Methods, apparatus, electronic devices, and storage media for determining the photolithography process window

Abstract

This application relates to a method, apparatus, electronic device, and storage medium for determining a lithography process window. The method includes: performing pattern segmentation on an acquired original lithography image to generate multiple standard pixel images; inputting the multiple standard pixel images into a trained target pattern extraction model for lithography pattern extraction processing, and outputting the required predicted lithography pattern; wherein the target pattern extraction model is generated by training an initial generative adversarial network based on preset random noise data; and performing lithography process window analysis processing based on the predicted lithography pattern and the original lithography image to determine the lithography process window. This application solves the problem of poor accuracy and reliability of analysis results due to limited sample data in lithography process window analysis, significantly reducing the requirement for a large amount of original data and improving the robustness and accuracy of the lithography process window analysis algorithm.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing processes, and in particular to methods, apparatus, electronic devices, and storage media for determining photolithography process windows. Background Technology

[0002] Photolithography is an indispensable part of semiconductor integrated circuit manufacturing. Its purpose is to precisely transfer patterns from a photomask onto a silicon wafer to form intricate circuit structures. The process window, as a crucial parameter in photolithography, is essential for ensuring production stability and product quality.

[0003] In terms of related technologies, traditional lithography process window analysis relies on large amounts of high-quality data for model training and validation. However, in actual production environments, due to factors such as high data acquisition costs, complex processes, and time constraints, it is often difficult to obtain a sufficient number of lithography image data. This problem of small sample datasets greatly limits the accuracy and reliability of process window analysis, thereby affecting the quality and production efficiency of semiconductor products.

[0004] Currently, no effective solution has been proposed for addressing the issue of poor accuracy and reliability of analysis results due to limited sample data in the analysis of photolithography process windows. Summary of the Invention

[0005] This application provides a method, apparatus, system, electronic device, and storage medium for determining a lithography process window, in order to at least solve the problem in the related art of how to address the poor accuracy and reliability of analysis results due to limited sample data in the analysis of lithography process windows.

[0006] In a first aspect, embodiments of this application provide a method for determining a photolithography process window, including:

[0007] The acquired raw photolithography image is subjected to pattern segmentation to generate multiple standard pixel images;

[0008] Multiple standard pixel images are input into a trained target pattern extraction model for lithographic pattern extraction and output the desired predicted lithographic pattern. The target pattern extraction model is generated by training an initial generative adversarial network based on preset random noise data.

[0009] Based on the predicted lithography pattern and the original lithography image, the lithography process window is analyzed and processed to determine the lithography process window.

[0010] In some embodiments, the step of performing pattern segmentation on the acquired raw photolithography image to generate multiple standard pixel images includes:

[0011] Based on the pattern type of the original photolithographic image, a pattern segmentation operation is performed on the original photolithographic image to obtain multiple single feature images;

[0012] Preprocessing operations are performed on multiple single-feature images to generate multiple standard pixel images.

[0013] In some embodiments, the preprocessing operation on the plurality of the single feature images to generate the plurality of the standard pixel images includes:

[0014] Multiple single-feature images are rotated and Gaussian blurred to generate multiple standard pixel images.

[0015] In some embodiments, the pattern segmentation operation on the original photolithographic image based on the pattern type of the original photolithographic image yields multiple single-feature images, including:

[0016] If the pattern type is a line-dense or space-dense type, the original lithographic image is segmented based on its position and key size data to obtain multiple single feature images.

[0017] If the pattern type is line sparse or spatial sparse, the lithographic pattern is normalized and segmented to obtain multiple single feature images;

[0018] If the pattern type is a hole-dense or diameter-dense type, the lithographic pattern is segmented based on the image side length of the original lithographic image and the distribution position of the holes in the original lithographic image to obtain multiple single feature images;

[0019] If the pattern type is a sparse hole type or a sparse diameter type, the photolithographic pattern is normalized and segmented to obtain multiple single feature images.

[0020] In some embodiments, the generative adversarial network includes a discriminator and a generator; the method further includes:

[0021] Acquire training lithography images and perform pattern segmentation on the training lithography images to generate multiple training segmented images;

[0022] The random noise data is input into the generator in the generative adversarial network, and the generator outputs a predicted training image.

[0023] The training segmented image and the predicted training image are input into the discriminator in the generative adversarial network, and the probability prediction result is output.

[0024] Based on the probability prediction results, the discriminator and the generator in the generative adversarial network are continuously iterated and optimized to finally generate the trained target pattern extraction model.

[0025] In some embodiments, the step of inputting the training segmented image and the predicted training image into the discriminator in the generative adversarial network and outputting a probability prediction result includes:

[0026] The training segmented image is input into the discriminator, and the discriminator is used to extract features from the training segmented image to obtain image feature data. Based on the image feature data, fixed network parameters for the discriminator are determined.

[0027] The predicted training image is input into a discriminator with the fixed network parameters for convolution calculation, and the probability prediction result is output.

