Network training method for generating EUV optimal mask

Abstract

The invention belongs to the technical field of computational lithography, and particularly discloses a network training method for generating an EUV optimal mask, which is characterized in that a DUV mask technology is applied to the generation of an EUV mask, and the mask generated by a DUV mask generation model guides a generator to correct in training, so that the training advances towards an optimized direction; adding a compensation part generated by an EUV mask in the generator so as to realize maximum optimization; a key graph screening technology is applied to a target image, an EUV mask technology and a DUV mask technology at the same time, and the speed and effect of full-chip mask optimization in training are guaranteed; according to the method, under the frame of the CGAN, a mask generated by a DUV mask generation model is used as additional information during training, an EUV compensation part is added on the basis of a general generator, and a model capable of generating an EUV optimal mask is obtained through training by a discriminator.

Description

Technical Field

[0001] The invention belongs to the technical field of computational lithography, and in particular relates to a network training method for generating an EUV optimal mask. Background Art

[0002] Optical lithography is a key step in the manufacture of semiconductor integrated circuits. It is used to accurately transfer microcircuit patterns onto silicon wafers, involving light source irradiation, photosensitive resist reaction and etching processes. As this technology advances, the technology nodes continue to shrink, and the market needs to develop EUV lithography machines with shorter wavelengths, that is, to use better resolution enhancement technology to meet the growing demand for chip performance.

[0003] Mask optimization technology is a type of resolution enhancement technology. The mask is the carrier of graphic design and process technology information in mask optimization technology. Its function is similar to the negative film of a traditional camera. It transfers the circuit pattern to the wafer of the downstream industry through exposure to achieve mass production. The accuracy and quality level of the mask directly affect the yield of the final product and is the key to connecting industrial design and process manufacturing. In other words, exploring mask optimization technology is an important means to promote the development of integrated circuit manufacturing.

[0004] As the critical dimensions of integrated circuits continue to shrink and move towards below 7nm, researchers have shifted their focus from deep ultraviolet (DUV) lithography to extreme ultraviolet (EUV) lithography, and mask optimization technology has also changed accordingly, gradually becoming a difficult part of integrated circuit manufacturing. This difficulty is mainly manifested in the following aspects: mask structure and mask materials that are different from DUV lithography, more complex imaging models and optimization algorithms, and lower mask defect tolerance. In response to these difficulties, the existing EUV mask optimization methods mainly include: studying the structure and materials of the EUV mask absorption layer, reducing the three-dimensional mask effect in EUV lithography, and improving the imaging quality of EUV lithography; further research on the imaging model, optimization algorithm, light source mask characterization and optimization strategy of source mask optimization (SMO); or optimizing the multilayer film defects of the EUV mask and defect detection and compensation technology based on the defect sensitivity of the EUV mask. However, the above methods all involve consideration of complex factors in EUV lithography, and each step will incur huge costs and need further improvement. Summary of the invention

[0005] The present invention provides a network training method for generating an EUV optimal mask, which aims to directly reduce the difference in photoresist patterns corresponding to the generated mask and train the network to generate the EUV optimal mask; under the framework of a conditional generative adversarial network (CGAN), the mask generated by a DUV mask generation model is used as additional information input during training, and an EUV compensation part is added to the general generator, and then a model capable of generating an EUV optimal mask is trained through a discriminator.

[0006] To solve the above problems, the technical solution provided by the present invention is as follows:

[0007] An embodiment of the present invention provides a network training method for generating an EUV optimal mask, comprising the following steps:

[0008] Step 1: Before formal training, data preprocessing and generator pretraining are performed. The generator is the most critical component in training and consists of two parts: the general generator G' and the compensator C. Among them, data preprocessing uses the algorithm based on breadth-first search in the key graph screening technology, including frequency extraction, frequency grouping and key graph screening, which is responsible for converting the obtained target image Z t And the mask m generated by the corresponding EUV and DUV mask generation models E *、m D * As the input of the CGAN network, a minimum set of key graphics is obtained to ensure the speed and effect of the training process;

[0009] Step 2: Training starts with m E * and Z t Composed of mask pairs (Z t ,m E *), by Z t and the mask G(Z) generated by the generator to form a mask pair (Z t ,G(Z)); where (Z t ,m E *) Mask pair represents m E *The difference between the photoresist pattern under the mask and the target image; (Z t ,G(Z)) mask pair represents the difference between the photoresist pattern under the G(Z) mask and the target image;

[0010] Step 3: (Z t ,m E *) and (Z t ,G(Z)) two sets of mask pair data are passed into the discriminator, which is a common convolutional neural network structure used to perform convergence, predict the difference between the target image and the photoresist pattern under the mask, and continuously move the generated EUV mask G(Z) to the given m E *near;

