A lithography process window detection method based on contrast learning

Abstract

The application provides a photolithography process window detection method based on contrast learning, comprising: obtaining a photolithography condition in a photolithography condition set and all photolithography images corresponding to the photolithography condition, i.e. a to-be-detected photolithography image and a target photolithography image; aligning the all photolithography images, i.e. the to-be-detected image and the target image; inputting the aligned photolithography image into a trained reconstruction neural network model; reconstructing the aligned photolithography image into a reconstructed photolithography image which retains contour details and removes background noise; inputting the reconstructed to-be-detected image and the reconstructed target image into a trained contrast learning network to calculate a similarity; when the similarity between the reconstructed to-be-detected image and the reconstructed target image meets the requirements and constraints, the to-be-detected image corresponding to the similarity is determined to be qualified, and the photolithography condition corresponding to all qualified to-be-detected images is the target photolithography process window. The application can effectively avoid overfitting and underfitting problems, and has high-precision detection and strong network interpretability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photolithography process in integrated circuit manufacturing and the field of computer image processing, and in particular to a photolithography process window detection method based on contrast learning. BACKGROUND

[0002] Photolithography is an important semiconductor manufacturing technology, also widely used in microelectronics, optoelectronics, nanotechnology and other fields. It uses photoresist and photomask to manufacture micron or even nanometer level patterns and structures on the surface of a semiconductor wafer, for manufacturing chips, sensors, displays, optical devices and other micro-nano processing applications. Photolithography is a micro-lithography technology that covers a chip surface with a layer of photoresist, aligns a photomask to the chip surface, and through exposure and chemical etching steps, transfers the photoresist pattern to the chip surface. The advantage of photolithography is that the manufacturing process is highly automated and has good repeatability, and can manufacture high-precision patterns and structures, where control of exposure conditions is an important challenge for photolithography.

[0003] The photolithography process window refers to a range of process parameters that can produce good pattern shape and size during photolithography. In photolithography, in order to obtain the desired pattern shape and size, important parameters such as exposure focal length and energy need to be controlled, and changes in these parameters will affect the chemical reaction and physical effect of the photoresist, and in turn affect the shape and size of the pattern. In practical applications, selecting the appropriate photolithography process window is very important for the success of photolithography and the improvement of product quality. Therefore, photolithography process window detection refers to the process of processing photolithography images to determine the range of changes in the focal length and energy parameters in the photolithography process. SUMMARY

[0004] The present application aims to solve the problems encountered in screening the process window of the photolithography process, and to improve the efficiency of the photolithography process. A photolithography process window detection method based on contrast learning is proposed. The similarity between the test photolithography image and the target photolithography image is calculated to detect the photolithography process window, ultimately meeting the quality requirements of actual photolithography process in production; this method can effectively avoid overfitting and underfitting problems, and has high precision detection and strong network interpretability.

[0005] To achieve the above purpose, the present application adopts the following specific technical solutions:

[0006] A photolithography process window detection method based on contrast learning, the method comprising:

[0007] The lithography conditions in the set of lithography conditions and all lithography images corresponding to the lithography conditions are acquired, the lithography conditions include lithography energy parameters and lithography focal length parameters to be set in a lithography process flow, and the all lithography images include a target lithography image and a to-be-tested lithography image; the all to-be-tested images are aligned with the target image; the aligned lithography images are input into the trained denoising neural network model to retain the contour details of the lithography images and remove background noise; the to-be-tested images after denoising and the target images after denoising are input into the trained contrast learning network to calculate the similarity, and when the similarity between the to-be-tested images after denoising and the target images after denoising meets the requirement, it is determined that the corresponding to-be-tested image is qualified, and the lithography conditions corresponding to all the qualified to-be-tested images are the target lithography process window, which meets the requirements of actual production.

[0008] The all to-be-tested lithography images are aligned with the target lithography image, and specifically includes:

[0009] The center region of the target lithography image is selected as the region of interest. The regions of interest in the two images are cross-correlated to calculate the similarity therebetween. According to the maximum value of the correlation measurement, it is determined how many pixels the to-be-tested lithography image should be moved to achieve the best matching. The to-be-tested lithography image is translated to align with the region of interest of the target lithography image.

[0010] The aligned lithography images are input into the trained denoising neural network model to obtain the denoised lithography images retaining the contour details and removing the background noise, and specifically includes:

[0011] The denoising neural network model includes an encoder module, a decoder module, a refinement module and a convolution layer; the encoder module and the decoder module adopt a symmetric design, the encoder module and the decoder module use a skip connection to pass the features extracted from the lithography image, and the decoder module restores the low-dimensional feature representation generated by the encoder module to a high-dimensional space; then, the refinement module is used to extract the detailed features of the lithography image; the convolution layer further processes the features through convolution operation to extract higher-level feature information; the output of the convolution layer is superimposed with the original input image to form a new feature map; the new feature map contains the combination of the original input image and the detailed features extracted by the refinement module, and improves the performance of the model in processing the lithography image task.

