Method of forming a shape on a mask based on deep learning, and mask manufacturing method using the method of forming a shape on a mask

Abstract

Provided herein are a method of forming a mask, a method of accurately and quickly restoring an image on a mask to a shape on the mask, and a mask manufacturing method using the method of forming a mask. The method of forming a mask includes: obtaining first images by performing rasterization and image correction on shapes on a mask corresponding to a first pattern on a wafer; obtaining second images by performing transformation on the shapes on the mask; performing deep learning based on a transformation relationship between one of the first images and one of the second images, wherein the second image corresponds to the first image; and forming a target shape on the mask corresponding to a target pattern on the wafer based on the deep learning. The mask is manufactured based on the target shape on the mask.

Description

[0001] Cross-references to related applications

[0002] This application is based on and claims priority to Korean Patent Application No. 10-2020-0094795, filed with the Korean Intellectual Property Office on July 29, 2020, the disclosure of which is incorporated herein by reference in its entirety. Technical Field

[0003] The present invention relates to a mask manufacturing method, and more specifically to a method for forming a shape on a mask, and a mask manufacturing method using the method of forming a shape on a mask. Background Technology

[0004] In semiconductor manufacturing, photolithography processes using masks can be performed to form patterns on semiconductor substrates such as wafers. A mask can be a pattern transfer body, in which a patterned shape of opaque material is formed on a transparent substrate. In a brief description of the mask fabrication process, firstly, the required circuitry is designed, and the layout of that circuitry is planned. Then, mask design data obtained via optical proximity correction (OPC) is sent as Mask Strip Output (MTO) design data. Afterward, Mask Data Preparation (MDP) is performed based on the MTO design data, and front-end lines (FEOL) for processes such as exposure and back-end lines (BEOL) for processes such as defect testing are executed to fabricate the mask. Summary of the Invention

[0005] The present invention provides a method for forming a mask, a method for accurately and quickly restoring an image on a mask to the shape on the mask, and a mask manufacturing method using the method for forming a mask.

[0006] According to one aspect of the present invention, a method for forming a mask is provided, the method comprising: obtaining a first image by performing rasterization and image correction on a shape on the mask corresponding to a first pattern on a wafer; obtaining a second image by transforming the shape on the mask; performing deep learning based on a transformation relationship between a first image in the first image and a second image in the second image, wherein the second image corresponds to the first image; forming a target shape on the mask corresponding to a target pattern on a wafer based on the deep learning; and manufacturing the mask based on the target shape on the mask.

[0007] According to another aspect of the inventive concept, there is provided a method of forming a mask, the method including: designing initial shapes corresponding to first patterns on a wafer on the mask; obtaining first images having a bitmap format by performing rasterization and image correction on the initial shapes; obtaining signed distance images by applying signed distance calculation to the initial shapes corresponding to the first patterns on the wafer on the mask; performing deep learning on a transformation relationship between one of the first images and one of the signed distance images corresponding to the first image; transforming one of the initial shapes corresponding to a target pattern on the wafer on the mask into a first target image by performing rasterization and image correction on the initial shape on the mask; transforming the first target image into a signed target distance image based on the deep learning; transforming the signed target distance image into a target shape on the mask; and manufacturing the mask including the target shape on the mask.

[0008] According to another aspect of the inventive concept, there is provided a mask manufacturing method, the method including: obtaining first images by performing first rasterization and first image correction on shapes corresponding to first patterns on a wafer on a mask; obtaining second images by transforming the shapes on the mask; performing deep learning on a transformation relationship between one of the first images and one of the second images corresponding to the first image; forming a target shape corresponding to a target pattern on the wafer on the mask based on the deep learning; transmitting the target shape as a mask tape out (MTO) design data; preparing mask data based on the MTO design data; and performing exposure on a mask substrate of the mask based on the mask data. BRIEF DESCRIPTION OF DRAWINGS

[0009] Example embodiments of the inventive concept will be more clearly understood from the following detailed description taken in conjunction with the accompanying drawings, and by comparison with the claims, in which:

[0010] Figure 1 is a flowchart of a method of forming shapes on a mask according to some embodiments of the inventive concept;

[0011] Figure 2A and Figure 2B is a conceptual diagram and a simulation diagram for explaining principles of the method of forming shapes on a mask in Figure 1

[0012] Figure 3A and Figure 3B is a flowchart of an operation of forming a target shape on a mask, the operation including in the method of forming shapes on a mask in Figure 1

[0013] Figure 4 ​​is a conceptual diagram for illustrating Figure 3A and Figure 3B is a conceptual diagram of an operation of forming a target shape on a mask in

[0014] Figures 5A to 5C is a simulation diagram for showing Figure 1 the effect of a method of forming a shape on a mask in

[0015] Figure 6A and Figure 6B is a simulation diagram and a graph for showing Figure 1 the effect of a method of forming a shape on a mask in

[0016] Figure 7 is a flowchart of a mask manufacturing method according to some embodiments of the inventive concept. DETAILED DESCRIPTION

[0017] Embodiments will now be described more fully with reference to the accompanying drawings. In the drawings, like reference numerals can indicate similar elements, and repetitive descriptions can be omitted.

