Photoetching mask method and device

Abstract

The embodiment of the invention discloses a photoetching mask method and device, and is applied to the field of semiconductors. The method comprises the following steps: acquiring an image to be photoetched, processing the image to be photoetched by using a target network to obtain a coarsened mask image, processing the coarsened mask image again by using an ILT to obtain a target mask image, and performing photolithography masking by using the target mask image. The target network is a conversion network trained based on a training sample, the training sample comprises a sample image and a reference image, and the reference image is an image which is synthesized by ILT and contains sub-resolution auxiliary features. Wherein the coarsened mask image is an image which is similar to an image which is synthesized by an ILT technology and contains sub-resolution auxiliary features. And refining processing is carried out on the coarsened mask image by using an ILT technology to obtain a more accurate target mask image. Therefore, on the basis of ensuring the precision of the photoetching mask, the calculation complexity is reduced, and the efficiency of the photoetching mask is improved.

Description

Technical Field

[0001] The present application relates to the field of semiconductor technology, and in particular to a photolithography mask method and device. Background Art

[0002] As chip size continues to shrink, the lithography wavelength is much larger than the characteristic size of the chip, which makes the optical proximity effect caused by diffraction, interference, etc. increasingly serious. Among them, the optical proximity effect refers to the transfer distortion between the mask image and the image actually lithographically transferred to the chip surface (i.e., the layout image). The optical proximity effect will affect the imaging quality of the chip surface, resulting in narrowing of line width, shrinkage of narrow line endpoints, and rounding of graphic corners. Among them, the smaller the layout image size, the more obvious the optical proximity effect, and the greater the image deviation after exposure. The limited numerical aperture of the lens is one of the causes of the optical proximity effect. Due to the limited numerical aperture of the lens, it is impossible to collect all the information of the spatial frequency, resulting in the loss of some spatial frequencies, which leads to the optical proximity effect.

[0003] At present, Optical Proximity Correction (OPC) technology is generally used to compensate for the problem of transfer distortion. Specifically, OPC technology is to modify the shape of the layout image and change the exposure intensity of the chip surface so that the image contour after exposure is close to the expected image contour (referred to as the target image contour). Currently, Inverse Lithography Technology (ILT) technology can be used to apply to mask design for process nodes corresponding to smaller chip sizes (for example, 7nm or below). Among them, ILT technology is a technology that optimizes mask design through inverse calculation, which can improve the efficiency of processing complex images and processing small feature sizes. However, ILT technology requires extremely complex calculations to obtain mask images, and the calculation cost and calculation time are relatively long, which leads to low mask efficiency of photolithography masks based on ILT technology. How to reduce the calculation amount of ILT technology and improve the mask efficiency of photolithography masks based on ILT technology has become a technical problem to be solved. Summary of the Invention

[0004] The embodiments of the present application provide a photolithography mask method and apparatus for reducing the computational complexity of obtaining a mask image using ILT technology and improving the mask efficiency of photolithography masking based on ILT technology.

[0005] In a first aspect, an embodiment of the present application provides a photolithography mask method, the method comprising:

[0006] Acquire an image to be photolithographically processed;

[0007] Processing the image to be photolithographically processed using a target network to obtain a roughened mask image;

[0008] The target network is a conversion network trained based on training samples, the training samples include sample images and reference images, and the reference image is an image synthesized by ILT and containing sub-resolution auxiliary features;

[0009] The roughened mask image is processed using ILT to obtain a target mask image, so as to use the target mask image for photolithography masking.

[0010] Optionally, the conversion network includes a first generator and a first discriminator, wherein the first generator is used to convert the image to be photolithographically processed into the roughened mask image; and the conversion network is trained in the following manner:

[0011] Obtaining sample images and reference images;

[0012] Processing the sample image using the first generator to generate an actual mask image;

[0013] Obtaining a first adversarial loss value using the first discriminator; the first adversarial loss value indicates a difference between a predicted probability obtained after the actual mask image is processed by the first discriminator and a preset ideal value;

[0014] Adjusting parameters of the first generator according to the first adversarial loss value;

[0015] Obtaining a second adversarial loss value using the first discriminator, where the second adversarial loss value is the sum of a correct recognition rate loss of the reference image and an incorrect recognition rate loss of the actual mask image;

[0016] Adjusting parameters of the first discriminator according to the second adversarial loss value;

[0017] Determining whether a sum of the first adversarial loss and the second adversarial loss is less than or equal to a preset threshold;

[0018] If not, re-execute the operation of generating the actual mask image, obtaining the first adversarial loss value, the second adversarial loss value, and the sum of the two loss values; wherein the adjusted first discriminator is used as the first discriminator to generate the actual mask image, and the adjusted second discriminator is used as the second discriminator to perform the operation of obtaining the first adversarial loss value, the second adversarial loss value, and the sum of the two loss values;

[0019] If so, stop training and obtain the target network.

[0020] Optionally, the conversion network further includes a second generator, and the method further includes:

[0021] Processing the actual mask image by the second generator to generate an actual lithography image;

[0022] Obtaining a cycle consistency loss value of the actual lithography pattern relative to the sample image;

[0023] The determining whether the sum of the first adversarial loss and the second adversarial loss is less than or equal to a preset threshold includes:

[0024] Determine whether the sum of the first adversarial loss value, the second adversarial loss value, and the cycle consistency loss value is less than or equal to a preset threshold.

[0025] Optionally, the method further includes:

[0026] obtaining an identity loss value, wherein the identity loss value is a first difference value of the actual mask image relative to the sample image, and obtaining a sum of a second difference value of the actual lithography image relative to the actual mask image;

[0027] The determining whether the sum of the first adversarial loss value, the second adversarial loss value, and the cycle consistency loss value is less than or equal to a preset threshold includes:

[0028] Determine whether the sum of the first adversarial loss value, the second adversarial loss value, the identity loss value, and the cycle consistency loss value is less than or equal to a first preset threshold.