[0012] The encoder module and the decoder module include a plurality of Transformer modules, the number of Transformer modules in each layer of the encoder module gradually increases, the resolution of the image gradually decreases, the number of Transformer modules in each layer of the decoder module gradually decreases, and the resolution of the image gradually increases; the Transformer modules in the encoder module convert the input image into a series of feature representations to capture local and global information of the image; the Transformer modules in the decoder gradually recover the spatial information of the image using the feature representations, thereby generating a more refined output; the Transformer modules use a self-attention mechanism (Self-Attention) to handle long-distance dependencies and complex structures; the self-attention mechanism assigns different weights to each element in the input sequence to focus on different parts of the image, thereby capturing local and global information; the Transformer modules use a point-wise feed-forward neural network (FFN) to enhance the expression ability of the model; the FFN performs a nonlinear transformation on the input features to capture complex image features; the Transformer modules use residual connection and layer normalization (Residual Connection & Layer Normalization) to stabilize the training process and improve the performance of the model; in the residual connection, information is directly transmitted between multiple layers to alleviate gradient vanishing; the layer normalization accelerates the convergence of the Transformer module and improves the generalization performance of the Transformer module.

[0013] The similarity is calculated by inputting the noise-reduced test image and the noise-reduced target image into the trained contrast learning network, specifically including: using a contrast learning model SimCLR (Simple Contrastive Learning of Visual Representations) of self-supervised image representation learning to learn lithography image features and calculate the similarity of the lithography image, the SimCLR model is a double-tower network, each tower contains a series of convolution and pooling layers, and finally outputs a high-dimensional vector representing the image, in the training process, two towers encode two similar images input into vectors and pass them to a contrast loss function; the contrast loss function is the core for training the SimCLR model; the contrast loss function compares the vector encoding of the input two vectors with the vector encoding of other images, so that the vectors from the same image are closer in the embedding space, while the vectors from different images are farther apart in the embedding space; the model is trained using a training data set, and the performance of the model is evaluated using a validation data set.

[0014] The contrast learning model SimCLR is trained, and the specific steps are:

[0015] i) Create dataset: Collect denoised lithography images to build a dataset, and divide it into training set and validation set;

[0016] ii) Data augmentation: Perform random data augmentation transformations on input images to generate a pair of similar images; The transformations include random cropping, horizontal flipping, and color jittering; The augmented image pair is considered as positive samples;

[0017] iii) Feature extraction: Use a convolutional neural network as a feature extractor to map input images to a feature space; The network is continuously updated during training to learn better feature representations; Add a projection head on top of the feature extractor, which is a simple feedforward neural network that maps the output of the feature extractor to a low-dimensional representation space; The low-dimensional representation space will be used to calculate similarity;

[0018] iv) Compute contrastive loss: In the representation space, calculate the similarity between positive sample pairs and compare them with other negative sample pairs; Negative sample pairs are generated from other images in the same batch; The goal of contrastive loss is to minimize the distance between positive sample pairs while maximizing the distance between negative sample pairs;

[0019] v) Backpropagation and optimization: Update the parameters of the feature extractor and projection head according to the contrastive loss to learn better image representations;

[0020] Use the trained contrastive learning model SimCLR: input the denoised test image and the denoised target image into the SimCLR model, encode them into vectors, and then calculate the cosine similarity between the vectors; The similarity between images is represented by the cosine similarity between the encoded vectors; When the similarity between the denoised test image and the denoised target image meets the requirements, the corresponding test image is determined to be qualified, and the lithography conditions corresponding to all qualified test images are the target lithography process window, which meets the requirements of actual production.

[0021] The beneficial effects of the present application are:

[0022] 1) The lithography image alignment method proposed in the present application corrects the offset of the lithography images collected in the lithography process flow, improving the accuracy of the lithography image similarity algorithm.

