Multi-scale information fusion method for photoetching mask correction

Abstract

The invention discloses a multi-scale information fusion method for photoetching mask correction. The method comprises the following steps: establishing a high-order multi-deformation interaction network to capture global optical information on a mask; capturing local optical proximity interaction between the line segments by establishing an interaction graph within an optical diameter range at corners of the polygon on the mask; by establishing sequence information of edge placement error degree, the overall variation trend of the polygon contour can be captured, including deviation scale and direction; the segmentation line segments are taken as basic units, so that the manufacturability of the mask is ensured, and the time efficiency of the mask optimization process is improved; and therefore, the manufacturability and the timeliness are ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithography masks, and particularly relates to a multi-scale information fusion method for lithography mask correction. Background Art

[0002] With the continuous reduction of the process feature size of integrated circuits, the existing manufacturing technologies are facing severe challenges. Since the circuit pattern size is close to the wavelength of the light source used in the lithography process, the diffraction effect inevitably occurs, resulting in the distortion of the lithography pattern, thereby affecting the manufacturing productivity of integrated circuits. To improve productivity, resolution enhancement technology has become the key, and optical proximity correction is one of the core technologies. During the optical proximity correction process, the distortion effect caused by optical scattering is compensated by adjusting the mask pattern. The existing methods are mainly divided into two categories: one is the design rule-based method, which is simple to operate and has a fast calculation speed, and is suitable for processing relatively simple designs. However, with the advancement of advanced technology nodes, the complexity of design rules has increased exponentially, resulting in such methods gradually being unable to cope in optimizing the mask quality. The other is the model-based method, whose advantage lies in having a larger solution space and can obtain higher-quality results, but the calculation time consumption increases significantly. With the increasing complexity of circuit design, both the design rule-based method and the model-based method are facing bottlenecks between efficiency and output quality. And machine learning-based methods are gradually emerging and becoming a new research and application trend.

[0003] Different from traditional methods, traditional methods require multiple iterations to optimize the mask, while machine learning-based methods can generate the mask in one forward pass by capturing and learning the complex optical interactions between different polygons and motion units on the mask, and improve the mask quality through several fine-tuning iterations, greatly reducing the time of optical proximity correction and showing great potential in promoting the efficient manufacturing of masks. Summary of the Invention

[0004] Aiming at the problems existing in the prior art, the present invention provides a multi-scale information fusion method for lithography mask correction, which uses machine learning-based methods for multi-scale information fusion to ensure the manufacturability of the mask and improve the time efficiency in mask optimization. This method has achieved good results in the local optimization of the photomask of 40nm technology node chips.

[0005] The technical solution of the present invention is realized as follows:

[0006] A multi-scale information fusion method for lithography 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, convert the polygons and the directed edges between the polygons to obtain hyperedges and edge nodes, construct a hypergraph, and use a dual hypergraph attention network to perform embedding learning on the hypergraph to obtain global optical information;

[0009] Represent the line segments of the divided polygons by the nodes at their centers. Taking the nodes of the line segments at the corners of the polygons as the centers, construct an interaction graph of the line segments at the corners of the polygons, and use a convolutional neural network to perform embedding learning on the interaction graph of the line segments at the corners of the polygons to obtain optical local adjacent information;

[0010] Through lithography simulation of the initial mask, construct an error sequence and substitute it into the sequence model for learning to obtain polygon contour change information;

[0011] S3. Use a multi-layer perceptron for information integration, predict the moving distance of each line segment of the polygon, and perform rounding and normalization processing;

[0012] S4. Re-convert the set of the moved line segments into the pattern of the mask;

[0013] S5. Perform lithography simulation on the pattern of the mask and calculate the error according to the loss function to form the gradient of backpropagation.

[0014] Further, in S1, set the line segment length, divide the edges of the polygons on the mask to obtain a line segment set S = {s1, s2,... s n}, where S represents the line segment set, and s1, s2,... s n represent the line segments in the line segment set;

[0015] Represent each line segment by the starting point coordinates and the ending point coordinates as s n = ((x startn , y startn ), (x endn , y endn ))), where (x startn , y startn ) represents the starting point coordinates, and (x endn , y endn ) represents the ending point coordinates; and record the length L of the line segment, the line segment direction and the ratio R of the line segment length to the edge length of the polygon where the line segment is located; obtain the line segment unit of each line segment and the information of the line segment unit; the line segment unit can move along its normal direction.

