A multi-scale information fusion method for photolithography mask correction

Abstract

This invention discloses a multi-scale information fusion method for photolithography mask correction. It establishes a high-order multi-deformation interaction network to capture global optical information on the mask; by establishing an interaction map within the optical diameter range at the corners of polygons on the mask, it captures local optical proximity interactions between line segments; by establishing sequential information on the degree of edge placement error, it can capture the overall change trend of the polygon contour, including the deviation scale and direction; using segmented line segments as basic units ensures the manufacturability of the mask and improves the time efficiency of the mask optimization process; thus ensuring manufacturability and timeliness.

Description

Technical Field

[0001] This invention belongs to the field of photolithography mask technology, and in particular relates to a multi-scale information fusion method for photolithography mask correction. Background Technology

[0002] As the feature size of integrated circuits continues to shrink, existing manufacturing technologies face severe challenges. Because the size of circuit patterns is close to the wavelength of the light source used in photolithography, diffraction effects inevitably occur, leading to distortion in the photolithographic patterns and thus affecting the manufacturing capacity of integrated circuits. To improve capacity, resolution enhancement technology has become crucial, with optical proximity correction (OPC) being one of the core technologies. During OPC, the distortion caused by optical scattering is compensated by adjusting the mask pattern. Existing methods are mainly divided into two categories: one is design rule-based methods, which are simple to operate and computationally fast, suitable for relatively simple designs. However, with the advancement of advanced technology nodes, the complexity of design rules increases exponentially, causing these methods to gradually become inadequate in optimizing mask quality. The other category is model-based methods, which have the advantage of a larger solution space and can obtain higher quality results, but the computation time increases significantly. With the increasing complexity of circuit designs, both design rule-based and model-based methods face bottlenecks in balancing efficiency and output quality. Machine learning methods are gradually emerging as a new research and application trend.

[0003] Unlike traditional methods, which require multiple iterations to optimize the mask, machine learning methods capture and learn the complex optical interactions between different polygons and motion units on the mask. This allows the mask to be generated in a single forward pass and its quality to be improved through several fine-tuning iterations, significantly reducing the time required for optical proximity correction. This demonstrates great potential for advancing the efficient manufacturing of masks. Summary of the Invention

[0004] To address the problems existing in current technologies, this invention provides a multi-scale information fusion method for photolithography mask correction. This method utilizes machine learning-based approaches for multi-scale information fusion, ensuring mask manufacturability and improving time efficiency in mask optimization. This method has achieved good results in the local optimization of photomasks for 40nm technology node chips.

[0005] The technical solution of this invention is implemented as follows:

[0006] A multi-scale information fusion method for photolithography mask correction includes the following steps:

[0007] S1. Extract the edges of the polygons on the mask, divide them into line segments to obtain line segment units, and save the information of the line segment units;

[0008] S2. Extract the polygons on the mask, transform the polygons and the directed edges between them to obtain hyperedges and edge nodes, and construct a hypergraph. Use a dual hypergraph attention network to perform embedding learning on the hypergraph to obtain global optical information.

[0009] The line segments of the divided polygon are represented by the nodes at their centers. The interaction graph of the line segments at the corners of the polygon is constructed with the nodes of the line segments at the corners of the polygon as the centers. The interaction graph of the line segments at the corners of the polygon is embedded and learned by a convolutional neural network to obtain optical local proximity information.

[0010] By performing photolithographic simulation on the initial mask, an error sequence is constructed and substituted into the sequence model for learning to obtain information on polygon contour changes.

[0011] S3. Use a multilayer perceptron to integrate information, predict the movement distance of each line segment of the polygon, and perform rounding and normalization processing.

[0012] S4. Convert the set of line segments after motion back into a mask pattern;

[0013] S5. Perform photolithographic simulation on the mask pattern and calculate the error based on the loss function to form the gradient for backpropagation.

[0014] Furthermore, in S1, the length of the line segments is set, and the edges of the polygons on the mask are divided to obtain a set of line segments S = {s1, s2, ... s}. n}, where S represents the set of line segments, s1, s2, ..., s2. n Represents the line segments in the set of line segments;

[0015] Represent each line segment using the coordinates of its starting point and ending point as s. n =((x) startn y startn ), (x endn y endn )), where (x startn y startn (x) represents the starting coordinates, (x) endn y endn () represents the coordinates of the endpoint; and records the length L of the line segment and the direction of the line segment. And the ratio R between the length of the line segment and the edge length of the polygon it belongs to; obtain the line segment unit and the information of the line segment unit for each line segment; the line segment unit can move along its normal direction.