[0029] Optionally, the first generator includes: an initialization module, a downsampling module, a self-attention residual module, an upsampling module and an output layer, and the processing of the image to be lithographically processed by the target network to obtain a coarsened mask image includes:

[0030] Creating an initial network image corresponding to the image to be photolithographically processed by the initialization module;

[0031] Sampling the initial network image using the downsampling module to obtain a first feature map;

[0032] Performing residual processing on the first feature map through the self-attention residual module to obtain a second feature map;

[0033] Sampling the second feature map using the upsampling module to obtain a third feature map;

[0034] The coarsening mask image corresponding to the third feature map is output through the output layer.

[0035] Optionally, the downsampling module includes a first feature extraction layer and a second feature extraction layer, and the first feature extraction layer and the second feature extraction layer both include: a convolution layer with a step size of 2, an instance normalization layer and a first activation function; the first activation function is an activation function that introduces nonlinear expression.

[0036] Optionally, outputting the coarsening mask image corresponding to the third feature map through the output layer includes:

[0037] The third feature map is processed by reflection filling, convolution and second activation function through the output layer to obtain the coarsening mask image.

[0038] Optionally, the first discriminator is a Markov discriminator structure, comprising a first convolutional layer, a second convolutional layer, a third convolutional layer and a fourth convolutional layer connected in sequence; spectral normalization processing is performed after the fourth convolutional layer;

[0039] The obtaining a first adversarial loss value by using the first discriminator includes:

[0040] Use the first convolution layer to downsample the actual mask image to obtain a fourth feature map; use the second convolution layer to downsample the fourth feature map to obtain a fifth feature map; use the third convolution layer to downsample the fifth feature map to obtain a sixth feature map; use the fourth convolution layer to reduce the number of channels on the sixth feature map to obtain the first adversarial loss value.

[0041] Optionally, the conversion network is a cycle-consistent adversarial network.

[0042] In a second aspect, an embodiment of the present application provides a photolithography mask device, comprising:

[0043] An acquisition unit, used for acquiring an image to be photolithographically processed;

[0044] A primary processing unit, configured to process the image to be photolithographically processed using a target network to obtain a roughened mask image;

[0045] The target network is a conversion network trained based on training samples, the training samples include sample images and reference images, and the reference image is an image synthesized by ILT and containing sub-resolution auxiliary features;

[0046] A post-processing unit is used to process the roughened mask image using ILT to obtain a target mask image; and to perform photolithography masking using the target mask image.

[0047] In a third aspect, an embodiment of the present application provides a computer program product, which includes: a computer program (also referred to as code, or instructions), which, when executed, enables a computer to execute a method in any possible implementation of any of the above aspects.

[0048] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program (also referred to as code, or instructions) that, when run on a computer, enables the computer to execute a method in any possible implementation of any of the above aspects.

[0049] In a fifth aspect, embodiments of the present application provide a chip system comprising one or more processors configured to retrieve and execute instructions stored in a memory, thereby executing the method of any of the above aspects or any possible implementations of each aspect. The chip system may be composed of a chip or may include a chip and other discrete devices.

[0050] The embodiment of the present application provides a photolithography mask method and device, wherein the method includes: obtaining an image to be photolithographically processed, processing the image to be photolithographically processed using a target network to obtain a coarsened mask image, and reprocessing the coarsened mask image using ILT to obtain a target mask image, so as to use the target mask image for photolithography masking. The target network is a conversion network trained based on training samples, wherein the training samples include sample images and reference images, and the reference image is an image synthesized by ILT containing sub-resolution auxiliary features. The coarsened mask image is similar to an image containing sub-resolution auxiliary features synthesized by ILT technology. The coarsened mask image is refined using ILT technology to obtain a more accurate target mask image. The present application first quickly obtains a coarsened mask image containing sub-resolution auxiliary features that is similar to the image synthesized by ILT technology through the target network. Since this part does not need to be processed by the physical model, the amount of calculation can be significantly reduced. The coarsened mask image is then fine-tuned using ILT technology, which not only improves the accuracy of the mask image but also significantly reduces computational complexity and improves lithography mask efficiency because it does not require a large search space and number of iterations. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0052] Figure 1A flow chart of a photolithography mask method provided in an embodiment of the present application;

[0053] Figure 2 A flow chart of a method for acquiring a target network provided in an embodiment of the present application;

[0054] Figure 3 A schematic diagram of the structure of a CycleGAN provided in an embodiment of the present application;

[0055] Figure 4 A schematic diagram of a U-Net structure provided in an embodiment of the present application;

[0056] Figure 5 A schematic diagram of an image processing method by a generator provided in an embodiment of the present application;

[0057] Figure 6 A schematic diagram of a method for image processing based on PatchGAN provided in an embodiment of the present application;

[0058] Figure 7 A schematic diagram of processing a photolithography mask provided in an embodiment of the present application;

[0059] Figure 8 A schematic structural diagram of a photolithography mask device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to help those skilled in the art better understand the present invention, the following will clearly and completely describe the technical solutions in this embodiment with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0061] As previously mentioned, the current use of ILT technology for lithography masks suffers from high computational complexity, long computational costs and time, and relatively low mask design efficiency. Further analysis reveals the reasons for these issues: ILT optimizes mask design through inverse calculations, often requiring complex physical models and high-precision iterative optimization algorithms to obtain the mask images required for lithography masks.