[0028] In some embodiments, the probability prediction result includes the similarity between the training switching image and the predicted training image; the step of iteratively optimizing the discriminator and the generator in the generative adversarial network based on the probability prediction result to finally generate the trained target pattern extraction model includes:

[0029] The similarity is compared with a preset confidence threshold to obtain a similarity comparison result. Based on the similarity comparison result, the generative adversarial network is iteratively optimized to generate the trained target pattern extraction model.

[0030] Secondly, embodiments of this application provide a device for determining a photolithography process window, comprising:

[0031] The preprocessing module is used to perform pattern segmentation on the acquired raw photolithography image to generate multiple standard pixel images;

[0032] The data augmentation module is used to input multiple standard pixel images into the trained target pattern extraction model for lithographic pattern extraction processing and output the required predicted lithographic pattern; wherein, the target pattern extraction model is generated by training an initial generative adversarial network based on preset random noise data;

[0033] The lithography analysis module is used to perform lithography process window analysis and processing based on the predicted lithography pattern and the original lithography image to determine the lithography process window.

[0034] Thirdly, embodiments of this application provide an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for determining the photolithography process window as described in the first aspect above.

[0035] Fourthly, embodiments of this application provide a storage medium storing a computer program that, when executed by a processor, implements the method for determining the photolithography process window as described in the first aspect above.

[0036] Compared to related technologies, the method, apparatus, electronic device, and storage medium for determining the lithography process window provided in this application collect and preprocess photoresist pattern measurement image data, construct a target pattern extraction model for training, and generate synthetic data that is highly similar to the real image, thereby expanding and balancing the dataset. This solves the problem of poor accuracy and reliability of analysis results due to limited sample data in lithography process window analysis, significantly reduces the need for the amount of original data, and improves the robustness and accuracy of the lithography process window analysis algorithm.

[0037] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0038] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0039] Figure 1 This is a hardware structure block diagram of the terminal of the photolithography process window determination method according to an embodiment of the present invention.

[0040] Figure 2 This is a flowchart of a method for determining a photolithography process window according to an embodiment of this application;

[0041] Figure 3 This is a schematic diagram of the structure of a depthwise convolutional generative adversarial network for determining the photolithography process window according to a preferred embodiment of this application.

[0042] Figure 4 This is a comparison image of a synthetic photoresist image generated during model training according to a preferred embodiment of this application and a real image;

[0043] Figure 5 This is a graph showing the relationship between the amount of raw data and the accuracy of synthesized image recognition according to a preferred embodiment of this application.

[0044] Figure 6This is a schematic diagram of a first segmentation method for dense patterns of line or space type according to a preferred embodiment of this application;

[0045] Figure 7 It is a segmented image of a first segmentation method for dense patterns of line or space type according to a preferred embodiment of this application;

[0046] Figure 8 This is a schematic diagram of a second segmentation method for dense patterns of line or space type according to a preferred embodiment of this application;

[0047] Figure 9 This is a schematic diagram of a third segmentation method for dense patterns of line or space type according to a preferred embodiment of this application;

[0048] Figure 10 This is a schematic diagram of a sparse pattern of line or space type according to a preferred embodiment of this application;

[0049] Figure 11 This is a schematic diagram of a dense pattern of diameter or hole type according to a preferred embodiment of this application;

[0050] Figure 12 It is a segmented image of a dense pattern of diameter or hole type according to a preferred embodiment of this application;

[0051] Figure 13 This is a schematic diagram of a sparse pattern of diameter or hole type according to a preferred embodiment of this application;

[0052] Figure 14 This is a structural block diagram of a photolithography process window determination apparatus according to an embodiment of this application. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application. Furthermore, it is understood that although the efforts made in such a development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, modifications to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0054] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0055] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application means two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The terms “first,” “second,” “third,” etc., used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0056] The method embodiments provided in this example can be executed on a terminal, computer, or similar computing device. Taking running on a terminal as an example, Figure 1 This is a hardware structure block diagram of the terminal of the method for determining the photolithography process window according to an embodiment of the present invention. For example... Figure 1 As shown, a terminal may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. Optionally, the terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the terminal described above. For example, the terminal may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1The different configurations shown.

[0057] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the method for determining the photolithography process window in this embodiment of the invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0058] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the terminal's communication provider. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0059] This embodiment provides a method for determining the photolithography process window. Figure 2 This is a flowchart of a method for determining a photolithography process window according to an embodiment of this application, such as... Figure 2 As shown, the process includes the following steps:

[0060] Step S201: Perform pattern segmentation on the acquired original photolithography image to generate multiple standard pixel images.

[0061] This step preprocesses the original lithographic image to prepare it for subsequent input into the deep learning model. Specifically, the image can be segmented based on its physical meaning and features. For example, for dense patterns of lines or spaces, the original image can be segmented into multiple smaller images containing only a single feature, based on the pitch and the width and position of the critical dimension bar (CD bar). The segmentation method should ensure that each smaller image accurately reflects the specific features in the original image. This segmentation operation significantly expands the data volume and reduces the adverse effects of excessive data classification; the preprocessing operation also standardizes the image data, contributing to improved training efficiency and accuracy of the deep learning model.