[0016] Further, in S2, obtaining the global optical information specifically includes:

[0017] Extract the geometric center of the Manhattan rectangle of the polygon on the mask as the central node, convert the central node of each polygon into a hyperedge, and convert the directed edge between polygons into an edge node; the initial features of the edge node include: the total number of line segments obtained by dividing the polygon, the distance between polygons, and the encoding of the 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} represents the set of edge nodes converted from the directed edges between polygons, and E = {e1, e2,... e n} represents the set of hyperedges converted from polygons. The initial feature of the edge node is X ∈ R n×d , where R represents the real number field, n is the number of rows of the matrix, and d is the number of columns of the matrix;

[0019] Then the topological structure of the hypergraph is represented by the incidence matrix as:

[0020]

[0021] where, v i represents the i-th node v i , e j represents the j-th hyperedge e j , that is, if the node e j belongs to the hyperedge e j , then is 1, otherwise it is 0.

[0022] Encode the direction between polygons through a combination of the numbers 0 or 1.

[0023] Furthermore, perform embedding learning on the hypergraph G H through a double hypergraph attention network. The specific method includes:

[0024] Transfer the information of the edge node to the hyperedge, and then converge the information on the hyperedge back to the edge node as one layer; then transfer the information of the edge node to the hyperedge. The embedding of the hyperedge e j in the l-th layer is represented as:

[0025]

[0026] where, represents the embedding of the node v i in the l-1 layer, and α E (e j , v i ) represents the hyperedge ej For each edge node v i the attention coefficient is calculated as follows:

[0027]

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

[0029] Then, the information on the hyper - edge is aggregated back to the edge node according to the attention coefficient of the hyper - edge, that is:

[0030]

[0031] where represents the embedding of the hyper - edge e at the (l - 1) - th layer, and α j (e N , v j ) represents the attention coefficient of the edge node v for each hyper - edge e after aggregation, and is calculated as follows: i i j For each hyper - edge e j the attention coefficient, calculated as follows:

[0032]

[0033] where exp() represents the exponential function, which is used to map values to the positive range and is used in the attention mechanism to ensure non - negative weights; a3 represents a learnable attention vector, which is a model parameter; T represents the transpose operation, which is used for dot - product operations. W E represents the learnable weight matrix after aggregation, and b E represents the bias vector after aggregation;

[0034] Through the l - layer double - hypergraph attention network, the embedding information h E L (e) of each hyper - edge is obtained; then, through the index information of the line segment unit saved in S1 and the polygon it corresponds to, the embedding information of the polygon is copied to obtain which is the global optical information, where R represents the real number field, n is the number of rows of the matrix, and d h is the number of columns of the matrix.

[0035] Further, in S2, for the straight-edge segments of the polygon, they are represented by the nodes at the centers of the segments; for the corner segments at the corners of the polygon, the centers of the segments at both ends of the polygon corner points are taken as the nodes for representation; with the nodes of the polygon corner segments as the centers, the nodes within the diameter range r are connected, and the value of r is:

[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 equals 1, it represents incoherent light;

[0038] Construct an interaction graph G of the segments at the corners of the polygon g ; where the initial features of the nodes include: the ratio R of the segment length to the edge length of the polygon where it is located, the segment direction, and the judgment identifier of the corner segment; the initial feature of the edge between nodes is the Manhattan distance between the nodes.

[0039] Further, use a convolutional graph neural network to perform embedding learning on the interaction graph G of the segments at the corners of the polygon g That is:

[0040]

[0041] where represents the graph adjacency matrix after adding a self-loop edge to each node, where represents the degree of each node;

[0042] Learn to obtain That is, the optical local neighborhood information of each segment, where R represents the real number field, n is the number of rows of the matrix, and d h is the number of columns of the matrix.