[0062] In view of this, an embodiment of the present application provides a lithography masking method, which obtains an image to be lithographically processed using a target network to obtain a coarsened mask image, and then further processes the coarsened mask image using ILT to obtain a target mask image, so as to use the target mask image for lithography masking. The target network is a conversion network trained based on training samples, wherein the training samples include sample images and reference images, and the reference image is an image synthesized by ILT and contains sub-resolution auxiliary features. It can be understood that the target network can process the image to be lithographically processed based on the correspondence between the learned sample images and the reference images, thereby obtaining a coarsened mask image containing sub-resolution auxiliary features that is similar to that synthesized by ILT. The coarsened mask image is then refined using ILT technology to obtain a more accurate target mask image. The present application first uses the target network to quickly obtain a coarsened mask image containing sub-resolution auxiliary features that is similar to that synthesized by ILT technology. Since this part does not require processing through a physical model, the amount of computation can be significantly reduced. The coarsened mask image is then fine-tuned using ILT technology, which not only improves the accuracy of the mask image but also significantly reduces computational complexity and improves lithography mask efficiency because it does not require a large search space and number of iterations.

[0063] The photolithography mask method provided in the embodiment of the present application is described in detail and completely below in conjunction with the accompanying drawings.

[0064] First, the photolithography mask method provided in the embodiments of the present application is applied to a computing device capable of performing ILT technology. The embodiments of the present application do not specifically limit the type of computing device. For example, the computing device can be a computer, a server, or a dedicated high-performance computing cluster.

[0065] The specific photolithography mask method is introduced below.

[0066] Attachment Figure 1 A flow chart of a photolithography mask method provided in an embodiment of the present application, the method includes the following contents:

[0067] S110, obtaining an image to be photolithographically processed.

[0068] The image to be photolithographically processed refers to an image that needs to be printed on the surface of the mask.

[0069] S120 , processing the to-be-photolithographic image using the target network to obtain a coarsened mask image.

[0070] In this embodiment of the present application, the target network is a conversion network trained based on training samples. The training samples include sample images and reference images. The sample images are historical images printed onto a reticle. The reference images are images synthesized using ILT technology and contain sub-resolution auxiliary features.

[0071] It should be noted that in the embodiments of the present application, the sample images and reference images may correspond one-to-one. For example, sample image A corresponds to reference image A, and sample image B corresponds to reference image B. Reference image A is an image obtained by processing sample image A using the ILT technique, and reference image B is an image obtained by processing sample image B using the ILT technique. The sample image and the reference image may also not correspond. For example, after training, the target network may learn that the correspondence between sample image A and reference image B corresponds. This is not specifically limited in the embodiments of the present application.

[0072] It should be noted that the target network provided in the embodiment of the present application learns the correspondence between the image before lithography and the image containing sub-resolution auxiliary features synthesized by ILT technology. Therefore, the target network can process the lithography image and use the learned correspondence to obtain the corresponding image containing sub-resolution auxiliary features synthesized by ILT technology.

[0073] In one example, the conversion network may include a first generator and a first discriminator. The first generator may convert the image to be lithographed into a coarsened mask image. In this embodiment, to ensure a more accurate coarsened mask image, the image to be lithographed is processed using ILT technology. The resulting image requires training of the conversion network.

[0074] Specifically, the computing device can train the conversion network based on the training samples, including: obtaining a sample image and a reference image. The sample image is processed by the first generator to obtain an actual mask image. In order to determine whether the actual mask image generated by the first generator is the same as or similar to the reference image corresponding to the sample image, the computing device can use the first discriminator to obtain an adversarial loss value. The adversarial loss value is used to maximize the probability that the image generated by the first generator is mistaken for a real image by the first discriminator. For the first generator, the smaller the adversarial loss value, the more the generated image can deceive the first discriminator, that is, the mask image generated by the first generator is closer to the distribution of the real image.

[0075] In the embodiment of the present application, the adversarial loss value is used to reflect the goals of the first generator and the first discriminator in the game process. In the embodiment of the present application, the adversarial loss value includes a first adversarial loss value and a second adversarial loss value.

[0076] The first adversarial loss, also referred to as the adversarial loss of the first generator, represents the difference between the predicted probability of the actual mask image generated by the first generator after post-processing by the first discriminator and a preset ideal value. The computing device can adjust the parameters of the first generator based on the first adversarial loss to reduce the first adversarial loss, thereby making the image generated by the first generator more consistent with the image containing sub-resolution auxiliary features synthesized by the ILT.

[0077] The second adversarial loss is also referred to as the adversarial loss of the first discriminator. The second adversarial loss is the sum of the first discriminator's correct recognition rate loss for the reference image and its false recognition rate loss for the actual mask image. The computing device can adjust the parameters of the first discriminator based on the second adversarial loss to reduce the second adversarial loss, thereby improving the first discriminator's discrimination capability.

[0078] The computing device then determines whether the sum of the first adversarial loss value and the second adversarial loss value is less than or equal to a first preset threshold. If not, this indicates that the image generated by the first generator is not realistic enough, and the first discriminator lacks the ability to distinguish between the actual mask image and the reference image. At this point, the sample image can be reprocessed using the adjusted parameters of the first generator, and the first adversarial loss value and the second adversarial loss value can be reacquired using the adjusted first discriminator until the sum of the first adversarial loss value and the second adversarial loss value is less than or equal to the first preset threshold, and training is stopped. This indicates that the image generated by the current first generator is more realistic, and the first discriminator has a strong ability to distinguish between the actual mask image and the reference image.

[0079] That is to say, the embodiment of the present application can alternately optimize the first discriminator parameters and the first generator parameters so that the image generated by the first generator is closer to the image containing sub-resolution auxiliary features synthesized by the real ILT technology, so that the first discriminator can be prepared to be the real image and the actual mask image.

[0080] In another example, the conversion network further includes a second generator. The second generator processes the actual mask image to generate an actual lithography image, and obtains a cycle consistency loss value for the actual lithography image relative to the sample image. It is understood that the cycle consistency loss value is the difference between the sample image and the actual lithography image after the first generator converts the sample image into the actual mask image and the second generator reversely reconstructs the actual mask image into the actual lithography image.