[0062] Step S202: Input multiple standard pixel images into the trained target pattern extraction model for lithographic pattern extraction processing, and output the required predicted lithographic pattern; wherein, the target pattern extraction model is generated by training the initial generative adversarial network based on preset random noise data.

[0063] The target pattern extraction model is a DCGAN (Deep Convolutional Generative Adversarial Networks) model. This step uses a trained DCGAN model to extract the lithographic pattern. Specifically, the generated standard pixel image is input into the DCGAN generator. The generator then generates computer-synthesized images similar to the real lithographic pattern based on the input random noise vector. These synthesized images predict the lithographic pattern and are output to subsequent analysis steps. This step, using the DCGAN model, can generate a large number of high-quality synthesized images, significantly reducing the amount of original data required and improving the efficiency and accuracy of lithographic pattern extraction, providing strong support for subsequent lithography process window analysis.

[0064] Step S203: Based on the predicted lithography pattern and the original lithography image, perform lithography process window analysis and processing to determine the lithography process window.

[0065] This step combines the predicted lithography pattern with the original lithography image to further train an advanced deep learning-based lithography process window analysis tool. Through data augmentation and dataset expansion, the lithography process window analysis model can better identify and predict features of exposure, development, or measurement process anomalies in the photoresist pattern measurement image. Ultimately, the model determines the lithography process window based on these features, i.e., the range of exposure energy and focus values ​​that meets the critical dimension (CD) and overlay requirements. This step, through data augmentation and expansion, improves the robustness and accuracy of the lithography process window analysis tool, reduces manual intervention, and enhances the efficiency and consistency of the analysis process.

[0066] Through the above steps, this application performs pattern segmentation on the acquired original lithographic image. Based on the pattern type of the lithographic image (such as dense lines, dense spaces, dense holes, etc.), the original image is segmented into multiple small images containing only a single feature, and then converted into standard pixel images suitable for input into DCGAN. This process not only expands the amount of data, but also improves the consistency of data through standardization. Next, the multiple standard pixel images are input into the trained target pattern extraction model for lithographic pattern extraction. This target pattern extraction model is generated by training the initial DCGAN based on preset random noise data. The DCGAN consists of two neural networks: a generator and a discriminator. The system is structured around a generator that, through optimized adversarial training, progressively generates images similar to real lithography pattern measurement images. These generated images possess features similar to real data, effectively supplementing the limitations of small sample datasets and balancing various data types. Finally, based on the predicted lithography pattern and the original lithography image, lithography process window analysis is performed. The computer-synthesized lithography pattern measurement image data generated by DCGAN is combined with the original data to further train an advanced deep learning-based lithography process window analysis tool. Through data augmentation and dataset expansion, the model can better identify and predict features in the lithography pattern measurement image, thereby accurately determining the lithography process window. Compared to related technologies, this application significantly expands the dataset and improves data balance through DCGAN technology. Using the augmented dataset for training allows for more thorough training of the deep learning model, enabling more accurate extraction of lithography pattern features. Therefore, this application reduces the amount of original data required while maintaining a performance level comparable to using only the original data, effectively solving the problem of poor accuracy and reliability of analysis results due to limited sample data.

[0067] In some embodiments, the acquired raw photolithographic image is subjected to pattern segmentation to generate multiple standard pixel images, including:

[0068] Based on the pattern type of the original lithographic image, a pattern segmentation operation is performed on the original lithographic image to obtain multiple single feature images;

[0069] Multiple single-feature images are preprocessed to generate multiple standard pixel images.

[0070] The process begins by identifying the pattern type of the original lithography image. These types may include dense line, dense spatial, sparse line, sparse spatial, dense aperture, and dense diameter patterns. Based on the pattern type, different pattern segmentation operations are performed on the original lithography image to obtain multiple single-feature images. These segmented single-feature images are then preprocessed. This embodiment, through precise pattern segmentation and preprocessing, significantly expands the dataset required for lithography process analysis, especially for categories with a smaller number of data points. This mitigates the adverse effects of imbalanced small-sample datasets on training deep learning networks. The automated pattern segmentation and preprocessing process reduces manual intervention, improves the efficiency and consistency of dataset processing, and replaces manual visual judgment with computer-standardized evaluation, ensuring the accuracy and reliability of data processing.

[0071] In some embodiments, preprocessing operations are performed on multiple single-feature images to generate multiple standard pixel images, including:

[0072] Multiple single-feature images are rotated and Gaussian blurred to generate multiple standard pixel images.