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

[0044] Further, in S2, perform lithography simulation on the initial mask to obtain a lithography pattern. If the lithography pattern is distorted, the distorted part is a protrusion or a depression; select the midpoint of the line segment unit as the detection point, and set a threshold th EPE , if the vertical distance from the detection point to the lithography pattern exceeds the threshold, a violation occurs, that is:

[0045]

[0046] Among them, EPE violation(x,y) represents the edge placement error violation at the coordinate (x,y). A value of 1 indicates that a violation has occurred, and a value of 0 indicates no violation. L2(x,y) represents the two-dimensional Euclidean distance between the lithography pattern and the lithography simulation pattern at the coordinate (x,y).

[0047] When multiple thresholds are set, for the midpoint of the line segment unit, there is F i = ±∑EPE violations, where ± represents the protrusion or depression of the lithography pattern, and F i represents the total number of mask edge placement errors in the i-th region. i corresponds to different protrusion or depression regions, and ∑EPE violations represents the sum of the total number of all edge placement errors.

[0048] Then, starting from the midpoint of the line segment unit at the lower right corner of the polygon, measure clockwise or counterclockwise to obtain the sequence information of the contour change of the corresponding polygon, construct an error sequence, and substitute it into the sequence model for training and learning to obtain that is, the polygon contour change information. Among them, R represents the real number field, n is the number of rows of the matrix, and d h is the number of columns of the matrix.

[0049] Furthermore, in S3, the global optical information, optical local neighboring information, and polygon contour change information obtained in S2 are input into a multi-layer perceptron for learning to predict the moving distance L of each line segment of the polygon along its normal direction M , and then perform rounding normalization, that is where Round() represents the rounding-down operation; then all line segments move along their normal directions to reduce the gap between the lithography pattern and the target pattern; calculate and update the reverse gradient, where the reverse gradient refers to a specific way of gradient calculation or use for optimizing the model or mask parameters.

[0050] Furthermore, in S4, by performing parallel ray projection on the line segments, the set of line segments is converted into the pattern of the mask.

[0051] Furthermore, in S5, a coherent system and solution of the wavelength system are used as the optical model for lithography modeling. The spatial light image intensity I is represented by the convolution H of the pixel mask M and a set of optical kernels. Perform an N k th th-order approximation on the coherent system. The specific formula is as follows:

[0052]

[0053] Among them, represents the convolution operation, and h i represents the i-th optical kernel after decomposition, and σi represent the corresponding weights of the coherent system; after obtaining the spatial light intensity image I of the lithography simulation, using a corrosion-resistant photoresist with a constant threshold I th as the reaction model, evaluate the lithography pattern Z, that is:

[0054]

[0055] where I(x, y) represents the light intensity emitted at the coordinate (x, y). When the light intensity I(x, y) is greater than the set threshold I th then lithography is performed at the coordinate (x, y), and Z represents the pattern after final lithography and etching;

[0056] Then calculate the gap Loss(Z, T) between the simulated lithography model and the target pattern through the mean squared error, and form a gradient for backpropagation, that is:

[0057]

[0058] where Loss() represents the loss function and T represents the target lithography pattern.

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

[0060] The present invention provides a machine learning technology for optimizing a mask by fusing multi-scale information. Machine learning methods can be divided into edge-level optical proximity correction methods and pixel-level optical proximity correction methods. Among them, the edge-level optical proximity correction method divides the edge contour of the mask pattern into several segments and iteratively optimizes the position of each segment along its normal direction to compensate for lithography imaging errors. The mask pattern generated by the pixel-level optical proximity correction method needs to decompose expensive rectangles into manufacturable Manhattan polygons. In addition, applying decomposition and mask rule checking methods to regularize the mask pattern may lead to a decline in the performance of machine learning-based mask optimization methods and introduce new hotspots, but the industry often tends to use edge-level optical proximity correction methods.

[0061] The present 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 to provide a global view for each divided line segment.

[0062] Establish a line segment interaction graph at the corners of the mask polygon. By establishing an interaction graph within the optical diameter range at the corners of the polygon on the mask, local optical proximity interactions between line segments can be captured. Thus, finer-grained information can be provided for the movement adjustment of the line segments at the corners.

[0063] Sequence information regarding the degree of edge placement error is established. Through the established timing information, the overall change trend of the polygon contour can be captured, including the deviation scale and direction, thereby providing information for the movement optimization of each edge.

