformation parameters may include an image processing identifier indicating the type of image processing and parameters of the image processing. For example, the deformation parameters may include an image processing identifier indicating a comprehensive suppression of pixel values, and may include a parameter indicating a ratio of the comprehensive suppression. Additionally, the deformation parameters may include an image processing identifier indicating jitter, and may include a parameter indicating a degree of jitter.

[0083] In some embodiments, a plurality of image processes may be sequentially applied to the original contour image. In this case, the deformation parameters may include both an identifier of each applied image process and a parameter of each applied image process.

[0084] In operation S250, the image of the original outline of the current photographed image is deformed using the deformation parameters. For example, in operation S250, the contour band image may be generated by performing a predetermined degree of dithering on the image of the original outline of the current photographed image.

[0085] In operation S260, an image resulting from deformation of the original contour image of operation S250 is overlapped with the current overlap image.

[0086] The operations from operation S230 to operation S260 will be repeated until the processing of all the plurality of acquired photographed images is completed (S270, S280).

[0087] Next, we will refer to Figure 8 The machine learning process of operation S300 is described in detail. The neighbor correction model 60 learned in operation S300 may be an artificial neural network model based on GAN. The neighbor correction model 60 may be configured to include a generator that is trained to make it difficult for the discriminator to distinguish between a true pattern image and a false pattern image generated by the generator. In addition, a plurality of convolutional layers may be used to train the neighbor correction model 60 based on a deep convolutional GAN ​​(DCGAN) configuring the generator. The learning method of the OPC model or PPC model based on DCGAN may refer to Korean Patent Publication No. 2023-0131715.

[0088] For ease of understanding, although Figure 8 A neighbor correction model 60 composed of a generator is shown, but in the machine learning process of the GAN technology, the generator and the discriminator will update each parameter through adversarial learning. The error back propagation process for updating the parameters will be described below.

[0089] First, the learning data will be described with respect to the proximity correction model 60 being an OPC model (however, the exemplary embodiment is not limited thereto). The learning data may include a mask layout image, an overlapped image obtained by overlapping each of a plurality of captured images, and a correct image as a representative image of the plurality of captured images. The correct image may be, for example, an average image of the original contour extracted from each of the plurality of captured images. The input layer 61 of the proximity correction model 60 as an OPC model receives an image 71 of a mask layout, and the output layer 62 outputs an image 72 of a photoresist pattern predicted as a result of performing a photolithography process using the mask layout.

[0090] The mask layout image 71a is input as an input image for learning the generator and discriminator constituting the proximity correction model 60, and the image 71b of the photoresist pattern inferred on the image of the mask layout can be output by the generator being learned. The difference in pixel level between the correct image 72a included in the learning data as the correct image of the mask layout image 71a and the result image 72b inferred by the generator is determined as loss data; so that in the present embodiment, the loss value constituting the loss data can be determined for each pixel. At this time, the weight 72c can be reflected in the pixel loss value.

[0091] The weight 72c can be determined using the overlapping image included in the learning data. The weight 72c can be determined using the pixel value of the overlapping image at the pixel position corresponding to the pixel loss value. At this time, the weight can be determined so that as the pixel value of the overlapping image approaches a value opposite to the background pixel value, the weight increases. The determination of the weight will be explained using the following example.

[0092] For example, if the pixel value of the first pixel on the correct image 72a is "100" and the pixel value of the first pixel on the inferred result image 72b is "102", the loss value of the first pixel becomes "2" corresponding to the difference between "102" and "100". In addition, the pixel value of the first pixel on the overlapping image is assumed to be "255" corresponding to the maximum sharpness. This may mean that the change of the first pixel of the correct image 72a is the minimum value, and the confidence of the pixel value of the first pixel of the correct image 72a is the maximum value. Therefore, the maximum value will be given as the weight 72c of the loss value of the first pixel.

[0093] As described above, a weighted reflection operation is performed on the loss value of each pixel, and the result 72 can be used to perform an error back-propagation process.

[0094] The neighbor correction model outputs an image that is as similar as possible to the correct image as a representative image of multiple captured images through the above-mentioned type of error back propagation process, and can be learned to be flexible for areas with large changes and low confidence among the areas of the correct image. Therefore, the method for generating a neighbor correction model according to the present embodiment provides an effect of being able to generate a neighbor correction model that is robust to changes in measurement results between multiple shots, and solves the performance problem of overfitting the correct image of the learning data and having poor accuracy for new input data.

