Graph modeling method and device, storage medium and computer equipment

Abstract

The invention provides a graph modeling method and device, a storage medium and computer equipment, and the method comprises the steps: carrying out the feature coding of each contour segment corresponding to a contour curve in a photoetching layout through a first neural network, and obtaining a segment structure feature vector corresponding to each contour segment; obtaining each arc length position parameter corresponding to the contour section; the arc length position parameter is used for indicating the corresponding arc length percentage of each sampling point in the contour section in the contour section; generating a sampling position vector corresponding to each sampling point by using a second neural network according to the arc length position parameter; determining a target coordinate corresponding to each sampling point in the contour segment by using a third neural network according to the segment structure feature vector and each sampling position vector; and according to each target coordinate corresponding to each contour section, determining a modeled curve corresponding to the contour curve.

Description

Technical Field

[0001] This application relates to the field of integrated circuit technology, and more specifically, to a graphical modeling method, apparatus, storage medium, and computer device. Background Technology

[0002] In the process of integrated circuit manufacturing and simulation, it is usually necessary to simulate the lithographic contours in the fabricated process lithography layout to simulate the contour reconstruction results. However, in the Open Artwork System Interchange Standard (OAS) data of lithography layouts, a closed lithographic contour is usually composed of hundreds to thousands of continuous control points, and the total number of control points in a micrometer-level local region can reach tens of thousands. The data structure is dense and non-redundant, and under the precision requirements of circuit structure, the subtle local structural changes between control points cannot be ignored. Therefore, extremely high requirements are placed on contour reconstruction.

[0003] Current contour modeling methods mainly focus on the structural or image representation of the design end, which is difficult to meet the accuracy and structural expression capabilities required for contour modeling after photolithography, resulting in poor contour reconstruction results. Summary of the Invention

[0004] In view of this, this application provides a graphical modeling method, apparatus, storage medium, and computer device to construct a contour expression mechanism that simultaneously possesses high-fidelity expression, continuous and controllable reconstruction, and structural semantic carrying capacity, thereby improving the accuracy and flexibility of contour modeling.

[0005] Specifically, this application is implemented through the following technical solution: In a first aspect, embodiments of this disclosure provide a graphical modeling method, including: Using a first neural network, feature encoding is performed on each contour segment corresponding to the contour curve in the photolithography pattern to obtain the segment structure feature vector corresponding to each contour segment. Obtain the arc length position parameters corresponding to each contour segment; the arc length position parameters are used to indicate the percentage of arc length corresponding to each sampling point in the contour segment. Using a second neural network, a sampling position vector corresponding to each sampling point is generated based on the arc length position parameter; Using a third neural network, the target coordinates corresponding to each sampling point in the contour segment are determined based on the segment structure feature vector and each sampling position vector; Based on the target coordinates corresponding to each contour segment, the modeled curve corresponding to the contour curve is determined.

[0006] In one possible implementation, the contour segment is obtained by dividing it according to the following steps: Based on the original coordinates of each position point on the contour curve, the position points are sorted in a consistent direction according to a preset direction, so that the sorted contour curve follows the left-hand or right-hand rule, and the sorted position points are obtained. Based on the target number of points, the sorted position points are divided into segments to obtain the contour segments corresponding to the contour curve.

[0007] In one possible implementation, the step of segmenting the sorted position points according to the target number to obtain the contour segments corresponding to the contour curve includes: If the number of position points in the last segment is less than the target number of points, then the number of position points in the last segment is supplemented to the target number of points using the end position points of the contour curve or preset completion points to obtain the last contour segment.

[0008] In one possible implementation, obtaining the position parameters of each arc length corresponding to the contour segment includes: The position parameters of each arc length corresponding to the contour segment are generated by using a uniform sampling method; or, the position parameters of each arc length corresponding to the contour segment are generated by using a random sampling method; or, based on the original coordinates of each position point on the contour segment, the percentage of each position point relative to the total length of the contour segment is determined, and the percentage corresponding to each position point is used as the position parameters of each arc length.

[0009] In one possible implementation, the step of using a first neural network to perform feature encoding on each contour segment corresponding to the contour curve in the photolithography pattern to obtain a segment structure feature vector corresponding to each contour segment includes: For any of the contour segments, the first neural network is used to construct the initial feature tensor corresponding to the contour segment based on the original coordinates of each position point in the contour segment and the number of position points. Based on the mask features corresponding to the state identifier information of each location point, the initial feature tensor is subjected to feature mapping processing to obtain an intermediate feature tensor; the state identifier information is used to indicate whether the location point is a completion location point. The intermediate feature tensor is subjected to feature compression processing to obtain the segment structure feature vector.

[0010] In one possible implementation, constructing the initial feature tensor corresponding to the contour segment based on the original coordinates of each position point in the contour segment and the number of position points includes: The original coordinates of each position point in the contour segment are normalized, and the normalized coordinates are encoded based on the number of position points to obtain a coordinate encoding tensor. Based on the original coordinates of each position point in the contour segment and the total length of the contour segment, determine the arc length scalar of each position point, and based on the number of position points and the arc length scalar, determine the arc length encoding tensor. The original coordinates of each point in the contour segment are encoded to obtain a position encoding tensor. The coordinate encoding tensor, the arc length encoding tensor, and the position encoding tensor are concatenated to obtain the initial feature tensor.

[0011] In one possible implementation, the step of performing feature compression processing on the intermediate feature tensor to obtain the segment structure feature vector includes: The intermediate feature vector is subjected to feature extraction processing to obtain the target feature tensor; Based on the status identifier information of each location point, determine the effective mask for each location point; Using the effective mask, the target feature tensor is compressed to obtain the segment structure feature vector.

[0012] In one possible implementation, generating the sampling position vector corresponding to each sampling point using a second neural network based on the arc length position parameter includes: Using the second neural network, feature encoding and feature extraction are performed on the arc length position parameter to obtain the sampling position vector corresponding to each sampling point.

[0013] In one possible implementation, determining the target coordinates corresponding to each sampling point in the contour segment using a third neural network based on the segment structure feature vector and each of the sampling position vectors includes: Using the third neural network, feature dimension alignment and feature concatenation are performed on the segment structure feature vector and the sampling position vector to obtain a concatenated feature vector; The spliced ​​feature vector is subjected to feature decoding processing to obtain the target coordinates corresponding to each sampling point in the contour segment.

[0014] In one possible implementation, the first neural network, the second neural network, and the third neural network are trained through the following steps: Obtain the sample arc length position parameters and label coordinates corresponding to each sample control point in the sample contour segment; Using the first neural network to be trained, the feature encoding of the sample contour segment is performed according to the label coordinates of each sample control point to obtain the predicted segment structure feature vector corresponding to the sample contour segment. Using the second neural network to be trained, a predicted sampling position vector corresponding to each sample control point is generated based on the sample arc length position parameter. Using the third neural network to be trained, the predicted coordinates and predicted arc length position parameters corresponding to each of the sample control points are determined based on the predicted segment structure feature vector and the predicted sampling position vector. A first network loss is determined based on the predicted coordinates and the label coordinates, and a second network loss is determined based on the sample arc length position parameters and the predicted arc length position parameters. The first neural network, the second neural network, and the third neural network are jointly trained using the first network loss and the second network loss to obtain the trained first neural network, the second neural network, and the third neural network.

[0015] Secondly, embodiments of this disclosure also provide a graphical modeling apparatus, the apparatus comprising: The encoding module is used to use the first neural network to encode the features of each contour segment corresponding to the contour curve in the photolithography pattern, and obtain the segment structure feature vector corresponding to each contour segment. The acquisition module is used to acquire the arc length position parameters corresponding to each contour segment; the arc length position parameters are used to indicate the percentage of arc length corresponding to each sampling point in the contour segment. The generation module is used to generate sampling position vectors corresponding to each sampling point based on the arc length position parameters using a second neural network. The first determining module is used to determine the target coordinates corresponding to each sampling point in the contour segment based on the segment structure feature vector and each sampling position vector using a third neural network. The second determining module is used to determine the modeled curve corresponding to the contour curve based on the target coordinates corresponding to each contour segment.

[0016] Thirdly, an optional implementation of this disclosure also provides a computer-readable storage medium having a computer program stored thereon, which, when run, implements the steps of the first aspect above, or any possible implementation of the first aspect.

[0017] Fourthly, an optional implementation of this disclosure also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, performs the steps of the first aspect above, or any possible implementation of the first aspect.

[0018] The graphic modeling method, apparatus, storage medium, and computer device provided in this disclosure utilize a first neural network to encode the features of each contour segment corresponding to the contour curve, extracting the local geometric and sequential features of the contour segments to obtain segment structure feature vectors representing the contour segment structure. A second neural network generates sampling position vectors corresponding to each sampling point based on arc length position parameters under different distributions, enabling the expression of sampling point positions under various sampling styles, thereby improving the diversity and flexibility of path segment structure expression. A third neural network generates control point coordinates through segment structure feature vectors and each sampling position vector, fully considering the overall structure of the contour segment and the point structure of each sampling point, thereby improving the accuracy of the generated control point coordinates (i.e., target coordinates). Furthermore, the target coordinates corresponding to the sampling points are used to remodel the contour curve, improving the consistency between the modeling result and the original contour curve, thus meeting the accuracy and structural expression capabilities required for contour modeling after photolithography, and improving the accuracy of the modeling result.