[0073] This method involves rotating each individual feature image by a certain angle (e.g., 0°, 90°, 180°, 270°, etc.) to simulate lithography patterns at different angles. Rotation maintains the integrity of the image content while increasing the diversity of the dataset. A Gaussian blur filter is applied to the rotated images, and the degree of blur is controlled by adjusting blur parameters (e.g., standard deviation). Gaussian blurring simulates image blurring phenomena that may occur during actual lithography, improving the model's ability to recognize blurred images. The rotated and Gaussian-blurred images are then resized to a standard size (e.g., 64×64 pixels) suitable for input into a deep convolutional generative adversarial network (DCGAN). This resizing ensures that all input images have consistent dimensions, facilitating model processing. This embodiment generates a large number of images with different feature combinations through rotation and Gaussian blurring, significantly enhancing the diversity and richness of the dataset. This helps the model learn more robust and generalized feature representations, improving its generalization ability. Training with pre-processed standard pixel images allows the model to better adapt to the diversity of actual lithography patterns, improving the model's accuracy and efficiency in lithography process window analysis.

[0074] In some embodiments, based on the pattern type of the original photolithographic image, a pattern segmentation operation is performed on the original photolithographic image to obtain multiple single-feature images, including:

[0075] If the pattern type is line-dense or space-dense, the original lithography image is segmented based on its position and key size data to obtain multiple single feature images.

[0076] If the pattern type is line-sparse or spatially sparse, the lithographic pattern is normalized and segmented to obtain multiple single feature images;

[0077] If the pattern type is a hole-dense or diameter-dense type, the lithographic pattern is segmented based on the image side length of the original lithographic image and the distribution position of the holes in the original lithographic image to obtain multiple single feature images;

[0078] If the pattern type is sparse in aperture or sparse in diameter, the photolithographic pattern is normalized and segmented to obtain multiple single feature images.

[0079] For line-dense or spatially-dense patterns, the segmentation operation considers the specific location information and key dimensions (such as line width and spacing) of the lithographic image. Based on the density of the pattern, the distribution of lines or space, and the specific values ​​of the key dimensions, the original image is precisely segmented to obtain multiple smaller images containing only a single feature. For line-sparse or spatially-sparse patterns, due to the large spacing between pattern elements, a normalized segmentation method is used to normalize the original lithographic image, ensuring that each segmented smaller image has a consistent size and features, facilitating subsequent processing and analysis. For hole-dense or diameter-dense patterns, the segmentation operation considers the overall size of the image and the specific distribution of the holes. Based on the density and distribution pattern of the holes, the original image is precisely segmented to obtain multiple smaller images containing a single hole feature. Similar to line-sparse or spatially-sparse patterns, hole-sparse or diameter-sparse patterns also use a normalized segmentation method to ensure that each segmented smaller image has a consistent size and features, facilitating subsequent data augmentation and deep learning model training.

[0080] This embodiment improves data quality and diversity: through precise segmentation, multiple standard pixel images containing a single feature can be obtained. These images have higher quality and diversity, which helps to improve the training effect of deep learning models. The automated segmentation method greatly reduces the degree of manual intervention and improves the efficiency and consistency of data processing. Using a precisely segmented and preprocessed dataset for training can enhance the robustness and accuracy of deep learning models in lithography process window analysis. Furthermore, specific segmentation methods are proposed for different types of lithography patterns, making this method more widely applicable.

[0081] In some embodiments, the generative adversarial network includes a discriminator and a generator; the method further includes:

[0082] Acquire training lithography images and perform pattern segmentation on the training lithography images to generate multiple training segmented images;

[0083] Random noise data is input into the generator in the generative adversarial network, and the output is a predicted training image;

[0084] The training segmented image and the predicted training image are input into the discriminator in the generative adversarial network, and the probability prediction result is output.

[0085] Based on the probability prediction results, the discriminator and generator in the generative adversarial network are continuously optimized through iterative cycles, and finally a well-trained target pattern extraction model is generated.

[0086] In this training process, a certain number of lithography images are selected from existing lithography datasets as training images. Appropriate segmentation methods are applied based on the pattern type of the lithography images, and the segmented images serve as training segmented images for subsequent GAN training. The generator receives a random noise vector as input and progressively transforms it into two-dimensional computer-synthesized lithography pattern measurement images through deconvolution layers. These generated images (i.e., predicted training images) have similar characteristics to real lithography images. The training segmented images and predicted training images are input to the discriminator in the generative adversarial network. The discriminator is a standard convolutional neural network used to distinguish whether the input image is a real training segmented image or a predicted training image generated by the generator. The discriminator differentiates between the training segmented images and the predicted training images and outputs a probability prediction result, i.e., the probability that the image is considered a real image. During training, the generator attempts to generate realistic images to deceive the discriminator, while the discriminator strives to identify whether the images are real or fake. By optimizing adversarial training and progressively adjusting the network parameters of the generator and discriminator, the quality of images generated by the generator is improved. After multiple iterations, the generator can produce high-quality images similar to real datasets. The training process ends when the images generated by the generator are realistic enough that the discriminator has difficulty distinguishing between real and fake images. At this point, the obtained generator is the trained target pattern extraction model, which can be used to generate synthetic images similar to real lithographic images.