[0016] Furthermore, in step S2, obtaining global optical information specifically includes:

[0017] The geometric centers of the Manhattan rectangles of the polygons on the mask are extracted as center nodes. Each polygon's center node is converted into a hyperedge, and the directed edges between polygons are converted into edge nodes. The initial features of the edge nodes include: the total number of line segments obtained from dividing the polygons, the distance between polygons, and the encoded direction between polygons. The converted hyperedges, edge nodes, and the initial features of the edge nodes constitute the hypergraph G. H ={V,E};

[0018] Where V = {v1, v2, ... v} n Let E = {e1, e2, ..., e} represent the set of edge nodes transformed from directed edges between polygons. n Let} represent the set of hyperedges transformed from polygons, where the initial characteristics of the edge nodes are X∈R. n×d , where R represents the real number field, n is the number of rows in the matrix, and d is the number of columns in the matrix;

[0019] The topology of the hypergraph is determined by the incidence matrix. Represented as:

[0020]

[0021] Among them, v i Represents the i-th node v i e j Represents the j-th superedge e j That is, if node e j Belongs to hyperedge e j ,but It is 1 if it is true, otherwise it is 0.

[0022] The direction between polygons is encoded by a combination of 0s or 1s.

[0023] Furthermore, the hypergraph G is treated using a dual hypergraph attention network. H Embedded learning can be performed using the following methods:

[0024] The information from the edge nodes is passed to the superedge, and then the information from the superedge is aggregated back to the edge nodes as a layer; thus, the superedge e of the l-th layer is passed to the superedge. j Embedded Represented as:

[0025]

[0026] in, Represents the node v at level l-1 i Embedding, α E (e j ,v i ) represents the superedge ej For each edge node v i The attention coefficient is calculated as follows:

[0027]

[0028] Here, exp() represents the exponential function, used to map values ​​to the positive range, and is used in the attention mechanism to ensure that the weights are non-negative; a2 represents a learnable attention vector, which are model parameters; T represents the transpose operation, used for dot product operations. W N b represents the learnable weight matrix before convergence. N This represents the bias vector before convergence;

[0029] Then, the information on the hyperedge is aggregated back to the edge node according to the hyperedge's attention coefficient, that is:

[0030]

[0031] in, Represents the (l-1)th layer hyperedge e j Embedding, α N (e j ,v i ) represents the edge node v after aggregation. i For each hyperedge e j The attention coefficient is calculated as follows:

[0032]

[0033] Here, exp() represents the exponential function, used to map values ​​to the positive range, and is used in the attention mechanism to ensure that the weights are non-negative; a3 represents a learnable attention vector, which are model parameters; T represents the transpose operation, used for dot product operations. W E Let b represent the converged, learnable weight matrix. E This represents the converged bias vector;

[0034] The embedding information h of each hyperedge is obtained through an l-layer dual hypergraph attention network. E L (e); Then, using the index information of the line segment unit and its corresponding polygon stored in S1, the embedding information of the polygon is copied to obtain... That is, global optical information, where R represents the real number field, n is the number of rows in the matrix, and d h It is the column number of the matrix.

[0035] Furthermore, in S2, for line segments of the straight edges of the polygon, the node at the center of the line segment is used as the representation; for corner line segments at the corners of the polygon, the centers of the line segments at both ends of the corner are taken as nodes; using the nodes of the corner line segments as the centers, nodes within a diameter range r are connected, where r takes the value of:

[0036] r = 20·(λ / NA) / (1+sigma) max );

[0037] Where λ represents the wavelength of light, NA represents the numerical aperture, and sigma max Represents the maximum normalized radius, when sigma max When the value is 1, it indicates incoherent light;

[0038] Construct the line segment interaction graph G at the polygon corner. g The initial features of a node include: the ratio R of the line segment length to the edge length of the polygon it belongs to, the line segment direction, and the identifier for the corner line segment; the initial feature of the edge between nodes is the Manhattan distance between nodes.

[0039] Furthermore, a convolutional graph neural network is used to analyze the interaction graph G of line segments at the corners of the polygon. g Embedded learning, i.e.:

[0040]

[0041] in, This represents the graph incidence matrix after adding a self-loop edge to each node. in This represents the degree of each node;

[0042] Learn That is, the optical local proximity information of each line segment, where R represents the real number field, n is the number of rows in the matrix, and d h It is the column number of the matrix.

