Rapid photoetching thick mask diffraction field calculation method and system

Abstract

The invention discloses a rapid photoetching thick mask diffraction field calculation method and system. A deep learning network with light source point coordinates and a thick mask pattern as input and a mask diffraction near field as output is constructed to complete prediction of a thick mask diffraction near field related to a whole light source. An existing deep learning-based method needs to establish a large number of deep networks according to light source sampling points, and according to the method, light source point coordinate information and thick mask graph fusion are integrated into an existing deep learning model, so that prediction precision similar to that of the existing method can be obtained within shorter prediction time by using fewer network parameters. According to the method, the calculation efficiency of the diffraction field of the photoetching thick mask can be improved, and the production efficiency of photoetching production under advanced nodes is improved.

Description

Technical Field

[0001] The present invention relates to the fields of thick mask models, integrated circuit manufacturing, and computational lithography, and particularly relates to a fast computational method and system for the diffraction field of a lithographic thick mask. Background Art

[0002] Lithography technology is one of the core processes in integrated circuit manufacturing and is a pattern transfer technology. In this technology, the circuit layout to be printed is first engraved on a mask, and then the mask is irradiated with a short-wavelength light source. The transmitted light is collected by a projection objective and projected onto a silicon wafer coated with photoresist on the surface. After processes such as exposure, development, and etching, the circuit layout is replicated on the silicon wafer.

[0003] Extreme ultraviolet (EUV) lithography is one of the most advanced chip mass production process technologies currently. Compared with deep ultraviolet lithography, EUV lithography can adopt a simpler process flow to achieve advanced process nodes of 7nm - 3nm and below. The EUV lithography system uses a reflective mask and optical system. Among them, the total thickness of the EUV mask reaches several hundred nanometers, which is several times the exposure wavelength of the EUV lithography system. This will lead to a significant thick mask effect (also known as the three-dimensional mask effect), that is, the EUV light emitted by the light source is obliquely incident on the mask surface, and the absorption layer with a high aspect ratio will generate a shadow on the mask, thereby changing the diffraction near-field (DNF) distribution formed after the EUV light passes through the mask, and ultimately affecting the imaging result. Therefore, when conducting research on EUV lithography modeling and related algorithms, the thick mask effect must be considered, otherwise serious deviations will occur in the calculation results.

[0004] The computational method for the thick mask diffraction field aims to accurately calculate the DNF of the thick mask by simulating and analyzing the propagation and interaction of light in the thick mask during the lithography process. It has gone through three development stages: the computational method based on strict electromagnetic field simulation, the computational method based on semi-strict electromagnetic field simulation, and the computational method based on machine learning and deep learning. Among them, the computational method for the thick mask DNF based on deep learning obtains the mapping relationship between the thick mask and the DNF by learning a large amount of historical data. This method has both computational speed and accuracy, and is very suitable for the simulation and iterative optimization of large-scale masks; at the same time, the characteristic of being able to calculate curved-edge masks also makes it more suitable for advanced lithography process nodes. This method has currently become an important method in related fields.

[0005] Deep learning-based methods also have some deficiencies. In existing fast calculation methods, generally only a specific light source point (corresponding to parallel light incident on the mask at a specific angle) can be used for deep learning modeling to predict the thick mask DNF corresponding to this specific light source point. Once the position of the light source point changes significantly, the network needs to be rebuilt and trained. This is because the diffraction spectrum of the mask changes with the position of the illumination light source point, and this change cannot be expressed as a clear functional relationship. The lithography system uses partially coherent illumination, and partially coherent illumination can be decomposed into numerous discrete light source points. Thus, even if a single light source point is used as a representative in the adjacent light source illumination area, many networks need to be established to complete the prediction of the thick mask DNF involving the entire light source, which greatly affects the calculation efficiency.