[0011] Step 4: The discriminator outputs the result to determine whether the discriminator can distinguish m E * and G(Z), the judgment result determines the subsequent training process of the model; the step 4 includes: step 41: the judgment result is yes, indicating that G(Z) and Z in the training result t If there is a large gap, go to step 5; Step 42: The judgment result is no, the EUV mask generated by the generator G(Z) satisfies the optimization result, and the training ends;

[0012] Step 5, judging whether the photoresist pattern Z under the G(Z) mask is closer to the target image, and the judgment result determines the subsequent training process of the model; the step 5 comprises: step 51: if the judgment result is yes, it means m E * and G(Z) have some differences, but the photoresist pattern Z under the G(Z) mask generated by the generator is closer to the target image, and such G(Z) also satisfies the optimization result, and the training ends; Step 52: If the judgment result is no, the EUV mask generated by the generator G(Z) is unqualified, and jump to step 6;

[0013] Step 6: Send the instruction to the DUV mask generation model to guide it to correct the mask G'(Z) generated in the previous round, and then further optimize it by the compensator to generate a new round of G(Z) and compare it with Z t To form a new mask pair (Z t ,G(Z)), jump to step 3.

[0014] In an optional embodiment of the present invention, the specific steps of frequency extraction, frequency grouping and key pattern screening in step 1 are:

[0015] Step 11: Perform Fourier transform on the mask pattern to obtain a diffraction spectrum, and extract the peak frequency position of each diffraction peak. and projection boundaries, and the size of the frequency contour is controlled by the growth factor γ;

[0016] Step 12, group the main frequencies to find the inclusion relationship between the mask patterns, so that the selected key patterns can represent the imaging quality of all mask patterns as much as possible. For example, if the diffraction peak F B The peak position In F A The outline of S A (S A =P A γ), then F B Belong to F A Grouping;

[0017] Step 13, use the breadth-first search graph screening method to complete the traversal and obtain the screening results of the key graph.

[0018] In an optional embodiment of the present invention, the specific steps of pre-training in step 1 are:

[0019] Step 14, perform mini-batch sampling on the target image and initialize the gradient of the general generator to zero;

[0020] Step 15, the small batch of images obtained in step 1 are sent to the general generator to obtain the mask G'(Z). The compensator C is an added part in the model generator, and its main function is to compensate for the mask G'(Z) generated by the general generator, including compensation for mask shadow effect, stray light effect and photoresist effect;

[0021] Step 16, the mask G'(Z) is transferred into the ILT program, and a corresponding photoresist pattern Z is generated using a photolithography simulator;

[0022] Step 17: Set the evaluated lithography error E as the objective function, and perform gradient calculation and parameter update.

[0023] In an optional embodiment of the present invention, in step 2, in the first round of formal training, the steps of generating two pairs of mask pairs are:

[0024] Step 21, from the target image Z t Guidance, the general generator G' approaches the EUV mask pattern according to the generation method of the DUV mask, denoted as G'(Z);

[0025] Step 22, G'(Z) is passed into the compensator C for further optimization to generate the initial mask G(Z) in training;

[0026] Step 23, m E * and Z t , G(Z) and Z t They are grouped together to form their own loss functions, which become mask pairs and are recorded as (Z t ,m E * )、(Z t ,G(Z)).

[0027] In an optional embodiment of the present invention, step 21 is represented by a loss function and an optimization algorithm. The MSE and DICE parameters are combined to consider the regional loss while taking into account the loss of a single pixel:

[0028] MSE is responsible for measuring the square of the error between each pixel and the true pixel, and taking the average according to the overall image size; n is the number of image pixels, y i Generate a photoresist pattern under the mask G(Z) for the generator, The photoresist pattern under the mask mE* generated by the EUV mask generation model; DICE evaluates the similarity of two samples, that is, the two masks are overlapped to calculate the ratio of the overlapping parts; the DICE loss is: ε in formula (4) is a very small number and serves as a smoothing coefficient;

[0029] The optimization algorithm selects the gradient descent algorithm. After the first round of formal training, if the discriminator can distinguish m E * and the photoresist pattern generated under G(Z), and the photoresist pattern Z under the mask G(Z) generated by the generator is not closer to the target image Z t , the network model passes the information from the output layer to the input layer to enter the next round of training, and the mask m generated by the existing DUV mask generation model D *Guidance, updating the parameters in the generator or compensator according to the gradient descent process to minimize the loss function.