[0064] Through a machine learning method of multi-scale information fusion, with segmented line segments as the basic unit, the manufacturability of the mask is ensured, and the time efficiency in the mask optimization process is improved, while having both manufacturability and timeliness. Brief Description of the Drawings

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

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

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

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

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

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

[0071] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0072] Embodiment

[0073] Such as Figures 1 to 6 , a multi-scale information fusion method for lithography mask correction, comprising 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] Set the minimum line segment length, divide the edges of the polygons on the mask, and obtain a set of line segments S = {s1, s2, …… s n}, where S represents the set of line segments, and s1, s2, …… s n represent the line segments in the set of line segments;

[0076] Each line segment is represented by the starting point coordinates and the ending point coordinates as s n = ((x startn , y startn ), (x endn , y endn ))), where (x startn , y startn ) represents the starting point coordinates, and (x endn , y endn ) represents the ending point coordinates; and record the length L of the line segment, the line segment direction and the ratio R of the line segment length to the edge length of the polygon where the line segment is located; obtain the line segment unit of each line segment and the information of the line segment unit; the line segment unit can be moved along its normal direction.

[0077] By adjusting the moving distance of each line segment, the optimization of the mask can be completed.

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

[0079] For example Figure 2 , extract the geometric center of the Manhattan rectangle of the polygon on the mask as the central node, convert the central node of each polygon into a hyperedge, and convert 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 polygon, the distance between the polygons, and the encoding of the direction between the polygons; through the above processing, the converted hyperedges, edge nodes, and the initial features of the edge nodes can form a hypergraph G H = {V, E};

[0080] where V = {v1, v2, …… v n}, represents the set of edge nodes converted from the directed edges between the polygons, E = {e1, e2, …… e n}, represents the set of hyperedges converted from the polygons, and the initial feature of the edge node is X ∈ R n×d , where R represents the real number field, n is the number of rows of the matrix, and d is the number of columns of the matrix;

[0081] Then the topological structure of the hypergraph is represented by the incidence matrix Expressed as:

[0082]

[0083] Wherein, v i represents the i-th node v i , e j represents the j-th hyperedge e j , that is, if the node e j belongs to the hyperedge e j , then is 1, otherwise it is 0.

[0084] Encode the direction between polygons through a combination of the numbers 0 or 1.

[0085] Perform embedding learning on the hypergraph G H through a dual hypergraph attention network. The specific method includes:

[0086] Transfer the information of the edge nodes to the hyperedges, and then converge the information on the hyperedges back to the edge nodes as one layer; then transfer the information of the edge nodes to the hyperedges, and the embedding of the hyperedge e j at the l-th layer is expressed as:

[0087]

[0088] Wherein, represents the embedding of the node v i at the l-1 layer, and α E (e j , v i ) represents the attention coefficient of each edge node v j on the hyperedge e i . The calculation method is:

[0089]

[0090] Wherein, exp() represents the exponential function, which is used to map the value to the positive range and is used to ensure non-negative weights in the attention mechanism; a2 represents a learnable attention vector, which is a model parameter; T represents the transpose operation, which is used for dot product operation. W N represents the learnable weight matrix before aggregation, and b N represents the bias vector before aggregation;

[0091] Then, converge the information on the hyperedges back to the edge nodes according to the attention coefficient of the hyperedges, that is:

[0092]

[0093] Wherein, Denote the embedding of the hyperedge \(e\) in the \((l - 1)\)-th layer as \(\alpha\) j j of N N (e j j , v i i ) represents the aggregated edge node \(v\) i i For each hyperedge \(e\) j j of, the attention coefficient is calculated as follows:

[0094]

[0095]

[0096] where \(\exp()\) represents the exponential function, which is used to map values to the positive range and is used in the attention mechanism to ensure non-negative weights; \(a_3\) represents a learnable attention vector, which is a model parameter; \(T\) represents the transpose operation, which is used for dot product operations. \(W\) E E represents the learnable weight matrix after aggregation, and \(b\) E E represents the bias vector after aggregation;

[0097] Through the above-designed \(l\)-layer dual hypergraph attention network, that is, the hypergraph neural network learning framework, finally, the embedding information \(h\) E E L (e) of each polygon can be learned, which represents the interaction information between each polygon and its surrounding polygons. Then, through the index information of the line segment unit saved in \(S1\) and the polygon where it is located, the embedding information of the polygon is copied to obtain i.e., the global optical information, where \(R\) represents the real number field, \(n\) is the number of rows of the matrix, and \(d\) h h is the number of columns of the matrix. To endow the embedding information of the polygon where each line segment is located.