[0006] In summary, it is necessary to invent a deep learning method that fuses the mask pattern and the light source point coordinates, so that it can predict the thick mask DNF corresponding to different illumination light source points and solve the application problems of the thick mask diffraction field calculation method in advanced lithography imaging and iterative optimization. Summary of the Invention

[0007] In view of this, the present invention provides a fast calculation method and system for the diffraction field of a thick lithography mask, which can predict the thick mask DNF corresponding to different illumination light source points, solve the problem that the current deep learning-based thick mask DNF calculation method needs to establish multiple deep learning networks for different illumination light source points to complete the prediction of the thick mask DNF involving the entire light source, and improves the operation efficiency.

[0008] To achieve the above object, a technical solution provided by an embodiment of the present invention - a fast calculation method for the diffraction field of a thick lithography mask includes the following steps:

[0009] S1: Set relevant parameters according to the actual usage scenario, including exposure parameters, mask thickness, and boundary conditions of the mask pattern.

[0010] S2: Establish a curvilinear mask DNF data set, build a corresponding DNF data set under a specific technology node, and the applicable range of the diffraction field calculation method depends on the built DNF data set; the data set contains the binary matrix or picture M corresponding to each mask pattern, the illumination light source point (x, y) corresponding to each mask pattern, and its corresponding diffraction near field E; considering the polarization state of the light source, the DNF of the mask is represented by 4 complex diffraction matrices, denoted as E UV , where U = X or Y, V = X or Y; the meaning of E UV is the complex amplitude generated by the polarization of the unit amplitude electric field incident in the V direction in the U direction.

[0011] S3: Using the illumination light source points and the mask patterns as inputs and the DNF as the output, construct a deep learning network and train it on the simulated DNF dataset to obtain a pre-trained deep learning network.

[0012] S4: For any given curved mask, perform preprocessing and then input it into the pre-trained deep learning network to calculate the DNF of the arbitrarily given curved mask.

[0013] Furthermore, establish a curved mask DNF dataset. The specific method is as follows:

[0014] On the normalized coordinate system of the light source plane, select light source sampling points according to certain rules. Each light source sampling point coordinate (x, y) corresponds to a parallel light incident on the thick mask at a specified angle; (x, y) satisfies x 2 +y 2 ≤ 1.

[0015] Collect typical mask layouts of specified process nodes and lithography layers.

[0016] Subsequently, use the rigorous electromagnetic field simulation method and, according to the specified simulation parameters, simulate each mask layout for different incident angles to calculate each component E of the corresponding DNF.

[0017] Furthermore, using the illumination light source points and the mask patterns as inputs and the DNF as the output, construct a deep learning network. Specifically:

[0018] The network consists of two structures, and their functions are respectively: fusing the mask pattern information and the illumination light source coordinate information to form an intermediate result, and then calculating the DNF from the intermediate result through an end-to-end network; this process is expressed as:

[0019] M f = F[M, (x, y)]

[0020]

[0021] where M represents the input curved mask; (x, y) represents the sampling light source point coordinates; F[·] represents the fusion function of the mask pattern and the coordinates; M f represents the intermediate result formed by fusion; G[·] represents the end-to-end network that calculates the DNF from the intermediate result; represents the estimated value of the DNF or each component of the DNF.

[0022] Furthermore, train on the simulated DNF dataset to obtain a pre-trained deep learning network:

[0023] First, determine the cost function, which evaluates the prediction performance of the invented method for a given DNF dataset. Its error is expressed as the error between the true value E calculated by the strict electromagnetic field simulation method and the estimated value obtained by this method between:

[0024]

[0025] where ||·||2 represents the L2 norm.

[0026] Put the training set into the network for training, and then adjust the parameters through the performance of the network on the validation set to obtain the optimal trained network, which is the pre-trained deep learning network.

[0027] Furthermore, in S4, for any given curved-edge mask, perform preprocessing. The specific method of preprocessing is as follows:

[0028] Read the curved-edge mask to be calculated, and adjust the sampling interval of the mask to the sampling interval of the mask in the training set by means of interpolation, filling, or deleting 0 elements, and then adjust the matrix size of the mask to meet the input requirements of the deep learning network.

[0029] Another embodiment of the present invention also provides a fast lithography thick mask diffraction field calculation system, including a parameter setting module, a curved-edge mask DNF dataset module, a deep learning network, and a training module.