[0019] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a graphical modeling method according to an exemplary embodiment of this application; Figure 2a This is a schematic diagram of various closed contour curves in a photolithographic pattern according to an exemplary embodiment of this application; Figure 2b This is an exemplary embodiment of the present application illustrating a contour segment where the number of position points contained in the contour curve is less than the number of points required as input to the first neural network. Figure 3 This is a schematic diagram illustrating a contour curve and a modeled curve according to an exemplary embodiment of this application; Figure 4 This is a schematic diagram illustrating another contour curve and a modeled curve as shown in an exemplary embodiment of this application; Figure 5 This is a schematic diagram illustrating the image modeling comparison results under a sampling method according to an exemplary embodiment of this application; Figure 6This is a schematic diagram illustrating the image modeling comparison results under another sampling method as shown in an exemplary embodiment of this application; Figure 7 This is a schematic diagram of a graphical modeling apparatus shown in an exemplary embodiment of this application; Figure 8 This is a schematic diagram of the structure of a computer device shown in an exemplary embodiment of this application. Detailed Implementation

[0021] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0022] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0023] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0024] Research has shown that in integrated circuit manufacturing and simulation, compressing a lithographic profile into a high-quality, structure-aware latent vector can not only achieve accurate profile reconstruction but also serve a wide range of tasks, including: generation-oriented profile style transfer, graphic transformation and layout enhancement (such as converting right-angled regions in a lithographic profile into rounded regions), and index-oriented graphic profile clustering and similar structure retrieval (such as using latent vectors of different lithographic profiles to determine profile similarity and clustering similar profiles together). These tasks place unified and stringent requirements on the latent vector's representation accuracy, structural reproducibility, and sampling flexibility, necessitating high-fidelity geometric compression capabilities to establish a usable intermediary between model compression and structure awareness. However, achieving such high-quality latent vector representation in lithographic profile data faces the following challenges: 1. Extremely high data density: In the OAS data of a lithographic layout, a closed profile is usually composed of hundreds to thousands of continuous control points. The total number of control points in a micrometer-level local area can reach tens of thousands, resulting in a dense data structure without redundancy. 2. No simplification allowed: The local structural changes between control points on the profile curve are subtle but significant. Directly downsampling the profile curve can easily destroy high-frequency features, leading to misjudgment of topological structure or distortion of geometric semantics, such as misjudging sharp corners as rounded corners or excessive smoothing of local jagged edges. 3. Requirements for expressing closure and internal / external relationships: The lithographic profile usually represents the boundary of circuit structure wires or the boundary of functional areas. Whether it is closed, its direction and order, and other attributes directly determine its physical meaning. Topological relationships and geometric closed-loop properties must be preserved.

[0025] To address the aforementioned challenges, constructing a contour representation mechanism that simultaneously possesses high-fidelity representation, continuous and controllable reconstruction, and structural semantic carrying capabilities has become a key technical issue in current graphical modeling. However, in existing technologies, for the integrated circuit (IC) field, the representation of lithographic contours mainly focuses on the design-end structure (layout) or image representation, which is insufficient to meet the accuracy and structural representation capabilities required for contour modeling after lithography. Design-end structure (layout) representation can be understood as representing regular contour curves, while image representation can be understood as converting both the lithographic layout and contour curves into images, and then representing the lithographic contour through image processing. Some common representative methods include the Squish Pattern method, rasterized image representation methods, and sequence representation methods.

[0026] The Squish Pattern method divides the entire design layout into fixed-size pattern cells. A topological matrix records the topological occupancy and geometric information of each pattern cell, while a geometric matrix records how the pattern cells were divided. This method is typically suitable for density and pattern modeling of regular layouts. However, it suffers from coarse-grained representation, failing to capture continuous changes in curve contours and lacking the ability to describe the "curved" and "dense" curves in real lithographic patterns. Therefore, it cannot meet the representation requirements of post-lithographic patterns.

[0027] Rasterized image representation methods extract each contour curve from the lithographic pattern individually, and different contour curves are modeled using graph structures. Each contour curve can be regarded as an image, and then modeling is performed using an image processing model. However, this method is heavily dependent on pixel resolution, suffers from interpolation errors and boundary blurring problems in high-density structures, cannot achieve continuous contour structure restoration, and cannot meet the coordinate-level accuracy requirements of lithographic patterns.

[0028] Sequential representation method: The entire lithography pattern is considered as a "segment," and each contour curve is considered as a "sentence" within that "segment." Each sentence is described using coordinates and direction markers to create a sequential representation. Finally, a custom Transformer (such as LayoutTransformer) is used for modeling. For example, the sequential representation can be broken down into three parts: starting point, direction, and distance, such as starting from (1,1) and moving towards 10nm. This method is essentially geared towards wiring and standard structure modeling, and can only describe standard rectangles. It is only suitable for manually controllable design drawings and lacks the ability to handle "irregular," "continuous curves," and "unstructured tokens" after lithography. It cannot be applied to shapes with slanted edges or curves and cannot be extended to real manufacturing contour scenarios.

[0029] Therefore, from an IC (Integrated Circuit) perspective, the aforementioned representative methods are ill-suited to the three key characteristics of contour modeling after lithography: high control point density (tens of thousands, which cannot be coarsely sampled); continuous curve structure (not simple rectangular blocks); and the geometric semantics of closed contours (requiring preservation of internal and external relationships). For fields outside IC, there are also vector curve representation methods, such as SketchRNN, DeepSVG, and SketchINR. These methods focus on the "action semantics of strokes," modeling graphics as drawing behaviors and often using stroke states and Bézier fragments to organize data. While they demonstrate good structural modeling capabilities in tasks such as icon generation and sketch recognition, their objectives are style generation, semantic alignment, and editability, neglecting control point-level geometric fitting and not requiring contour closure or high-precision coordinate reconstruction. Their modeling logic fundamentally differs from our needs in lithography scenarios. Therefore, currently, there is a lack of a modeling method that can simultaneously achieve the following goals: preserving the complete closed structure and topological relationships of the contour; handling a large number of dense control points without simplifying sampling; supporting continuous sampling and flexible style reconstruction; and achieving structure-aware expression and precise control point regression.

[0030] Based on the above research, this disclosure provides a graphical modeling method, apparatus, storage medium, and computer device. Utilizing a first neural network, feature encoding is performed on each contour segment corresponding to the contour curve, extracting local geometric and sequential features of the contour segments to obtain segment structure feature vectors representing the contour segment structure. Using a second neural network, sampling position vectors corresponding to each sampling point are generated based on arc length position parameters under different distributions, enabling the expression of sampling point positions under various sampling styles, thereby improving the diversity and flexibility of path segment structure expression. Using a third neural network, control point coordinates are generated through the segment structure feature vectors and each sampling position vector, fully considering the overall structure of the contour segment and the point structure of each sampling point, thus improving the accuracy of the generated control point coordinates (i.e., target coordinates). Furthermore, the target coordinates corresponding to the sampling points are used to remodel the contour curve, improving the consistency between the modeling result and the original contour curve, thereby meeting the accuracy and structural expression capabilities required for contour modeling after photolithography and improving the accuracy of the modeling results.

[0031] The shortcomings of the above solutions are the result of the inventor's practical experience and careful research. Therefore, the discovery process of the above problems and the solutions proposed in this disclosure below are all contributions made by the inventor to this disclosure.

[0032] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0033] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0034] To facilitate understanding of this embodiment, a detailed description of the graphics modeling method disclosed in this disclosure is provided first. The execution subject of the graphics modeling method provided in this disclosure is generally a terminal device or other processing device with certain computing capabilities. The terminal device can be a user equipment (UE), a mobile device, a terminal, a personal digital assistant (PDA), a handheld device, a computer device, etc. In some possible implementations, the graphics modeling method can be implemented by the processor calling computer-readable instructions stored in the memory.

[0035] The following describes the graphical modeling method provided in this disclosure using a computer device as an example.

[0036] like Figure 1 The flowchart shown is a graphical modeling method provided in an embodiment of this disclosure, which may include the following steps: S101: Using the first neural network, feature encoding is performed on each contour segment corresponding to the contour curve in the photolithography pattern to obtain the segment structure feature vector corresponding to each contour segment.

[0037] Here, the photolithography pattern can be a pre-constructed process pattern, which may include at least one closed contour curve. For example... Figure 2a The diagram shown is a schematic representation of various closed contour curves in a photolithographic pattern according to an embodiment of this application. Wherein, Figure 2a The blue closed curves in Figure (1) are the closed contour curves.

[0038] Each contour curve can correspond to a continuous function. A contour curve includes multiple location points, each with an initial coordinate (x, y). Connecting these location points sequentially forms a contour curve. For example, the contour curve to be modeled may consist of a series of closed paths, each path represented by a sequentially arranged sequence of two-dimensional location points, which can be formalized as shown in the following formula: (Formula 1) in, A two-dimensional sequence of position points representing the contour curve. This represents the two-dimensional coordinates of the Nth position point, where the number of N is the same as the number of position points on the contour curve. This means that the paths of the contour curves are connected end to end to form a closed structure.

[0039] A contour segment can be obtained by dividing the contour curve based on the original coordinates of each position point. A contour curve can be divided into one or more contour segments. The number of contour segments into which a contour curve is divided can be determined based on the number of input points required by the first neural network and the number of position points of the contour curve.

[0040] The first neural network can be a pre-trained neural network. Its input can be the original coordinate sequence of each point in the contour segment, and its output can be the segment structure feature vector corresponding to the entire contour segment. The segment structure feature vector can be a latent vector used to characterize the local geometric features and positional order features of the contour segment. Specifically, the segment structure feature vector can be used for subsequent sampling point reconstruction and sampling modeling.

[0041] In practice, for any contour curve in the photolithography pattern, the contour curve can be divided into at least one contour segment according to the original coordinates of each position point in the contour curve. The original coordinates of each position point in each contour segment are used as the contour segment information. Then, for each contour segment, the contour segment information can be input into the first neural network to extract and encode the local set features and position point sequence features to obtain the segment structure feature vector of the contour segment.