[0098] The line segments of the divided polygons are represented by the nodes at their centers. Taking the nodes of the line segments at the corners of the polygon as the centers, an interaction graph of the line segments at the corners of the polygon is constructed, and a convolutional neural network is used to perform embedding learning on the interaction graph of the line segments at the corners of the polygon to obtain the optical local neighborhood information; here, it is mainly to capture the optical proximity influence at the corners of the polygon, specifically including:

[0099] For the line segments of the straight edges of the polygon, they are represented by the nodes at the centers of the line segments; for the corner line segments at the corners of the polygon, the centers of the line segments at both ends of the polygon corner points are taken respectively as the node representations; taking the nodes of the corner line segments of the polygon as the centers, the nodes within the diameter range \(r\) are connected. Considering the partial coherence effect of light, the value of \(r\) is:

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

[0101] Among them, λ represents the wavelength of light, NA represents the numerical aperture, and sigma max represents the maximum normalized radius. When sigma max equals 1, it represents incoherent light;

[0102] For example Figure 3 , construct the line segment interaction graph G at the polygon corners g ; among them, the initial features of the nodes include: the ratio R of the line segment length to the edge length of the polygon where it is located, the line segment direction, and the judgment identifier of the corner line segment; the initial feature of the edge between nodes is the Manhattan distance between nodes.

[0103] Then use the convolutional graph neural network to perform embedding learning on the line segment interaction graph G at the polygon corners g , that is:

[0104]

[0105] Among them, represents the graph adjacency matrix after adding a self-loop edge to each node, where represents the degree of each node;

[0106] Learn to obtain that is, the optical local proximity information of each line segment. Among them, R represents the real number field, n is the number of rows of the matrix, and d h is the number of columns of the matrix.

[0107] Use the number 0 or 1 as the judgment identifier to judge whether it is a corner line segment.

[0108] By performing lithography simulation on the initial mask, constructing an error sequence, and substituting it into the sequence model for learning, the polygon contour change information is obtained; here, it is mainly to capture the overall contour change information of the polygon. Specifically include:

[0109] Perform lithography simulation on the initial mask to obtain a lithography pattern. If the lithography pattern is distorted, the distorted part may be a bulge or a depression; select the midpoint of the line segment unit as the detection point, and set a threshold th EPE , if the vertical distance from the detection point to the lithography pattern exceeds the threshold, a violation occurs, that is:

[0110]

[0111] Among them, EPE violation(x,y) represents the edge placement error violation at the coordinate (x,y). A value of 1 indicates that a violation has occurred, and a value of 0 indicates no violation. L2(x,y) represents the two-dimensional Euclidean distance between the lithography pattern and the lithography simulation pattern at the coordinate (x,y).

[0112] When a series of thresholds are set, for the midpoints of the line segments, there is F i = ±∑EPE violations, where ± represents the protrusion or depression of the lithography pattern, F i represents the total number of mask edge placement errors in the i-th region. i corresponds to different protrusion or depression regions, and ∑EPE violations represents the sum of the total number of all edge placement errors.

[0113] Such as Figure 4 , and then starting from the midpoint of the line segment at the lower right corner of the polygon, measure clockwise to obtain the sequence information of the contour change of the corresponding polygon, construct an error sequence, and substitute it into the LSTM 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 of the matrix, and d h is the number of rows of the matrix.

[0114] The entire S2 mainly models the multi-scale information on the mask and performs embedding learning by designing a neural network to capture the optical system information.

[0115] S3. Use a multi-layer perceptron for information integration, predict the moving distance of each line segment of the polygon, and perform rounding and normalization processing; such as Figure 5 , S3 mainly synthesizes the multi-scale information learned in S2.