[0030] The parameter setting module is used to receive the input parameter setting values, which are the exposure parameters, mask thickness, and boundary conditions of the mask pattern set according to the actual usage scenario.

[0031] The curved-edge mask DNF dataset module is used to establish a curved-edge mask DNF dataset, build a corresponding DNF dataset under a specific technology node, and the applicable range of the diffraction field calculation method depends on the established DNF dataset; the dataset contains the binary matrix or picture M corresponding to each mask pattern, the illumination light source points (x, y) corresponding to each mask pattern, and their corresponding diffraction near fields E; considering the polarization state of the light source, the DNF of the mask is represented by 4 complex diffraction matrices, denoted as E UV , where U = X or Y, V = X or Y; the meaning of E UV is the complex amplitude generated by the unit amplitude electric field incident in the V direction and polarized in the U direction.

[0032] The deep learning network is constructed with the illumination light source points and the mask pattern as the input and the DNF as the output.

[0033] The training module is trained on the simulated DNF dataset to obtain a pre-trained deep learning network.

[0034] A preprocessing module, which, for any given curved-edge mask, after preprocessing, inputs the preprocessed mask pattern into a pre-trained deep learning network to calculate the DNF of any given curved-edge mask.

[0035] Furthermore, a curved-edge mask DNF dataset module that establishes a curved-edge mask DNF dataset. The specific method is as follows:

[0036] On the light source plane normalization coordinate system, light source sampling points are selected according to certain rules. Each light source sampling point coordinate (x, y) corresponds to a parallel light incident on the thick mask at a specified angle; (x, y) satisfies x 2 +y 2 ≤1.

[0037] Collect typical mask layouts of specified process nodes and photolithography layers.

[0038] Subsequently, using the rigorous electromagnetic field simulation method and based on the specified simulation parameters, simulate each mask layout incident at different angles to calculate each component E of the corresponding DNF.

[0039] Furthermore, the deep learning network includes two structures, and their functions are respectively: fusing the mask pattern information and the illumination light source coordinate information to form an intermediate result, and then calculating the DNF from the intermediate result through an end-to-end network; this process is expressed as:

[0040] M f = F[M, (x, y)]

[0041]

[0042] where M represents the input curved-edge mask; (x, y) represents the sampling light source point coordinates; F[·] represents the fusion function of the mask pattern and the coordinates; M f represents the intermediate result formed by fusion; G[·] represents the end-to-end network that calculates the DNF from the intermediate result; represents the estimated value of the DNF or each component of the DNF.

[0043] Furthermore, train on the simulated DNF dataset to obtain a pre-trained deep learning network:

[0044] First, determine the cost function. The cost function evaluates the prediction performance of the invented method under the given DNF dataset. Its error is expressed as the error between the true value E calculated by the rigorous electromagnetic field simulation method and the estimated value obtained by this method :

[0045]

[0046] Among them, ||·||2 represents the L2 norm.

[0047] Put the training set into the network for training, and then adjust the parameters through the performance of the network on the validation set to obtain the optimal trained network, which is the pre-trained deep learning network.

[0048] Furthermore, in the preprocessing module, for any given curved-edge mask, preprocessing is performed. The specific method of preprocessing is as follows:

[0049] Read the curved-edge mask to be calculated, and adjust the sampling interval of the mask to the sampling interval of the mask in the training set by means of interpolation, filling, or deleting 0 elements, and then adjust the matrix size of the mask to meet the input requirements of the deep learning network.