[0030] In an optional embodiment of the present invention, before step 1, the following steps are further included:

[0031] Step 101, setting initial light source parameters: determining the light source wavelength, light source type, light source power, conversion efficiency and light source generation method for EUV lithography;

[0032] Step 102: Before formal training, the target image Z in the training library is filtered using key image screening technology. t Make a selection, and then use the selected Z t Generate mask m for the target into EUV and DUV models E *、m D *;

[0033] Step 103, before formal training, pre-train the generator in combination with the ILT program, that is, use the lithography simulator to generate the photoresist pattern Z corresponding to G'(Z), and use the evaluation model to compare it with the target image Z t Compare until the evaluation quality of the photoresist pattern reaches the standard and the pre-training is completed;

[0034] Step 104: generate the EUV mask model generated by E *As a data sample, the existing DUV mask generation model generates m D * as random noise, target image Z t Used to guide the generator to generate G(Z).

[0035] An embodiment of the present invention further provides an electronic device, comprising:

[0036] at least one memory for storing a computer program;

[0037] At least one processor is used to execute the program stored in the memory. When the program stored in the memory is executed, the processor is used to execute a network training method for generating an EUV optimal mask as described in the above embodiment.

[0038] An embodiment of the present invention further provides a computer-readable storage medium storing a computer program, wherein when the computer program runs on a processor, the processor executes a network training method for generating an EUV optimal mask as described in the above embodiment.

[0039] Beneficial effect: The method for generating the EUV optimal mask provided by the present invention is a simpler and more efficient method. Before training, the key graphic screening technology is used to concentrate the training part on a small number of graphics with all the features, and then the DUV mask generation model is used for guidance and further optimized with a compensator, thereby avoiding the use of a complex EUV mask generation model in each round of training and improving efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 Flow chart of a network training method for generating EUV optimal mask provided in an embodiment of the present application

[0042] Figure 2 A CGAN training flowchart provided in an embodiment of the present application.

[0043] Figure 3 A schematic diagram of pre-training of a generator under an ILT program provided in an embodiment of the present application. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0045] The method of the present invention aims to directly reduce the difference in photoresist patterns corresponding to the generated mask and train the network to generate the optimal mask for EUV. Technical solution: Under the framework of conditional generative adversarial network (CGAN), the mask generated by the DUV mask generation model is used as additional information input during training, and an EUV compensation part is added to the general generator, and then through the discriminator, a model that can generate the optimal EUV mask is trained.

[0046] like Figure 1 As shown, an embodiment of the present invention provides a network training method for generating an EUV optimal mask, comprising the following steps:

[0047] Step 1: Before formal training, data preprocessing and generator pretraining are performed. The generator is the most critical component in training and consists of two parts: the general generator G' and the compensator C. Among them, data preprocessing uses the algorithm based on breadth-first search in the key graph screening technology, including frequency extraction, frequency grouping and key graph screening, which is responsible for converting the obtained target image Z t And the mask m generated by the corresponding EUV and DUV mask generation models E *、m D * As the input of the CGAN network, a minimum set of key graphics is obtained to ensure the speed and effect of the training process;

[0048] Step 2: Training starts with m E * and Z t Composed of mask pairs (Z t ,m E *), by Z t and the mask G(Z) generated by the generator to form a mask pair (Z t ,G(Z)); where (Z t ,m E *) Mask pair represents m E *The difference between the photoresist pattern under the mask and the target image; (Z t ,G(Z)) mask pair represents the difference between the photoresist pattern under the G(Z) mask and the target image; assuming that there is enough model capacity and training time to converge, the discriminator can compare the two sets of data and make G(Z)≈m after training. E * ;

[0049] Step 3: (Z t ,m E *) and (Z t,G(Z)) two sets of mask pair data are passed into the discriminator, which is a common convolutional neural network structure used to perform convergence, predict the difference between the target image and the photoresist pattern under the mask, and continuously move the generated EUV mask G(Z) to the given m E * near;

[0050] Step 4: The discriminator outputs the result to determine whether the discriminator can distinguish m E * and G(Z); Step 4 includes: Step 41: The judgment result is yes, indicating that G(Z) and Z in the training result t If there is a large gap, go to step 5; Step 42: The judgment result is no, the EUV mask generated by the generator G(Z) satisfies the optimization result, and the training ends;

[0051] Step 5, judging whether the photoresist pattern Z under the G(Z) mask is closer to the target image; said step 5 comprises: step 51: if the judgment result is yes, it means m E * and G(Z) have some differences, but the photoresist pattern Z under the G(Z) mask generated by the generator is closer to the target image, and such G(Z) also satisfies the optimization result, and the training ends; Step 52: If the judgment result is no, the EUV mask generated by the generator G(Z) is unqualified, and jump to step 6;

[0052] Step 6: Send the instruction to the DUV mask generation model to guide it to correct the mask G'(Z) generated in the previous round, and then further optimize it by the compensator to generate a new round of G(Z) and compare it with Z t To form a new mask pair (Z t ,G(Z)), jump to step 3.