[0087] This embodiment utilizes GAN technology to generate a large amount of high-quality synthetic data using a small amount of real data, thereby reducing reliance on large amounts of original data. The generated synthetic data can balance the quantity of various types of data, mitigating the adverse effects of imbalanced small sample datasets on deep learning network training. GAN technology can automatically generate highly similar synthetic images, reducing the workload of manual intervention and data annotation, and improving the efficiency of data collection and processing. Furthermore, by using data-enhanced datasets for training, the deep learning network can more accurately extract the features of photoresist patterns, improving the accuracy and robustness of photolithography process window analysis tools.

[0088] In some embodiments, the training segmented image and the predicted training image are input into the discriminator in the generative adversarial network, and the output probability prediction result includes:

[0089] The training segmented image is input into the discriminator, which extracts features from the training segmented image to obtain image feature data. Based on the image feature data, fixed network parameters for the discriminator are determined.

[0090] The predicted training image is input into a discriminator with fixed network parameters for convolutional computation, and the probability prediction result is output.

[0091] In this process, multiple training segmented images obtained through pattern segmentation are input into a discriminator, which is a standard convolutional neural network with multiple convolutional layers for feature extraction. The discriminator performs convolution operations on the input training segmented images to extract image feature data. Through training, the discriminator gradually learns how to distinguish features between real and generated images. When the discriminator's feature extraction of the training segmented images reaches a stable state, the network parameters of the discriminator are fixed for subsequent use in determining the authenticity of generated images. Random noise data is input into the generator in the generative adversarial network to generate predictive training images. These predictive training images are then input into the discriminator with fixed network parameters for convolutional processing. The discriminator judges the predictive training images based on the learned features and outputs a probability prediction result, i.e., the probability that the image is considered a real image.

[0092] This embodiment extracts features from the training segmented images using a discriminator. The discriminator can learn the key features of real images, which helps it to more accurately identify the authenticity of generated images in subsequent tests, thereby improving image recognition capabilities. During training, the generator attempts to generate realistic images to deceive the discriminator. The discriminator provides feedback to the generator by outputting probability prediction results. The generator continuously adjusts its generation strategy based on the discriminator's feedback, thereby gradually improving the quality of its generated images. Fixing the discriminator's network parameters reduces the computational load during training, helping to accelerate the training process and improve training efficiency. Furthermore, through continuous interaction and optimization between the discriminator and the generator, the final model can better adapt to various image data, enhancing its robustness and enabling it to perform better in practical applications.

[0093] In some embodiments, the probabilistic prediction results include the similarity between the training switching image and the predicted training image; based on the probabilistic prediction results, the discriminator and generator in the generative adversarial network are iteratively optimized to finally generate a trained target pattern extraction model, including:

[0094] The similarity is compared with a preset similarity threshold to obtain the similarity comparison result. Based on the similarity comparison result, the generative adversarial network is iteratively optimized to generate a trained target pattern extraction model.

[0095] During training, the discriminator distinguishes between the input training segmented image and the predicted training image generated by the generator, outputting a similarity score. This similarity score is typically a probability value, representing the likelihood that the discriminator considers the input image to be a real image (i.e., similar to images in the training dataset). To evaluate the quality of the images generated by the generator, a confidence threshold for similarity is preset. This threshold determines whether the generated image is realistic enough and can be considered valid data for subsequent iterative optimization. The similarity score output by the discriminator is compared with the preset confidence threshold to obtain a similarity comparison result. If the similarity score of the generated image is higher than the confidence threshold, it indicates that the generator's performance is good and can be further fine-tuned to pursue higher image quality. If the similarity score of the generated image is lower than the confidence threshold, the generator's parameters need to be adjusted to improve its image generation ability. This comparison and adjustment process is continuously iterated until the generator can generate synthetic images that are highly similar to real images. After multiple iterations of optimization, the generator will be able to stably generate high-quality synthetic images. At this point, the trained generator can be used as a target pattern extraction model for subsequent lithography process window analysis tasks. This embodiment compares similarity with a preset confidence threshold and iteratively optimizes the generative adversarial network, which can significantly improve the quality of the generated images, making the synthesized images more realistic and better able to reflect the characteristics of real photoresist patterns.

[0096] The embodiments of this application will be described and illustrated below through preferred embodiments.

[0097] Step 1: Acquisition, preprocessing, and annotation of photoresist pattern measurement image data.

[0098] (1) Create an automatic measurement program (recipe) in the critical dimension scanning electron microscope, set the IDW and IDP parameters of the relevant recipe, and put in the wafer to be measured. The machine will execute the steps set in the recipe, measure the photoresist pattern image and its corresponding critical dimension (CD) of the wafer, and then collect the photoresist pattern image data.

[0099] (2) The original image of the acquired photoresist pattern is segmented according to the corresponding method to obtain a small image of 120×120 pixels containing only a single feature. These small images are preprocessed by rotation and Gaussian blur, and then converted into a standard size of 64×64 pixels suitable for input to DCGAN by an algorithm.

[0100] (3) According to the classification criteria applicable to the analysis of the photolithography process window, manually label each image as category 0, 1, 2, and move the image file to the corresponding location.