[0116] Input the global optical information, optical local adjacent information, and polygon contour change information obtained in S2 into the multi-layer perceptron for learning, predict the moving distance L M of each line segment of the polygon along its normal direction, and then perform rounding and normalization, that is where Round() represents the rounding-down operation; then all line segments move along their normal directions to reduce the gap between the lithography pattern and the target pattern; it should be noted that the movement of the line segments at the corners may cause the line segments to intersect, then the starting and ending point coordinates of the line segments need to be readjusted. Calculate and update the reverse gradient, and the reverse gradient refers to a specific way of gradient calculation or use for optimizing the model or mask parameters.

[0117] The Multilayer Perceptron (MLP) is a classic artificial neural network model widely used in the fields of machine learning and deep learning. It is a feedforward neural network composed of multiple neuron layers, including an input layer, hidden layers, and an output layer, and is used to handle non-linear problems such as classification and regression.

[0118] S4. Re-convert the set of post-movement line segments into a mask pattern; by performing parallel ray projection on the line segments, convert the set of line segments into a mask pattern.

[0119] The lithography simulation model only accepts mask inputs at the pixel level. Lithography simulation or lithography emulation means the same thing. Therefore, it is necessary to pixelize the moved line segments and convert the line segment set into a mask at the pixel level, and this process needs to be differentiable to allow the gradient to flow from the pixel mask to the line segments, thereby updating the parameters that control the movement of the line segments. For example, Figure 6 , here use Figure 6 the method 1 shown in

[0120] to perform parallel ray projection on the edges to convert the line segment set into a pixel mask.

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

[0122] Common examples can use the coherent system and solution of the 193nm wavelength system as the optical model for lithography modeling. The spatial light image intensity I is represented by the convolution H of the pixel mask M and a set of optical kernels. Perform an N k th order approximation on the coherent system, and the specific formula is expressed as:

[0123]

[0124] where, represents the convolution operation, h i represents the i-th optical kernel of the decomposition, and σ i represents the corresponding weight of the coherent system; after obtaining the spatial light intensity image I of the lithography simulation, use the anti-corrosion photoresist with a constant threshold I th as the reaction model to evaluate the lithography pattern Z, that is:

[0125]

[0126] where, I(x, y) represents the light intensity emitted at the coordinate (x, y). When the light intensity I(x, y) is greater than the set threshold Ith When it is [time], photolithography is performed at the coordinates (x, y), and Z represents the pattern after final photolithographic etching;

[0127] Then, the difference Loss(Z,T) between the simulated photolithography model and the target pattern is calculated through the square difference error, and a gradient is formed for backpropagation, that is:

[0128]

[0129] Among them, Loss() represents the loss function, and T represents the target photolithography pattern.

[0130] The present invention provides a multi-scale information fusion method for photolithography mask correction. By establishing a high-order mask polygon interaction network, a global view is provided for each divided line segment; a line segment interaction graph at the corners of the mask polygon is established, so as to provide more fine-grained information for the movement adjustment of the line segments at the corners; sequence information about the degree of edge placement error is established, so as to provide information for the movement optimization of each edge; taking the segmented line segment as the basic unit, the manufacturability of the mask is ensured, and at the same time, the time efficiency of the mask optimization process is improved.

[0131] According to the disclosure and teachings of the above specification, those skilled in the art to which the present invention pertains can also 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. In addition, although some specific terms are used in this specification, these terms are only for convenience of description and do not constitute any limitation to the present invention.

Claims

1. A multi-scale information fusion method for lithography mask correction, characterized in that The following steps are involved: S1, extracting the edges of the polygons on the mask and dividing them into line segments to obtain line segment units, and saving the information of the line segment units; S2, extract the polygons on the mask, convert the polygons and the directed edges between the polygons to obtain hyperedges and edge nodes, and construct a hypergraph. Use the dual hypergraph attention network to embed the hypergraph and obtain global optical information. The line segments of the divided polygons are represented by the nodes at their centers, and the line segment interaction graph at the corners of the polygons is constructed with the nodes of the line segments at the corners of the polygons as the center. The line segment interaction graph at the corners of the polygons is embedded and learned using a convolutional neural network to obtain the optical local proximity information. By simulating the initial mask through lithography, the error sequence is constructed and substituted into the sequence model for learning to obtain the polygon contour change information; S3, use a multi-layer perceptron to integrate information, predict the moving distance of each line segment of the polygon, and perform rounding and normalization processing; S4, converting the set of moved line segments back into a mask pattern; S5. Perform photolithography simulation on the pattern of the mask and calculate the error according to the loss function to form a gradient for back propagation.