[0043] The number 0 or 1 is used as a judgment identifier to determine whether it is a corner segment.

[0044] Furthermore, in step S2, a photolithographic simulation is performed on the initial mask to obtain a photolithographic pattern. If the photolithographic pattern is distorted, the distorted part is a protrusion or a depression. The midpoint of the line segment unit is selected as the detection point, and a threshold th is set. EPE If the vertical distance from the detection point to the photolithographic pattern exceeds the threshold, a violation occurs, i.e.:

[0045]

[0046] Wherein, EPE violation(x,y) represents an edge placement error violation at coordinates (x,y), with a value of 1 indicating a violation and a value of 0 indicating no violation, and L2(x,y) represents the two-dimensional Euclidean distance between the lithographic pattern and the lithographic simulation pattern at coordinates (x,y).

[0047] When multiple thresholds are set, for the midpoint of a line segment unit, F... i =±∑EPE violations, where ± represents the protrusions or depressions in the lithographic pattern, F i The total number of mask edge placement error violations in the i-th region is represented by ∑EPE violations, where i corresponds to different raised or recessed regions.

[0048] Then, starting from the midpoint of the lower right corner line segment unit of the polygon, measurements are taken clockwise or counterclockwise to obtain the sequence information of the corresponding polygon contour changes. An error sequence is constructed and substituted into the sequence model for training and learning to obtain... That is, the polygon contour change information, where R represents the real number field, n is the number of rows in the matrix, and d h It is the column number of the matrix.

[0049] Furthermore, in step S3, the global optical information, local optical proximity information, and polygon contour change information obtained in step S2 are input into a multilayer perceptron for learning, predicting the movement distance L of each line segment of the polygon along its normal direction. M Then perform rounding and normalization, i.e. Here, Round() represents the round-down operation; then all line segments are moved along their normal direction to reduce the gap between the lithographic pattern and the target pattern; the reverse gradient is calculated and updated. The reverse gradient refers to a specific gradient calculation or usage method used to optimize model or mask parameters.

[0050] Furthermore, in S4, the set of line segments is transformed into a mask pattern by projecting parallel rays onto the line segments.

[0051] Furthermore, in S5, the coherent system and solution of the wavelength system are used as the optical model for photolithography modeling. The spatial light image intensity I is represented by the convolution H of the pixel mask M and a set of optical kernels. The coherent system is subjected to N... k th The order approximation, specifically expressed by the formula:

[0052]

[0053] in, H represents the convolution operation. i σ represents the i-th optical nucleus decomposed.i The corresponding weights of the coherent system are represented; after obtaining the spatial light intensity image I simulated by photolithography, a constant threshold I is used. th Using an anti-corrosion photoresist as a reaction model, the photolithographic pattern Z was evaluated, namely:

[0054]

[0055] Where I(x, y) represents the light intensity emitted at coordinates (x, y). When the light intensity I(x, y) is greater than a set threshold I... th When the time is right, photolithography is performed at coordinates (x, y), and Z represents the final pattern after photolithography etching.

[0056] Then, the difference Loss(Z,T) between the simulated lithography model and the target pattern is calculated using the squared difference error, and the gradient is generated for backpropagation, i.e.:

[0057]

[0058] Where Loss() represents the loss function and T represents the lithographic pattern of the target.

[0059] Compared with the prior art, the present invention achieves the following beneficial effects:

[0060] This invention provides a machine learning technique for optimizing masks by fusing multi-scale information. Machine learning methods can be divided into edge-level optical proximity correction (OPC) methods and pixel-level OPC methods. Edge-level OPC methods divide the edge contour of the mask pattern into several segments and iteratively optimize the position of each segment along its normal direction to compensate for lithographic imaging errors. In contrast, pixel-level OPC methods require decomposing expensive rectangles into manufacturable Manhattan polygons to generate mask patterns. Furthermore, applying decomposition and mask rule checking methods to regularize the mask pattern may lead to performance degradation and the introduction of new hotspots in machine learning mask optimization methods. However, the industry tends to favor edge-level OPC methods.

[0061] This invention establishes a high-order mask polygon interaction network. By establishing a high-order polygon interaction network, global optical information on the mask can be captured, providing a global field of view for each segmented line.

[0062] An interaction map of line segments at the corners of the mask polygons is constructed. By creating an interaction map within the optical diameter range at the corners of the polygons on the mask, local optical proximity interactions between line segments are captured. This provides finer-grained information for adjusting the movement of line segments at the corners.