[0018] Figure 1 is a flowchart of a method of forming a shape on a mask according to some embodiments of the inventive concept.

[0019] Referring to Figure 1 , in a method of forming a shape on a mask, a first image for a shape on a mask is obtained through rasterization and image correction (S110). The shape on a mask (hereinafter referred to as a "mask shape") can refer to a pattern on a mask for an exposure process represented in a vector format (e.g., GDSⅡ (Graphic Data SystemⅡ) or OASIS (Open Artwork System Interchange Standard)) so as to form a desired pattern on a wafer, for example, a first pattern. The mask shape in a vector format (e.g., GDSⅡ / OASIS) can generally have a resolution of about 0.1 nm. For example, the mask shape can have a format of a resolution of about 0.05 nm or about 0.1 nm.

[0020] For reference, a pixel-based optical proximity correction (OPC) or inverse lithography technology (ILT) employed in a mask manufacturing process can include a formation or restoration technique of a mask shape. The formation technique of a mask shape can refer to a technique of performing optimization on an image on a mask (hereinafter referred to as a mask image) by using an ILT, and then reconstructing a mask shape from the mask image in a vector format such as GDSⅡ / OASIS without loss of accuracy. The optimization is a process of correcting a mask image to a mask image optimized for a desired target pattern, and is also referred to as an image correction process. Hereinafter, the image correction process is unified and used as optimization.

[0021] In the optimization of a mask image by using the ILT, the optimization is usually performed after pixelizing a mask shape stored in a format such as GDS II / OASIS and parameterizing the pixelized mask shape by transmittance to change the mask shape into a mask image having a bitmap format. As described above, a format such as GDS II / OASIS is a format having a resolution of 0.05 nm or 0.1 nm, and a mask shape represented in a format such as GDS II / OASIS can be one of: a value of 1 for a portion having a polygon, and a value of 0 for a portion not having a polygon. However, a pixelized mask image has a bitmap format, and can usually have a resolution of several nanometers or more due to a limitation of a storage capacity of a memory. Therefore, from a viewpoint of digital signal processing, the conversion of a mask shape having a format such as GDS II / OASIS into a bitmap format mask image can correspond to down-sampling.

[0022] In order to optimize or improve a mask image, it is very important to accurately calculate an intensity of light passing through a mask on a wafer. For this, it is necessary to accurately calculate a spatial frequency of a mask shape within a certain frequency band. Therefore, an aliasing effect due to a high spatial frequency can be eliminated by applying a low-pass filter to a mask image before down-sampling the mask shape. Finally, after applying a low-pass filter to a mask shape having a format such as GDS II / OASIS, a pixelized mask image can be obtained by down-sampling.

[0023] Thereafter, as described above, an initial mask image can be generated by parameterizing a pixelized mask image. A final mask image can be generated by performing optimization on the parameterized initial mask image. Both the initial mask image and the final mask image can have a bitmap format, similar to the pixelized mask image. The first image can correspond to the final mask image, and thus can have a bitmap format. For reference, the low-pass filtering, the down-sampling, and / or the parameterization can be referred to as rasterization.

[0024] After generating the final mask image, it is necessary to restore the final mask image into a mask shape such as having a GDS II / OASIS format. Such restoration can be very difficult or almost impossible. In other words, as described above, since a process of creating a bitmap format mask image from a mask shape such as having a GDS II / OASIS format employs "low-pass filtering + down-sampling", the process can be a non-reversible conversion process. Therefore, independent iterative optimization processes are respectively required to accurately restore a mask shape having a resolution of about 0.1 nm from a mask image having a resolution of about several nanometers or more. Such independent iterative optimization processes can be difficult to implement, and can not be practically applied.

[0025] In the existing ILT method of directly parameterizing transmittance, there is no method of quickly and accurately restoring the mask shape from the mask image, so it can not be possible to restore the precise curved shape. Therefore, the following method can be used: by guessing an initial condition from a rectangular shape, a simple Manhattan shape, or the like, and then changing the parameterization method based on transmittance to a parameterization method based on segments, to perform additional optimization or improvement. As a result, the following three problems can occur.

[0026] The first problem is that the ILT optimization performance can be reduced due to the inability to restore the curved shape. The second problem is that the difficulty of optimization can increase due to the use of a parameterization method based on segments for the mask shape. The third problem is that the restoration speed can be reduced due to the additional optimization.

[0027] In the mask shape forming method according to some embodiments of the inventive concept, the mask shape can be accurately and quickly restored by obtaining a second image capable of being reversibly transformed by a transformation algorithm to be described below, and performing deep learning on the transformation relationship between the first image and the second image. Therefore, the problems of the existing ILT method can be solved. The second image obtained by the transformation algorithm and deep learning will now be described in more detail.

[0028] After obtaining the first image, a second image capable of being reversibly transformed can be obtained by applying a transformation algorithm to the mask shape (S130). The transformation algorithm can be, for example, a Euclidean distance calculation or a signed distance calculation (SDC) algorithm. The second image can be a signed distance image (SDI) generated by the SDC algorithm. The signed distance image can correspond to a mask image having a certain bitmap format and capable of being reversibly transformed. Being capable of being reversibly transformed can refer to being easily and accurately transformed from the mask image to the mask shape by interpolation. For example, the signed distance image can be easily and accurately restored to the mask shape by zero-level set interpolation.