[0050] Beneficial effects:

[0051] The present invention provides a fast method and system for calculating the diffraction field of a lithographic thick mask, which takes the fused mask pattern and the light source point coordinates together as the input of the deep learning network, aiming to replace a large number of networks that need to be established in the existing methods for DNF calculation involving the entire light source with a deep learning network. This method significantly reduces the parameters required by the network, improves the calculation efficiency, and helps to improve the actual production efficiency. Brief Description of the Drawings

[0052] Figure 1 It is a flowchart of a fast method and system for calculating the diffraction field of a lithographic thick mask provided by the present invention;

[0053] Figure 2 It is the light source sampling points of the training set and the test set light source sampling points provided by the present invention;

[0054] Figure 3 It is the mask layout and DNF example in the DNF dataset provided by the present invention;

[0055] Figure 4 It is the deep learning network model structure used in the embodiment provided by the present invention;

[0056] Figure 5 It is the cost function convergence curve of the network on the training set when training the deep learning network provided by the present invention for a specific mask dataset;

[0057] Figure 6 It is the comparison of the prediction accuracy of DNF at different light source points in the test set between the deep learning network model used in the embodiment provided by the present invention and two existing models. Detailed Embodiment

[0058] The following combines the drawings and gives embodiments to describe the present invention in detail.

[0059] Please refer to Figure 1 , Figure 1 which shows a flowchart of a fast calculation method for the diffraction field of a lithographic thick mask, specifically including the following steps:

[0060] S1: Set relevant parameters according to the actual usage scenario, including light source parameters, mask structure, boundary conditions of mask patterns, etc. In this embodiment, a mask layout with feature sizes of 12 nm - 15 nm is demonstrated. The illumination wavelength used is 13.5 nm; a non-doubly telecentric optical design is adopted, so there is a 6° angle between the incident principal ray and the optical axis. The lithographic mask mainly consists of a substrate, a multi-layer film, and an absorption layer. The thickness of the substrate is in millimeters and can be considered infinitely thick, and the material is silicon dioxide. The material of the absorption layer is tantalum nitride, with a thickness of 60 nm. The material of the multi-layer film is composed of multiple 3-nm molybdenum layers stacked with 4-nm silicon layers. The boundary conditions of the mask pattern are set to be spatially periodic.

[0061] S2: Establish a curvilinear mask DNF data set, build a corresponding DNF data set under a specific technology node, and the applicable range of the diffraction field calculation method depends on the established DNF data set. The data set contains the binary matrix or picture M corresponding to each mask pattern, the illumination light source points (x, y) corresponding to each mask pattern, and their corresponding diffraction near fields E. Considering the polarization state of the light source, the DNF of the mask can be represented by 4 complex diffraction matrices, denoted as E UV , where U = X or Y, V = X or Y. The meaning of E UV is the complex amplitude generated by the polarization of the unit amplitude electric field incident in the V direction in the U direction.

[0062] In step S2, the specific method for establishing the curvilinear mask DNF data set is as follows:

[0063] S201: Select light source sampling points on the normalized coordinate system of the light source plane according to certain rules. Each light source sampling point coordinate (x, y) corresponds to a parallel light incident on the thick mask at a specified angle. (x, y) should satisfy x 2 +y 2 ≤1. In this embodiment, a total of nine light source points are selected to make a training set, and the positions of the above sampling points are as shown by the circular points in Figure 2 .

[0064] S202: Collect typical mask layouts of the specified process node and lithography layer. Specifically, in the process node for DNF calculation, some representative mask patterns can be selected, such as square holes, lines, L-shaped patterns, etc. Then, the obtained set of curvilinear masks is divided into a training set and a test set according to a ratio of 2:1. Figure 3 (a) and Figure 3(b) shows some mask layout examples in two datasets respectively.

[0065] S203: Using the strict electromagnetic field simulation method, the strict electromagnetic field simulation method referred to in the embodiments of the present invention refers to the electromagnetic simulation method using Maxwell's calculation, and simulates the mask layouts in each training set according to the simulation parameters specified in step S1 and the illumination light source points specified in step S201, and calculates each component E of the corresponding DNF. Figure 3 (c) shows the real and imaginary parts of the 4 complex diffraction matrices of a curvilinear mask DNF in the example.

[0066] S3: Design the network model structure, determine the network depth, and train on the simulated DNF dataset. Train a deep learning network with the illumination light source points and mask patterns as inputs and the DNF as the output.