[0063] By establishing a sequence of information on the degree of edge placement error, the overall trend of polygon contour changes, including deviation scale and direction, can be captured through the established temporal information, thereby providing information for the optimization of the movement of each edge.

[0064] By employing a multi-scale information fusion machine learning method, using segmented line segments as the basic unit, the manufacturability of the mask is guaranteed, and the time efficiency in the mask optimization process is improved, thus achieving both manufacturability and timeliness. Attached Figure Description

[0065] Figure 1 This is a flowchart of a multi-scale information fusion method for photolithography mask correction provided in an embodiment of the present invention;

[0066] Figure 2 This is a schematic diagram of constructing a hypergraph when capturing global optical information in an embodiment of the present invention;

[0067] Figure 3 This is a schematic diagram of constructing an interaction graph when capturing local optical proximity information in an embodiment of the present invention;

[0068] Figure 4 This is a schematic diagram illustrating the capture of polygon contour change information in an embodiment of the present invention;

[0069] Figure 5 This is a schematic diagram illustrating the integration of multi-scale information in an embodiment of the present invention;

[0070] Figure 6 This is a schematic diagram of the method for converting a set of line segments into a mask in an embodiment of the present invention. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0072] Example

[0073] like Figures 1 to 6 A multi-scale information fusion method for photolithography mask correction includes the following steps:

[0074] S1. Extract the edges of the polygons on the mask, divide them into line segments to obtain line segment units, and save the information of the line segment units;

[0075] By setting a minimum line segment length, the edges of the polygons on the mask are divided to obtain a set of line segments S = {s1, s2, ..., s}. n}, where S represents the set of line segments, s1, s2, ..., s2. n Represents the line segments in the set of line segments;

[0076] Represent each line segment using the coordinates of its starting point and ending point as s. n =((x) startn y startn ), (x endn y endn )), where (x startn y startn (x) represents the starting coordinates, (x) endn y endn () represents the coordinates of the endpoint; and records the length L of the line segment and the direction of the line segment. And the ratio R between the length of the line segment and the edge length of the polygon it belongs to; obtain the line segment unit and the information of the line segment unit for each line segment; the line segment unit can move along its normal direction.

[0077] The mask can be optimized by adjusting the movement distance of each line segment.

[0078] S2. Extract the polygons from the mask, transform the polygons and the directed edges between them to obtain hyperedges and edge nodes, and construct a hypergraph. Use a dual hypergraph attention network to perform embedding learning on the hypergraph to obtain global optical information; this is mainly to capture the global optical information of the mask, specifically including:

[0079] like Figure 2 The geometric center of the Manhattan rectangle of the polygons on the mask is extracted as the center node. Each polygon's center node is converted into a hyperedge, and the directed edges between polygons are converted into edge nodes. The initial features of the edge nodes include: the total number of line segments obtained from dividing the polygons, the distance between polygons, and the encoded direction between polygons. Through the above processing, the converted hyperedges, edge nodes, and the initial features of the edge nodes can be used to construct the hypergraph G. H ={V,E};

[0080] Where V = {v1, v2, ... v} n Let E = {e1, e2, ..., e} represent the set of edge nodes transformed from directed edges between polygons. n Let} represent the set of hyperedges transformed from polygons, where the initial characteristics of the edge nodes are X∈R. n×d , where R represents the real number field, n is the number of rows in the matrix, and d is the number of columns in the matrix;

[0081] The topology of the hypergraph is determined by the incidence matrix. Represented as:

[0082]

[0083] Among them, v i Represents the i-th node v i e j Represents the j-th superedge e j That is, if node e j Belongs to hyperedge e j ,but It is 1 if it is true, otherwise it is 0.

[0084] The direction between polygons is encoded by a combination of 0s or 1s.

[0085] Hypergraph G is studied using a dual hypergraph attention network. H Embedded learning can be performed using the following methods:

[0086] The information from the edge nodes is passed to the superedge, and then the information from the superedge is aggregated back to the edge nodes as a layer; thus, the superedge e of the l-th layer is passed to the superedge. j Embedded Represented as:

[0087]

[0088] in, Represents the node v at level l-1 i Embedding, α E (e j ,v i ) represents the superedge e j For each edge node v i The attention coefficient is calculated as follows:

[0089]

[0090] Here, exp() represents the exponential function, used to map values ​​to the positive range, and is used in the attention mechanism to ensure that the weights are non-negative; a2 represents a learnable attention vector, which are model parameters; T represents the transpose operation, used for dot product operations. W N b represents the learnable weight matrix before convergence. N This represents the bias vector before convergence;