[0042] In one embodiment, the contour segment of the contour curve can be obtained by dividing it according to the following steps A~B: Step A: Based on the original coordinates of each position point on the contour curve, sort the position points in a consistent direction according to a preset direction, so that the sorted contour curve follows the left-hand or right-hand rule, and obtain the sorted position points.

[0043] Here, the preset direction can be set based on experience, and this application does not impose specific limitations. Sorting the various position points according to the preset direction ensures that the sorted contour curves follow the left-hand or right-hand rule, guaranteeing that the subsequent contour segments are sorted in a unified direction. This effectively maintains the geometric closure of the contour and the consistency of the internal and external topological relationships, so as to determine the internal and external information enclosed by the contour.

[0044] Because each closed contour curve has a clear geometric distinction between its internal region and external background—for example, in photolithography, this distinction corresponds to the boundary recognition of physical structures such as metallic and insulating regions—it is necessary to preprocess the original coordinates of the contour curve's points after acquisition. This preprocessing ensures that the order of the points across all contours is consistent, maintaining semantic consistency between internal and external regions. Specifically, for any contour curve, the sequence of its points can be obtained first, and the original coordinates of each point can be extracted from this sequence. Then, based on the original coordinates of each point on the contour curve, the order of the points in a preset direction can be determined. Finally, the points are sorted according to this preset order to obtain the sorted points. This sorting of points in a preset direction ensures that the sorted contour curve follows either a left-hand or right-hand rule, facilitating the determination of the internal and external information enclosed by the contour segment.

[0045] For example, the preset direction can be clockwise or counterclockwise. The positions can be sorted in a consistent direction according to the clockwise or counterclockwise direction to obtain the sorted positions.

[0046] Step B: Divide the sorted position points into segments according to the target number of points to obtain the contour segments corresponding to the contour curve.

[0047] Here, the target number of points can be the number of points required as input to the first neural network.

[0048] In practice, the sorted position points can be segmented according to the number of target points and the order of the sorted position points, thus dividing the target point count into contour segments. For example, if the number of target points is L, the first L position points can be segmented to obtain the first contour segment, the L+1 to 2L position points can be segmented to obtain the second contour segment, and so on, until all contour segments are obtained.

[0049] like Figure 2a As shown in Figure (2), this is the... Figure 2a The contour segments are obtained by dividing the contour curves corresponding to the middle (1) figure. Specifically, Figure 2a The closed curves of different colors in Figure (2) are the contour segments.

[0050] In one embodiment, since the number of position points in different contour curves varies, the number of position points may be an integer multiple of the target number, or it may not be an integer multiple of the target number. When the number of position points is an integer multiple of the target number, it can be guaranteed that each segment divided using the sorted position points has the target number of points. When the number of position points is not an integer multiple of the target number, the number of position points in the last segment may be less than the target number. Since the first neural network requires that the number of points in each input contour segment be the target number, it is necessary to pad the number of position points in the last segment to ensure that the number of position points in the last segment is also the target number. Therefore, this embodiment also provides a method for padding position points. Specifically, if the number of position points in the last segment is less than the target number, the position points in the last segment are padded to the target number using the end position points of the contour curve or preset padding points to obtain the last contour segment.

[0051] Here, the preset completion point can be a pre-set point without gradient participation. For example, the preset completion point can be 0 or a preset sign.

[0052] The end point is the last point in the sorted sequence of all points corresponding to the entire contour curve.

[0053] In practical implementation, assuming the target number of points is L and the contour curve includes N position points, the contour curve will be divided into ceil(N / L) segments. Here, ceil represents a "round up" mathematical function, used to ensure that even if the last segment has fewer than L points, it can still be retained as a separate contour segment for modeling. The last segment may originate from the end of the contour curve path, where the number of position points is less than L. Before modeling, the number of position points will be padded to L. For example, for segments less than L, padded points can be created by copying the last point, or by adding preset padded points in the form of 0 or preset symbols. Here, each position point in the contour curve can be defined as a valid position point, and each padded point can be defined as a padded position point. To eliminate the impact of padded points on subsequent model processing, a status identifier penstate can be added to each padded position point. Penstate indicates whether it is a padded position point. For example, each position point in a contour segment can be represented in three dimensions as [x, y, penstate]. Penstate takes different values ​​to characterize whether a position point is a padded position point or a valid position point. For example, the penstate value rule is: 0 indicates a valid control point, and 2 indicates a complete control point.

[0054] For example, setting the target number of points L=256, a certain contour curve contains N=1020 control points. Before segmentation, the complete path is represented as [ , ..., The length is 1020. After segmentation, there are 4 segments: 3 complete segments containing 256 valid position points and 1 last segment containing only the last 252 valid position points. The 3 complete segments are directly used as contour segments. In the last segment, the 1019th position point... Repeat the process four times to obtain the last contour segment with four completion points. Set the penstate of the four completion points to 2. The structure of the last contour segment is as follows: [ …, , , ..., There are a total of L location points.

[0055] like Figure 2b The diagram shown is a schematic representation of each contour segment in an embodiment of this application when the number of position points contained in the contour curve is less than the number of input points required by the first neural network. Figure 2b Figure (3) shows the various contour curves. Figure 2b Figure (4) shows the contour segments corresponding to each contour curve, indicating that each contour curve corresponds to a contour segment.

[0056] In this way, by sorting, segmenting, and completing the location points, the consistency of all contour segments in the data dimension can be guaranteed, which facilitates batch processing by the first neural network. At the same time, the penstate information can effectively distinguish between valid location points and completed location points, ensuring the accuracy of subsequent feature processing and loss calculation.

[0057] S102: Obtain the position parameters of each arc length corresponding to the contour segment; the arc length position parameters are used to indicate the percentage of arc length corresponding to each sampling point in the contour segment.

[0058] Here, for each contour segment, this application supports contour modeling under different sampling location distributions, thereby guiding reconstruction strategies of different styles during the inference stage. Here, different style sampling does not refer to traditional image style transfer, but rather to the ability to reconstruct the structure of the same path latent vector (i.e., the segment structure feature vector mentioned above) under different sampling location distributions, manifested as adjustable control over sampling density, distribution pattern, and the degree of local geometry preservation. Sampling location distributions include uniform distribution, random distribution, or the original distribution determined based on the original coordinates of each location point on the contour segment, or a non-uniform distribution predicted by the arc length position decoding network in the third neural network described later. A sampling location can be understood as each location point sampled from the contour segment; this location point is also called a sampling point. Under different sampling point distributions, the final modeled curve will be different.

[0059] The arc length position parameter t is a normalized parameter used to indicate the percentage of arc length corresponding to each sampling point in the contour segment. Here, considering that contour curves have a finite length, defining the length percentage position (i.e., the arc length position parameter t) as the sampling position is a more appropriate approach. For example, a length percentage position of 0.1 means that for a contour curve, the corresponding coordinate value is sampled at the position where the arc length percentage is 0.1 as the coordinate value of the sampling point.

[0060] In one embodiment, the method for obtaining the arc length position parameter in S102 above can be any one of the following methods A to C: Method A: Use uniform sampling to generate the position parameters of each arc length corresponding to the contour segment.

[0061] For example, the interval [0, 1] can be divided into equal intervals using dense uniform sampling or sparse uniform sampling to obtain the normalized arc length position parameters. These normalized arc length position parameters are then used as the normalized arc length parameters of the contour segment.

[0062] Method B: Use random sampling to generate the position parameters of each arc length corresponding to the contour segment.

[0063] For example, a certain number of normalized arc length position parameters can be randomly selected from the interval [0, 1] using dense non-uniform sampling, sparse non-uniform sampling, dense non-uniform random sampling, or sparse non-uniform random sampling. These normalized arc length position parameters can then be used as the normalized arc length parameters of the contour segment.

[0064] Method C: Based on the original coordinates of each position point on the contour segment, determine the percentage of each position point relative to the total length of the contour segment, and use the percentage corresponding to each position point as the position parameter of each arc length.

[0065] For example, the original coordinate sampling method can be used to take each position point on the contour segment as a sampling point. Then, based on the original coordinates of each position point and the total length of the contour segment, the percentage of the position point relative to the total length of the contour segment can be determined, and this percentage can be used as a normalized arc length parameter corresponding to the position point.

[0066] Alternatively, an arc length position decoding network can be used to generate a set of arc length position parameters corresponding to the contour segment based on the segment structure feature vector corresponding to the contour segment.

[0067] For example, the segment structure feature vector corresponding to the contour segment can be input into a pre-trained arc length position decoding network. Then, the arc length position decoding network can predict the non-uniformly normalized arc length position parameters during reconstruction based on the latent expression of the contour segment (i.e., the segment structure feature vector), in order to simulate the point density distribution of the contour segment and support a contour expression that is closer to the original style of the contour segment.

[0068] In practical implementation, any of the methods described above can be used to obtain the arc length position parameters corresponding to each contour segment. Furthermore, based on the total arc length of the contour segment and the arc length position parameters, the sampling point position corresponding to each arc length position parameter within the contour segment can be determined. Understandably, if the sampling point position corresponding to a certain arc length position parameter exceeds the length of the contour segment, a preset value can be added to indicate that the sampling point corresponding to that arc length position parameter is an invalid sampling point exceeding the contour segment. The preset value can be 0, 1, or any value greater than 1.

[0069] S103: Using the second neural network, generate the sampling position vector corresponding to each sampling point based on the arc length position parameter.

[0070] Here, the sampling position vector is a query vector representing the sampling points corresponding to the arc length position parameter, used to achieve continuous position indexing of the contour segment. The sampling position vector and the segment structure feature vector can be used together for implicit regression of the sampling point coordinates.

[0071] The second neural network is a pre-trained neural network. The input of this neural network can be the arc length position parameter, and the output can be the sampling position vector of the sampling point corresponding to each arc length position parameter.