[0053] Specifically, step 1, before formal training, data preprocessing and generator pre-training are performed. The generator part of this embodiment is composed of a general generator G' and a compensator C. Among them, the general generator is based on the Abbe lithography spatial imaging model, and is mainly composed of an encoder and a decoder, which is a stacked convolution architecture. Before formal training, we first pre-train it to reduce the possibility of it getting stuck in the local minimum area and "guide" the generator to reach the optimal. After pre-training is completed, a small number of target images are sampled, and then the encoder performs hierarchical layout feature abstraction, and the decoder operates in the opposite way to complete the mask correction relative to the target. In addition, the general generator G' is responsible for the mask m D *Under the guidance of , a pattern close to the EUV mask is generated and recorded as G'(Z) and transmitted to the compensator. The compensator is composed of a compensation model, including compensation for mask shadow effect, stray light effect and photoresist effect. Using the lithography imaging model that takes these three effects into account, this generator can further correct the mask G'(Z) generated by the general generator, which is equivalent to further optimizing the EUV mask, recorded as G(Z). The optimized mask G(Z) will be consistent with the target image Z t As a set of data sent to the discriminator.

[0054] Data preprocessing and generator pre-training: Using the key graph screening method (breadth-first search algorithm), we can get the key graph group with the least number of key graphs representing the entire target image Zt. mE* and mD* generate masks based on this minimum key graph group, that is, a small number of graphs are used to represent all the graphic structure features of each, so as to ensure the speed and effect of full-chip mask optimization. The generator is the most critical component in the network model. The generator in this model consists of two parts: a general generator and a compensator. In order to train a satisfactory mask generation network model, the joint training algorithm guided by the ILT program is used to initialize the general generator part so that it can converge better in the subsequent optimization process. Before training the mask generation for a certain target image, the generator must be pre-trained.

[0055] The specific steps of frequency extraction, frequency grouping and key graph screening in step 1 are:

[0056] Step 11: Perform Fourier transform on the mask pattern to obtain a diffraction spectrum, and extract the peak frequency position of each diffraction peak. and projection boundaries, and the size of the frequency contour is controlled by the growth factor γ;

[0057] Step 12, group the main frequencies to find the inclusion relationship between the mask patterns, so that the selected key patterns can represent the imaging quality of all mask patterns as much as possible. For example, if the diffraction peak F B The peak position In F A The outline of S A (S A =P A γ), then F B Belong to F A Grouping;

[0058] Step 13, use the breadth-first search graph screening method to complete the traversal and obtain the screening results of the key graph.

[0059] Generator pre-training is performed before the model is formally trained, such as Figure 3This is a schematic diagram of generator pre-training, which is completed by the joint training algorithm guided by the model's generator and the ILT program. The object of pre-training is the general generator part. The specific steps of pre-training in step 1 are:

[0060] Step 14, perform mini-batch sampling on the target image and initialize the gradient of the general generator to zero;

[0061] Step 15, sending the small batch of images obtained in step 1 to the general generator to obtain the mask G'(Z); the compensator C is an added part in the model generator, and its main function is to compensate for the mask G'(Z) generated by the general generator, including compensation for mask shadow effect, stray light effect and photoresist effect;

[0062] Step 16, the mask G'(Z) is transferred into the ILT program, and a corresponding photoresist pattern Z is generated using a photolithography simulator;

[0063] Step 17, setting the evaluated lithography error E as the objective function, performing gradient calculation and parameter update.

[0064] The photolithography error is:

[0065] Siomoid function:

[0066] Incomplete binary mask:

[0067] Gradient representation:

[0068] Among them, α and β represent the steepness factor, H represents the original convolution kernel, and H * represents the conjugate matrix of H.

[0069] Specifically, the compensator is an added part in the model generator, and its main function is to compensate for the mask G'(Z) generated by the general generator, including the compensation of mask shadow effect, stray light effect and photoresist effect. The specific methods used are:

[0070] The mask shadow effect compensation adopts a simplified mask shadow effect model, which is equivalent to introducing an edge error with a specific width in the photoresist imaging. The shadow width B and the incident angle α s The form is:

[0071]

[0072] Among them, B max is the maximum shadow width, α s' is the azimuth angle of the mask pattern edge to the x-axis of the annular sector exposure field of the EUV lithography system, W is the width of the annular sector exposure field, F is the opening angle of the annular sector exposure field, and x is the x-axis coordinate of the exposure field position where the edge is located.