[0029] For reference, the SDC algorithm is an algorithm that calculates the absolute value of the shortest distance from the mask shape contour with each pixel value and a sign indicating the inside or outside of the mask shape. The signed distance image can refer to an image in which the calculated sign and the calculated absolute value are stored in each pixel value. Therefore, when the sign and the absolute value stored in the signed distance image are used, the mask shape can be accurately restored. A representative example of a method of restoring the mask shape from the signed distance image is a two-dimensional (2D) marching squares algorithm. For example, the aforementioned zero-level set interpolation can be based on the marching squares algorithm.

[0030] Although the SDC algorithm has been explained and described above as a transformation algorithm, the transformation algorithm is not limited to the SDC algorithm. For example, the transformation algorithm can include all types of algorithms capable of transforming a mask shape into a mask image in a bitmap format capable of reversible transformation.

[0031] The order of obtaining the first image and obtaining the second image can be reversed. The obtaining of the first image and the obtaining of the second image can be performed simultaneously. In addition, as will be described later with reference to Figure 2B The first image and the second image can be obtained from a known mask shape. However, the first image can be obtained from an unknown mask shape. When the first image is obtained from an unknown mask shape, the obtaining of the first image and the mask shape according to the first image can be determined through additional iterative optimization. The second image corresponding to the first image can be obtained by applying the transformation algorithm to the mask shape determined as described above.

[0032] After obtaining the second image, a transformation relationship between each first image and each second image corresponding to the first image can be performed through deep learning (S150). A transformation model between the first image and the second image can be inferred through deep learning. Deep learning can use, for example, a generative neural network algorithm.

[0033] When no answer is given, the generative neural network algorithm can include an unsupervised learning algorithm in which learning is performed. As a reference, a representative example of unsupervised learning is clustering, which refers to determining the characteristics of data and grouping similar data even when no answer is given. For example, a person A unfamiliar with animals can be shown animal pictures including a cat and a giraffe. Even when person A is not told which picture is a cat and which picture is a giraffe, person A can distinguish the animal with a spotted pattern and a long neck from other animals. In other words, unsupervised learning refers to determining a pattern by itself by learning the characteristics of data for which no answer is given.

[0034] A generative neural network enables estimation of a probability distribution having original data and enables an artificial neural network to create a probability distribution. Therefore, these neural networks are different from unsupervised learning based on simple clustering. To understand the generative neural network, the concept of a probability distribution needs to be understood because a plurality of data processed by the generative neural network is a random variable having a probability distribution. When an unknown quantity x in a second-order equation is a variable, the unknown quantity x becomes a specific number by solving the second-order equation. A random variable produces a different value in each measurement, but produces a number following a specific probability distribution. Therefore, knowing the probability distribution of a random variable can be the same as being able to understand the entire random variable (i.e., the entire data).

[0035] For example, when a probability distribution is known, a predicted expectation value of data, a dispersion of data, etc. can be directly determined, and thus a statistical characteristic of data can be directly analyzed. Further, when data is arbitrarily generated to follow a given probability distribution, the data can have a similar characteristic to original data used to obtain the probability distribution. In other words, when a probability distribution of data is modeled by an algorithm capable of unsupervised learning (e.g., a generative neural network), an infinite number of data accurately sharing the probability distribution with the original data can be re-generated.

[0036] For example, in the mask shape forming method according to some embodiments, a probability distribution of a transformation relationship between the first image and the second image can be modeled by learning the transformation relationship between the first image and the second image using deep learning of a generative neural network algorithm. Thus, the second image can be directly generated based on the first image and the modeled probability distribution. The probability distribution can correspond to a transformation model between the first image and the second image. In other words, by learning the transformation relationship between the first image and the second image using deep learning of a generative neural network algorithm, a transformation model is inferred and a second image corresponding to a new first image is directly generated by using the inferred transformation model. In addition, as the number of first images and the number of second images corresponding to the first images for deep learning increases, the inferred transformation model can more accurately represent the transformation relationship between the first image and the second image. Also, the second image corresponding to a new first image can be more accurately generated based on the inferred transformation model.

[0037] Although the generative neural network algorithm has been described above as an algorithm for deep learning, the algorithm for deep learning in the mask shape forming method according to some embodiments is not limited thereto. For example, according to some embodiments, various types of algorithms for inferring a transformation relationship model between the first image and the second image can be used in deep learning of the mask shape forming method.