[0091] Then, the information on the hyperedge is aggregated back to the edge node according to the hyperedge's attention coefficient, that is:

[0092]

[0093] in, Represents the (l-1)th layer hyperedge e j Embedding, α N (e j ,v i ) represents the edge node v after aggregation. i For each hyperedge e j The attention coefficient is calculated as follows:

[0094]

[0095]

[0096] Here, exp() represents the exponential function, used to map values ​​to the positive range, and is used in the attention mechanism to ensure that the weights are non-negative; a3 represents a learnable attention vector, which are model parameters; T represents the transpose operation, used for dot product operations. W E Let b represent the converged, learnable weight matrix. E This represents the converged bias vector;

[0097] Through the l-layer dual hypergraph attention network designed above, which is a hypergraph neural network learning framework, it is ultimately able to learn each polygon, that is, obtain the embedding information h of each hyperedge. E L (e) represents the interaction information between each polygon and its surrounding polygons. Then, using the line segment unit stored in S1 and its corresponding polygon index information, the polygon embedding information is copied to obtain... That is, global optical information, where R represents the real number field, n is the number of rows in the matrix, and d h It is the number of columns in the matrix. This is used to assign embedding information to the polygon containing each line segment.

[0098] The line segments of the divided polygon are represented by their central nodes. An interaction graph of the line segments at the polygon corners is constructed, centered on the nodes of the line segments at the polygon corners. A convolutional neural network is then used to embed this interaction graph to obtain optical local proximity information. This is primarily to capture the optical proximity effects at the polygon corners, specifically including:

[0099] For line segments along the straight edges of a polygon, the node at the center of the line segment represents the segment. For corner line segments at the corners of the polygon, the centers of the line segments at both ends of the corner are taken as nodes. Using the nodes of the corner line segments as centers, connect the nodes within a diameter range r. Considering the partial coherence effect of light, the value of r is:

[0100] r = 20·(λ / NA) / (1+sigma) max );

[0101] Where λ represents the wavelength of light, NA represents the numerical aperture, and sigma max Represents the maximum normalized radius, when sigma max When the value is 1, it indicates incoherent light;

[0102] like Figure 3 Construct the interaction graph G of line segments at the corner of the polygon. g The initial features of a node include: the ratio R of the line segment length to the edge length of the polygon it belongs to, the line segment direction, and the identifier for the corner line segment; the initial feature of the edge between nodes is the Manhattan distance between nodes.

[0103] Then, a convolutional neural network is used to process the interaction map G of the line segments at the corners of the polygon. g Embedded learning, i.e.:

[0104]

[0105] in, This represents the graph incidence matrix after adding a self-loop edge to each node. in This represents the degree of each node;

[0106] Learn That is, the optical local proximity information of each line segment, where R represents the real number field, n is the number of rows in the matrix, and d h It is the column number of the matrix.

[0107] The number 0 or 1 is used as a judgment identifier to determine whether it is a corner segment.

[0108] By performing photolithographic simulation on the initial mask, an error sequence is constructed and substituted into the sequence model for learning to obtain polygon contour change information; the main purpose here is to capture the overall contour change information of the polygon. Specifically, this includes:

[0109] A photolithographic simulation is performed on the initial mask to obtain the photolithographic pattern. If the photolithographic pattern is distorted, the distorted part may be a bump or a depression. The midpoint of the line segment unit is selected as the detection point, and a threshold th is set. EPE If the vertical distance from the detection point to the photolithographic pattern exceeds the threshold, a violation occurs, i.e.:

[0110]

[0111] Wherein, EPE violation(x,y) represents an edge placement error violation at coordinates (x,y), with a value of 1 indicating a violation and a value of 0 indicating no violation, and L2(x,y) represents the two-dimensional Euclidean distance between the lithographic pattern and the lithographic simulation pattern at coordinates (x,y).

[0112] When a series of thresholds are set, then for the midpoint of a line segment unit, F... i =±∑EPE violations, where ± represents the protrusions or depressions in the lithographic pattern, F i The total number of mask edge placement error violations in the i-th region is represented by ∑EPE violations, where i corresponds to different raised or recessed regions.