[0073] The stray light effect is determined by the stray light point spread function PSF f And the total integrated scattering factor TIS is characterized by:

[0074]

[0075] in, is the spatial coordinate of the mask surface, n f is the spectral index, r min is the boundary range between low-frequency and high-frequency phase errors.

[0076] TIS is obtained by integrating the scattered light from all possible directions.

[0077] The compensation of photoresist effect adopts a simplified photoresist model of point spread function, using photoresist point spread function PSF r ;

[0078]

[0079] Among them, σ PSF It is the blurring effect of photoresist on the final image.

[0080] After the above three compensation models, the mask G'(Z) generated by the general generator is further optimized and recorded as G(Z).

[0081] Before step 1, also include:

[0082] Step 101, setting initial light source parameters: determining the light source wavelength, light source type, light source power, conversion efficiency and light source generation method for EUV lithography;

[0083] Step 102: Before formal training, the target image Z in the training library is filtered using key image screening technology. t Make a selection, and then use the selected Z t Generate mask m for the target into EUV and DUV models E *、m D *;

[0084] Step 103, before formal training, pre-train the generator in combination with the ILT program, that is, use the lithography simulator to generate the photoresist pattern Z corresponding to G'(Z), and use the evaluation model to compare it with the target image Z t Compare until the evaluation quality of the photoresist pattern reaches the standard and the pre-training is completed;

[0085] Step 104: generate the EUV mask model generated by E *As a data sample, the existing DUV mask generation model generates m D * as random noise, target image Z t Used to guide the generator to generate G(Z).

[0086] Step 2: Training starts with m E * and Z t Composed of mask pairs (Z t ,m E * ), Z t and the mask G(Z) generated by the guiding generator form a mask pair (Z, G(Z)).

[0087] Step 2 In the first round of formal training, the steps for generating two mask pairs are:

[0088] Step 21: From the target image Z t Guidance, the general generator G' approaches the EUV mask pattern according to the generation method of the DUV mask, denoted as G'(Z);

[0089] Step 22: Pass G'(Z) into the compensator C for further optimization to generate the initial mask G(Z) in training;

[0090] Step 23: Place m E * and Z t , G(Z) and Z t They are grouped together to form their own loss functions, which become mask pairs and are recorded as (Z t ,m E * )、(Z t ,G(Z)).

[0091] Specifically, as a bridge connecting model prediction, training objectives, and optimization algorithms, the loss function directly affects the learning effect and final performance of the model.

[0092] Target image Z t It is expressed by the combination of mean square error MSE and DICE loss function:

[0093] MSE is responsible for measuring the square of the error between each pixel and the true pixel, and taking the average based on the overall image size.

[0094] Where n is the number of image pixels, y i Generate a photoresist pattern under the mask G(Z) for the generator, Mask m generated for EUV mask generation model E * The photoresist pattern below.

[0095] The DICE loss function is responsible for focusing on the edge of the mask and evaluating the similarity of two samples, that is, overlapping the two masks to calculate the ratio of the overlapping parts.

[0096] The DICE loss is:

[0097] Among them, ε is a very small number, which serves as a smoothing coefficient.

[0098] The combination of the above two parameters can take into account the loss of a single pixel while considering the regional loss.

[0099] The optimization algorithm selects the gradient descent algorithm. After the first round of formal training, if the discriminator can distinguish m E * and the photoresist pattern generated under G(Z), and the photoresist pattern Z under the mask G(Z) generated by the generator is not closer to the target image Z t , the network model passes the information from the output layer to the input layer to enter the next round of training, and the mask m generated by the existing DUV mask generation model D * Guide, update the parameters in the general generator according to the gradient descent process to minimize the loss function.

[0100] Step 3: (Z t ,m E * ) and (Z t ,G(Z)) two sets of mask pair data are passed into the discriminator. Specifically, the discriminator in this model is a common convolutional neural network structure, which is mainly used to perform convergence, predict the difference between the target image and the photoresist pattern under the mask, and continuously move the generated EUV mask G(Z) to the given m E * near.

[0101] Step 4: The discriminator outputs the result to determine whether the discriminator can distinguish m E * and G(Z), the judgment result determines the subsequent training process of the model, which is divided into two cases: Step 41: The judgment result is yes, indicating that G(Z) and Z in the training result t If there is a large gap, go to step 5; Step 42: The judgment result is no, the EUV mask generated by the generator G(Z) satisfies the optimization result, and the training ends.