[0038] After performing deep learning, a target shape corresponding to a target pattern on a wafer is formed on a mask based on the deep learning (S170). In detail, by applying the transformation model inferred through deep learning to a first target image corresponding to a target pattern on a wafer, a second target image is generated. The second target image is restored as a target shape on a mask, and thus the target shape on the mask can be quickly and accurately restored without additional iterative optimization. In addition, the target shape on the mask can include a curved shape. Thus, in the mask shape forming method according to some embodiments, a mask image corresponding to a target pattern on a wafer can be accurately and quickly restored as a mask shape through the above-described process. When the mask shape forming method according to some embodiments is applied to a mask shape forming process for a mask shape including a curved shape, the mask shape forming process can be performed more quickly and accurately than in the related art. Figure 3A and Figure 3BThe operation S170, which forms a target shape on a mask that corresponds to a target pattern on a wafer, will be described in more detail.

[0039] In a mask shape forming method according to some embodiments of the present invention, the target shape (i.e., mask shape) on the mask can be accurately and quickly recovered by: obtaining a first image through rasterization and optimization; obtaining a second image capable of reversible transformation through a transformation algorithm; inferring a transformation model of the transformation relationship between the first and second images through deep learning; and generating a second target image by applying the transformation model to a mask image corresponding to the target pattern. In other words, in a mask shape forming method according to some embodiments of the present invention, the mask shape can be accurately and quickly recovered from the mask image by inferring a model using deep learning, without additional iterative optimization. Therefore, the mask shape forming method according to some embodiments can solve all three problems that arise when using existing ILT methods that directly parameterize transmittance.

[0040] Figure 2A and Figure 2B It is used for explanation Figure 1 Conceptual diagrams and simulation diagrams of the principle of the mask shape formation method are presented in this paper. This paper briefly gives or omits related information. Figure 1 The same description Figure 2A and Figure 2B The description.

[0041] Reference Figure 2A When the left image is a mask shape such as one in GDSII / OASIS format, as shown by the right arrow above, a rasterized image can be generated through rasterization and optimization. The process of generating a rasterized image can correspond to... Figure 1 The operation S110 generates the first image. Therefore, the first image can correspond to a rasterized image. As described above, rasterization may include low-pass filtering, downsampling, and / or parameterization. Generating a rasterized image may include optimization. However, for ease of explanation, Figure 2A Low-pass filtering is only shown on the right-pointing arrow above.

[0042] As mentioned above, rasterized images have an irreversible bitmap format, and therefore it may be difficult or impossible to recover the mask shape from a rasterized image. Therefore, recovering the mask shape from a rasterized image may require additional iterative operations to optimize or improve the recovered mask shape.

[0043] As shown by the right arrow below, a signed distance image is generated from the mask shape using the SDC algorithm. The process of generating the signed distance image can be described as follows: Figure 1The operation S130 of generating a second image is performed. Accordingly, the second image can correspond to a signed distance image. As described above, the signed distance image can have a bit map format of reversible transformation, and thus the signed distance image can be easily and / or accurately restored to a mask shape.

[0044] For a new mask shape that is unknown, a signed distance image can not be generated, but only a rasterized image can be generated. In other words, when a new target pattern is to be formed on a wafer, a mask shape corresponding to the target pattern can not be directly determined, but only a final mask image corresponding to a rasterized image that is rasterized by rasterization and optimization can be generated. The rasterized image cannot be directly restored to a mask shape, and additional iterative optimization or improvement can be required to restore the mask shape.

[0045] Assuming that there is a transformation model between a rasterized image and a signed distance image, the rasterized image can be transformed or mapped to the signed distance image through the transformation model, and then the mask shape can be easily and accurately restored from the signed distance image. For example, the mask shape can be accurately restored by applying a zero level set interpolation to the signed distance image.

[0046] In a mask shape forming method according to some embodiments, deep learning can be performed on a transformation relationship between a rasterized image and a signed distance image corresponding thereto to infer a transformation model. The process of performing deep learning on a transformation relationship between a rasterized image and a signed distance image corresponding thereto can correspond to Figure 1 The operation S150 of performing deep learning on a transformation relationship between the first image and the second image is performed. Through this deep learning, a transformation model between a rasterized image and a signed distance image can be inferred. Although Figure 2A One rasterized image and one signed distance image are shown, but several hundred to several thousand or more rasterized images and several hundred to several thousand or more signed distance images corresponding thereto can be required for deep learning and a more accurate transformation model.

[0047] After the transformation model is inferred through deep learning, a rasterized image corresponding to a target pattern on a wafer can be transformed into a signed distance image by applying the transformation model, and then a mask shape can be restored from the signed distance image. The process of transforming a rasterized image into a signed distance image and restoring a mask shape from the signed distance image can correspond to Figure 1 The operation S170 of forming a target shape on a mask is performed.

[0048] Referring to Figure 2BAs described above, multiple first images and multiple second images can be needed to infer the transformation model by deep learning. For example, the first image on the top right can be generated by rasterization and optimization on a known mask shape (e.g., the simulated image on the left), and the second image on the bottom right can be generated by the SDC algorithm on the known mask shape. The mask shape can have a vector format such as GDS II / OASIS. The first image is a rasterized image and can have a bitmap format, and the second image is a signed distance image and can have a bitmap format. Since the mask shape is known, in some embodiments, the optimization can be omitted during the generation of the first image. As above with reference to Figure 2A Figure 2B Low-pass filtering is shown, but according to some embodiments, rasterization can include downsampling and parameterization in addition to low-pass filtering.