[0067] In step S3, the design and training method of the deep learning network with the illumination light source point coordinates and mask patterns as inputs and the DNF as the output is as follows:

[0068] S301: The network includes two structures, and their functions are respectively: fusing the mask pattern information and the illumination light source coordinate information to form an intermediate result, and then calculating the DNF from the intermediate result through an end-to-end network. This process can be expressed as:

[0069] M f = F[M, (x, y)]

[0070]

[0071] Among them, M represents the input curvilinear mask. (x, y) represents the sampling light source point coordinates. F[·] represents the fusion function of the mask pattern and coordinates. In this embodiment, first, stack the mask pattern and the light source coordinates to form three-channel input data: the first channel of the input data is still the mask pattern; the second channel is a matrix with the same size as the mask pattern and all elements being the abscissa x of the light source point; the third channel is a matrix with the same size as the mask pattern and all elements being the ordinate y of the point light source. After obtaining the three-channel input data, then realize the fusion of the mask pattern and coordinates through three-dimensional convolution to form the intermediate result M f . G[·] represents the end-to-end network for calculating the DNF from the intermediate structure. In this embodiment, it is realized through a U-shaped network (abbreviated as U-Net). At this time, represents the estimated value of the curvilinear mask DNF, which is one of the real and imaginary parts of the 4 complex diffraction matrices. The deep learning network model structure used in this embodiment is as Figure 4 shown. It should be noted that F[·] and G[·] can be realized by different functions.

[0072] S302: Determine the cost function, which evaluates the prediction performance of the invented method for a given DNF dataset. The error is expressed as the error between the true value E calculated by the strict electromagnetic field simulation method and the estimated value obtained by this method: where:

[0073]

[0074] Here, ||·||2 represents the L2 norm. It should be noted that the embodiments of the present invention are not limited to specifying the cost function as the function shown above, but there can be various variations. For example, the Mean Squared Error (MSE) or the Mean Absolute Error (MAE), etc., can be set as the cost function or a part of the cost function.

[0075] S303: Put the training set obtained in step S2 into the network for training and debug the parameters to obtain the optimal trained network. In this embodiment, the network results corresponding to the 4 more important components (i.e., the real and imaginary parts of E XX and E YY 's real and imaginary parts, and the values of these 4 components are much larger than the remaining 4 components) among the real and imaginary parts of the 4 complex diffraction matrices of DNF (a total of 8 components) are shown. The network was optimized by 500 iterations respectively, and the convergence curve of the obtained cost function is as shown in Figure 5 Figure.

[0076] S4: Calculate the arbitrarily given curvilinear mask DNF. For an arbitrarily given curvilinear mask, after preprocessing, it is input into the deep learning network pre-trained by the present invention, and the arbitrarily given curvilinear mask DNF can be calculated. In this embodiment, the DNF of the mask layout in each test set is calculated when the light source coordinates are (0,0), (0,0.3), (0,0.5), (0,0.6), (0,0.7), and (0,0.9), and the positions of the above sampling points are as shown by the red crosses in Figure 2 Figure.

[0077] In step S4, the specific method for calculating the curvilinear mask DNF in the test set is as follows:

[0078] S401: Preprocess the data in the test set. Read the curvilinear mask to be calculated, and adjust the sampling interval of the mask to the sampling interval of the mask in the training set by means of interpolation, filling, or deleting 0 elements, and then adjust the matrix size of the mask to meet the input requirements of the network.

[0079] S402: Input the preprocessed mask data and the light source point coordinates into the trained deep learning network model to obtain the DNF formed by the mask at this illumination angle. Figure 6 For the real and imaginary parts of E in the DNF XX and E YY The calculation results of the four components are shown, and the relative RMSE between the method proposed in the present invention and the rigorous electromagnetic field simulation method is calculated, defined as:

[0080]

[0081] where E and respectively represent the DNFs obtained by the rigorous electromagnetic field simulation method and the method proposed in the present invention, R and C represent the number of rows and columns of the test mask matrix, and max(E) is the maximum absolute value among the true values obtained by the rigorous electromagnetic field simulation method.