[0023] 2) The present application uses a denoising neural network to reduce noise in lithography images, making the lithography images clearer, more accurate and easier to identify. By reducing the noise in the lithography images, the signal-to-noise ratio of the lithography images is improved, making the lithography images easier to process and analyze, ensuring the accuracy and reliability of the final results

[0024] 3) The application uses a contrast learning network to calculate the similarity of lithography images, which has the following advantages: a relatively simple structure ensures computational efficiency, the network has better data utilization, and it has wider applicability and excellent performance for lithography images produced by different lithography machines. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a flowchart of an embodiment of the application;

[0026] Figure 2 is an image alignment schematic diagram of an embodiment of the application;

[0027] Figure 3 is a denoising neural network structure diagram of an embodiment of the application;

[0028] Figure 4 is a contrast learning neural network training schematic diagram of an embodiment of the application;

[0029] Figure 5 is a contrast learning neural network application schematic diagram of an embodiment of the application. DETAILED DESCRIPTION

[0030] To make the objectives, technical solutions and advantages of the application clearer, the technical solutions of the application will be described clearly and completely below in conjunction with embodiments and drawings. Obviously, the described embodiments are only some of the embodiments of the application, not all. The components of the embodiments of the application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the application provided in the drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the application. Based on the embodiments of the application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the application.

[0031] A lithography process window detection method based on contrast learning of the application comprises the following steps:

[0032] Step S1: Obtain lithography conditions in a set of lithography conditions and all lithography images corresponding to the lithography conditions, wherein the lithography conditions include lithography energy parameters and lithography focal length parameters that need to be set in a lithography process flow, and the all lithography images include a to-be-detected lithography image and a target lithography image.

[0033] Step S2: First, all the test lithography images are aligned with the target lithography image, that is, each test lithography image is moved a certain distance in the horizontal and vertical directions respectively. The matching degree between each translated test lithography image and the target lithography image is calculated. The specific method is to use the normalized cross-correlation coefficient to measure the matching degree between two images based on the cross-correlation principle. The translation result with the highest matching degree is taken as the final alignment result.

[0034] Step S3: Input the aligned lithography image into the trained denoising neural network model to obtain a denoised lithography image. The denoising neural network model is pre-trained using a large number of natural image pairs with and without noise as input and output, with the purpose of learning how to remove noise from the noisy lithography image and restore a clear lithography image. The denoising neural network includes an encoder module, a decoder module and a refinement module; the encoder module and the decoder module adopt a symmetrically designed network structure, the encoder and the decoder use a skip connection to pass the features extracted from the lithography image, and a refinement module is used after the decoder to extract the detailed features of the lithography image and input the convolution layer and the input image for superposition, obtaining the denoising result of the lithography image. The encoder and the decoder include multiple Transformer modules, which are responsible for converting the lithography image data. In the image encoder, the number of Transformer modules in each layer gradually increases, while the image resolution gradually decreases, and the dimension of the data becomes smaller and smaller, so more Transformer modules are needed to process the data. At the same time, with the decrease of resolution, the encoder can process the data more finely, so as to obtain more accurate features. Conversely, in the image decoder, the number of Transformer modules in each layer gradually decreases, while the image resolution gradually increases. This is because in the image decoder, the output of the encoder needs to be converted into the original image, so more Transformer modules are needed to remap the features of the encoder to the size and dimension of the original image. With the increase of resolution, the decoder can more easily generate more accurate images. Therefore, the number of Transformer modules and the resolution in the encoder and the decoder are different in different image processing processes, which also shows that they undertake different tasks and functions when processing images.

[0035] Step S4: input the aligned and denoised test lithography image and the target lithography image into the trained contrast learning network to calculate the similarity and determine whether the test image meets the lithography process production requirements. The present application adopts a double-tower convolutional neural network SimCLR. SimCLR is a self-supervised learning method that uses a large amount of unlabeled data to learn features. Each tower contains multiple layers of convolution and pooling layers, and finally outputs a high-dimensional vector representing the image. By comparing the distance of the vectors in the tower, the feature representation of the denoised image can be learned. Specifically, by inputting the denoised and enhanced lithography image into two towers respectively, two high-dimensional vectors can be obtained. These vectors are subjected to a contrast loss function to learn how to cluster vectors from the same image together and separate vectors from different images in the embedding space. The loss function is the core of model training. By using the data in the training set to train the model, the parameters of the model can be optimized. The validation set is used to evaluate the performance of the model to determine whether the model is overfitting or underfitting. Finally, the trained model is used to calculate the similarity of the denoised lithography image.