[0049] By this process, first images and second images can be generated for multiple known mask shapes and constitute independent image domains, respectively, and then deep learning (e.g., deep learning using a generative neural network algorithm) can be performed between the image domains, thereby inferring the transformation model between the first images and the second images corresponding thereto.

[0050] Figure 3A and Figure 3B is Figure 1 is a flowchart of the operation S170 of forming a target shape on a mask in Figure 4 is a conceptual diagram for explaining the operation of forming a target shape on a mask in Figure 3A and Figure 3B The description of Figures 1 to 2B is omitted herein briefly or identically with the description of Figure 3A and Figure 3B .

[0051] Referring to Figure 3A First, a mask shape corresponding to a target pattern on a wafer can be transformed into a first target image by rasterization and image correction (S172). The first target image can be, for example, a rasterized image and can have a bitmap format. As described above, the first target image can be an irreversible transformed mask image, and thus it can be impossible or can be very difficult to recover to the mask shape from the first target image.

[0052] Next, based on deep learning, the first target image can be transformed into a second target image (S174). In detail, the first target image can be directly transformed or mapped into the second target image by using a transformation model between the first image and the second image corresponding thereto, which has been described above with reference to Figure 1 ​The transformation relationship between the first image and the second image is inferred in operation S150 in which deep learning is performed. The second target image can correspond to, for example, a signed distance image, and can have a bitmap format.

[0053] Thereafter, the second target image can be transformed into a target pattern on a mask (S176). For example, when the second target image is a signed distance image, the second target image can be restored to a target shape on a mask by zero-level set interpolation. The target shape on a mask can eventually correspond to a mask shape such as a GDS II / OASIS format to form a target pattern on a wafer.

[0054] Referring to Figure 3B In describing operation S172 of transforming a mask shape into a first target image by rasterization and image correction in more detail, an initial shape on a mask corresponding to a target pattern on a wafer is designed (S171) before operation S172. Designing the initial shape on a mask can correspond to an initial process performed in manufacturing a mask for forming a new target pattern on a wafer.

[0055] Thereafter, the initial shape on a mask is low-pass filtered, down-sampled, and / or parameterized to generate a first initial image (S172a). Each of the processes of low-pass filtering, down-sampling, and / or parameterization is the same as those described above with reference to Figure 1 As described above, the set of low-pass filtering, down-sampling, and / or parameterization is referred to as rasterization.

[0056] After generating the first initial image, the first initial image is image corrected to generate a first target image (S172b). As described above, image correction can refer to optimization. The first target image can be a rasterized image, and can have a bitmap format. However, as described above, the first target image can be an irreversible transformed mask image, and thus it can be impossible, or can be very difficult, to restore from the first target image to the mask shape.

[0057] Referring to Figure 4 When describing operation S170 of forming a target shape on a mask, an initial shape on a mask corresponding to a target pattern on a wafer can be first designed. For example, when the target pattern on a wafer is a rectangle, the initial shape on a mask can also have a rectangular shape. Thereafter, the initial shape on a mask can be rasterized. Rasterization can include low-pass filtering, down-sampling, and / or parameterization. However, for ease of explanation, only low-pass filtering is shown. By this rasterization, an initial first image Ii can be generated.

[0058] After the first initial image Ii is generated, optimization is performed on the first initial image Ii. By the optimization, a first target image It1 can be generated. The first target image It1 is a rasterized image, and thus has a bitmap format, and can be a mask image that is not invertible. Successively, a transformation model inferred by deep learning is applied to the first target image It1. The first target image It1 can be mapped or transformed by applying the transformation model, and thus a second target image It2 can be generated. The second target image It2 is, for example, a signed distance image, and thus has a bitmap format, and can be a mask image that is invertible.

[0059] Thereafter, restoration can be performed on the second target image It2. By the restoration, a target shape S on a mask can be generated. For example, the target shape S on the mask can have a vector format such as GDS II / OASIS, and can be used in mask manufacturing to form a target pattern on a wafer.

[0060] Figures 5A to 5C is a simulation diagram for illustrating Figure 1 the effect of the mask shape forming method in Figure 5A is a simulation diagram of a rasterized image RI obtained by rasterizing and optimizing a mask shape, Figure 5B is a simulation diagram of a first signed distance image SDI1 obtained by applying a transformation model inferred by deep learning to the rasterized image RI, and Figure 5C is a simulation diagram of a second signed distance image SDI2 obtained by applying a SDC algorithm to the mask shape.

[0061] Referring to Figures 5A to 5C , the rasterized image RI, the first signed distance image SDI1, and the second signed distance image SDI2 can all have a bitmap format. However, the rasterized image RI can be a mask image that is not invertible to the mask shape, and the first signed distance image SDI1 and the second signed distance image SDI2 can be mask images that are invertible to the mask shape.

[0062] From Figure 5B and Figure 5C , it can be seen that the first signed distance image SDI1 is substantially the same as the second signed distance image SDI2, to a level that cannot be distinguished by the human eye. Now referring to Figure 6A and Figure 6B By comparing the first signed distance image SDI1 and the second signed distance image SDI2 with the mask shape restored therefrom at a more precise level, the effect of the mask shape forming method according to some embodiments will be described.