[0081] In an embodiment of the present application, a computing device comprehensively calculates the sum of a first adversarial loss value, a second adversarial loss value, and a cycle consistency loss value, and uses this sum value to determine whether it is less than or equal to a first preset threshold value. By introducing a second generator, an embodiment of the present application enables the conversion network to achieve a cyclic conversion from a sample image to an actual mask image, and then from the actual mask image to the sample image, thereby obtaining a cycle consistency loss value. Since the calculation of the cycle consistency loss value is based on the consistency of the content before and after the image conversion, by minimizing the cycle consistency loss value, the network can learn the bidirectional rule of image conversion, thereby ensuring the accuracy of the image conversion.

[0082] In another example, a first difference value of the actual mask image relative to the sample image and a second difference value of the actual lithography image relative to the actual mask image can also be obtained; based on the sum of the first difference value and the second difference value, an identity loss value is determined; the sum of the first adversarial loss value, the second adversarial loss value, the cycle consistency loss value and the identity loss value is obtained, and the sum is used to determine whether it is less than or equal to a preset threshold. This embodiment introduces an identity loss value by calculating the first difference value and the second difference value, which is used to evaluate the stability of the first generator when processing images in the same field. The identity loss value can be used to determine the similarity of the image before and after conversion in the same field. Therefore, by minimizing the identity loss value, the first generator can learn the rules for maintaining stable output in the same field, thereby avoiding information loss during image conversion and further optimizing the performance of the network.

[0083] It should be noted that the embodiment of the present application can also obtain the target network through other methods, which are not specifically limited in the embodiment of the present application.

[0084] In the embodiment of the present application, the computing device can use the target network to process the image to be lithographed to obtain a coarsened mask image. It can be understood that the target network has learned the image to be lithographed and the image containing sub-resolution assist features synthesized using the ILT synthesis technique corresponding to the image to be lithographed. In other words, the coarsened mask image is similar to the image containing sub-resolution assist features synthesized using the ILT technique.

[0085] S130 , using ILT to process the roughened mask image to obtain a target mask image, and using the target mask image to perform photolithography masking.

[0086] Although the coarsened mask image obtained using the target network approximates a mask image containing sub-resolution assist features synthesized using ILT technology, its accuracy is still insufficient to meet the requirements of actual lithography. This is primarily because the coarsened mask image obtained using the target network does not focus on image details, resulting in inaccurate details. To address this, the computing device uses ILT technology to refine the coarsened mask image, adjusting the details of the coarsened mask image to improve the accuracy of the obtained target mask image.

[0087] In summary, the embodiments of the present application can obtain an image to be photolithographically processed using a target network to obtain a coarsened mask image, and further process the coarsened mask image using ILT to obtain a target mask image, so as to use the target mask image for photolithographic masking. The target network is a conversion network trained based on training samples, wherein the training samples include sample images and reference images, and the reference image is an image synthesized by ILT containing sub-resolution auxiliary features. It can be understood that the target network can process the image to be photolithographically processed based on the correspondence between the learning sample image and the reference image, thereby obtaining a coarsened mask image containing sub-resolution auxiliary features that is similar to that synthesized by ILT. The coarsened mask image is refined using ILT technology to obtain a more accurate target mask image. Moreover, the use of ILT technology to fine-tune the coarsened mask image does not require a large search space and number of iterations, thereby significantly reducing the computational complexity and thus improving the efficiency of the photolithographic mask.

[0088] In actual use, the conversion network can be specifically a cycle-consistent generative adversarial network (Cycle-Consistent Generative Adversarial Network, CycleGAN). CycleGAN is a model that realizes cross-domain conversion of images based on adversarial training. Among them, cross-domain conversion refers to the conversion of images from one domain to another. In other words, the images converted across domains have styles from different domains, but the content of the images is consistent. CycleGAN is an unsupervised learning that can realize image-to-image conversion in an unpaired dataset, while the content of the two images is consistent. The following is a detailed explanation with reference to the accompanying drawings.

[0089] Attachment Figure 2 A flow chart of a method for acquiring a target network provided in an embodiment of the present application, the method comprising the following contents:

[0090] S210: Acquire a sample image set and a reference image set.

[0091] The sample image set and the reference image set do not need to correspond. The sample image set includes multiple sample images, and the reference image set includes multiple reference images, wherein the samples in the sample image set do not need to correspond to the reference images in the reference image set.

[0092] Furthermore, to reduce computational complexity, an interpolation method can be used to process the sample images in the sample image set to obtain a target lithography image, and to process the reference images in the reference image set to obtain a target reference image. The target lithography image and the target reference image have the same resolution. For example, the resolution of the target lithography image and the target reference image are both 256×256.

[0093] It should be noted that the interpolation method provided in the embodiment of the present application may be a nearest neighbor interpolation method, or a bilinear interpolation method, etc., which is not specifically limited in the embodiment of the present application.

[0094] S220: Train CycleGAN using the sample image set and the reference image set to obtain a target network.

[0095] In one specific implementation, a computing device first constructs a CycleGAN. CycleGAN is described below.

[0096] For example, the Figure 3 This is a schematic diagram of the structure of a CycleGAN provided in an embodiment of the present application. CycleGAN includes a first generator G, a first discriminator DY, a second generator F, and a second discriminator DX. The first generator G is used to convert images in domain X to images in domain Y, and the second generator F is used to convert images in domain Y to images in domain X. The first discriminator Y is used to process images in domain Y and determine the adversarial loss value, and the second discriminator DX is used to process images in domain X.

[0097] The first generator G and the second generator F have the same structure. Both adopt the U-Net structure. Figure 4 A schematic diagram of a U-Net structure provided in an embodiment of the present application. The generator includes an initialization module 401, a downsampling module 402, a self-attention residual module 403, an upsampling module 404, and an output layer 405.

[0098] Specifically, initialization module 401 obtains an image to be processed, processes the image to be processed, and establishes an initial network image. It should be noted that the image to be processed can be an image to be photolithographically processed, a sample image for training, or a reference image, and this is not specifically limited in this embodiment of the application. For ease of explanation, the following description uses the sample image as an example.