[0113] like Figure 4 Then, starting from the midpoint of the line segment unit in the lower right corner of the polygon, measurements are taken clockwise to obtain the sequence information of the corresponding polygon contour changes. An error sequence is constructed and substituted into the LSTM sequence model for training and learning to obtain... That is, the polygon outline change information, where R represents the real number field, n is the number of rows in the matrix, and d h It is the row number of the matrix.

[0114] The entire S2 mainly involves modeling multi-scale information on the mask and embedding learning through the design of neural networks to capture information from the optical system.

[0115] S3. Use a multilayer perceptron to integrate information, predict the movement distance of each line segment of the polygon, and perform rounding and normalization processing; for example... Figure 5 S3 mainly integrates the multi-scale information learned in S2;

[0116] The global optical information, local optical proximity information, and polygon contour change information obtained from S2 are input into a multilayer perceptron for learning, predicting the distance L that each line segment of the polygon moves along its normal direction. M Then perform rounding and normalization, i.e. Here, `Round()` represents a round-down operation; then all line segments are moved along their normal direction to reduce the gap between the lithographic pattern and the target pattern; it should be noted that the line segments at the corners may intersect after being moved, so the coordinates of the start and end points of the line segments need to be readjusted. The backpropagation gradient is calculated and updated. The backpropagation gradient refers to a specific gradient calculation or usage method used to optimize model or mask parameters.

[0117] A multilayer perceptron (MLP) is a classic artificial neural network model widely used in machine learning and deep learning. It is a feedforward neural network composed of multiple neuron layers, including input, hidden, and output layers, used to handle nonlinear problems such as classification and regression.

[0118] S4. Transform the set of line segments after motion back into a mask pattern; transform the set of line segments into a mask pattern by projecting parallel rays onto the line segments.

[0119] Photolithography simulation models only accept pixel-level mask input. "Photolithography simulation" and "photolithography modeling" mean the same thing. Therefore, the moved line segments need to be pixelated, converting the line segment set into a pixel-level mask. This process needs to be differentiable to allow gradients to flow from the pixel mask to the line segments, thereby updating the parameters controlling the line segment movement. For example... Figure 6 Use here Figure 6 Method 1 shown performs parallel ray projection on the edges to convert the line segment set into a pixel mask.

[0120] During reverse gradient propagation, interpolation methods can be used, or the gradient can be directly taken at the corresponding position on the pixel mask gradient based on the coordinates of the midpoint of the line segment, thereby constructing the gradient flow in its direction.

[0121] S5. Perform photolithographic simulation on the mask pattern and calculate the error based on the loss function to form the gradient for backpropagation;

[0122] A common example is to use the coherent system and solution of a 193nm wavelength system as the optical model for photolithography modeling. The spatial light image intensity I is represented by the convolution H of the pixel mask M and a set of optical kernels. The coherent system is then subjected to N... k th The order approximation, specifically expressed by the formula:

[0123]

[0124] in, H represents the convolution operation. i σ represents the i-th optical nucleus decomposed. i The corresponding weights of the coherent system are represented; after obtaining the spatial light intensity image I simulated by photolithography, a constant threshold I is used. th Using an anti-corrosion photoresist as a reaction model, the photolithographic pattern Z was evaluated, namely:

[0125]

[0126] Where I(x, y) represents the light intensity emitted at coordinates (x, y). When the light intensity I(x, y) is greater than a set threshold I...th When the time is right, photolithography is performed at coordinates (x, y), and Z represents the final pattern after photolithography etching.

[0127] Then, the difference Loss(Z,T) between the simulated lithography model and the target pattern is calculated using the squared difference error, and the gradient is generated for backpropagation, i.e.:

[0128]

[0129] Where Loss() represents the loss function and T represents the lithographic pattern of the target.

[0130] This invention provides a multi-scale information fusion method for photolithography mask correction. It establishes a high-order mask polygon interaction network to provide a global view for each segmented line; it establishes an interaction graph of line segments at the corners of the mask polygons to provide finer-grained information for the movement and adjustment of line segments at the corners; it establishes sequential information about the degree of edge placement error to provide information for the movement optimization of each edge; and by using segmented line segments as the basic unit, it ensures the manufacturability of the mask while improving the time efficiency of the mask optimization process.

[0131] Based on the disclosure and teachings of the foregoing specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the present invention should also fall within the protection scope of the claims of the present invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the present invention.