[0095] Already referenced above Figures 6 to 8 The method for generating a proximity correction model according to the present embodiment is described. The proximity correction model generated by the method for generating a proximity correction model according to the present embodiment can be used to perform proximity correction of a manufacturing process. This will be described in detail below. For example, the proximity correction can be an OPC process or a PPC process.

[0096] A proximity correction method for a manufacturing process according to at least one embodiment of the present disclosure will be described. Figure 1 The described proximity correction system is used to implement this embodiment.

[0097] First, the image data of the first layout is input to the proximity correction model that has undergone machine learning. The computing system on which the proximity correction model is deployed can input the image data of the first layout to the proximity correction model.

[0098] The computing system then uses the output data of the proximity correction model to generate a predicted image of a pattern formed after performing the first process using the first layout.

[0099] Next, the proximity correction system performs a proximity correction process using the difference between the predicted image and the target image. For example, when an adjustment command of a first layout for reducing the difference between the target image and the predicted image is input from a user terminal, such a task can be repeated: the first layout is adjusted according to the adjustment command, and the computing system inputs the image data of the adjusted first layout into the proximity correction model, thereby checking the predicted image of the updated pattern.

[0100] At this time, the proximity correction model may be a model that performs machine learning using an overlapped image obtained by overlapping a plurality of captured images of the wafer after the first process is performed. As mentioned above, the proximity correction model is a model generated by performing error back propagation using a weighted result of loss data indicating a difference between a correct image that is a measurement result of a pattern formed as a result of performing the first process using the first layout and an image output by the proximity correction model, and the weight may be a value determined using the overlapped image.

[0101] For example, the proximity correction method of the manufacturing process according to the present embodiment may be a method for performing PPC. At this time, the first layout is a layout of a photoresist pattern, the first process may be an etching process, the pattern formed after the first process may be an etching pattern, and the target image may be an after-cleaning inspection (ACI) target image. At this time, performing the proximity correction process may include adjusting the layout of the photoresist pattern using the difference between the ACI target image and the predicted image of the etching pattern. Then, a pattern may be generated on a wafer based on the adjusted photoresist pattern.

[0102] Fig. 9 is a hardware configuration diagram of a computing system 1000 according to some embodiments of the present disclosure. In at least one embodiment, reference Figure 1 At least one of the described proximity correction model trainer 100 and / or the service server 200 may include Fig. 9 The computing system 1000 may include one or more processors 1100 , a system bus 1600 , a communication interface 1200 , a memory 1400 for loading a computer program 1500 to be executed by the processor 1100 , and a storage device 1300 for storing the computer program 1500 .

[0103] The processor 1100 is configured to control the overall operation of each component of the computing system 1000. The processor 1100 may perform calculations for at least one application or program for performing the method / operation according to various embodiments of the present disclosure. The memory 1400 is configured to store various types of data, instructions and / or information. The memory 1400 may load one or more computer programs 1500 from the storage device 1300 to perform the method / operation according to various embodiments of the present disclosure. The bus 1600 provides a communication function between the components of the computing system 1000.

[0104] The communication interface 1200 is configured to support communication with an external device using, for example, Internet communication, a wired local area network (LAN), a wireless local area network (WLAN) such as wireless fidelity (Wi-Fi), a wireless personal area network (WPAN) such as Bluetooth, a universal serial bus (USB), Zigbee, near field communication (NFC), radio frequency identification (RFID), power line communication (PLC), and a communication interface capable of connecting to a mobile cellular network such as a third generation (3G), a fourth generation (4G), a fifth generation (5G), long term evolution (LTE), and / or the like. For example, the communication interface 1200 may receive at least one of image data of a first layout and data of a plurality of captured images of a pattern formed as a result of performing a first process using the first layout.

[0105] The storage device 1300 may non-temporarily store one or more computer programs 1500. Additionally, the storage device 1300 may store learning data 1600 and executable data of the computer program.