[0063] Figure 6A and Figure 6B are simulation graphs and a curve graph for showing effects of the mask shape forming method in Figure 1 . Figure 6A is a simulation graph of a mask shape recovered via interpolation by applying a marching squares algorithm to a signed distance image. In Figure 6A , a white image can correspond to a signed distance image, and a tiny dot around an edge of the white image can correspond to a mask shape recovered via interpolation.

[0064] Referring to Figure 6A , the points can include points corresponding to a first mask shape recovered via interpolation from a first signed distance image SDI1 and points corresponding to a second mask shape recovered via interpolation from a second signed distance image SDI2. However, in Figure 6A , the simulation graphs are shown in black and white, and the points corresponding to the first mask shape and the points corresponding to the first mask shape appear to overlap each other at substantially the same location, and thus are not distinguished from each other.

[0065] Figure 6B is a curve graph for comparing how much the first mask shape and the second mask shape differ from each other. Referring to Figure 6B , an edge placement error (EPE) representing how much the points differ in a normal direction of the mask shape can be calculated and checked. In Figure 6B , the x-axis represents the EPE in nm, and the y-axis represents the number of points having a corresponding EPE value. Figure 5B and Figure 5C Each of the first signed distance image SDI1 and the second signed distance image SDI2 can have a resolution of several nanometers (nm) or more, and each of the first mask shape and the second mask shape recovered therefrom can have a resolution of about 0.1 nm.

[0066] As can be seen from the graphs, the root mean square (RMS) EPE between the first mask shape and the second mask shape is about 0.05 nm, and the maximum EPE is about 0.26 nm. The difference reaching this level means that the first mask shape and the second mask shape are substantially the same as each other, and thus the mask shape forming method according to some embodiments is proven to be a very effective mask shape recovery method.

[0067] In other words, using Figure 5BThe mask shape can be easily and accurately recovered from the first-symbol distance image SDI1 in the rasterized image RI, and the first-symbol distance image SDI1 can be easily and quickly obtained by applying a transformation model inferred through deep learning to the rasterized image RI. Therefore, in the mask shape forming method according to some embodiments, the first-symbol distance image SDI1 is generated from the rasterized image RI based on deep learning, and the mask shape is recovered from the first-symbol distance image SDI1 via interpolation, thereby easily, quickly and accurately recovering the mask shape. In addition, the process of recovering the mask shape from the rasterized image RI may not require additional iterative optimization or improvement.

[0068] Figure 7 This is a flowchart of a mask manufacturing method according to some embodiments of the present invention. The details herein are briefly given or omitted. Figures 1 to 6B The same description Figure 7 The description.

[0069] Reference Figure 7 In a mask manufacturing method according to some embodiments, operations from obtaining a first image (S210) to forming a target shape on a mask (S270) are performed sequentially. The operations from obtaining the first image (S210) to forming the target shape on the mask (S270) are... Figure 1 Operation S110 to obtain the first image Figure 1 The operation S170, which forms the target shape on the mask, is the same. Therefore, a detailed description of each operation will be omitted.

[0070] Subsequently, the target shape is sent as Mask Belt Output (MTO) design data (S280). Typically, MTO can refer to requesting mask manufacturing by transferring data of the final mask shape obtained via OPC or ILT to the mask manufacturing team. Therefore, the MTO design data can ultimately correspond to data regarding the target shape on the mask. Since the target shape on the mask has a vector format such as GDS2Ⅱ / OASIS, the MTO design data may also have a vector format such as GDS2Ⅱ / OASIS. For reference, a vector format such as GDS2Ⅱ / OASIS can correspond to a graphic data format used in electronic design automation (EDA) software, etc.

[0071] After sending the MTO design data, mask data preparation (MDP) is performed (S285). MDP may include, for example, format transformations known as segmentation, barcode augmentation for mechanical reading, standard mask patterns for inspection, work platforms, etc., and automatic and manual verification. The work platform may refer to the creation of a text file relating to a series of instructions, such as layout information for multiple mask files, reference doses, and exposure speeds or methods.

[0072] The format conversion (i.e., segmentation) can refer to a process of segmenting the MTO design data of each region and changing a format of the segmented MTO design data to a format of an electron beam exposure device. The segmentation can include, for example, data processing (e.g., scaling), data size adjustment, data rotation, pattern reflection, or color inversion. During the conversion through the segmentation, data about many systematic errors that can be generated anywhere during the transfer from the design data to the image on the wafer can be corrected. The process of correcting the data about the systematic errors is referred to as mask process correction (MPC), and can include, for example, line width control referred to as CD control and an operation of improving pattern layout accuracy. Accordingly, the segmentation can contribute to improving the quality of the final mask, and can be a process that is performed in advance in order to implement the MPC. The systematic errors can be caused by distortions that occur in the exposure process, mask development and etching processes, and wafer imaging processes.