[0082] Figure 6 At the same time, the curves of the real and imaginary parts corresponding to the method proposed in the present invention, the "nearby substitution method" (that is, first train the deep learning network for some reference light source points, and then use the DNF of the reference point closest to the light source point to be calculated as the DNF of the target point), and the "near point combination method" (that is, first train the deep learning network for some reference light source points, and then calculate the DNF of this point by weighted summation of the DNFs of several reference points closest to the light source point to be calculated according to the distance) are shown. The deep learning models involved in the two existing methods are both U-Net, and the network structure is the same as the U-Net in the method proposed in the embodiment (that is, the U-Net involved in step S301 of the embodiment). Therefore, the network parameters required for the two existing methods are approximately 9 times that of the proposed method. Figure 6 shows the curves of the real and imaginary parts corresponding to the method proposed in the present invention, the "nearby substitution method", and the "near point combination method".

[0083] From Figure 6 it can be seen that since the relative RMSE of the diffraction matrix of the method proposed in the present invention is close to 4% and the relative RMSE of the "near point combination method" and is significantly lower than that of the "nearby substitution method", it has high accuracy in the DNF prediction task. In addition, the average prediction time of the deep learning network constructed by the proposed method on the test set is 0.0116 s, and the average prediction time of a single U-Net in the two existing common methods is 0.0110 s. Then, if you want to complete the prediction of the thick mask DNF involving the entire light source, the running time of the two existing methods is about 9 times that of the proposed method, which shows that the proposed method significantly improves the operation efficiency.

[0084] Another embodiment of the present invention further provides a fast lithography thick mask diffraction field calculation system, including a parameter setting module, a curved mask DNF dataset module, a deep learning network, and a training module;

[0085] The parameter setting module is used to receive the input parameter setting values, which are the exposure parameters, mask thickness, and boundary conditions of the mask pattern set according to the actual usage scenario;

[0086] The curved mask DNF dataset module is used to establish a curved mask DNF dataset, build a corresponding DNF dataset under a specific technology node, and the applicable range of the diffraction field calculation method depends on the established DNF dataset; the dataset contains the binary matrix or picture M corresponding to each mask pattern, the illumination light source points (x, y) corresponding to each mask pattern, and their corresponding diffraction near fields E; considering the polarization state of the light source, the DNF of the mask is represented by 4 complex diffraction matrices, denoted as E UV , where U = X or Y, V = X or Y; E UV means the complex amplitude generated by the unit amplitude electric field incident in the V direction and polarized in the U direction;

[0087] The deep learning network is constructed with the illumination light source points and the mask pattern as inputs and the DNF as the output;

[0088] The training module is trained on the simulated DNF dataset to obtain a pre-trained deep learning network;

[0089] The preprocessing module is used to preprocess any given curved mask, and then input the preprocessed mask pattern into the pre-trained deep learning network to calculate the DNF of any given curved mask.

[0090] The curved mask DNF dataset module establishes a curved mask DNF dataset, and the specific method is as follows:

[0091] Select light source sampling points on the normalized coordinate system of the light source plane according to certain rules. Each light source sampling point coordinate (x, y) corresponds to a parallel light incident on the thick mask at a specified angle; (x, y) satisfies x 2 +y 2 ≤1;

[0092] Collect typical mask layouts of the specified process node and lithography layer;

[0093] Subsequently, use the rigorous electromagnetic field simulation method and simulate each mask layout at different incident angles according to the specified simulation parameters to calculate each component E of the corresponding DNF.

[0094] A deep learning network, comprising two structures, whose functions are respectively: fusing mask pattern information and illumination light source coordinate information to form an intermediate result, and then calculating the DNF from the intermediate result through an end-to-end network; this process is expressed as:

[0095] M f = F[M, (x, y)]

[0096]

[0097] where M represents the input curved-edge mask; (x, y) represents the coordinates of the sampled light source point; F[·] represents the fusion function of the mask pattern and the coordinates; M f represents the intermediate result formed by fusion; G[·] represents the end-to-end network for calculating the DNF from the intermediate result; represents the estimated value of the DNF or each component of the DNF.