[0106] The computer program 1500 may include one or more instructions for implementing the methods / operations according to various embodiments of the present disclosure therein. When the computer program 1500 is loaded into the memory 1400, the processor 1100 may execute the methods / operations according to various embodiments of the present disclosure by executing one or more instructions.

[0107] In some embodiments, reference Fig. 9 The computing system 100 described may be configured using one or more physical servers (such as virtual machines) included in a server farm based on cloud technology. Fig. 9 Among the components shown, at least some of the processor 1100 , the memory 1400 , and the storage 1300 may be virtual hardware, and the communication interface 1200 may also be configured as a virtualized networking element such as a virtual switch.

[0108] The computer program 1500 may include instructions for acquiring a plurality of captured images on a wafer after performing a first process using a first layout; instructions for generating an overlapped image by overlapping a plurality of captured images; and instructions for performing machine learning on a neighbor correction model using the image of the first layout and the overlapped image. In the neighbor correction model 1700 generated as a result of machine learning, at least some layers are stored in the memory 1400 and may be deployed to the storage device 1300 or an external computing device. It should be clearly understood that the computer program 1500 may be configured to include instructions for implementing methods that may be understood through some of the above-described embodiments.

[0109] Already referenced Figures 1 to 9Various embodiments of the present disclosure and effects according to the embodiments are described. Effects according to the technical concept of the present disclosure are not limited to the above-mentioned effects, and other effects not described can be clearly understood by those skilled in the art from the above description.

[0110] The technical ideas of the present disclosure described so far can be implemented as computer-readable codes on a computer-readable medium. The computer program recorded on the computer-readable recording medium can be sent to another computing device through a network such as the Internet, installed on another computing device, and thus used on another computing device.

[0111] Although operations are shown in a specific order in the drawings, it should be understood that, unless otherwise indicated and / or implied, operations are not limited to a specific order or sequential order and / or when all operations must be performed. For example, in some cases, multitasking and parallel processing may be advantageous. Although embodiments of the present disclosure have been described above with reference to the drawings, it should be understood by those skilled in the art that the present disclosure can be implemented in other specific forms without changing the technical concept or essential features. The above embodiments should be understood in all aspects to be illustrative rather than restrictive. The scope of protection of the present disclosure should be interpreted according to the attached claims, and all technical concepts within the equivalent scope should be interpreted as being included within the scope of rights of the technical concepts defined in the present disclosure.

Claims

1. A method for generating a proximity correction model performed by a computing system, the method comprising: acquiring a plurality of captured images of the wafer after performing a first process on the wafer using the first layout; generating an overlapped image obtained by overlapping the plurality of captured images; as well as Machine learning is performed on the proximity correction model using the images of the first layout and the overlapping images.

2. The method for generating a proximity correction model according to claim 1, wherein: The first region of the overlapped image has a first sharpness indicative of a first measurement confidence, The second region of the overlapped image has a second sharpness indicative of a second measurement confidence, The first area and the second area are areas that do not overlap each other, and The first sharpness is different from the second sharpness.

3. The method for generating a proximity correction model according to claim 2, wherein: The first area is a corner area of ​​a pattern included in the field of view of the wafer, The second area is a line area included in the field of view, The first sharpness is lower than the second sharpness, and The first measurement reliability is lower than the second measurement reliability.

4. The method for generating a proximity correction model according to claim 2, wherein: The first area is a line end area of ​​a pattern included in the field of view of the wafer, The second area is a line area included in the field of view, The first sharpness is lower than the second sharpness, and The first measurement reliability is lower than the second measurement reliability.

5. The method for generating a proximity correction model according to claim 2, wherein: The first region is a convex region of the pattern included in the field of view of the wafer, The second area is a line area included in the field of view, The first sharpness is lower than the second sharpness, and The first measurement reliability is lower than the second measurement reliability.

6. The method for generating a proximity correction model according to claim 1, wherein: The proximity correction model is an image-to-image type model configured to receive an image of the first layout and output a predicted image of a pattern formed corresponding to a shape of the first layout after performing the first process.

7. The method for generating a proximity correction model according to claim 1, wherein: Generating the overlapping image comprises: extracting a contour of each of the plurality of captured images; dithering the contour of each of the extracted plurality of captured images; and The overlapped image is generated by overlapping a shaken profile of each of the plurality of captured images.