[0073] The MDP can include the MPC. The MPC refers to a process of correcting errors (i.e., systematic errors) that are generated during the exposure process as described above. The exposure process can be a concept including all of electron beam writing, development, etching, and baking. Data processing can be further performed before the exposure process. The data processing is a process of preprocessing the mask data, and thus can include syntax checking of the mask data, exposure time prediction, etc.

[0074] After the MDP, an exposure (S290) is performed on a substrate of the mask based on the mask data. The exposure can refer to, for example, electron beam writing. The electron beam writing can be performed according to, for example, a gray scale writing method using a multi-beam mask writer (MBMW). The electron beam writing can be performed using a variable shaped beam (VSB) exposure device.

[0075] After the MDP, a process of converting the mask data to pixel data can be performed before the exposure process. The pixel data is directly used for the actual exposure, and can include data about an image to be exposed and data about a dose allocated to the image. The data about the image can be bitmap data, which is converted from shape data as vector data through rasterization, etc.

[0076] After the exposure process, a series of processes can be performed to complete the mask. The series of processes can include, for example, development, etching, and cleaning. The series of processes for mask manufacturing can include a measurement process, a defect inspection, a defect repair process, etc. The series of processes for mask manufacturing can also include a pellicle coating process. The pellicle coating process refers to a process of attaching a pellicle to protect a mask surface from subsequent contamination during the mask delivery and the service life of the mask when no contaminants or chemical stains are confirmed through the final cleaning and inspection.

[0077] In the mask manufacturing method according to some embodiments, the mask can be an extreme ultraviolet (EUV) mask. However, the mask in the mask manufacturing method according to some embodiments is not limited to the EUV mask. For example, in the mask manufacturing method according to some embodiments, a mask for other wavelengths, for example, a deep ultraviolet (DUV) mask, can be manufactured.

[0078] In the mask manufacturing method according to some embodiments, the transformation model for the transformation relationship between the first image (e.g., the rasterized image) and the second image (e.g., the signed distance image) can be inferred by deep learning, and by using the transformation model, the mask shape can be easily, quickly, and accurately recovered from the mask image corresponding to the target pattern. For example, from the mask image to the mask shape, the design of the initial shape corresponding to the target pattern on the mask, the transformation through rasterization and optimization or other improvements to the first target image, the transformation to the second target image by using the transformation model, and the recovery of the second target image to the target shape (i.e., the mask shape) on the mask through interpolation.

[0079] In the mask manufacturing method according to some embodiments, the optimal mask data can be generated by transmitting the mask shape quickly and accurately recovered through the above process as MTO design data. In addition, based on the optimal mask data, the mask is manufactured through an exposure process, and a good mask capable of optimally forming a target pattern on a wafer can be implemented.

[0080] The various operations described herein (e.g., performing rasterization, image correction, transformation, deep learning, etc.) can be performed by a processor circuit including one or more processors. The processor circuit can include one or more data processing circuits such as general purpose processors and / or special purpose processors (e.g., microprocessors and / or digital signal processors). The processor circuit can include one or more processors embodied by hardware, software, firmware, microcode, etc. that support the operation of one or more processors. The processor circuit is configured to execute computer-readable program code in memory to perform at least some of the operations and methods described herein. According to some embodiments, the memory can include a non-transitory computer-readable storage medium having computer-readable program code stored therein, which can be executed by the processor circuit to perform various operations.

[0081] This document describes exemplary embodiments with reference to block diagrams and / or flowcharts illustrating computer-implemented methods, apparatus (systems and / or devices), and / or computer program products. It should be understood that the blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by computer program instructions executed by one or more computer circuits. These computer program instructions can be provided to processor circuitry of general-purpose computer circuitry, special-purpose computer circuitry, and / or other programmable data processing circuitry to generate a machine, such that the instructions, executed by the processor of the computer and / or other programmable data processing apparatus, transform and control transistors, values ​​stored in memory locations, and other hardware components within such circuitry to implement the functions / actions specified in the block diagrams and / or flowcharts, thereby creating functional bodies and / or structures for implementing the functions / actions specified in the block diagrams and / or flowcharts.

[0082] Although the inventive concept has been specifically shown and described with reference to embodiments thereof, it should be understood that various changes in form and detail may be made therein without departing from the scope of the appended claims.

Claims

1. A method of forming a mask, the method comprising: obtaining a first image by performing rasterization and image correction on shapes on the mask corresponding to a first pattern on a wafer; obtaining a second image by applying a transformation to the shapes on the mask, the transformation including a signed distance computation; performing deep learning based on a transformation relationship between the first image and the second image, wherein the second image corresponds to the first image; forming a target shape on the mask corresponding to a target pattern on the wafer based on the deep learning; and manufacturing the mask based on the target shape on the mask; wherein the image correction includes correcting the mask to a mask image corresponding to a desired target pattern.

2. The method of claim 1, wherein, forming the target shape on the mask includes: transforming one of the shapes on the mask corresponding to the target pattern on the wafer to a first target image by rasterization and image correction; transforming the first target image to a second target image based on the deep learning; and transforming the second target image to the target shape on the mask.

3. The method of claim 2, further comprising: designing an initial shape on the mask corresponding to the target pattern prior to transforming one of the shapes to the first target image, wherein transforming the shape to the first target image includes: generating a first initial image by low-pass filtering, down-sampling, and parameterization based on the initial shape; and generating the first target image by performing image correction on the first initial image.