[0099] The initialization module 401 sends the initial network image to the downsampling module 402. Correspondingly, the downsampling module 402 obtains the initial network structure and performs multiple convolution downsampling on the initial network structure to obtain a first feature map. The downsampling module 402 sends the first feature map to the self-attention residual module 403. For example, the downsampling module includes two convolution layers, and each convolution layer includes a convolution with a stride of 2. This stage is mainly used to extract high-level features of the image to be processed. Specifically, the downsampling module 402 performs downsampling processing through a series of convolution layers, which can reduce the size of the feature map and increase the number of feature channels to capture more complex image features.

[0100] Exemplary description, attached Figure 5This is a schematic diagram of a generator provided in an embodiment of the present application processing an image. Figure 5 As shown, first, the initialization module 401 receives a 3-channel, 256×256 pixel image to be processed (i.e., 3×256×256). Next, the initialization module 401 processes the image to be processed to obtain an initial network image, specifically 64×256×256. This means that the initial network image has the same image size as the image to be processed, but the number of channels has increased to 64. The initialization module 401 sends the initial network image to the downsampling module 402.

[0101] Correspondingly, after receiving the initial network image, the downsampling module 402 downsamples the initial network image. Figure 5 The initial network image is convolved twice. The first convolution process produces a 128×128×128 feature map. The second convolution process reprocesses the 128×128×128 feature map to produce a first feature map of 256×64×64. Compared to the initial network image, the number of channels in the first feature map has increased to 256, and the size has changed. This means that the first feature map can learn more complex image features compared to the initial network image. The downsampling module 402 sends the first feature map to the self-attention residual module.

[0102] Furthermore, in the downsampling module, instance normalization and the first activation function are also used. Instance normalization is used to accelerate the training process, and the first activation function is used to introduce nonlinearity and improve the network's expressiveness. For example, the first activation function is the ReLU activation function.

[0103] Correspondingly, the self-attention residual module 403 can perform residual processing on the first feature map to obtain a second feature map. In an embodiment of the present application, the self-attention residual module includes a plurality of sub-residual modules, and the sub-residual modules are jump-connected so that the residual module allows the conversion to learn more complex features at a deep level, while maintaining the information consistency between the input and output through the jump connection, avoiding the loss of information and the problem of gradient disappearance during training. Exemplarily, the self-attention residual module 403 includes 9 sub-residual modules. Each sub-residual module includes two 3×3 convolutional layers. Among them, the self-attention mechanism is added to the sub-residual modules of the 4th, 5th, and 6th layers to enhance the generator's ability to capture key features and enhance the ability to retain details of complex graphics. The combination of the attention mechanism and the residual network can improve the performance of the generator, enabling it to better learn the task of generating complex mask patterns.

[0104] For example, see Figure 5As shown, the self-attention residual module 403 performs residual processing on the first feature map. After multiple residual processing, a second feature map of 256×64×64 is obtained. The second feature map of 256×64×64 includes more complex features and is consistent with the information of the input first feature map. The self-attention residual module sends the second feature map of 256×64×64 to the upsampling module.

[0105] Correspondingly, the upsampling module 404 upsamples the second feature map to obtain a third feature map, and sends the third feature map to the output layer 405. The goal of upsampling is to gradually restore the size of the image until the output is the same size as the input. This process is achieved through deconvolution or upsampling convolution, while reducing the number of feature channels to produce a more refined and clear image. Similar to the downsampling module 402, the upsampling module 404 typically uses the ReLU activation function and may utilize skip connections at certain levels to fuse the features obtained by downsampling with the features obtained by upsampling to restore finer details.

[0106] For example, referring to FIG. 5 , the upsampling module 404 performs a first convolution on the second feature map of 256×64×64 to obtain a feature map of 128×128×128, and performs a second convolution on the feature map of 128×128×128 to obtain a third feature map of 64×256×256. The upsampling module sends the third feature map to the output layer.

[0107] Correspondingly, the output layer 405 obtains the third feature map and outputs the actual mask image corresponding to the third feature map. In the embodiment of the present application, the output layer 405 is specifically used to: perform re-reflection filling, convolution, and a second activation function on the third feature map, mapping the output to a preset range to obtain the actual mask image. For example, the second activation function is a Tanh function, which helps the image output by the generator to be numerically consistent with the actual mask pattern to adapt to subsequent inversion lithography processing.

[0108] It should be noted that the preset range is a range set by those skilled in the art as needed, such as the preset range [-1, 1].

[0109] For example, see Figure 5 As shown in Figure 1, the output layer receives the third feature map of 64×256×256 and processes it through reflection padding, convolution and the second activation function to obtain the actual mask image of 3×256×256.

[0110] In the embodiment of the present application, the first discriminator DY and the second discriminator DX have the same structure. In one example, both structures are Markov discriminator (PatchGAN) structures. This structure is completely processed by convolutional layers.

[0111] For example, the Figure 6 A schematic diagram of a method for image processing based on PatchGAN provided in an embodiment of the present application, wherein the method includes three convolution downsampling processes and one spectral normalization process. Specifically, the discriminator includes a first convolutional layer, a second convolutional layer, a third convolutional layer, and a fourth convolutional layer, wherein the first, second, and third convolutional layers perform downsampling processing, and the fourth convolutional layer is used for channel number reduction processing.

[0112] For example, see Figure 6 As shown in the figure, after the discriminator receives the actual mask image of 3×256×256, it is processed by the first convolution layer to obtain a fourth feature map of 64×128×128. The second convolution layer performs a second convolution on the fourth feature map of 64×128×128 to obtain a fifth feature map of 128×64×64. The third convolution layer processes the fifth feature map of 128×64×64 to obtain a sixth feature map of 256×32×32. The fourth convolution layer processes the sixth feature map, and the output feature map of 1×32×32 is a 32×32 matrix. Each point (True / False) in the matrix represents the evaluation value of a small area in the generated actual mask image.