[0072] There is no strict restriction on the execution order between the second neural network and the first neural network; they can be executed in parallel or asynchronously. That is, S101 and S103 can be executed synchronously, and the arc length position parameter needs to be obtained before synchronous execution.

[0073] For example, a second neural network can be used to perform feature mapping on each arc length position parameter to obtain the sampling position vector corresponding to each sampling point.

[0074] S104: Using the third neural network, the target coordinates corresponding to each sampling point in the contour segment are determined based on the segment structure feature vector and each sampling position vector.

[0075] Here, the target coordinates can be the coordinates of each point used for modeling. By connecting the target coordinates corresponding to each contour segment in sequence, the modeled curve corresponding to the contour curve can be obtained.

[0076] The third neural network is a pre-trained neural network. Its input can include a joint input sampling position vector and a segment structure feature vector. The output can be the target coordinates of each sampling point in the contour segment. The location of the target coordinates may be the same as the sampling point, or it may deviate from the sampling point's location, but each sampling point corresponds to one target coordinate. Typically, the locations of each target coordinate can be referred to as the control points of the contour curve.

[0077] In practice, for any contour segment, the sampling position vector output by the second neural network and the segment structure feature vector related to that contour segment can be jointly input into a pre-trained third neural network. The third neural network can perform feature recognition and extraction on the sampling position vector and the segment structure feature vector to obtain the target coordinates corresponding to each sampling point.

[0078] Because this application supports modeling using different arc length position parameters obtained through different sampling methods under the same segment structural feature vector, it can achieve the expression of structural diversity of contour segments. Since the sampling distribution obtained by each sampling method represents a specific sampling style, using different arc length position parameters can reflect the importance of different structural features or geometric regions.

[0079] S105: Determine the modeled curve corresponding to the contour curve based on the target coordinates corresponding to each contour segment.

[0080] In practice, after obtaining the target coordinates corresponding to each contour segment, the target coordinates can be connected in the order of the target coordinates to obtain the modeled curve corresponding to the contour curve.

[0081] like Figure 3 The diagram shown is a schematic representation of a contour curve and a modeled curve provided in an embodiment of this application. Figure 3 In the diagram, 'a' represents a schematic representation of each contour curve. Different colored line segments in the contour curves represent different initial contour segments. Figure 3 In the diagram, 'b' represents the modeled curve corresponding to each contour curve, and each modeled curve corresponding to a contour curve is represented by a color.

[0082] like Figure 4 The diagram shown is a schematic representation of another contour curve and a modeled curve provided in an embodiment of this application. Figure 4 In the diagram, 'c' represents a schematic representation of each original contour curve. Figure 4 In this context, d represents the model reconstruction result (i.e., the predicted curve) corresponding to each contour curve. This model reconstruction result is generated based on the target coordinates predicted by the third neural network.

[0083] Thus, the implicit representation and multi-style control point reconstruction method for lithographic pattern contours proposed in this application is particularly suitable for high-density, high-continuity and geometrically semantic post-lithographic contour modeling tasks, and can overcome the shortcomings of existing technologies in terms of expression granularity, structure preservation and sampling style consistency.

[0084] Optionally, the graphical modeling method provided in this application can be encapsulated into a model module with a calling interface. Then, when there is a need for image modeling, the model module can be scheduled by calling the interface, and the target coordinates of each sampling point can be obtained by inputting the contour segments of the contour curve and the arc length position parameters under any sampling position distribution. This can be used for application scenarios such as graphical modeling, simulation alignment, or structural reconstruction.

[0085] Optionally, the first, second, and third neural networks of this application can be individually encapsulated and provided with calling interfaces. In this way, different neural networks can be selectively invoked using the provided calling interfaces when performing image modeling.

[0086] In one embodiment, the above-described S101 can be implemented according to the following steps: S101-1: For any contour segment, the first neural network is used to construct the initial feature tensor corresponding to the contour segment based on the original coordinates of each position point in the contour segment and the number of position points.

[0087] Here, the first neural network encodes each contour segment into a fixed-length latent representation (i.e., a segment structure feature vector). Its input can be a tensor of the form (B, L, 3), where B represents the batch size (which can be a pre-defined value); L represents the number of points in the contour segment, i.e., the target number of points; and 3 can specifically include the coordinates (x, y) and penstate information of each position point. The processing flow of the first neural network can be divided into three parts: the first part is input feature construction, which constructs the initial feature tensor corresponding to the contour segment to enhance the first neural network's ability to perceive the geometric structure of the contour segment and the order of its position points; the second part is feature mapping to a unified dimension, which maps the output of the first part to a unified dimension, laying the foundation for subsequent temporal modeling and compressed representation; the third part is feature compression extraction, which extracts the segment structure feature vector of the contour segment as the basic representation of the contour segment structure in subsequent query and decoding stages. The number of position points in each contour segment is the target number of points.

[0088] In practical implementation, for any contour segment, an input tensor of the form (B, L, 3) can be constructed based on the original coordinates, target point count, and penstate of each position point in the contour segment, and then input into the first neural network. The first neural network can construct input features from the input tensor. For example, at least one auxiliary encoding method can be used to encode the original coordinates and target point count of each position point in the input tensor to obtain the contour segment corresponding to the form (B, L, 3). The initial feature tensor of ). Where, This represents the feature dimension of the initial feature tensor at each location point.

[0089] In one embodiment, the auxiliary encoding method may include three types: coordinate Fourier encoding, arc length position encoding, and sequential position encoding. For S101-1 above, it can be implemented according to the following steps: S101-1-1: Normalize the original coordinates of each position point in the contour segment, and encode the normalized coordinates based on the number of position points to obtain the coordinate encoding tensor.

[0090] Here, encoding the normalized coordinates is called coordinate encoding. Specifically, coordinate encoding can be done using coordinate Fourier encoding, and the coordinate encoding tensor is of the form (B, L, ...). ), This refers to the feature dimension of the coordinate encoding tensor of each location point, i.e., the Fourier encoding dimension.

[0091] In specific implementation, the original coordinates (x, y) of each position point in the contour segment can be normalized to obtain the normalized coordinates of the position points. Here, since the original coordinates of each position point are defined within the photolithography pattern, and the area of ​​the photolithography pattern is relatively large, it is necessary to normalize the original coordinates of each position point to a smaller target range to facilitate processing by the first neural network. The target range can be set empirically, and this embodiment does not impose a specific limitation; for example, the target range can be 512nm*512nm. After normalization, Fourier encoding can be performed on each normalized coordinate based on the number of position points L to construct position features containing different frequency components, thus obtaining a coordinate encoding tensor. The coordinate encoding process can be denoted as... .

[0092] S101-1-2: Based on the original coordinates of each position point in the contour segment and the total length of the contour segment, determine the arc length scalar of each position point, and based on the number of position points and the arc length scalar, determine the arc length encoding tensor.

[0093] Here, the arc length scalar is used to characterize the curve position corresponding to the location point in the contour segment. The arc length encoding tensor is the tensor obtained after encoding the arc length position of each location point.

[0094] In practice, based on the original coordinates of each point in the contour segment and the total length of the contour segment, the percentage of each point relative to the total length can be determined. This percentage can then be normalized to the interval [0, 1] to obtain the arc length scalar for each point. In this context, the arc length scalar of the starting point of the contour segment is 0. Then, arc length encoding can be performed on the number of position points and the arc length scalar to obtain an arc length encoded tensor. For example, the arc length scalar... The input is fed into a fully connected mapping layer. This layer performs fully connected mapping based on the number of location points, obtaining the low-dimensional embedding features corresponding to each location point, i.e., the arc-length encoded tensor. The arc-length encoding process can be denoted as: ArcEmbed( ) ,in, This refers to the feature dimension of the arc length encoding tensor at each location point, and ArcEmbed represents scalar embedding feature extraction.

[0095] S101-1-3: Perform position encoding on the original coordinates of each point in the contour segment to obtain the position encoding tensor.

[0096] In practical implementation, a standard sine / cosine position encoder can be used to sequentially encode the original coordinates of each position point in the contour segment, thereby constructing a periodic representation for each position point and obtaining a position encoding tensor. The sequential position encoding process can be denoted as: ;in, The feature dimension of the position encoding tensor for each location point.

[0097] S101-1-4: Concatenate the coordinate encoding tensor, arc length encoding tensor, and position encoding tensor to obtain the initial feature tensor.

[0098] In practice, the coordinate encoding tensor, arc length encoding tensor, and position encoding tensor can be concatenated according to the following formula to obtain the initial feature tensor of the contour segment: (Formula 2) in, That is, in the shape of The initial feature tensor, = + + ; For tensor concatenation, it means concatenating multiple tensors along their last dimension (i.e., the dimension where dim=-1) to form a new channel dimension.

[0099] Understandably, when the first neural network encodes each location point, it can encode only the valid location points based on the penstate of each location point, and mask the padding location points, thereby avoiding the influence of the padding location points on the encoded features.

[0100] S101-2: Based on the mask features corresponding to the state identifier information of each location point, perform feature mapping processing on the initial feature tensor to obtain the intermediate feature tensor; the state identifier information is used to indicate whether the location point is a completion location point.

[0101] Here, the state identifier information is the penstate of the location point. The mask feature corresponding to the penstate can be determined based on the penstate, and different penstates correspond to different masks.

[0102] Due to the three tensors in the initial feature tensor (i.e. , , Since the feature dimensions and feature distributions of the initial feature tensors are different, in order to achieve a fused feature representation, it is necessary to map the concatenated initial feature tensors to a unified intermediate feature dimension `latent_dim`. Specifically, the feature mapping process can be implemented using the following formula: (Formula 3) in, This represents the fused feature, which is the intermediate feature tensor, in the form of... ;mask represents the mask feature corresponding to the status identification information of each location point; This describes a channel attention fusion operation that, based on the input tensor (in this application, the input tensor is the Encoded type corresponding to Formula 2 above) and the mask features of each location point, performs weighted fusion, nonlinear transformation, or linear projection on the initial feature tensors of each location point, thereby outputting an embedded representation with a unified feature dimension. In this way, multi-source input information (i.e., , , Mapping them to the same dimension provides a foundation for subsequent time series modeling and compressed representation.