4. The method of claim 2, wherein the rasterization includes low-pass filtering, down-sampling, and parameterization, and wherein no additional image correction is performed after transforming to the target shape.

5. The method of claim 2, wherein, the first target image has a bitmap format, and wherein the second target image has a bitmap format with interpolation applied.

6. The method of claim 5, wherein, the second target image includes a signed distance image.

7. The method of claim 1, wherein, the first image and the second image have a resolution of several nanometers (nm) or greater, and wherein the target shape has a resolution smaller than the resolution of the first image and the second image.

8. The method of claim 1, wherein, the target shape has a vector format and includes a curvilinear shape.

9. The method of claim 1, wherein, the deep learning uses a generative neural network algorithm.

10. The method of claim 1, wherein, the target shape is sent as a mask tape-out (MTO) design data and used for manufacturing of the mask.

11. A method of forming a mask, the method comprising: designing an initial shape on the mask corresponding to a first pattern on a wafer; obtaining a first image having a bitmap format by performing rasterization and image correction on the initial shape; obtaining a signed distance image by applying a signed distance computation to the initial shape on the mask corresponding to the first pattern on the wafer, wherein the signed distance computation includes a sign indicating an interior or an exterior of a mask shape; performing deep learning on a transformation relationship between the first image and the signed distance image, wherein the signed distance image corresponds to the first image; transforming an initial shape on the mask corresponding to a target pattern on the wafer to a first target image by performing rasterization and image correction on one of the initial shapes on the mask corresponding to the target pattern on the wafer; transforming the first target image to a signed target distance image based on the deep learning; transforming the signed target distance image to a target shape on the mask; and manufacturing a mask including the target shape on the mask.

12. The method of claim 11, the rasterization includes low pass filtering, down sampling, and parameterization, and wherein wherein no additional image correction is performed after transforming the signed target distance image to the target shape.

13. The method of claim 11, the signed target distance image has a bitmap format with interpolation applied, and wherein wherein the target shape has a vector format and includes a curved shape. the deep learning uses a generative neural network algorithm.

14. The method of claim 11, wherein, 15. A mask manufacturing method, comprising: obtaining a first image by performing first rasterization and first image correction on a shape on a mask corresponding to a first pattern on a wafer, wherein the image correction includes correcting the mask to a mask image corresponding to a desired pattern; obtaining a second image by applying a transformation to the shape on the mask; performing deep learning on a transformation relationship between the first image and the second image, wherein the second image corresponds to the first image; forming a target shape on the mask corresponding to a target pattern on the wafer based on the deep learning; sending the target shape as a mask tape out (MTO) design data; preparing mask data based on the MTO design data; and performing exposure on a mask substrate of the mask based on the mask data. forming the target shape on the mask includes:

16. The mask manufacturing method according to claim 15, wherein transforming one of the shapes on the mask corresponding to the target pattern on the wafer to a first target image by second rasterization and second image correction; transforming the first target image to a second target image based on the deep learning; and transforming the second target image to the target shape on the mask.

17. The mask manufacturing method of claim 16, the second rasterization includes low pass filtering, down sampling, and parameterization, and wherein, wherein no additional image correction is performed after transforming the second target image to the target shape.

18. The mask manufacturing method of claim 16, the first target image has a bitmap format, wherein wherein the second target image has a bitmap format with interpolation applied, wherein the transformation of the shape is a signed distance calculation, and wherein the second target image is a signed distance image. the target shape has a vector format and includes a curved shape.

19. The mask manufacturing method according to claim 15, wherein the deep learning uses a generative neural network algorithm.

20. The mask manufacturing method according to claim 15, wherein 19. A mask manufacturing method, comprising: obtaining a first image by performing first rasterization and first image correction on a shape on a mask corresponding to a first pattern on a wafer, wherein the image correction includes correcting the mask to a mask image corresponding to a desired pattern; obtaining a second image by applying a transformation to the shape on the mask; performing deep learning on a transformation relationship between the first image and the second image, wherein the second image corresponds to the first image; forming a target shape on the mask corresponding to a target pattern on the wafer based on the deep learning; sending the target shape as a mask tape out (MTO) design data; preparing mask data based on the MTO design data; and performing exposure on a mask substrate of the mask based on the mask data. forming the target shape on the mask includes: transforming one of the shapes on the mask corresponding to the target pattern on the wafer to a first target image by second rasterization and second image correction; transforming the first target image to a second target image based on the deep learning; and transforming the second target image to the target shape on the mask.

21. The mask manufacturing method of claim 19, the second rasterization includes low pass filtering, down sampling, and parameterization, and wherein no additional image correction is performed after transforming the second target image to the target shape.

22. The mask manufacturing method of claim 19, the first target image has a bitmap format, wherein the second target image has a bitmap format with interpolation applied, wherein the transformation of the shape is a signed distance calculation, and wherein the second target image is a signed distance image. the target shape has a vector format and includes a curved shape. the deep learning uses a generative neural network algorithm.

Images