[0113] It should be noted that the discriminator provided in the embodiment of the present application can constrain the weights of the convolutional layer through spectral normalization, stabilize the adversarial training process, and prevent mode collapse.

[0114] The following introduces the training of CycleGAN using a sample image set and a reference image set as training sample sets. Specifically: the first generator G is used to convert images from domain X into images from domain Y, and the second generator F is used to convert images from domain Y into images from domain X. Among them, the images from domain X are sample images, and the images from domain Y are actual mask images corresponding to the sample images synthesized based on the ILT set. First, after the first generator G converts the sample images from domain X into actual mask images from domain Y, the first discriminator DY obtains the adversarial loss value through the following adversarial loss formula (1) to evaluate the realism of the actual mask image generated by the first generator, thereby adjusting the first generator G so that the image G(x) generated by the first generator G is closer to the real image y. Among them, the real image Y is the image synthesized from the sample image x based on the ILT technology.

[0115]

[0116] Among them, L GAN (G,D Y ,X,Y) is the adversarial loss value, is the second adversarial loss value, E(*) represents the expected value of the distribution function, P data(y)represents the distribution of real images y, P data(x) represents the distribution of the real sample image x, is the first adversarial loss value. Y (y) represents the probability estimate of the first discriminator Dy that the input real image y belongs to the real distribution of the reference image set Y. Y (G(x)) represents the probability estimate of the first discriminator Dy that the image G(x) generated by the first generator G belongs to the true distribution of the reference image set Y.

[0117] In the embodiment of the present application, it is possible to minimize L during the training process. GAN (G,D Y ,X,Y), so that the image G(x) is closer to the real image y.

[0118] Furthermore, the image can be reversely reconstructed from the Y domain to the X domain through the second generator F. That is, the second generator F can process the actual mask image G(x) to obtain the actual lithography image F(G(x)). The second generator F can determine the difference between the actual lithography image F(G(x)) and the sample image X through the cycle consistency loss formula (3) to ensure that the actual lithography image F(G(x)) is consistent with the sample image X:

[0119]

[0120] Among them, L cycle (G,F) is the cycle consistency function, ||F(G(x))-x|| l is the l1 norm of F(G(x)) and x, which is used to express the difference between F(G(x)) and x, ||G(F(y))-y|| l is the l1 norm of G(F(y)) and y, which is used to represent the difference between G(F(y)) and y.

[0121] In the embodiment of the present application, by minimizing L cycle (G,F), ensuring that the image generated by the first generator is consistent with the real image in content.

[0122] The second generator G is also used to obtain the identity loss value L Identity (G, F). The specific acquisition formula is shown in formula (4):

[0123]

[0124] It can be understood that identity loss means that when the input image to be processed already belongs to the target domain (ie, Y domain) image, then the image generated by the first generator should be close to the input image. In this way, the consistency of the sample characteristics can be enhanced, and the first generator can be helped to learn the mapping that maintains the image structure, that is, G(y) = y, F(x) = x. If they are not equal, the identity loss will correct the generator by penalizing the modification.

[0125] The computing device can obtain the total loss function value, that is, refer to the acquisition formula (5):

[0126] L(G,F,D X ,D Y )=L GAN (G,D Y ,X,Y)+L GAN (F,D X ,Y,X)+λ cyc L cyc (G,F) (5)

[0127] +λ id L Identity (G,F)

[0128] Among them, λcyc and λid are weight parameters used to adjust the cycle consistency loss and identity loss. These three loss functions work together in the training process of CycleGAN, enabling the generator to effectively convert between different domains while maintaining the cycle consistency and identity consistency of the image. It is through these three loss functions that the CycleGAN model can still generate high-quality images in the case of unpaired training samples.

[0129] In an embodiment of the present application, the discriminator can alleviate mode collapse by adopting a historically generated image buffer pool, and iteratively update the parameters of the first generator, the second generator, the first discriminator and / or the second discriminator through the Adam optimizer to achieve high-quality image conversion across domains.

[0130] For example, during training, the generator learning rate can be set to 0.0001, the discriminator learning rate to 0.00001, and the Adam optimizer hyperparameters β1 = 0.5 and β2 = 0.999. During data loading, data augmentation is performed on the data, and the loaded images are transformed, including zooming, cropping, horizontal flipping, and normalization. This enhances the model's ability to generalize image transformations during training. After training, the trained model weights for the curve mask are obtained, which constitute the target network.

[0131] In the embodiment of the present application, the computing device takes the image to be lithographically processed as the input of the first generator G, and can directly generate a roughened mask pattern, which is then refined by the ILT technology to obtain a target mask pattern. Figure 7 A schematic diagram of processing a photolithography mask provided in an embodiment of the present application, wherein: Figure 7 (a) is the image to be photolithographically processed. Figure 7 (b) in the figure is the coarsened mask image obtained by processing with the target network. Figure 7 (c) in FIG. 5 is a target mask image obtained by refining the coarsening mask image using the ILT technology.

[0132] To sum up, the target network of this application is a cycle-consistent generative adversarial network. By training the cycle-consistent generative adversarial network through a sample image set and a reference image set, it can be ensured that the obtained coarsened mask image is closer to the image containing sub-resolution auxiliary features synthesized by the ILT technology corresponding to the actual image to be photolithography.

[0133] In addition, an embodiment of the present application also provides a photolithography mask device.

[0134] Attachment Figure 8 This is a schematic structural diagram of a photolithography mask device provided in an embodiment of the present application. The device 800 includes:

[0135] An acquisition unit 801 is used to acquire an image to be photolithographically processed;

[0136] The initial processing unit 802 is used to process the image to be photolithographically processed using a target network to obtain a roughened mask image;

[0137] The target network is a conversion network trained based on training samples, the training samples include sample images and reference images, and the reference image is an image synthesized by inversion lithography technology (ILT) and containing sub-resolution auxiliary features;

[0138] The post-processing unit 803 is configured to process the roughened mask image using ILT to obtain a target mask image; and use the target mask image to perform photolithography masking.