[0103] S101-3: Perform feature compression on the intermediate feature tensor to obtain the segment structure feature vector.

[0104] Here, the intermediate feature tensor of the above output... ∈ As a sequence feature on the contour segment, there are L rows of features. Each row of features represents the local embedding information of a certain point in the contour segment. In order to represent the contour segment result, the sequence feature needs to be compressed into a structural representation vector at the path segment level.

[0105] Specifically, a preset dimension feature compression module can be used to compress intermediate feature tensors. ∈ Compression yields a segment structure feature vector of the form (B, latent_dim). .

[0106] In one embodiment, S101-3 described above can be implemented according to the following steps: S101-3-1: Perform feature extraction processing on the intermediate feature vector to obtain the target feature tensor.

[0107] Here, feature extraction is used to further extract deeper feature information from the intermediate feature vectors. Feature extraction can be performed using various deep network modules, such as deep convolutional residual modules, deep transformer modules, deep fully connected modules, and normalization modules. In practice, any one or more combinations of deep network modules can be flexibly selected for feature extraction based on experience; this embodiment does not impose specific limitations. For example, the following description uses a deep convolutional residual module to perform convolutional residual processing on the intermediate feature vectors to achieve feature extraction.

[0108] Among them, convolutional residual processing can be implemented by several convolutional residual blocks. Each convolutional residual block can be a stacked module consisting of multiple one-dimensional convolutions, nonlinear activations, and residual connections.

[0109] In practice, to capture the structural relationships and semantic context information between the position points in the contour segment, a one-dimensional convolution module can be used to model the intermediate feature vector along the L direction. This process can be regarded as encoding the morphological changes of the contour segment curve along the sequence direction to extract the relative geometric distribution features between different control points.

[0110] Specifically, the intermediate feature vector can be adjusted to a convolutional input format according to the feature dimension, as shown in Formula 4 below: (Formula 4) in, This indicates the adjusted convolution input format. This represents the transpose function. ( , ) indicates that the intermediate feature vector ∈ The positions of the second axis (axes) and the first axis (axes) are transformed. The intermediate feature vector has three axes: the 0th axis, the first axis, and the second axis. The 0th axis is a parameter. The axis in question, the first axis is a parameter. The axis in question, the second axis is a parameter. The axis in which it is located.

[0111] Then, the adjusted convolution input format The input is fed into several convolutional residual blocks for convolutional residual processing, resulting in a form like... The target feature tensor. For example, the convolution residual can be achieved using the following formula: (Formula 5) in, ) represents a convolutional residual block, which uses a receptive field expansion mechanism (such as kernel size > 1) to capture short-range and medium-range dependencies between location points. This represents the convolution residual output of the convolution residual block.

[0112] Then, the format of the convolution residual output can be restored to its original format, resulting in a format like... The target feature tensor. For example, the format recovery can be achieved according to the following formula six: (Formula Six) in, Represents the target feature tensor. ( , This indicates that the convolution residual will be output. ∈ The positions of the second axis (axes) and the first axis (axes) are transformed, where the 0th axis of the convolution output is a parameter. The axis in question, the first axis is a parameter. The axis in question, the second axis is a parameter. The axis in which it is located.

[0113] S101-3-2: Determine the effective mask for each location point based on the status identifier information of each location point.

[0114] Here, after completing the structural modeling using Formula 6 above and obtaining the target feature tensor, the enhanced sequence features can be further compressed into a unified representation at the contour segment level. Specifically, to avoid the completion points in the contour segments affecting the accuracy of the structural representation, an effective mask for each position point can be constructed based on the penstate of each position point. .

[0115] For example, the effective mask for each location point can be determined using the following method based on point seven: (Formula 7) in, The effective mask is represented by Formula 7 above, which means that for positions where penstate is not equal to 2, the effective mask is 1, and for positions where penstate is equal to 2, the effective mask is 0.

[0116] S101-3-3: Using an effective mask, the target feature tensor is compressed to obtain the segment structure feature vector.

[0117] In practice, the effective mask at each location point can be used to compress the target feature tensor into a single vector based on a weighted averaging strategy. For example, the segment structure feature vector can be determined according to the following formula: (Formula 8) in, The form obtained after feature compression is as follows The segment structure feature vector, This means that on the dimension L of the target points, only the valid location points are weighted and summed before normalization to ensure that the completed location points are not included in the feature statistics.

[0118] Based on Formula 8 above, the output can be in the form of... The segment structure feature vector serves as the basic expression of the contour segment structure in the subsequent query and decoding stages.

[0119] In one embodiment, S103 described above can be implemented according to the following steps: Using a second neural network, feature encoding and feature extraction are performed on the arc length position parameter to obtain the sampling position vector corresponding to each sampling point.

[0120] Here, the second neural network receives a set of normalized arc length position parameters t corresponding to the contour segment and maps the arc length position parameters t to the corresponding query vector (i.e., the sampling position vector). The query vector can be jointly constructed with the segment structure feature vector to construct the implicit function query conditions.

[0121] For example, in this application, the arc length position parameter = / The total arc length of the contour segment, where The value of can be in the range of [0, 1]. This represents the i-th sampling point. This represents the arc length corresponding to the i-th sampling point. That is, each... The position of a sampling point on the corresponding contour segment can be used as an implicit function in the subsequent coordinate decoding module. Index variables.

[0122] For the second neural network, the arc length position parameter t input during the training phase can be the true coordinates of each real control point (i.e., the real sampling point) in the sample contour segment. Assuming a contour segment contains L effective control points, the cumulative arc length of each effective control point can be calculated and normalized to obtain... , of which each (i∈[1,L]) corresponds to a supervision coordinate, which can be: The true coordinates of the corresponding effective control points. In the application stage of the second neural network, the arc length position parameter t can be specified by the user. Specifically, the arc length position parameter t can be determined using any of the uniform sampling method, random sampling method, original coordinate sampling method, and arc length position decoding network prediction method mentioned above.

[0123] For example, whether in the training or application phase, t can always represent the normalized arc length position parameter, with dimensions ( ),in, This represents the number of sampling points / control points, and also the number of arc length position parameters.

[0124] Specifically, the second neural network may include an encoding module and a feedforward decoding module. The encoding module is used to encode the arc length position parameters to obtain the encoding results corresponding to each arc length position parameter. The feedforward decoding module is used to extract features from the output of the encoding module to obtain the sampling position vector of the sampling point.

[0125] For example, the encoding module can obtain the arc length position parameter t. And according to the following formulas nine to twelve, t is repeatedly projected onto the target frequency domain to form a periodic code: (Formula Nine) (Formula 10) (Formula Eleven) ; (Formula 12) The target frequency domain can be a pre-defined frequency domain, typically a high-frequency domain. This represents the frequency band, where i is the frequency basis function label used to construct the frequency domain features, and F is the preset number of frequency basis functions. This means projecting t onto the frequency basis functions (i.e., the frequency band). This indicates the encoding result. For tensor splicing operations, it means splicing along the last dimension. This represents the 2F+1 dimension frequency feature obtained after feature transformation of the original coordinates of the sampling points.

[0126] Then, the encoding result output by the encoding module can be input into the feedforward network module in the second neural network. The feedforward network module is used to extract features from the encoding result to obtain the sampling position vector corresponding to each sampling point. Here, the feedforward network module can specifically be a multilayer perceptron (MLP).

[0127] For example, the feedforward network module can determine the sampling position vector according to the following formula thirteen: (Formula Thirteen) Here, the query represents the sampling location vector, in the form of... .

[0128] In one embodiment, S104 described above can be implemented according to the following steps: S104-1: Using the third neural network, feature dimension alignment and feature concatenation are performed on the segment structure feature vector and the vectors at each sampling position to obtain the concatenated feature vector.

[0129] Here, the third neural network can be a pre-trained decoding network, which may include a coordinate decoding network. The coordinate decoding network can be used to jointly decode the sampling position vector output by the feedforward decoding module and the segment structure feature vector output by the first neural network to predict the coordinates of the control point of the contour segment at the corresponding position t, thus realizing the implicit function expression of the contour segment. The core idea of ​​this module is to construct an implicit function f(t), which is jointly determined by the latent structural information of the contour segment (i.e., the segment structure feature vector) and the position index (i.e., the sampling position vector), and is used for continuous spatial interpolation of arbitrary arc length parameter positions.

[0130] In practice, the sampling position vectors output by the feedforward network module and the segment structure feature vector output by the first neural network can be jointly input into the coordinate decoding network for feature dimension alignment and feature concatenation to obtain the concatenated feature vector.

[0131] Here, the segment structure feature vector ∈ Corresponding to the global structural representation of the contour segment, the sampling position vector Query∈ For each The vector query indicates that the feature dimensions of the two vectors do not match. Therefore, to construct the complete query conditions for each sampling point, it is necessary to... Align the feature dimensions of the query with the feature dimensions of the query and then concatenate them.

[0132] Specifically, feature alignment and concatenation can be performed using the input combination and dimension alignment network layer in the coordinate decoding network, according to the following formulas fourteen and fifteen: (Formula Fourteen) (Formula Fifteen) in, Represents the latent vector of branches aligned with the query dimension. ; Indicates the first axis Perform dimensional alignment on top; Indicates to In dimensions The tensor is copied to achieve dimension alignment; This indicates that on dimension -1, ... and Perform channel splicing. This indicates the concatenation of feature vectors.

[0133] S104-2: Perform feature decoding on the spliced ​​feature vector to obtain the target coordinates corresponding to each sampling point in the contour segment.