[0036] Step S5: detect the lithography process window, use the trained SimCLR model to calculate the similarity between the test image and the target image to determine whether the test image is a qualified image, and thus determine the corresponding lithography conditions. Specifically, by passing the test denoised lithography image and the target denoised lithography image to the trained SimCLR model, they can be encoded into high-dimensional vectors, and the cosine similarity between them can be calculated. If the similarity between the two lithography images meets certain requirements, the test image is determined to be a qualified image, otherwise it is determined to be an unqualified image. Using the SimCLR model can improve the accuracy and efficiency of denoised lithography image processing. Because the SimCLR model can learn the feature representation of the lithography image, by calculating the cosine similarity between the vectors, the similarity between the two lithography images can be more accurately represented. Compared with traditional similarity algorithms, using the SimCLR model can better grasp the image features, thus more accurately determining whether the test image meets the requirements. When the similarity between the test image and the target image meets the requirements, the corresponding test image can be determined to be qualified. For all qualified test images, the corresponding lithography conditions can be determined to meet the requirements of actual production. This method has high practicality and precision, and can help improve the quality and production efficiency of lithography images.

[0037] Embodiment

[0038] Reference Figure 1 The present embodiment provides a lithography process window detection method based on contrast learning, comprising the following steps:

[0039] Step S1: Obtain the lithography conditions and all corresponding lithography images from a lithography condition set. The lithography conditions include the lithography energy parameters and lithography focal length parameters required for the lithography process flow. The lithography images include the test lithography image and the target lithography image. Lithography energy and focal length parameters are fundamental control parameters in the lithography process flow. By processing the lithography images and selecting these parameters, important features such as line width, shape, position, and depth can be effectively controlled, resulting in high-quality lithography results.

[0040] Step S2: See Figure 2 As shown in the figure, first, all the lithography images to be tested are aligned with the target lithography image. This means that each lithography image to be tested is moved a certain distance horizontally and vertically. The degree of match between each shifted lithography image to be tested and the target lithography image is calculated. Specifically, based on the cross-correlation principle, a normalized cross-correlation coefficient is used to measure the degree of match between the two images. The translation result with the highest degree of match is taken as the final alignment result. For a pair of images of sizes P×P and Q×Q, respectively, the cross-correlation coefficient is expressed as:

[0041]

[0042] Where X and Y are two images, μ is the average value. The normalized cross-correlation coefficient normalizes them all to the range [0,1]. Where 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no correlation. X and Y represent two images, μ is the average value. X and μ Y Represent the mean of images X and Y, respectively. x and y represent the horizontal and vertical coordinates of a pixel, respectively. The numerator of the formula calculates the covariance of images X and Y, representing the correlation between pixel values ​​at the same location. The denominator is the product of the standard deviations of images X and Y, which is used to normalize the result. The resulting value is a value between -1 and 1, representing the linear correlation between images X and Y. This formula is used to calculate the similarity between the image to be tested and the target image. The alignment result is then found by finding the translated image to be tested with the highest similarity.

[0043] Step S3: Input the aligned lithography image into the trained denoising neural network model to obtain a denoised lithography image. The denoising neural network model is pre-trained using a large number of noisy and noise-free natural image pairs as input and output, so that the model can learn how to remove noise from noisy lithography images and restore clear lithography images. Figure 3As shown, the noise reduction neural network is a neural network model capable of performing noise reduction on lithography images. It mainly consists of an encoder module, a decoder module, and a refinement module. First, there are the encoder module and the decoder module. These two modules use a symmetric network structure, and the encoder and the decoder use skip connections to pass the features extracted from the lithography image. The encoder is responsible for encoding the input lithography image into a feature vector, and the decoder decodes the feature vector into the output image after noise reduction. Meanwhile, multiple Transformer modules are used in both the encoder and the decoder, which are relatively new in convolutional neural networks and are responsible for transforming image data. In the image encoder, the number of Transformer modules in each layer gradually increases, while the resolution of the image gradually decreases, and the dimension of the data becomes smaller and smaller, so more Transformer modules are needed to process the data. As the resolution decreases, the encoder can process the data more finely, resulting in more accurate features. Conversely, in the image decoder, the number of Transformer modules in each layer gradually decreases, while the image resolution gradually increases. This is because in the image decoder, the output of the encoder needs to be converted into the original image, so more Transformer modules are needed to map the encoder's features back to the original image's size and dimension. As the resolution increases, the decoder can more easily generate more accurate images. Therefore, the number of Transformer modules in the encoder and the decoder and the resolution vary with the image processing process. After the decoder, a refinement module is used to extract the detailed features of the lithography image and input them into a convolutional layer and the input image for superposition, resulting in the noise reduction result of the lithography image. The purpose of this refinement module is to further improve the clarity and quality of the image based on the image generated by the decoder, enhancing the details and texture of the image. In summary, the noise reduction neural network is a layer-by-layer progressive image processing model composed of an encoder, a decoder, and a refinement module. Among them, the encoder and the decoder use a symmetric network structure and multiple layers of Transformer modules to process lithography image data, and the refinement module is used to improve the clarity and details of the image.