[0102] Specifically, the loss function of the model is given by m E* and G(Z), which can be regarded as trying to make the mask generated by the network model close to the mask generated by the existing EUV mask generation model. It is still represented by the combination of the mean square error MSE and the DICE loss function.

[0103] Step 5: Determine whether the photoresist pattern Z under the G(Z) mask is closer to the target image. The determination result determines the subsequent training process of the model, which is divided into two cases: Step 51: The determination result is yes, indicating that although m E * There is a certain difference between the generator G(Z) and G(Z), but the photoresist pattern Z under the G(Z) mask generated by the generator is closer to the target image. Such G(Z) also meets the optimization result, and the training is completed; Step 52: If the judgment result is no, the EUV mask generated by the generator G(Z) is unqualified, and jump to step 6.

[0104] Step 6, send the instruction to the DUV mask generation model and jump to step 3; Step 6 is not the last step of the training, but a bridge to the next round of training, and is also a key part of the present invention. The DUV mask generation model mainly corrects the mask G'(Z) generated in the previous round, and then the compensator further optimizes it to generate a new round of G(Z), and compares it with Z t To form a new mask pair (Z t ,G(Z)).

[0105] An embodiment of the present invention also provides an electronic device, which includes: at least one memory for storing computer programs; at least one processor for executing the programs stored in the memory, and when the program stored in the memory is executed, the processor is used to execute a network training method for generating EUV optimal masks as described in the above embodiment.

[0106] An embodiment of the present invention further provides a computer-readable storage medium storing a computer program, wherein when the computer program runs on a processor, the processor executes a network training method for generating an EUV optimal mask as described in the above embodiment.

[0107] Example 1

[0108] This embodiment provides a network training method for generating an EUV optimal mask, comprising the following steps:

[0109] Step 1: During data preprocessing, frequency extraction, frequency grouping, and key graph screening are completed using an algorithm based on breadth-first search to obtain the Z of the input layer. t 、m E * and m D * ;

[0110] Step 2: pre-train the model generator, mainly combining the general model generator with the training algorithm of the ILT program, and performing gradient calculation and parameter update based on the evaluated lithography error;

[0111] Step 3: The preprocessed data m E * and Z t As a set of inputs, we get a mask pair (Z t ,m E * ), and then through the compensation model, Z t And G(Z) generated by the generator as a set of inputs to get the mask pair (Z, G(Z));

[0112] Step 4: The discriminator discriminates the mask pair and obtains E * and the loss function determined by G(Z);

[0113] Step 5, follow Figure 2 The flowchart of CGAN training in the figure judges in turn, uses the gradient descent method to update the parameters in the generator, and minimizes the loss function of the model;

[0114] Step 6: Repeat the above steps until the training is completed.

[0115] In step 1, the breadth-first search-based algorithm groups the graphs by extracting the peak frequency positions and projection boundaries and setting growth factors, thereby finding the key graph to represent the whole.

[0116] In step 2, the pre-training of the generator is mainly the pre-training of the general generator, such as Figure 3 As shown, the general generator is completed together with the joint training algorithm guided by the ILT program.

[0117] The ILT program consists of a lithography simulator and an evaluation model. The lithography simulator is mainly based on the Abbe imaging principle.

[0118] Use the square of the second norm of the difference as the objective function:

[0119]

[0120] The sigmoid function and the incomplete binarization function are used to describe the photolithography process:

[0121]

[0122] Among them, α and β are steepness factors, I this the threshold, usually 0.5 or the position of the maximum steepness point. The larger the steepness factor, the closer the function is to the ideal photoresist function.

[0123] Through this procedure, the mask G'(Z) generated by the general generator is converted into the corresponding photoresist pattern Z.

[0124] In step 3, the training officially begins:

[0125] Generate the target image Z from the existing EUV mask generation model t The mask with Z t Together they form a mask pair (Z t ,m E * ), directly fed into the discriminator;

[0126] Generate Z from a pre-trained general generator t The reference DUV mask, denoted as G'(Z), is fed into the compensator part of the generator to compensate the DUV mask in turn:

[0127] Simplified mask shadow effect:

[0128]

[0129] Among them, B max is the maximum shadow width, α s ' is the azimuth angle of the mask pattern edge to the x-axis of the annular sector exposure field of the EUV lithography system, W is the width of the annular sector exposure field, F is the opening angle of the annular sector exposure field, and x is the x-axis coordinate of the exposure field position where the edge is located.

[0130] Stray light effect:

[0131]

[0132] Among them, PSF f represents the stray light point spread function, TIS represents the total integrated scattering factor, is the spatial coordinate of the mask surface, n f is the spectral index, r min is the boundary range between low-frequency and high-frequency phase errors.