[0098] Train on the simulated DNF dataset to obtain a pre-trained deep learning network:

[0099] First, determine the cost function, which evaluates the prediction performance of the invented method under a given DNF dataset. Its error is expressed as the error between the true value E calculated by the strict electromagnetic field simulation method and the estimated value obtained by this method between:

[0100]

[0101] where ||·||2 represents the L2 norm;

[0102] Put the training set into the network for training, and then adjust the parameters through the performance of the network on the validation set to obtain the optimal trained network, which is the pre-trained deep learning network.

[0103] In the preprocessing module, for any given curved-edge mask, perform preprocessing. The specific preprocessing method is: read the curved-edge mask to be calculated, and adjust the sampling interval of the mask to the sampling interval of the mask in the training set by means of interpolation, filling, or deleting 0 elements, and then adjust the matrix size of the mask to meet the input requirements of the deep learning network.

[0104] In summary, the above is only a preferred embodiment of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A fast method for calculating the diffraction field of a thick mask for lithography, characterized in that: The steps include: S1: Set relevant parameters according to the actual usage scenario, including exposure parameters, mask thickness, and boundary conditions of the mask pattern; S2: Establish a curved edge mask DNF dataset. Build a corresponding DNF dataset under a specific technology node. The scope of application of the diffraction field calculation method depends on the constructed DNF dataset. The dataset contains a binary matrix or image M corresponding to each mask pattern, an illumination light source point (x, y) corresponding to each mask pattern, and its corresponding diffraction near field E. Considering the polarization state of the light source, the DNF of the mask is represented by four complex diffraction matrices, denoted as E UV , where U = X or Y, V = X or Y; E UV The meaning of is the complex amplitude generated by the incident unit amplitude electric field in the V direction polarized in the U direction; S3: constructing a deep learning network with the illumination light source point and the mask pattern as input and the DNF as output, and training on the simulated DNF data set to obtain a pre-trained deep learning network; S4: For any given curved edge mask, preprocessing is performed, and then the pre-trained deep learning network is input to calculate any given curved edge mask DNF.

2. A fast method for calculating the diffraction field of a thick mask for lithography as claimed in claim 1, characterized in that: The specific method of establishing the curved edge mask DNF data set is as follows: Select light source sampling points on the normalized coordinate system of the light source plane according to certain rules. Each light source sampling point coordinate (x, y) corresponds to a beam of parallel light incident on the thick mask at a specified angle. (x,y) satisfies x 2 +y 2 ≤1; Collect typical mask layouts for specified process nodes and photolithography layers; Then, a rigorous electromagnetic field simulation method is used to simulate each mask layout at different incident angles according to the specified simulation parameters, and each component E of the corresponding DNF is calculated.

3. A fast method for calculating the diffraction field of a thick lithography mask as claimed in claim 1, characterized in that: The deep learning network is constructed with the illumination light source point and the mask pattern as input and DNF as output, specifically: The network includes two structures, whose functions are: fusing the mask pattern information with the illumination light source coordinate information to form an intermediate result, and then calculating the DNF from the intermediate result through an end-to-end network; the process is expressed as: M f =F[M,(x,y)] Where M represents the input curved edge mask; (x, y) represents the coordinates of the sampled light source point; F[·] represents the fusion function of the mask pattern and coordinates; M f represents the intermediate result formed by fusion; G[·] represents the end-to-end network that calculates DNF from the intermediate result; Represents the estimated value of DNF or each component of DNF.

4. A fast method for calculating the diffraction field of a thick mask for lithography as claimed in claim 3, characterized in that: Training is performed on the simulated DNF dataset to obtain a pre-trained deep learning network: First, a cost function is determined, which evaluates the prediction performance of the invented method under a given DNF data set; its error is expressed as the difference between the true value E calculated by the rigorous electromagnetic field simulation method and the estimated value obtained by this method The error between: Among them, ||·||2 represents the L2 norm; The training set is put into the network for training, and then the parameters are debugged according to the performance of the network on the validation set to obtain the optimal trained network, which is the pre-trained deep learning network.