2. The multi-scale information fusion method for lithography mask correction according to claim 1, wherein In the above S1, set the line segment length, divide the edges of the polygon on the mask, and obtain a line segment set S = {s1, s2,... s n}, where S represents the line segment set, and s1, s2,... s n represent the line segments in the line segment set; Each line segment is represented by the starting point coordinates and the ending point coordinates as s n = ((x startn , y startn ), (x endn , y endn ))), where (x startn , y startn ) represents the starting point coordinates, and (x endn , y endn ) represents the ending point coordinates; and record the length L of the line segment, the line segment direction and the ratio R of the line segment length to the edge length of the polygon where the line segment is located; Get the line segment unit and line segment unit information of each line segment; Line segment elements can be moved along their normal direction.

3. A multi-scale information fusion method for lithography mask correction according to claim 1, characterized in that In S2, obtaining global optical information specifically includes: Extract the geometric center of the Manhattan rectangle of the polygon on the extraction mask as the central node, convert the central node of each polygon into a hyperedge, and convert the directed edge between polygons into an edge node; the initial features of the edge node include: the total number of line segments obtained by dividing the polygon, the distance between polygons, and the encoding of the direction between polygons; the converted hyperedges, edge nodes, and the initial features of the edge nodes constitute a hypergraph G H ={V, E}; where \(V = \{v_1, v_2, \ldots v\) n \}\), representing the set of edge nodes converted from the directed edges between polygons, \(E=\{e_1, e_2, \ldots e\) n \}\), representing the set of hyperedges converted from polygons, the initial feature of the edge node is \(X\in\mathbb{R}\) n×d where \(\mathbb{R}\) represents the real number field, \(n\) is the number of rows of the matrix, and \(d\) is the number of columns of the matrix; Then the topological structure of the hypergraph is represented by the incidence matrix as follows: Among them, v i represents the i-th node v i , e j represents the j-th hyperedge e j , that is, if the node e j belongs to the hyperedge e j , then is 1, otherwise it is 0.

4. A multi-scale information fusion method for lithography mask correction according to claim 3, characterized in that, Embedding learning is performed on the hypergraph G through a dual hypergraph attention network H The specific method includes: Transfer the information of edge nodes to hyperedges, and then aggregate the information on hyperedges back to edge nodes as one layer; then transfer the information of edge nodes to hyperedges, and the hyperedge \(e\) at the \(l\)th layer j Embedding is expressed as: Among them, represents the embedding of the node v at the (l - 1)-th layer i , and α E (e j , v i ) represents the attention coefficient for each edge node v j on the hyperedge e i , and the calculation method is as follows: Among them, exp() represents the exponential function, which is used to map values to the positive range and is used in the attention mechanism to ensure non - negative weights; a2 represents a learnable attention vector, which is a model parameter; T represents the transpose operation, which is used for dot - product operations. W N represents the learnable weight matrix before pooling, b N represents the bias vector before pooling; Then the information on the hyperedge is gathered back to the edge node according to the attention coefficient of the hyperedge, that is: Among them, represents the embedding of the (l-1)-th layer hyperedge e j , and 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: Among them, exp() represents the exponential function, which is used to map values to the positive range and is used in the attention mechanism to ensure non - negative weights; a3 represents a learnable attention vector, which is a model parameter; T represents the transpose operation, which is used for dot - product operations. W E represents the learnable weight matrix after pooling, b E represents the bias vector after pooling; Through the l-layer dual hypergraph attention network, the embedding information h of each hyperedge is obtained E L (e); then, through the index information of the line segment unit saved in S1 and the polygon where it is located, the embedding information of the polygon is copied to obtain That is, the global optical information, where R represents the real number field, n is the number of rows of the matrix, and d h is the number of columns of the matrix.