[0133] Photoresist effect:

[0134] Among them, PSF r represents the photoresist point spread function, σ PSF It is the blurring effect of photoresist on the final image.

[0135] After the above three compensation models, the mask G'(Z) generated by the general generator is further optimized and recorded as G(Z).

[0136] In step 4, the mask pair (Z t ,m E * ) and (Z, G(Z)) are passed into the discriminator, which distinguishes m E * With G(Z), its loss function is expressed as the union of mean square error MSE and DICE loss function:

[0137]

[0138] Where n is the number of image pixels, y i Generate a photoresist pattern under the mask G(Z) for the generator, The mask m generated by the EUV mask generation model E * The photoresist pattern below.

[0139] The DICE loss is:

[0140] Among them, ε is a very small number, which serves as a smoothing coefficient.

[0141] The optimization algorithm selects the gradient descent algorithm.

[0142] After the first round of formal training, if the discriminator can distinguish m E * and the photoresist pattern generated under G(Z), and the photoresist pattern Z under the mask G(Z) generated by the generator is not closer to the target image Z t , the network model passes the information from the output layer to the input layer to enter the next round of training, and the mask m generated by the existing DUV mask generation model D * Guide, update the parameters in the general generator according to the gradient descent process to minimize the loss function.

[0143] In step 5, the discriminator outputs the result according to Figure 2 The training can be divided into the following situations:

[0144] Case 1: If the discriminator cannot distinguish m E * and G(Z), indicating that the photoresist pattern under the G(Z) mask is very close to the pattern given by the model, that is, the EUV mask generated by the generator is qualified, and the training is completed;

[0145] Case 2: If the discriminator can distinguish m E *and G(Z), but with m E * In comparison, the photoresist pattern under the G(Z) mask is closer to the target image Z t , indicating that the EUV mask generated by the generator is better, and the training is completed;

[0146] Case 3: If the discriminator can distinguish m E * and G(Z), and with m E * Compared with the original image, the photoresist pattern under the G(Z) mask is not closer to the target image Z. t , indicating that the EUV mask generated by the generator is not trained enough. At this time, the information is passed back and the parameters in the generator are updated using the gradient descent method to minimize the loss function of the model until the training is completed.

[0147] In summary, although the present invention has been disclosed as above in terms of preferred embodiments, the above preferred embodiments are not intended to limit the present invention. A person skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope defined in the claims.

Claims

1. A network training method for generating EUV optimal mask, characterized in that: The following steps are involved: Step 1: Before formal training, data preprocessing and generator pretraining are performed. The generator is the most critical component in training and consists of two parts: the general generator G' and the compensator C. Among them, data preprocessing uses the algorithm based on breadth-first search in the key graph screening technology, including frequency extraction, frequency grouping and key graph screening, which is responsible for converting the obtained target image Z t And the mask m generated by the corresponding EUV and DUV mask generation models E *、m D * As the input of the CGAN network, a minimum set of key graphics is obtained to ensure the speed and effect of the training process; Step 2: Training starts with m E * and Z t Composed of mask pairs (Z t ,m E *), by Z t and the mask G(Z) generated by the generator to form a mask pair (Z t ,G(Z)); where (Z t ,m E *) Mask pair represents m E *The difference between the photoresist pattern under the mask and the target image; (Z t ,G(Z)) mask pair represents the difference between the photoresist pattern under the G(Z) mask and the target image; Step 3: (Z t ,m E *) and (Z t ,G(Z)) two sets of mask pair data are passed into the discriminator, which is a common convolutional neural network structure used to perform convergence, predict the difference between the target image and the photoresist pattern under the mask, and continuously move the generated EUV mask G(Z) to the given m E * near; Step 4: The discriminator outputs the result to determine whether the discriminator can distinguish m E * and G(Z), the judgment result determines the subsequent training process of the model; the step 4 includes: step 41: the judgment result is yes, indicating that G(Z) and Z in the training result t If there is a large gap, go to step 5; Step 42: The judgment result is no, the EUV mask generated by the generator G(Z) satisfies the optimization result, and the training ends; Step 5, judging whether the photoresist pattern Z under the G(Z) mask is closer to the target image, and the judgment result determines the subsequent training process of the model; the step 5 comprises: step 51: if the judgment result is yes, it means m E * and G(Z) have some differences, but the photoresist pattern Z under the G(Z) mask generated by the generator is closer to the target image, and such G(Z) also satisfies the optimization result, and the training ends; Step 52: If the judgment result is no, the EUV mask generated by the generator G(Z) is unqualified, and jump to step 6; Step 6: Send the instruction to the DUV mask generation model to guide it to correct the mask G'(Z) generated in the previous round, and then further optimize it by the compensator to generate a new round of G(Z) and compare it with Z t To form a new mask pair (Z t ,G(Z)), jump to step 3.