[0134] In practice, the concatenated feature vector can be input into the coordinate decoding layer of the coordinate decoding network to perform coordinate decoding processing and obtain the target coordinates corresponding to each sampling point in the contour segment.

[0135] For example, the coordinate decoding layer can be composed of several implicit modeling sub-modules. Each sub-module can be constructed empirically, and this application does not impose specific limitations. For instance, each sub-module may include an input normalization and adjustment sub-module, a multi-layer sinusoidal activation network (SIREN) sub-module, and a gated residual fusion sub-module. The input normalization and adjustment sub-module can be used to perform layer normalization (LayerNorm) and learnable scaling on the input vector; the multi-layer SIREN sub-module is used to build high-frequency response capabilities on the output of the input normalization and adjustment sub-module, used to model fine-grained changes in coordinates; the gated residual fusion sub-module is used to fuse multiple feature branches at different depths on the output of the multi-layer SIREN sub-module to improve the stability and expressive power of the coordinate decoding network. The final output of the coordinate decoding layer is mapped to two-dimensional coordinates through a coordinate header. For example, the processing of the coordinate decoding layer can be represented by the following formula sixteen: (Formula Sixteen) in, Indicates the target coordinates. Represents a coordinate decoding network. The 2 in the figure represents the x and y coordinates of the target.

[0136] against In other words, this coordinate represents the location at the specified sampling position. The predicted control point positions on the curve form the implicit representation of the contour segment at that position. The second neural network module supports arbitrarily distributed input t and outputs a sampling position vector that matches the input. The third neural network combines the sampling position vector and a general segment structure feature vector to generate contour reconstruction results with consistent style but flexible sampling strategies.

[0137] Optionally, the third neural network can include not only a coordinate decoding network but also an arc length position decoding network (t_decoder). The arc length position decoding network is used to predict the non-uniform sampling position distribution of the contour segment during reconstruction based on its latent representation. This is to support contour representations that more closely resemble the original style of the contour curve. Therefore, the output of the third neural network can include the target coordinates output by the coordinate decoding network and the non-uniform sampling position distribution output by the arc length position decoding network. The non-uniform sampling position distribution output by the arc length position decoding network. This can be simulated by performing dense sampling at locations with high curvature in the contour segment to preserve geometric details at inflection points, and sparse sampling at locations with low curvature. The input to the arc length position decoding network is the segment structure feature vector output by the first neural network. ∈ The output is a set of normalized sampling location parameters: ,in This indicates the number of sampling points (which is also the number of arc length position parameters). Each item in the range is [0, 1]. This indicates that prediction sampling is performed on T.

[0138] During the training phase, The goal is to reproduce the sampling distribution of the original contour segment sampling points in the arc-length space as accurately as possible, such as the locations of dense and sparse regions of sampling points, and to capture the importance distribution of local contour details. Training loss can be metric such as distribution distance (e.g., Earth Mover Distance), point ranking loss, or kernel density matching.

[0139] In the application phase, It can be used as input to a second neural network to replace manually specified sampling strategies such as uniform sampling or random sampling, achieving contour reconstruction with higher style consistency. The network structure of the arc length position decoding network can be set empirically, and this application does not impose specific limitations. For example, the arc length position decoding network may include an input normalization and scaling layer for normalization and scaling processing, a multi-layer SIREN network layer for constructing non-linear expressive power, a gated residual fusion layer for enhancing structure awareness, and a Sigmoid activation layer for outputting normalized position values.

[0140] Optionally, during the application phase, It can also be used to input the discrete modeled curves together into other neural networks to fit the discrete modeled curves into a closed contour curve.

[0141] In one embodiment, since the present application provides a first neural network, a second neural network and a third neural network, the present application also provides a method for end-to-end joint training of the three networks. The training objective is to minimize the coordinate decoding error and the sampling prediction error in order to obtain a more stable and expressive structural embedding.

[0142] Specifically, the neural network can be jointly trained according to the following steps S1~S6: S1: Obtain the sample arc length position parameters and label coordinates corresponding to each sample control point in the sample contour segment.

[0143] Here, the sample contour segment is the actual acquired contour segment, and the sample control points are the pre-determined valid control points. The sample arc length position parameter can be determined based on the label coordinates of the sample control points and the length of the contour segment. The label coordinates are the actual coordinates of the sample control points, and the label coordinates can be obtained using... express, Indicates the number of sample control points. The 2 in the label coordinates represents the x and y coordinates.

[0144] In practice, the sample arc length position parameters and label coordinates corresponding to each sample control point in any sample contour segment can be obtained.

[0145] S2: Using the first neural network to be trained, the feature encoding of the sample contour segment is performed according to the label coordinates of each sample control point to obtain the predicted segment structure feature vector corresponding to the sample contour segment.

[0146] Here, the predicted segment structure feature vector is the segment structure feature vector output by the first neural network to be trained for the sample contour segment during the training phase.

[0147] For example, after sorting the sample control points according to orientation consistency, the label coordinates of each sorted sample control point can be input into the first neural network to be trained to obtain the predicted segment structure feature vector corresponding to the sample contour segment. .

[0148] S3: Using the second neural network to be trained, generate the predicted sampling position vector corresponding to each sample control point based on the sample arc length position parameter.

[0149] Here, the predicted sampling position vector is the second neural network to be trained, which uses the sample arc length position parameter to predict the sampling position vector for the sample control point.

[0150] S4: Using the third neural network to be trained, determine the predicted coordinates and predicted arc length position parameters corresponding to each sample control point based on the predicted segment structure feature vector and the predicted sampling position vector.

[0151] Here, the predicted coordinates are the target coordinates corresponding to each sample control point output by the coordinate decoding network in the third neural network to be trained. The predicted arc length position parameters are the arc length position decoding network (t_decoder) in the third neural network to be trained, which outputs the arc length position parameters for the sample contour segment using the predicted segment structure feature vector. For each sample contour segment, the arc length position corresponding to its sample control point is normalized to obtain the sample arc length position parameters, which can constitute the target distribution. .

[0152] In practice, for the effective control point positions of the sample contour segment, the corresponding normalized sample arc length position parameters are as follows: Where i ∈ (1, N), and N is the number of sample control points. The sample arc length position parameters are input into the second neural network to be trained to obtain the predicted sampling position vector Trunk( Then, the predicted sampling position vector and the predicted segment structure feature vector of the contour segment are input together into the third neural network to be trained for processing to obtain the predicted coordinates and predicted arc length position parameters.

[0153] The process of generating the predicted coordinates can be achieved using the following formula (Equation 17): (Formula 17) in, Indicates the predicted coordinates. Indicates the number of sample control points. The 2 in the predicted coordinates represents the x and y coordinates, Trunk ( () represents the preset sampling position vector corresponding to the arc length position parameter of the i-th sample. This represents the coordinate decoding network in the third neural network to be trained.

[0154] The various predicted arc length position parameters can form a prediction distribution. The process of generating the predicted distribution can be represented by the following formula (Equation 18): (Formula 18) in, t represents the predicted arc length position parameter, and tDecoder represents the arc length position decoding network in the third neural network to be trained.

[0155] Understandably, after tDecoder is trained, the arc length position parameter output by tDecoder can be used to fit various discrete modeled curves, or it can be used as the input of a second neural network.

[0156] S5: Determine the first network loss based on the predicted coordinates and label coordinates, and determine the second network loss based on the sample arc length position parameters and the predicted arc length position parameters.

[0157] Here, the first network loss is used to represent the coordinate decoding loss of the second neural network when predicting coordinates. This loss can include two parts: one part is the mean-square error (MSE) loss, and the other part is the average distance loss between the two nearest points.

[0158] Among them, the coordinate mean square error loss It can be determined according to the following formula nineteen: (Formula 19) in, The number of sample control points, This represents the predicted coordinates corresponding to the i-th sample control point. This represents the label coordinates corresponding to the i-th sample control point.

[0159] Bidirectional nearest point distance average loss It can be determined according to the following formula: (Formula 20) Where CD represents the average distance between the two nearest points. These are the predicted coordinates corresponding to each sample control point; These are the label coordinates corresponding to each sample control point.

[0160] In obtaining and Then, both can be considered together as the first network loss.

[0161] The second network loss characterizes the sampling distribution prediction loss of the arc length position decoding network when predicting arc length position parameters. For example, the second network loss... Mean squared error loss can be used. The second network loss is determined using the following formula (21): (Formula 21) in, This represents the ratio between 1 and the number of predicted arc length position parameters. This represents the mean square error between the predicted arc length position parameters and the sample arc length position parameters for each sample control point.

[0162] S6: Using the first network loss and the second network loss, jointly train the first neural network, the second neural network and the third neural network to obtain the trained first neural network, the second neural network and the third neural network.

[0163] In practical implementation, the following formula twenty-two can be used to determine the total network loss based on the first network loss and the second network loss. : (Formula 22) in, , and These are adjustable hyperparameters used to balance the modeling capability of reconstruction accuracy and sampling distribution. Each hyperparameter can be set based on experience, and the embodiments in this application do not impose any limitations.

[0164] Based on Formula 3 above, the first network loss and the second network loss are weighted and summed to obtain the total network loss. Then, the total network loss can be used to jointly iteratively train the first, second, and third neural networks until a preset cutoff condition is met, resulting in well-trained first, second, and third neural networks. The preset cutoff condition can be that the prediction accuracy of each trained neural network reaches a set accuracy and / or the number of iterations reaches a preset number.

[0165] Optionally, during joint iterative training, the total network loss can be used, and the standard gradient descent optimization method can be employed, based on the Adam optimizer. In this way, using coordinate prediction error and sampling distribution error as supervision signals to jointly train the three neural networks can effectively improve reconstruction accuracy and representation consistency.