[0139] Optionally, the conversion network includes a first generator and a first discriminator, wherein the first generator is used to convert the image to be photolithographically processed into the roughened mask image; the apparatus 800 further includes a training unit, wherein the training unit is used to train the conversion network, and the training unit is specifically used to:

[0140] Obtaining sample images and reference images;

[0141] Processing the sample image using the first generator to generate an actual mask image;

[0142] Obtaining a first adversarial loss value using the first discriminator; the first adversarial loss value indicates a difference between a predicted probability obtained after the actual mask image is processed by the first discriminator and a preset ideal value;

[0143] Adjusting parameters of the first generator according to the first adversarial loss value;

[0144] Obtaining a second adversarial loss value using the first discriminator, where the second adversarial loss value is the sum of a correct recognition rate loss of the reference image and an incorrect recognition rate loss of the actual mask image;

[0145] Adjusting parameters of the first discriminator according to the second adversarial loss value;

[0146] Determining whether a sum of the first adversarial loss and the second adversarial loss is less than or equal to a preset threshold;

[0147] If not, re-execute the operation of generating the actual mask image, obtaining the first adversarial loss value, the second adversarial loss value, and the sum of the two loss values; wherein the adjusted first discriminator is used as the first discriminator to generate the actual mask image, and the adjusted second discriminator is used as the second discriminator to perform the operation of obtaining the first adversarial loss value, the second adversarial loss value, and the sum of the two loss values;

[0148] If so, stop training and obtain the target network.

[0149] Optionally, the conversion network further includes a second generator, and the training unit is specifically configured to:

[0150] Processing the actual mask image by the second generator to generate an actual lithography image;

[0151] Obtaining a cycle consistency loss value of the actual lithography pattern relative to the sample image;

[0152] The determining whether the sum of the first adversarial loss and the second adversarial loss is less than or equal to a preset threshold includes:

[0153] Determine whether the sum of the first adversarial loss value, the second adversarial loss value, and the cycle consistency loss value is less than or equal to a preset threshold.

[0154] The training unit is also used to:

[0155] obtaining an identity loss value, wherein the identity loss value is a first difference value of the actual mask image relative to the sample image, and obtaining a sum of a second difference value of the actual lithography image relative to the actual mask image;

[0156] The determining whether the sum of the first adversarial loss value, the second adversarial loss value, and the cycle consistency loss value is less than or equal to a preset threshold includes:

[0157] Determine whether the sum of the first adversarial loss value, the second adversarial loss value, the identity loss value, and the cycle consistency loss value is less than or equal to a first preset threshold.

[0158] Optionally, the first generator includes: an initialization module, a downsampling module, a self-attention residual module, an upsampling module and an output layer, and the pre-processing unit 802 is specifically used to:

[0159] Creating an initial network image corresponding to the image to be photolithographically processed by the initialization module;

[0160] Sampling the initial network image using the downsampling module to obtain a first feature map;

[0161] Performing residual processing on the first feature map through the self-attention residual module to obtain a second feature map;

[0162] Sampling the second feature map using the upsampling module to obtain a third feature map;

[0163] The coarsening mask image corresponding to the third feature map is output through the output layer.

[0164] Optionally, the downsampling module includes a first feature extraction layer and a second feature extraction layer, and the first feature extraction layer and the second feature extraction layer both include: a convolution layer with a step size of 2, an instance normalization layer and a first activation function; the first activation function is an activation function that introduces nonlinear expression.

[0165] Optionally, outputting the coarsening mask image corresponding to the third feature map through the output layer includes:

[0166] The third feature map is processed by reflection filling, convolution and second activation function through the output layer to obtain the coarsening mask image.

[0167] Optionally, the first discriminator is a Markov discriminator structure, comprising a first convolutional layer, a second convolutional layer, a third convolutional layer and a fourth convolutional layer connected in sequence; the fourth convolutional layer is used for channel number reduction processing;

[0168] The obtaining a first adversarial loss value by using the first discriminator includes:

[0169] downsampling the actual mask image using the first convolutional layer to obtain a fourth feature map;

[0170] Downsampling the fourth feature map using the second convolutional layer to obtain a fifth feature map;

[0171] Downsampling the fifth feature map using the third convolutional layer to obtain a sixth feature map;

[0172] The fourth convolutional layer is used to reduce the number of channels of the sixth feature map to obtain the first adversarial loss value.

[0173] Optionally, the conversion network is a cycle-consistent adversarial network.

[0174] The lithography mask device provided in embodiments of the present application can obtain an image to be lithographically processed using a target network to obtain a coarsened mask image, and then further process the coarsened mask image using ILT to obtain a target mask image, which can then be used for lithography masking. The target network is a transformer network trained based on training samples, where the training samples include sample images and reference images, with the reference image being an ILT-synthesized image containing sub-resolution auxiliary features. It can be understood that the target network can process the image to be lithographically processed based on the correspondence between the learned sample images and the reference images to obtain a coarsened mask image containing sub-resolution auxiliary features that is similar to the ILT-synthesized one. The coarsened mask image is then refined using ILT to obtain a more accurate target mask image. The coarsened mask image is then fine-tuned using ILT without requiring a large search space or number of iterations. This significantly reduces computational complexity while maintaining lithography mask accuracy, thereby improving lithography mask efficiency.

[0175] Based on the methods provided in the embodiments of the present application, the present application also provides a chip system, which includes one or more processors configured to retrieve and execute instructions stored in a memory, thereby executing the methods of the embodiments of the present application. The chip system can be composed of a chip or can include a chip and other discrete devices.

[0176] Among them, the chip system may include an input circuit or interface for sending information or data, and an output circuit or interface for receiving information or data.

[0177] According to the method provided in the embodiments of the present application, the present application also provides a computer program product, which includes: computer program code, which, when running on a computer, enables the computer to execute the various steps or processes executed by the network device and terminal device in any of the aforementioned method embodiments.