Claims

1. A multi-scale information fusion method for lithographic mask correction, characterized in that, The method comprises the following steps: S1, extracting the edges of the polygons on the mask and dividing them into line segments to obtain line segment units and save the information of the line segment units; S2, extracting the polygons on the mask, converting the polygons and the directed edges between the polygons into hyperedges and edge nodes, constructing a hypergraph, using a double hypergraph attention network to perform embedding learning on the hypergraph to obtain global optical information; Specifically comprising: extracting the geometric center of the Manhattan rectangle of the polygon on the mask as the center node, converting the center node of each polygon into a hyperedge, and converting the directed edge between the polygons into an edge node; the initial features of the edge node include: the total number of line segments obtained by dividing the polygons, the distance between the polygons, and the coding of the direction between the polygons; the converted hyperedge, edge node and initial features of the edge node constitute a hypergraph G H ={V,E}. where V = {v1, v2,... v n}, represents a set of edge nodes converted from the directed edges between the polygons, E = {e1, e2,... e n}, represents a set of hyper-edges converted from the polygons, the initial feature of the edge node is where R represents a real number field, n is the number of rows of the matrix, and d is the number of columns of the matrix; The topology of the hypergraph is then given by the incidence matrix is represented as: ； where v i represents the i-th node v i , e j represents the j-th hyperedge e j , i.e., if the node e j belongs to the hyperedge e j , then is 1, otherwise 0; The divided line segments of the polygons are represented by the nodes at the centers of the line segments, and the line segment interaction graph at the corners of the polygons is constructed with the nodes at the centers of the line segments at the corners of the polygons as the center, and a convolutional neural network is used to perform embedding learning on the line segment interaction graph at the corners of the polygons to obtain local optical information near the corners; By performing photolithography simulation on the initial mask, an error sequence is constructed, and the sequence model is learned to obtain polygon contour change information; S3, using a multilayer perceptron to integrate information, predicting the movement distance of each line segment of the polygon, and performing integer normalization processing; S4, converting the set of moved line segments into a mask pattern; S5, performing photolithography simulation on the mask pattern, and calculating the error according to the loss function to form the gradient of back propagation.

2. The multi-scale information fusion method for photomask correction according to claim 1, wherein, In the S1, set the line segment length, divide the edge of the polygon on the mask to obtain the line segment set S={s1, s2, ……s n}, wherein S represents the line segment set, s1, s2, ……s n represent the line segment in the line segment set; Each line segment is expressed as a start point coordinate and an end point coordinate, s n = ( (x startn , y startn ), (x endn , y endn ) ), where (x startn , y startn ) represents the start point coordinate and (x endn , y endn ) represents the end point coordinate; and the length L of the line segment, the direction of the line segment, and the ratio R of the length of the line segment to the length of the edge of the polygon in which the line segment is located are recorded; The line segment units of each line segment and the information of the line segment units are obtained; The line segment units can move along their normal direction.

3. The multi-scale information fusion method for photomask correction of claim 1, wherein, Hypergraph G is studied using a dual hypergraph attention network. H Embedded learning can be performed using the following methods: The information of the edge node is transmitted to the hyperedge, and the information on the hyperedge is gathered back to the edge node as a layer; the information of the edge node is transmitted to the hyperedge, and the information on the hyperedge is gathered back to the edge node as a layer j Embedding is represented as: ； wherein, denotes the embedding of the l-1 layer node v i , denotes the attention coefficient for each edge node v j on the hyperedge e i , which is calculated as: ； ； where exp() represents an exponential function, used to map values to a positive number range, in the attention mechanism to ensure the weight non-negative; a2 represents a learnable attention vector, which is a model parameter; T represents a transpose operation, used for dot product operation; W N represents a learnable weight matrix before convergence, b N represents a bias vector before convergence; Then the information on the hyperedge is gathered back to the edge node according to the attention coefficient of the hyperedge, that is: ； wherein, denotes the l-1th layer hyper-edge e j embedding of v denotes the aggregated edge node v i for each hyper-edge e j of v the attention coefficient of v ； ； where exp() represents an exponential function, used to map values to a positive number range, in the attention mechanism to ensure the weight non-negative; a3 represents a learnable attention vector, which is a model parameter; T represents a transpose operation, used for dot product operation; W E represents the converged learnable weight matrix, b E represents the converged bias vector; Through the l-layer double hypergraph attention network, the embedding information of each hyperedge is obtained ; then the embedding information of the polygon is copied through the index information of the line segment unit saved by S1 and the polygon corresponding to the line segment unit, to obtain , that is, global optical information, wherein R represents a real number field, n is the number of rows of the matrix, and d h is the number of columns of the matrix.