[0166] This application achieves structured processing of contour curves by sorting the position point sequence of each closed contour curve in a uniform direction (e.g., clockwise) to ensure geometric semantic consistency, dividing it into several contour segments of uniform length, and using a completion method to unify the structure when the last segment is insufficient, and adding a mask to assist subsequent training. A first neural network is constructed to receive the control point sequence of each contour segment and its auxiliary codes (including coordinate Fourier encoding, arc length position encoding, and sequential position encoding), extracting structural expressions and generating latent vector representations of the contour segments, thus completing the modeling of contour segment structural feature vectors. A second neural network is constructed to map the normalized arc length position parameters within the contour segments into query vectors, which, together with the segment structural feature vector modeling, are input into the coordinate decoding network to achieve continuous coordinate generation of the contour segments, i.e., the implicit function expression f(t), thus completing the reconstruction of control point coordinates. An arc length position decoding network is constructed to output non-uniformly normalized arc length position parameters after inputting the segment structural feature vectors of the contour segments, used to simulate the control point density distribution of the original contour segments, and can guide different reconstruction strategies during the inference stage.

[0167] The graphical modeling method provided in this application not only achieves a consistent structural representation of high-density location point sequences, preserving the topological and geometric closure of the contour, but also supports reconstruction output under various sampling strategies, adapting to the density and style requirements of different downstream modeling tasks (such as post-lithography modeling tasks). Furthermore, a unified structural latent vector representation is constructed using a first neural network, facilitating contour style transfer, similar structure retrieval, and high-level semantic modeling. It also possesses good scalability, being encapsulated as a model module with an input interface for use in graphical simulation, alignment, reconstruction, and other scenarios. Therefore, the graphical modeling method of this application is suitable for manufacturing graphical modeling tasks under lithography processes and can also be extended to other high-precision graphical modeling fields with closed contour structures, such as microstructure simulation, vector graphic generation, and geometric compression.

[0168] In summary, the method proposed in this application, which combines contour segment modeling and query-based implicit function decoding for the first time, integrates structure-aware encoding, continuous position indexing, and controllable style expression. It fills the technical gap in high-precision manufacturing contour modeling scenarios and has significant engineering value and broad model adaptability.

[0169] Furthermore, during the inference phase, this application allows users to freely specify the sampling style to generate the t input, thereby obtaining contour descriptions with multiple distribution styles under the same segment structure feature vector expression. For example: uniform sampling generates structurally complete, isodense contours; non-uniform sampling positions predicted by tDecoder can restore the original sampling style during training, representing dense keypoint regions; random perturbation sampling can be used for robustness testing or to enhance data diversity. In addition, the segment structure feature vector can be used as a structure encoding vector for high-level graphics tasks, such as segment structure feature vectors between different contour segments can be used to measure contour structure similarity, can be used as conditional embedding input to achieve "style-preserving" contour transfer or synthesis tasks; and can be used for graphics analysis tasks such as clustering and classification to improve the model's structural understanding ability. Moreover, the concatenated feature vector composed of segment structure feature vectors and various sampling position vectors can be used as a conditional control branch and embedded into other graphics translation models or structural reconstruction models to improve their structural understanding ability and sampling flexibility, realizing the generation from contour coordinate points to images.

[0170] Since the arc length position parameters in this application can be obtained using different sampling methods, this application also provides, to facilitate understanding the process of obtaining arc length position parameters using different sampling methods for graphical modeling, methods such as... Figure 5 and Figure 6 The diagram shows a comparison of image modeling results under different sampling methods. Figure 5 In this context, 'e' represents the positional parameters of each arc length obtained using a dense non-uniform sampling method, and the modeling results obtained by image modeling using the first to third neural networks. Figure 5 In this context, f represents the positional parameters of each arc length obtained using a sparse uniform sampling method, and the modeling results are obtained by using the first to third neural networks for image modeling. Figure 6 In this context, g represents the positional parameters of each arc length obtained using a dense non-uniform random sampling method, and the modeling results obtained by image modeling using the first to third neural networks. Figure 6 In this context, h represents the positional parameters of each arc length obtained using a sparse, non-uniform random sampling method, and the modeling results obtained by using the first to third neural networks for image modeling.

[0171] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0172] Based on the same inventive concept, this disclosure also provides a graphics modeling device corresponding to the graphics modeling method. Since the principle of the device in this disclosure for solving the problem is similar to the graphics modeling method described above in this disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0173] like Figure 7 The diagram shown is a schematic representation of a graphical modeling apparatus provided in an embodiment of this disclosure, comprising: The encoding module 701 is used to use the first neural network to perform feature encoding on each contour segment corresponding to the contour curve in the photolithography pattern, and obtain the segment structure feature vector corresponding to each contour segment. The acquisition module 702 is used to acquire the arc length position parameters corresponding to each contour segment; the arc length position parameters are used to indicate the percentage of arc length corresponding to each sampling point in the contour segment. The generation module 703 is used to generate a sampling position vector corresponding to each sampling point based on the arc length position parameter using a second neural network. The first determining module 704 is used to determine the target coordinates corresponding to each sampling point in the contour segment based on the segment structure feature vector and each sampling position vector using a third neural network. The second determining module 705 is used to determine the modeled curve corresponding to the contour curve based on the target coordinates corresponding to each contour segment.

[0174] In one possible implementation, the device further includes a segmentation module 706 for dividing the contour segment in the following step: Based on the original coordinates of each position point on the contour curve, the position points are sorted in a consistent direction according to a preset direction, so that the sorted contour curve follows the left-hand or right-hand rule, and the sorted position points are obtained. Based on the target number of points, the sorted position points are divided into segments to obtain the contour segments corresponding to the contour curve.

[0175] In one possible implementation, the segmentation module 706, when segmenting the sorted position points according to the target number of points to obtain the contour segments corresponding to the contour curve, is used to: If the number of position points in the last segment is less than the target number of points, then the number of position points in the last segment is supplemented to the target number of points using the end position points of the contour curve or preset completion points to obtain the last contour segment.

[0176] In one possible implementation, the acquisition module 702, when acquiring the arc length position parameters corresponding to the contour segment, is used to: The position parameters of each arc length corresponding to the contour segment are generated by using a uniform sampling method; or, the position parameters of each arc length corresponding to the contour segment are generated by using a random sampling method; or, based on the original coordinates of each position point on the contour segment, the percentage of each position point relative to the total length of the contour segment is determined, and the percentage corresponding to each position point is used as the position parameters of each arc length.

[0177] In one possible implementation, the encoding module 701, when using the first neural network to perform feature encoding on each contour segment corresponding to the contour curve in the photolithography pattern to obtain the segment structure feature vector corresponding to each contour segment, is used to: For any of the contour segments, the first neural network is used to construct the initial feature tensor corresponding to the contour segment based on the original coordinates of each position point in the contour segment and the number of position points. Based on the mask features corresponding to the state identifier information of each location point, the initial feature tensor is subjected to feature mapping processing to obtain an intermediate feature tensor; the state identifier information is used to indicate whether the location point is a completion location point. The intermediate feature tensor is subjected to feature compression processing to obtain the segment structure feature vector.

[0178] In one possible implementation, the encoding module 701, when constructing the initial feature tensor corresponding to the contour segment based on the original coordinates of each position point in the contour segment and the number of position points, is used to: The original coordinates of each position point in the contour segment are normalized, and the normalized coordinates are encoded based on the number of position points to obtain a coordinate encoding tensor. Based on the original coordinates of each position point in the contour segment and the total length of the contour segment, determine the arc length scalar of each position point, and based on the number of position points and the arc length scalar, determine the arc length encoding tensor. The original coordinates of each point in the contour segment are encoded to obtain a position encoding tensor. The coordinate encoding tensor, the arc length encoding tensor, and the position encoding tensor are concatenated to obtain the initial feature tensor.

[0179] In one possible implementation, the encoding module 701, when performing feature compression processing on the intermediate feature tensor to obtain the segment structure feature vector, is used to: The intermediate feature vector is subjected to feature extraction processing to obtain the target feature tensor; Based on the status identifier information of each location point, determine the effective mask for each location point; Using the effective mask, the target feature tensor is compressed to obtain the segment structure feature vector.

[0180] In one possible implementation, the generation module 703, when generating the sampling position vector corresponding to each sampling point using the second neural network based on the arc length position parameter, is used to: Using the second neural network, feature encoding and feature extraction are performed on the arc length position parameter to obtain the sampling position vector corresponding to each sampling point.

[0181] In one possible implementation, the first determining module 704, when determining the target coordinates corresponding to each sampling point in the contour segment based on the segment structure feature vector and each sampling position vector using the third neural network, is configured to: Using the third neural network, feature dimension alignment and feature concatenation are performed on the segment structure feature vector and the sampling position vector to obtain a concatenated feature vector; The spliced ​​feature vector is subjected to feature decoding processing to obtain the target coordinates corresponding to each sampling point in the contour segment.

[0182] In one possible implementation, the device further includes a training module 707 for training the first neural network, the second neural network, and the third neural network through the following steps: Obtain the sample arc length position parameters and label coordinates corresponding to each sample control point in the sample contour segment; Using the first neural network to be trained, the feature encoding of the sample contour segment is performed according to the label coordinates of each sample control point to obtain the predicted segment structure feature vector corresponding to the sample contour segment. Using the second neural network to be trained, a predicted sampling position vector corresponding to each sample control point is generated based on the sample arc length position parameter. Using the third neural network to be trained, the predicted coordinates and predicted arc length position parameters corresponding to each of the sample control points are determined based on the predicted segment structure feature vector and the predicted sampling position vector. A first network loss is determined based on the predicted coordinates and the label coordinates, and a second network loss is determined based on the sample arc length position parameters and the predicted arc length position parameters. The first neural network, the second neural network, and the third neural network are jointly trained using the first network loss and the second network loss to obtain the trained first neural network, the second neural network, and the third neural network.