[0101] Step 2: Construct a deep convolutional generative adversarial network (DCGAN).

[0102] Figure 3 This is a schematic diagram of the structure of a depthwise convolutional generative adversarial network for determining the photolithography process window according to a preferred embodiment of this application, as shown below. Figure 3 As shown, DCGAN consists of two neural networks: a generator and a discriminator.

[0103] (1) Generator: The generator takes a 1-dimensional, 100-bit depth random vector as input (encoded by a random function in the code, which generates 100 random number sequences and transposes them into a 1-dimensional, 100-bit vector). Through three deconvolutional layers, the random noise vector is progressively converted into a two-dimensional computer-synthesized photoresist pattern measurement image. The deconvolutional layers gradually decode the noise vector data into image data until a 64×64 pixel image similar to the real photoresist pattern measurement image is generated. The similarity criteria are obtained by training the discriminator.

[0104] (2) Discriminator: The discriminator is a standard convolutional neural network with three convolutional layers. Its input is a 64×64 pixel image. During discriminator training, a certain class of images from the pre-labeled dataset are input, and the three convolutional layers extract features from the input images, fixing the corresponding network parameters. When using the discriminator to make judgments, the image generated by the generator is input into the discriminator. The convolutional neural network with fixed parameters will perform convolution operations on the generated image and output whether the image is considered to be an image with high similarity to images in the dataset (real image) or an image with low similarity (fake image). DCGAN utilizes the competition between the discriminator and the generator, and through optimized adversarial training, it gradually improves the quality of the images generated by the generator.

[0105] Step 3, the training process of DCGAN.

[0106] (1) During training, the generator will try to generate realistic images in an attempt to deceive the discriminator, while the discriminator will try to identify the authenticity of the images based on the features extracted during training. Figure 4 This is a comparison image of a synthetic photoresist image and a real image generated during the model training process according to a preferred embodiment of this application, such as... Figure 4 As shown, as training progresses, the generator gradually becomes able to generate high-quality images similar to those in the real dataset.

[0107] (2) The specific training parameters are as follows: the input image size is 64×64 pixels, the batch size is 8, the random noise vector depth is 100, the training is 300 epochs, and a checkpoint is set to prevent too many epochs from being set.

[0108] The model uses the Adam optimizer with a learning rate of 0.0002 (the weights are adjusted by an increment of 0.0002 each time the weights are updated, based on the calculated gradient), and a β1 parameter of 0.5 (used to calculate the decay rate of the first moment estimate (i.e., the exponential moving average of the gradient), controlling the exponential decay rate of past gradients and determining the weights of the current gradient when updating parameters).

[0109] (3) During training, the generator and discriminator continuously engage in a game until the image generated by the generator is visually highly similar to the real image. By periodically checking the performance of the discriminator, the quality of the image generated by the generator is ensured to gradually improve.

[0110] Step 4: Generate synthetic data and expand the dataset.

[0111] (1) Using a trained DCGAN model, 1000 sets of 1-dimensional 100-bit random noise vectors were input to generate 1000 computer-synthesized photoresist pattern measurement image data. These images have features similar to real data, which can effectively supplement the deficiencies of small sample datasets and balance various types of data.

[0112] (2) The discriminator is trained by inputting different numbers of original photoresist pattern measurement images (such as 1, 10, 100, 1000 and 10000 images), and 1000 computer-synthesized photoresist pattern measurement images are generated using the DCGAN trained under the above conditions. Figure 5 This is a graph showing the relationship between the amount of raw data and the accuracy of synthesized image recognition according to a preferred embodiment of this application. Figure 5 As shown, the similarity of these synthesized images was tested using existing deep learning-based advanced lithography process window analysis tools. The results showed that when training with 1,000 real images, the similarity between the generated images and the real images reached 73.1%; when using 10,000 real images, the similarity increased to 85.2%.

[0113] Step 5, applied to the analysis of the photolithography process window.

[0114] (1) The computer-synthesized photoresist pattern measurement image data generated by DCGAN is combined with the original photoresist pattern measurement image data to further train the advanced deep learning-based photolithography process window analysis tool. Through data augmentation and dataset expansion, the model can better identify and predict the features of exposure, development or measurement process anomalies in the photoresist pattern measurement images, and accurately determine the photolithography process window.

[0115] (2) Actual experimental results show that by using DCGAN technology to expand the dataset and perform data augmentation, with a confidence threshold of 70%, the existing advanced lithography process window analysis tools based on deep learning require approximately 13 times less raw photoresist pattern measurement image data, while the performance of the lithography process window analysis tools is comparable to that when only raw data is used.

[0116] Furthermore, the above-mentioned method for segmenting the original image is as follows:

[0117] ① For dense patterns of line or space type, there are two cases: the left and right ends of the measurement image are located within the pitch and the critical dimension bar (CD bar).

[0118] For cases where the left and right ends of the measurement image are within the spacing, there are two scenarios: the center of the measurement image is within the spacing and the center is within the critical dimension line.