4. The multi-scale information fusion method for photomask correction of claim 2, wherein, In S2, for the line segments of the straight edges of the polygons, the nodes at the centers of the line segments are represented; for the corner line segments at the corners of the polygons, the centers of the line segments at both ends of the corner points of the polygons are taken as nodes to represent; taking the nodes of the corner line segments of the polygons as the center, connecting the nodes within the diameter range r, the value of r is: ； where λ denotes a wavelength of light, and NA denotes a numerical aperture, denotes a maximum normalized radius, and when denotes incoherent light when equal to 1. Constructing a line segment interaction graph G at a polygon corner g ; wherein the initial features of the nodes include a ratio R of the line segment length to the edge length of the polygon in which it is located, the line segment direction, and a judgment identifier of the corner line segment; and the initial features of the edges between nodes are the Manhattan distance between the nodes.

5. The multi-scale information fusion method for photomask correction according to claim 4, wherein, Using convolutional graph neural networks for line segment interaction graphs G at polygon corners g performing embedding learning, i.e.: ； wherein, denotes the graph adjacency matrix after adding a self-loop edge to each node, wherein denotes the degree of each node; learned i.e. the optical local neighborhood information of each line segment, where R denotes the real number field, n is the number of rows of the matrix, and d h is the number of columns of the matrix.

6. The multi-scale information fusion method for photomask correction of claim 1, wherein, In S2, a photolithography simulation is performed on the initial mask to obtain a photolithography pattern. If the photolithography pattern is distorted, the distorted part is a protrusion or a depression. A midpoint of a line segment unit is selected as a detection point, and a threshold value th is set EPE . If a vertical distance from the detection point to the photolithography pattern exceeds the threshold value, a violation occurs, that is: ； Wherein, EPE violation(x,y) represents the edge placement error violation at coordinate (x,y), the value is 1 indicating that the violation occurs, and the value is 0 indicating that there is no violation, L2(x,y) represents the two-dimensional Euclidean distance between the lithography pattern and the lithography simulation pattern at coordinate (x,y); When multiple threshold values are set, then for the midpoint of the line segment element, there is where ± denotes a protrusion or a recess of the lithographic pattern, F i denotes the total number of mask edge placement error violations for the i-th region, i corresponds to different protrusion or recess regions, denotes the summation over all edge placement error violations; Then the sequence information of the corresponding polygon contour change is obtained by measuring clockwise or counterclockwise from the midpoint of the right lower corner of the line segment unit of the multi-deformation, an error sequence is constructed, and the sequence model is substituted into the training learning to obtain , that is, the polygon contour change information, wherein R represents a real number field, n is the number of rows of the matrix, and d h is the number of columns of the matrix.

7. The multi-scale information fusion method for photomask correction of claim 1, wherein, In the S3, the global optical information, the optical local proximity information and the polygon profile variation information obtained in the S2 are input into a multi-layer perception machine for learning to predict the moving distance L of each line segment of the polygon along the normal direction thereof M Then, rounding normalization is performed, i.e. wherein Round() represents a rounding down operation; then all the line segments are moved along the normal direction thereof to reduce the difference between the lithography pattern and the target pattern; and the back gradient is calculated and updated, which is used for optimizing the model or the mask parameter.

8. The multi-scale information fusion method for photomask correction of claim 1, wherein, In S4, the set of line segments is converted into a mask pattern by parallel ray projection on the line segments.

9. The multi-scale information fusion method for photomask correction of claim 1, wherein, In step S5, the coherent system and solution of the wavelength system are used as the optical model for photolithography modeling. The spatial light image intensity I is represented by the convolution H of the pixel mask M and a set of optical kernels. The coherent system is then subjected to... The order approximation, specifically expressed by the formula: ； wherein, denotes a convolution operation, h i denotes the i-th optical kernel of the decomposition, denotes the respective weight of the coherent system; after obtaining the aerial intensity image I of the lithography simulation, the spatial intensity image I is obtained with a constant threshold I th The etch-resistant photoresist as a reaction model, evaluate the lithography pattern Z, that is: ； where I(x, y) represents the light intensity emitted at coordinates (x, y), and when the light intensity I(x, y) is greater than a set threshold value I th , then photoetching is performed at coordinates (x, y), and Z represents the pattern after final photoetching and etching. Then the difference between the simulated photolithography model and the target pattern is calculated by the square error, and the gradient is formed to perform back propagation, that is: ； Wherein, Loss() represents the loss function, and T represents the target lithography pattern.

Images