2. The network training method for generating EUV optimal mask according to claim 1, characterized in that: The specific steps of frequency extraction, frequency grouping and key graph screening in step 1 are: Step 11: Perform Fourier transform on the mask pattern to obtain a diffraction spectrum, and extract the peak frequency position of each diffraction peak. and projection boundaries, and the size of the frequency contour is controlled by the growth factor γ; Step 12: Group the main frequencies to find the inclusion relationship between the mask patterns, so that the selected key patterns can represent the imaging quality of all mask patterns as much as possible. B The peak position In F A The outline of S A (S A =P A γ), then F B Belong to F A Grouping; Step 13, use the breadth-first search graph screening method to complete the traversal and obtain the screening results of the key graph.

3. The network training method for generating EUV optimal mask according to claim 1, characterized in that: The specific steps of pre-training in step 1 are: Step 14, perform mini-batch sampling on the target image and initialize the gradient of the general generator to zero; Step 15, the small batch of images obtained in step 1 are sent to the general generator to obtain the mask G'(Z). The compensator C is an added part in the model generator, and its main function is to compensate for the mask G'(Z) generated by the general generator, including compensation for mask shadow effect, stray light effect and photoresist effect; Step 16, the mask G'(Z) is transferred into the ILT program, and a corresponding photoresist pattern Z is generated using a photolithography simulator; Step 17: Set the evaluated lithography error E as the objective function, and perform gradient calculation and parameter update.

4. The network training method for generating EUV optimal mask according to claim 1, characterized in that: Step 2 In the first round of formal training, the steps for generating two mask pairs are: Step 21, from the target image Z t Guidance, the general generator G' approaches the EUV mask pattern according to the generation method of the DUV mask, denoted as G'(Z); Step 22, G'(Z) is passed into the compensator C for further optimization to generate the initial mask G(Z) in training; Step 23, m E * and Z t , G(Z) and Z t They are grouped together to form their own loss functions, which become mask pairs and are recorded as (Z t ,m E * )、(Z t ,G(Z)).

5. A network training method for generating EUV optimal mask according to claim 4, characterized in that: Step 21 is represented by a loss function and an optimization algorithm. The MSE and DICE parameters can take into account the regional loss while taking into account the loss of a single pixel: MSE is responsible for measuring the square of the error between each pixel and the true pixel, and taking the average according to the overall image size; n is the number of image pixels, y i Generate a photoresist pattern under the mask G(Z) for the generator, Mask m generated for EUV mask generation model E *The photoresist pattern under; DICE evaluates the similarity of two samples, that is, the two masks are overlapped to calculate the ratio of the overlapping parts; the DICE loss is: ε in formula (4) is a very small number and serves as a smoothing coefficient; The optimization algorithm selects the gradient descent algorithm. After the first round of formal training, if the discriminator can distinguish m E * and the photoresist pattern generated under G(Z), and the photoresist pattern Z under the mask G(Z) generated by the generator is not closer to the target image Z t , the network model passes the information from the output layer to the input layer to enter the next round of training, and the mask m generated by the existing DUV mask generation model D *Guidance, updating the parameters in the generator or compensator according to the gradient descent process to minimize the loss function.

6. A network training method for generating EUV optimal mask according to claim 1, characterized in that: Before step 1, also include: Step 101, setting initial light source parameters: determining the light source wavelength, light source type, light source power, conversion efficiency and light source generation method for EUV lithography; Step 102: Before formal training, the target image Z in the training library is filtered using key image screening technology. t Make a selection, and then use the selected Z t Generate mask m for the target into EUV and DUV models E *、m D *; Step 103, before formal training, pre-train the generator in combination with the ILT program, that is, use the lithography simulator to generate the photoresist pattern Z corresponding to G'(Z), and use the evaluation model to compare it with the target image Z t Compare until the evaluation quality of the photoresist pattern reaches the standard and the pre-training is completed; Step 104: generate the EUV mask model generated by E *As a data sample, the existing DUV mask generation model generates m D * as random noise, target image Z t Used to guide the generator to generate G(Z).

7. An electronic device, characterized in that: include: at least one memory for storing a computer program; At least one processor is used to execute the program stored in the memory. When the program stored in the memory is executed, the processor is used to execute a network training method for generating an EUV optimal mask as described in any one of claims 1-6.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program runs on a processor, the processor is enabled to execute a network training method for generating an EUV optimal mask as described in any one of claims 1 to 6.

Images