[0119] i. For cases where the center of the measurement image is located within the spacing. Figure 6 This is a schematic diagram of a first segmentation method for dense patterns of line or space type according to a preferred embodiment of this application, such as... Figure 6 As shown, the measured image is a square image with a side length (in pixels) of v. The width of the key dimension lines corresponds to pixels c, and the width of the spacing corresponds to pixels p. There are n key dimension lines on each side of the center spacing, where n is a positive integer. Then, starting from the top left corner (point A), the original image is divided into 2n segments along the x-direction and [v / (c+p)] segments along the y-direction. Figure 7 This is a segmented image of a first segmentation method for dense patterns of line or space type according to a preferred embodiment of this application, such as... Figure 7 As shown, the image is divided into square images of (c+p) pixels.

[0120] ii. For cases where the center of the measurement image is located within the critical dimension line, Figure 8 This is a schematic diagram of a second segmentation method for dense patterns of line or space types according to a preferred embodiment of this application, such as... Figure 8 As shown, the measured image is a square image with a side length (pixels) of v. The width value of the key dimension line corresponds to pixels c, and the spacing width value corresponds to pixels p. There are n key dimension lines on each side of the center key dimension line, where n is a positive integer. Then, starting from the top left corner (point A), the original image is divided into (2n+1) segments along the x direction and into [v / (c+p)] segments along the y direction. The divided image is a square image with (c+p) pixels.

[0121] For cases where the left and right ends of the measurement image are within the critical dimension lines. Figure 9This is a schematic diagram of a third segmentation method for dense patterns of line or space types according to a preferred embodiment of this application, such as... Figure 9 As shown, the measurement image is a square image with a side length (pixels) of v. The width of the key dimension line corresponds to pixels c, and the width of the spacing corresponds to pixels p. When the center of the measurement image is within the spacing, we only need to move the starting point A of the segmentation from the upper left corner to the right to the first segment spacing. That is, we add an offset of (c+p / 2) to the x coordinate of all segmentation points. The other segmentation methods are consistent with the two cases where the left and right ends of the measurement image are within the spacing.

[0122] ② For sparse patterns of line or space type, Figure 10 This is a schematic diagram of a sparse pattern of line or space type according to a preferred embodiment of this application, such as... Figure 10 As shown, the known measurement image is a square image with a side length (pixels) of v, and the key dimension line width value corresponds to pixels of c. Normalization is performed directly.

[0123] ③ For dense patterns of diameter or hole type, Figure 11 This is a schematic diagram of a dense pattern of diameter or hole type according to a preferred embodiment of this application, such as... Figure 11 As shown, the measured image is a square image with a side length (in pixels) of v, where each row has h holes and each column has t holes. The original image is uniformly divided into h segments along the x-direction and t segments along the y-direction, starting from the top left corner (point A). Figure 12 It is a segmented image of a dense pattern of diameter or hole type according to a preferred embodiment of this application, such as Figure 12 As shown, the image is segmented into rectangles with length and width (or width and length) of v / h pixels and v / t pixels, respectively.

[0124] ④ For sparse patterns of diameter or hole type, Figure 13 This is a schematic diagram of a sparse pattern of diameter or hole type according to a preferred embodiment of this application, such as... Figure 13 As shown, the known measurement image is a square image with a side length (pixels) of v, and it is directly normalized.

[0125] This embodiment also provides a device for determining a photolithography process window. This device is used to implement the above embodiments and preferred embodiments, and will not be repeated as already described. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0126] Figure 14 This is a structural block diagram of the photolithography process window determination apparatus according to an embodiment of this application, such as... Figure 14As shown, the device includes:

[0127] The preprocessing module 10 is used to perform pattern segmentation on the acquired raw photolithography image to generate multiple 1-standard pixel images.

[0128] The data augmentation module 20 is used to input multiple standard pixel images into the trained target pattern extraction model for lithographic pattern extraction processing and output the required predicted lithographic pattern; wherein, the target pattern extraction model is generated by training an initial generative adversarial network based on preset random noise data.

[0129] The lithography analysis module 30 is used to perform lithography process window analysis and processing based on the predicted lithography pattern and the original lithography image to determine the lithography process window.

[0130] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.

[0131] This embodiment also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0132] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0133] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0134] S1, perform pattern segmentation on the acquired original photolithography image to generate multiple standard pixel images.

[0135] S2, input multiple standard pixel images into the trained target pattern extraction model for lithographic pattern extraction processing, and output the required predicted lithographic pattern; wherein, the target pattern extraction model is generated by training the initial generative adversarial network based on preset random noise data.

[0136] S3, based on the predicted lithography pattern and the original lithography image, performs lithography process window analysis and processing to determine the lithography process window.

[0137] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0138] Furthermore, in conjunction with the photolithography process window determination method in the above embodiments, this application embodiment can provide a storage medium for implementation. This storage medium stores a computer program; when executed by a processor, the computer program implements any of the photolithography process window determination methods in the above embodiments.