5. A multi-scale information fusion method for lithography mask correction according to claim 2, characterized in that In S2, for the straight edge line segments of the polygon, the nodes at the center of the line segments are used to represent them; for the corner line segments at the corners of the polygon, the centers of the line segments at both ends of the corner points of the polygon are used to represent them as nodes; with the nodes of the corner line segments of the polygon as the center, the nodes within the diameter range r are connected, and the value of r is: r = 20·(λ / NA) / (1 + sigma max ); where λ represents the wavelength of light, NA represents the numerical aperture, and sigma max represents the maximum normalized radius. When sigma max equals 1, it represents incoherent light; Construct the interaction graph G of line segments at the corners of the polygon g ; among them, the initial features of the nodes include: the ratio R of the line segment length to the edge length of the polygon where it is located, the line segment direction, and the judgment identifier of the corner line segment; the initial feature of the edge between nodes is the Manhattan distance between the nodes.

6. A multi-scale information fusion method for lithography mask correction according to claim 5, characterized in that Use a convolutional graph neural network to perform embedding learning on the line segment interaction graph G at the polygon corners, that is: g Perform embedding learning, namely: Among them, represents the graph incidence matrix after adding a self-loop edge to each node, where represents the degree of each node; Learned That is, the optical local neighborhood information of each line segment, where R represents the real number field, n is the number of rows of the matrix, and d h is the number of columns of the matrix.

7. A multi-scale information fusion method for lithography mask correction according to claim 1, characterized in that, In S2, perform lithography simulation on the initial mask to obtain a lithography pattern. If the lithography pattern is distorted, the distorted part is a protrusion or a depression; select the midpoint of the line segment unit as the detection point and set a threshold th EPE , if the vertical distance from the detection point to the lithography pattern exceeds the threshold, a violation occurs, that is: Where, EPE violation (x, y) represents the edge placement error violation at the coordinate (x, y), a value of 1 indicates a violation occurs, and a value of 0 indicates no violation, L2 (x, y) represents the two-dimensional Euclidean distance between the lithography pattern and the lithography simulation pattern at the coordinate (x, y); When multiple thresholds are set, for the midpoint of the line segment unit, there is F i = ±∑EPE violations, where ± indicates the protrusion or depression of the lithographic pattern, and F i represents the total number of mask edge placement error violations in the i-th region, i corresponds to different protrusion or depression regions, and ∑EPE violations represents the summation of the total number of all edge placement error violations; Then, starting from the midpoint of the line segment unit at the lower right corner of the polygon, measure clockwise or counterclockwise to obtain the sequence information of the contour change of the corresponding polygon, construct an error sequence, and substitute it 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 of the matrix, and d h is the number of columns of the matrix.

8. A multi-scale information fusion method for lithography mask correction according to claim 1, characterized in that In S3, the global optical information, optical local adjacent information, and polygon contour change information obtained in S2 are input into a multi-layer perceptron for learning to predict the moving distance L of each line segment of the polygon along its normal direction. M , and then rounding normalization is performed, that is, where Round() represents the floor operation; then all line segments move along their normal directions to reduce the gap between the lithography pattern and the target pattern; calculate and update the reverse gradient, which is used to optimize the model or mask parameters.

9. A multi-scale information fusion method for lithography mask correction according to claim 1, characterized in that In S4, the set of line segments is converted into a mask pattern by performing parallel ray projection on the line segments.

10. A multi-scale information fusion method for lithography mask correction according to claim 1, characterized in that In S5, a coherent system solution using a wavelength system is used as an optical model for lithography modeling. The spatial light image intensity I is represented by the convolution H of the mask M of pixels and a set of optical kernels. An N k th -order approximation is performed on the coherent system, and the specific formula is expressed as: Among them, represents a convolution operation, and h i represents the i-th optical core of the decomposition, and σ i represents the corresponding weight of the coherent system; after obtaining the spatial light intensity image I of the lithography simulation, using a constant-threshold anti-corrosion photoresist with I th as the reaction model, the lithography pattern Z is evaluated, that is: Among them, I(x, y) represents the light intensity emitted at the coordinates (x, y). When the light intensity I(x, y) is greater than the set threshold I th then photolithography is performed at the coordinates (x, y), and Z represents the final pattern after photolithography and etching; Then the difference Loss(Z,T) between the simulated lithography model and the target pattern is calculated by the square error to form a gradient for back propagation, that is: Among them, Loss() represents the loss function, and T represents the target lithography pattern.

Images