5. A fast method for calculating the diffraction field of a thick mask for lithography as claimed in claim 4, characterized in that: In S4, preprocessing is performed for any given curved edge mask, and the specific method of preprocessing is: The curved edge mask to be calculated is read, and the sampling interval of the mask is adjusted to the sampling interval of the mask in the training set by means of interpolation, padding or deletion of 0 elements, and then the matrix size of the mask is adjusted to meet the input requirements of the deep learning network.

6. A fast photolithography thick mask diffraction field calculation system, characterized in that: It includes parameter setting module, curved edge mask DNF data set module, deep learning network and training module; The parameter setting module is used to receive input parameter setting values, where the parameter setting values ​​are exposure parameters, mask thickness, and boundary conditions of mask patterns set according to actual usage scenarios; The curved edge mask DNF data set module is used to establish a curved edge mask DNF data set. A corresponding DNF data set is built under a specific technology node. The applicable scope of the diffraction field calculation method depends on the built DNF data set. The data set contains a binary matrix or image M corresponding to each mask pattern, an illumination light source point (x, y) corresponding to each mask pattern and its corresponding diffraction near field E. Considering the polarization state of the light source, the DNF of the mask is represented by four complex diffraction matrices, recorded as E UV , where U = X or Y, V = X or Y; E UV The meaning of is the complex amplitude generated by the incident unit amplitude electric field in the V direction polarized in the U direction; The deep learning network is constructed by taking the illumination light source point and the mask pattern as input and taking DNF as output; The training module performs training on the simulated DNF data set to obtain a pre-trained deep learning network; The preprocessing module is used for inputting the preprocessed mask pattern into a pre-trained deep learning network for calculating any given curved edge mask DNF after preprocessing.

7. A fast photolithography thick mask diffraction field calculation system as claimed in claim 6, characterized in that: The curved edge mask DNF data set module establishes a curved edge mask DNF data set, and the specific method is as follows: Select light source sampling points on the normalized coordinate system of the light source plane according to certain rules. Each light source sampling point coordinate (x, y) corresponds to a beam of parallel light incident on the thick mask at a specified angle. (x,y) satisfies x 2 +y 2 ≤1; Collect typical mask layouts for specified process nodes and photolithography layers; Then, a rigorous electromagnetic field simulation method is used to simulate each mask layout at different incident angles according to the specified simulation parameters, and each component E of the corresponding DNF is calculated.

8. A fast photolithography thick mask diffraction field calculation system as claimed in claim 6, characterized in that: The deep learning network includes two structures, whose functions are: fusing the mask pattern information with the illumination light source coordinate information to form an intermediate result, and then calculating the DNF from the intermediate result through an end-to-end network; the process is expressed as: M f =F[M,(x,y)] Where M represents the input curved edge mask; (x, y) represents the coordinates of the sampled light source point; F[·] represents the fusion function of the mask pattern and coordinates; M f represents the intermediate result formed by fusion; G[·] represents the end-to-end network that calculates DNF from the intermediate result; Represents the estimated value of DNF or each component of DNF.

9. A fast photolithography thick mask diffraction field calculation system as claimed in claim 6, characterized in that: Training is performed on the simulated DNF dataset to obtain a pre-trained deep learning network: First, a cost function is determined, which evaluates the prediction performance of the invented method under a given DNF data set; its error is expressed as the difference between the true value E calculated by the rigorous electromagnetic field simulation method and the estimated value obtained by this method The error between: Among them, ||·||2 represents the L2 norm; The training set is put into the network for training, and then the parameters are debugged according to the performance of the network on the validation set to obtain the optimal trained network, which is the pre-trained deep learning network.

10. A fast photolithography thick mask diffraction field calculation system as claimed in claim 6, characterized in that: In the preprocessing module, preprocessing is performed for any given curved edge mask, and the specific method of preprocessing is: The curved edge mask to be calculated is read, and the sampling interval of the mask is adjusted to the sampling interval of the mask in the training set by means of interpolation, padding or deletion of 0 elements, and then the matrix size of the mask is adjusted to meet the input requirements of the deep learning network.

Images