[0139] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0140] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for determining a photolithography process window, characterized in that, include: The acquired raw photolithography image is subjected to pattern segmentation to generate multiple standard pixel images; Multiple standard pixel images are input into a trained target pattern extraction model for lithographic pattern extraction and output the desired predicted lithographic pattern. The target pattern extraction model is generated by training an initial generative adversarial network based on preset random noise data. Based on the predicted lithography pattern and the original lithography image, the lithography process window is analyzed and processed to determine the lithography process window; The step of performing pattern segmentation on the acquired original photolithography image to generate multiple standard pixel images includes: Based on the pattern type of the original lithographic image, a pattern segmentation operation is performed on the original lithographic image to obtain multiple single feature images; a preprocessing operation is performed on the multiple single feature images to generate multiple standard pixel images; The generative adversarial network includes a discriminator and a generator; the method further includes: Acquire training lithography images and perform pattern segmentation on the training lithography images to generate multiple training segmented images; The random noise data is input into the generator in the generative adversarial network, and the generator outputs a predicted training image. The training segmented image and the predicted training image are input into the discriminator in the generative adversarial network, and a probability prediction result is output; wherein, the probability prediction result includes the similarity between the training segmented image and the predicted training image; Based on the probability prediction results, the discriminator and the generator in the generative adversarial network are continuously iterated and optimized to finally generate the trained target pattern extraction model. The step of iteratively optimizing the discriminator and generator in the generative adversarial network based on the probability prediction results to ultimately generate the trained target pattern extraction model includes: The similarity is compared with a preset confidence threshold to obtain a similarity comparison result. Based on the similarity comparison result, the generative adversarial network is iteratively optimized to generate the trained target pattern extraction model.

2. The method for determining the photolithography process window according to claim 1, characterized in that, The preprocessing operation on the multiple single-feature images to generate multiple standard pixel images includes: Multiple single-feature images are rotated and Gaussian blurred to generate multiple standard pixel images.

3. The method for determining the photolithography process window according to claim 1, characterized in that, Based on the pattern type of the original photolithographic image, a pattern segmentation operation is performed on the original photolithographic image to obtain multiple single-feature images, including: If the pattern type is a line-dense or space-dense type, the original lithographic image is segmented based on its position and key size data to obtain multiple single feature images. If the pattern type is line sparse or spatial sparse, the lithographic pattern is normalized and segmented to obtain multiple single feature images; If the pattern type is a hole-dense or diameter-dense type, the lithographic pattern is segmented based on the image side length of the original lithographic image and the distribution position of the holes in the original lithographic image to obtain multiple single feature images; If the pattern type is a sparse hole type or a sparse diameter type, the photolithographic pattern is normalized and segmented to obtain multiple single feature images.

4. The method for determining the photolithography process window according to claim 1, characterized in that, The step of inputting the training segmented image and the predicted training image into the discriminator in the generative adversarial network and outputting the probability prediction result includes: The training segmented image is input into the discriminator, and the discriminator is used to extract features from the training segmented image to obtain image feature data. Based on the image feature data, fixed network parameters for the discriminator are determined. The predicted training image is input into a discriminator with the fixed network parameters for convolution calculation, and the probability prediction result is output.

5. A device for determining a photolithography process window, characterized in that, include: The preprocessing module is used to perform pattern segmentation on the acquired raw lithography image to generate multiple standard pixel images; wherein, the process of performing pattern segmentation on the acquired raw lithography image to generate multiple standard pixel images includes: performing pattern segmentation on the raw lithography image based on the pattern type of the raw lithography image to obtain multiple single feature images; and performing preprocessing on the multiple single feature images to generate multiple standard pixel images. A data augmentation module is used to input multiple standard pixel images into a trained target pattern extraction model for lithographic pattern extraction processing and output the required predicted lithographic pattern. The target pattern extraction model is generated by training an initial generative adversarial network based on preset random noise data. The generative adversarial network includes a discriminator and a generator. The training process of the target pattern extraction model includes: Acquire training lithography images and perform pattern segmentation on the training lithography images to generate multiple training segmented images; The random noise data is input into the generator in the generative adversarial network, and the generator outputs a predicted training image. The training segmented image and the predicted training image are input into the discriminator in the generative adversarial network, and a probability prediction result is output; wherein, the probability prediction result includes the similarity between the training segmented image and the predicted training image; Based on the probability prediction results, the discriminator and the generator in the generative adversarial network are continuously iterated and optimized to finally generate the trained target pattern extraction model. The step of iteratively optimizing the discriminator and generator in the generative adversarial network based on the probability prediction results to ultimately generate the trained target pattern extraction model includes: The similarity is compared with a preset confidence threshold to obtain a similarity comparison result. The generative adversarial network is iteratively optimized based on the similarity comparison result to generate the trained target pattern extraction model. The lithography analysis module is used to perform lithography process window analysis and processing based on the predicted lithography pattern and the original lithography image to determine the lithography process window.

6. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method for determining the photolithography process window according to any one of claims 1 to 4.

7. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method for determining the photolithography process window according to any one of claims 1 to 4 when it runs.

Images