8. The method for generating a proximity correction model according to claim 7, wherein: Dithering the extracted contours includes: The extracted contour is dithered to the same extent for each of the plurality of captured images.

9. The method for generating a proximity correction model according to claim 1, wherein: Performing machine learning involves: The machine learning is performed on the proximity correction model by further using a representative image representing the plurality of captured images as a correct image.

10. The method for generating a proximity correction model according to claim 9, wherein: The representative image is an image representing an average profile of the plurality of captured images.

11. The method for generating a proximity correction model according to claim 9, wherein: Performing the machine learning on the proximity correction model by further using the representative image representing the plurality of captured images includes: Parameters of the neighbor correction model are updated by performing error back propagation using loss data representing the difference between an image output by the neighbor correction model and the representative image.

12. The method for generating a proximity correction model according to claim 11, in, The loss data includes pixel loss values ​​in pixel units, and Wherein, performing the error back propagation includes: Determining a weight for each of the pixel loss values ​​using the overlapped image; and Back propagation is performed to update parameters of the neighbor correction model using per-pixel loss values ​​reflecting the determined weights.

13. The method for generating a proximity correction model according to claim 12, wherein: Using the overlapped image to determine the weight of each pixel loss value includes: The weights are determined using pixel values ​​of the overlapping images for pixel positions corresponding to the pixel loss values.

14. The method for generating a proximity correction model according to claim 13, wherein: Using the pixel values ​​of the overlapping images to determine the weights includes: The weight is determined such that the weight increases as the pixel value of the overlapping image approaches a value opposite to a pixel value of a background pixel.

15. The method for generating a proximity correction model according to claim 1, wherein: The first layout is a layout of a photoresist pattern, The first process is an etching process, The plurality of captured images on the wafer are a plurality of images of etching patterns included in a field of view area of ​​the wafer, The etching pattern is a pattern formed on the wafer by the etching process, and The method further includes adjusting a layout of the photoresist pattern by performing a process proximity correction process that utilizes a difference between an output image of the proximity correction model and a post-cleaning inspection target image.

16. The method for generating a proximity correction model according to claim 1, wherein: The first layout is a mask layout, The first process is a photolithography process, The plurality of captured images of the wafer are a plurality of images of a photoresist pattern included in a field of view region of the wafer, The photoresist pattern is a pattern developed on the wafer by the photolithography process, and The method further includes adjusting a layout of the photoresist pattern by performing an optical proximity correction process using a difference between an output image of the proximity correction model and a post-development inspection target image.

17. A proximity correction method for a manufacturing process performed by a computing system, the proximity correction method comprising: Inputting the image data of the first layout into the proximity correction model subjected to machine learning; generating a predicted image of a pattern formed after performing a first process using the first layout using output data of the proximity correction model; as well as performing a proximity correction process based on the difference between the predicted image and the target image, The proximity correction model is machine-learned using an overlapped image obtained by overlapping a plurality of captured images of the wafer after performing the first process on the wafer.

18. The proximity correction method for a manufacturing process according to claim 17, wherein: The neighbor correction model is generated by error back propagation using a weighted result of loss data indicating the difference between a correct image and an image output by the neighbor correction model, The correct image includes a measurement result of a pattern formed as a result of performing the first process using the first layout, and The weight is a value determined using the overlapping image.

19. The proximity correction method for a manufacturing process according to claim 17, wherein: The first layout is a layout of a photoresist pattern, The first process is an etching process, The pattern formed after performing the first process is an etched pattern, The target image is a post-cleaning inspection target image, and Performing the proximity correction process includes adjusting a layout of the photoresist pattern based on a difference between the post-cleaning inspection target image and a predicted image of the etched pattern.

20. A proximity correction system, comprising: a communication interface configured to receive image data of a first layout; A storage device storing executable data of a computer program; Memory; as well as a processor configured to execute a computer program loaded into the memory, wherein the computer program includes instructions that, when executed by the processor, cause the proximity correction system to: input the received image data of the first layout into the proximity correction model, use output data of the proximity correction model to generate a predicted image of a pattern formed after performing a first process using the first layout, and perform a proximity correction process based on a difference between the predicted image and a target image, The proximity correction model is machine-learned using an overlapping image obtained by overlapping a plurality of captured images on a wafer after performing the first process.

Images