[0183] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0184] Based on the same technical concept, embodiments of this application also provide a computer device. (Refer to...) Figure 8 The diagram shown is a structural schematic of a computer device provided in an embodiment of this application, comprising: The system comprises a processor 801, a memory 802, and a bus 803. The memory 802 stores machine-readable instructions executable by the processor 801. The processor 801 executes these machine-readable instructions, performing the following steps: S101: Using a first neural network, feature encoding is performed on each contour segment corresponding to the contour curve in the photolithography pattern to obtain a segment structure feature vector corresponding to each contour segment; S102: Obtaining arc length position parameters corresponding to each contour segment; the arc length position parameters indicate the percentage of arc length corresponding to each sampling point in the contour segment; S103: Using a second neural network, a sampling position vector corresponding to each sampling point is generated based on the arc length position parameters; S104: Using a third neural network, the target coordinates corresponding to each sampling point in the contour segment are determined based on the segment structure feature vector and the sampling position vectors; and S105: Based on the target coordinates corresponding to each contour segment, the modeled curve corresponding to the contour curve is determined.

[0185] The aforementioned memory 802 includes a main memory 8021 and an external memory 8022. The main memory 8021, also known as internal memory, is used to temporarily store the computational data in the processor 801, as well as the data exchanged with external memory such as a hard disk 8022. The processor 801 exchanges data with the external memory 8022 through the main memory 8021. When the computer device is running, the processor 801 and the memory 802 communicate through the bus 803, so that the processor 801 executes the execution instructions mentioned in the above method embodiments.

[0186] The embodiments of the subject matter and functional operation described in this specification can be implemented in the following ways: digital electronic circuits, tangibly embodied computer software or firmware, computer hardware including the structures disclosed in this specification and their structural equivalents, or combinations thereof. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible, non-transitory program carrier for execution by a data processing apparatus or for controlling the operation of a data processing apparatus. Alternatively or additionally, the program instructions may be encoded on artificially generated propagation signals, such as machine-generated electrical, optical, or electromagnetic signals, which are generated to encode information and transmit it to a suitable receiving device for execution by the data processing apparatus. The computer storage medium may be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or combinations thereof.

[0187] The processing and logic flows described in this specification can be executed by one or more programmable computers that execute one or more computer programs to perform corresponding functions by operating on input data and generating output. The processing and logic flows can also be executed by dedicated logic circuits—such as field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs), and the device can also be implemented as dedicated logic circuits.

[0188] Suitable computers for executing computer programs include, for example, general-purpose and / or special-purpose microprocessors, or any other type of central processing unit. Typically, the central processing unit receives instructions and data from read-only memory and / or random access memory. The basic components of a computer include a central processing unit for implementing or executing instructions and one or more memory devices for storing instructions and data. Typically, a computer will also include one or more mass storage devices for storing data, such as disks, magneto-optical disks, or optical disks, or the computer will be operatively coupled to such mass storage devices to receive data from or transfer data to them, or both. However, a computer is not required to have such devices. Furthermore, a computer can be embedded in another device, such as a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device such as a Universal Serial Bus (USB) flash drive, to name just a few.

[0189] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and memory devices, such as semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal hard disks or removable disks), magneto-optical disks, and CD-ROM and DVD-ROM disks. Processors and memory may be supplemented by or incorporated into dedicated logic circuitry.

[0190] While this specification contains numerous specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily intended to describe features of specific embodiments of a particular invention. Certain features described in the various embodiments herein may also be implemented in combination in a single embodiment. Conversely, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation thereof.

[0191] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0192] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings are not necessarily shown in a specific order or sequence to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0193] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A graphical modeling method, characterized in that, The method includes: Using a first neural network, feature encoding is performed on each contour segment corresponding to the contour curve in the photolithography pattern to obtain the segment structure feature vector corresponding to each contour segment. Obtain the arc length position parameters corresponding to each contour segment; the arc length position parameters are used to indicate the percentage of arc length corresponding to each sampling point in the contour segment. Using a second neural network, a sampling position vector corresponding to each sampling point is generated based on the arc length position parameter; Using a third neural network, the target coordinates corresponding to each sampling point in the contour segment are determined based on the segment structure feature vector and each sampling position vector; Based on the target coordinates corresponding to each contour segment, the modeled curve corresponding to the contour curve is determined.

2. The method according to claim 1, characterized in that, The contour segment is obtained by dividing it according to the following steps: Based on the original coordinates of each position point on the contour curve, the position points are sorted in a consistent direction according to a preset direction, so that the sorted contour curve follows the left-hand or right-hand rule, and the sorted position points are obtained. Based on the target number of points, the sorted position points are divided into segments to obtain the contour segments corresponding to the contour curve.

3. The method according to claim 2, characterized in that, The step of dividing the sorted position points into segments according to the target number to obtain the contour segments corresponding to the contour curve includes: If the number of position points in the last segment is less than the target number of points, then the number of position points in the last segment is supplemented to the target number of points using the end position points of the contour curve or preset completion points to obtain the last contour segment.

4. The method according to claim 1, characterized in that, The step of obtaining the position parameters of each arc length corresponding to the contour segment includes: The position parameters of each arc length corresponding to the contour segment are generated by using a uniform sampling method; or, the position parameters of each arc length corresponding to the contour segment are generated by using a random sampling method; or, based on the original coordinates of each position point on the contour segment, the percentage of each position point relative to the total length of the contour segment is determined, and the percentage corresponding to each position point is used as the position parameters of each arc length.

5. The method according to claim 1, characterized in that, The first neural network is used to encode the features of each contour segment corresponding to the contour curve in the photolithography pattern to obtain the segment structure feature vector corresponding to each contour segment, including: For any of the contour segments, the first neural network is used to construct the initial feature tensor corresponding to the contour segment based on the original coordinates of each position point in the contour segment and the number of position points. Based on the mask features corresponding to the state identifier information of each location point, the initial feature tensor is subjected to feature mapping processing to obtain an intermediate feature tensor; the state identifier information is used to indicate whether the location point is a completion location point. The intermediate feature tensor is subjected to feature compression processing to obtain the segment structure feature vector.

6. The method according to claim 5, characterized in that, The step of constructing the initial feature tensor corresponding to the contour segment based on the original coordinates of each position point in the contour segment and the number of position points includes: The original coordinates of each position point in the contour segment are normalized, and the normalized coordinates are encoded based on the number of position points to obtain a coordinate encoding tensor. Based on the original coordinates of each position point in the contour segment and the total length of the contour segment, determine the arc length scalar of each position point, and based on the number of position points and the arc length scalar, determine the arc length encoding tensor. The original coordinates of each point in the contour segment are encoded to obtain a position encoding tensor. The coordinate encoding tensor, the arc length encoding tensor, and the position encoding tensor are concatenated to obtain the initial feature tensor.

7. The method according to claim 5, characterized in that, The step of performing feature compression processing on the intermediate feature tensor to obtain the segment structure feature vector includes: The intermediate feature vector is subjected to feature extraction processing to obtain the target feature tensor; Based on the status identifier information of each location point, determine the effective mask for each location point; Using the effective mask, the target feature tensor is compressed to obtain the segment structure feature vector.

8. The method according to claim 1, characterized in that, The step of using a second neural network to generate sampling position vectors corresponding to each sampling point based on the arc length position parameter includes: Using the second neural network, feature encoding and feature extraction are performed on the arc length position parameter to obtain the sampling position vector corresponding to each sampling point.

9. The method according to claim 1, characterized in that, The step of using a third neural network to determine the target coordinates corresponding to each sampling point in the contour segment based on the segment structure feature vector and each sampling position vector includes: Using the third neural network, feature dimension alignment and feature concatenation are performed on the segment structure feature vector and the sampling position vector to obtain a concatenated feature vector; The spliced ​​feature vector is subjected to feature decoding processing to obtain the target coordinates corresponding to each sampling point in the contour segment.

10. The method according to claim 4, characterized in that, The first neural network, the second neural network, and the third neural network are trained through the following steps: Obtain the sample arc length position parameters and label coordinates corresponding to each sample control point in the sample contour segment; Using the first neural network to be trained, the feature encoding of the sample contour segment is performed according to the label coordinates of each sample control point to obtain the predicted segment structure feature vector corresponding to the sample contour segment. Using the second neural network to be trained, a predicted sampling position vector corresponding to each sample control point is generated based on the sample arc length position parameter. Using the third neural network to be trained, the predicted coordinates and predicted arc length position parameters corresponding to each of the sample control points are determined based on the predicted segment structure feature vector and the predicted sampling position vector. A first network loss is determined based on the predicted coordinates and the label coordinates, and a second network loss is determined based on the sample arc length position parameters and the predicted arc length position parameters. The first neural network, the second neural network, and the third neural network are jointly trained using the first network loss and the second network loss to obtain the trained first neural network, the second neural network, and the third neural network.

11. A graphic modeling device, characterized in that, The device includes: The encoding module is used to use the first neural network to encode the features of each contour segment corresponding to the contour curve in the photolithography pattern, and obtain the segment structure feature vector corresponding to each contour segment. The acquisition module is used to acquire the arc length position parameters corresponding to each contour segment; the arc length position parameters are used to indicate the percentage of arc length corresponding to each sampling point in the contour segment. The generation module is used to generate sampling position vectors corresponding to each sampling point based on the arc length position parameters using a second neural network. The first determining module is used to determine the target coordinates corresponding to each sampling point in the contour segment based on the segment structure feature vector and each sampling position vector using a third neural network. The second determining module is used to determine the modeled curve corresponding to the contour curve based on the target coordinates corresponding to each contour segment.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 10.

13. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 10.

Images