[0178] According to the method provided in the embodiments of the present application, the present application also provides a computer-readable storage medium, which stores program code. When the program code runs on a computer, the computer executes the various steps or processes performed by the network device and terminal device in any of the aforementioned method embodiments.

[0179] The computer-readable storage medium may be the aforementioned volatile memory or non-volatile memory, or may include both volatile memory and non-volatile memory.

[0180] In the embodiments of this application, each term and English abbreviation is provided for convenience of description and shall not constitute any limitation to this application. This application does not exclude the possibility of defining other terms that can achieve the same or similar functions in existing or future agreements.

[0181] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part.

[0182] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

Claims

1. A photolithography mask method, characterized in that: The method comprises: Acquire an image to be photolithographically processed; Processing the image to be photolithographically processed using a target network to obtain a roughened mask image; The target network is a conversion network trained based on training samples, the training samples include sample images and reference images, and the reference image is an image containing sub-resolution auxiliary features synthesized by inversion lithography technology (ILT); The roughened mask image is processed using ILT to obtain a target mask image, so as to use the target mask image for photolithography masking.

2. The mask method according to claim 1, wherein: The conversion network includes a first generator and a first discriminator, wherein the first generator is used to convert the image to be photolithography into the roughening mask image; the conversion network is trained in the following manner: Obtaining sample images and reference images; Processing the sample image using the first generator to generate an actual mask image; Obtaining a first adversarial loss value using the first discriminator; the first adversarial loss value indicates a difference between a predicted probability obtained after the actual mask image is processed by the first discriminator and a preset ideal value; Adjusting parameters of the first generator according to the first adversarial loss value; Obtaining a second adversarial loss value using the first discriminator, where the second adversarial loss value is the sum of a correct recognition rate loss of the reference image and an incorrect recognition rate loss of the actual mask image; Adjusting parameters of the first discriminator according to the second adversarial loss value; Determining whether a sum of the first adversarial loss and the second adversarial loss is less than or equal to a preset threshold; If not, re-execute the operation of generating the actual mask image, obtaining the first adversarial loss value, the second adversarial loss value, and the sum of the two loss values; wherein the adjusted first discriminator is used as the first discriminator to generate the actual mask image, and the adjusted second discriminator is used as the second discriminator to perform the operation of obtaining the first adversarial loss value, the second adversarial loss value, and the sum of the two loss values; If so, stop training and obtain the target network.

3. The mask method according to claim 2, wherein: The conversion network further includes a second generator, and the method further includes: Processing the actual mask image by the second generator to generate an actual lithography image; Obtaining a cycle consistency loss value of the actual lithography pattern relative to the sample image; The determining whether the sum of the first adversarial loss and the second adversarial loss is less than or equal to a preset threshold includes: Determine whether the sum of the first adversarial loss value, the second adversarial loss value, and the cycle consistency loss value is less than or equal to a preset threshold.

4. The mask method according to claim 3, wherein: The method further comprises: obtaining an identity loss value, wherein the identity loss value is a first difference value of the actual mask image relative to the sample image, and obtaining a sum of a second difference value of the actual lithography image relative to the actual mask image; The determining whether the sum of the first adversarial loss value, the second adversarial loss value, and the cycle consistency loss value is less than or equal to a preset threshold includes: Determine whether the sum of the first adversarial loss value, the second adversarial loss value, the identity loss value, and the cycle consistency loss value is less than or equal to a first preset threshold.

5. The masking method according to any one of claims 1 to 4, characterized in that: The first generator includes: an initialization module, a downsampling module, a self-attention residual module, an upsampling module and an output layer. The target network is used to process the image to be lithographically processed to obtain a coarsened mask image, including: Creating an initial network image corresponding to the image to be photolithographically processed by the initialization module; Sampling the initial network image using the downsampling module to obtain a first feature map; Performing residual processing on the first feature map through the self-attention residual module to obtain a second feature map; Sampling the second feature map using the upsampling module to obtain a third feature map; The coarsening mask image corresponding to the third feature map is output through the output layer.

6. The masking method according to claim 5, characterized in that: The downsampling module includes a first feature extraction layer and a second feature extraction layer, and the first feature extraction layer and the second feature extraction layer both include: a convolution layer with a step size of 2, an instance normalization layer and a first activation function; the first activation function is an activation function that introduces nonlinear expression.

7. The mask method according to claim 6, wherein: Outputting the coarsening mask image corresponding to the third feature map through the output layer includes: The third feature map is processed by reflection filling, convolution and second activation function through the output layer to obtain the coarsening mask image.

8. The mask method according to claim 2, wherein: The first discriminator is a Markov discriminator structure, comprising a first convolutional layer, a second convolutional layer, a third convolutional layer and a fourth convolutional layer connected in sequence; the fourth convolutional layer is weighted by spectral normalization; The obtaining a first adversarial loss value by using the first discriminator includes: Use the first convolution layer to downsample the actual mask image to obtain a fourth feature map; use the second convolution layer to downsample the fourth feature map to obtain a fifth feature map; use the third convolution layer to downsample the fifth feature map to obtain a sixth feature map; use the fourth convolution layer to reduce the number of channels on the sixth feature map to obtain the first adversarial loss value.

9. The mask method according to claim 1, wherein: The conversion network is a cycle-consistent adversarial network.

10. A photolithography mask device, characterized in that: The device comprises: An acquisition unit, used for acquiring an image to be photolithographically processed; A primary processing unit, configured to process the image to be photolithographically processed using a target network to obtain a roughened mask image; The target network is a conversion network trained based on training samples, the training samples include sample images and reference images, and the reference image is an image containing sub-resolution auxiliary features synthesized by inversion lithography technology (ILT); A post-processing unit is used to process the roughened mask image using ILT to obtain a target mask image; and to perform photolithography masking using the target mask image.

Images