[0044] Step S4: In this embodiment, the SimCLR double-tower convolutional neural network is used to input the to-be-tested noise-reduced lithography image and the target lithography image into the trained contrast learning network for similarity calculation to determine whether the to-be-tested image meets the production requirements of the lithography process. The specific process is as follows:

[0045] 1. Preprocess the to-be-tested lithography image and the target lithography image, including alignment and noise reduction operations;

[0046] 2. Input the preprocessed to-be-tested lithography image and the target lithography image into the SimCLR double-tower convolutional neural network, respectively;

[0047] 3. Each tower in the dual-tower convolutional neural network contains multiple layers of convolution and pooling, and finally outputs a high-dimensional vector representing the image.

[0048] 4. By comparing the high-dimensional vectors in the two towers through a contrastive loss function, the model learns how to cluster vectors from the same image together and separate vectors from different images. The loss function is the core of model training, and the contrastive loss function of SimCLR is an unsupervised learning method similar to dimension compression and metric learning. This method learns representation learning algorithms on large-scale datasets to obtain efficient image representations. This method uses the interaction between different augmented versions of the same image as positive example samples and different image pairs as negative example samples. The goal is to maximize the similarity of positive example samples and minimize the similarity of negative example samples in the embedding vector space. The contrastive loss function of SimCLR consists of two parts, the first part is the comparison of positive samples, and the second part is the comparison of negative samples. Among them, for each sample i, it can be regarded as a mapping function from data space to embedding space, that is, z(i) = f(x(i)), where x(i) represents the input sample, and z(i) represents the output embedding vector. For each sample i, the SimCLR model produces two embedding vectors, denoted as z(i)1 and z(i)2. They can be used to calculate the similarity of sample i as a positive sample, that is, they share the same label. SimCLR uses cosine similarity to measure the distance between two embedding vectors, so the contrastive loss function of positive samples can be represented as follows:

[0049]

[0050] where N represents the size of the training batch, sim(z(i)1, z(i)2) represents the cosine similarity between the two embedding vectors, and the exponent is divided by the temperature parameter τ, which is used to adjust the range of similarity values, so that the loss function is easier to optimize, which is equivalent to reducing the distance between embedding vectors; j is the index of the sample, and [j≠i] represents other samples except sample i, which calculates all negative samples. For each sample i, it can be regarded as a sample pair of positive samples and a sample pair of potential negative samples. is the similarity between positive sample pairs, usually calculated using cosine similarity. and is the similarity between negative sample pairs. Next, SimCLR uses different sample pairs to calculate different negative samples and positive samples in the algorithm, so as to effectively process all sample pairs in the loss function. Since the number of sample pairs is too large, SimCLR can make the algorithm more efficient and scalable by partitioning all sample pairs and isolating these sample pairs. SimCLR can learn effective image representations on a large amount of lithography image data, thereby achieving efficient image similarity calculation tasks. In this embodiment, the contrast loss function is one of the keys to learning useful embedding representations, which captures meaningful feature representations by learning to generalize samples by constraining the similarity and difference of samples in the embedding space;

[0051] 5. Using the data in the training set to train the model, optimizing the parameters of the model, evaluating the performance of the model through the validation set, determining whether the model is overfitting or underfitting;

[0052] 6. Using the trained model to calculate the similarity of the to-be-tested lithography image and determine whether it meets the production requirements of the lithography process.

[0053] SimCLR is a self-supervised learning method that can learn features using a large amount of unlabeled data. In this invention, SimCLR double-tower convolutional neural network is used to extract and learn features from denoised and enhanced lithography images, which can effectively solve the problem of insufficient labeled data or high labeling cost. Compared with traditional image processing methods based on rules or feature engineering, SimCLR double-tower convolutional neural network can more adaptively and flexibly process image data, with higher accuracy and efficiency.

[0054] Step S5: Detecting the lithography process window, using the trained SimCLR model to calculate the similarity between the to-be-tested image and the target image to determine whether the to-be-tested image is a qualified image, thereby determining the corresponding lithography conditions. Specifically, by passing the to-be-tested denoised lithography image and the target denoised lithography image to the trained SimCLR model, its training process is referred to Figure 4 As shown in the figure, they can be encoded as high-dimensional vectors, and the cosine similarity between them can be calculated. If the similarity between the two lithography images meets certain requirements, the to-be-tested image is determined to be a qualified image, otherwise it is determined to be an unqualified image. Using the SimCLR model can improve the accuracy and efficiency of denoised lithography image processing. Because the SimCLR model can learn the feature representation of the lithography image, by calculating the cosine similarity between the vectors, the similarity between the two lithography images can be more accurately represented. Refer to Figure 5As shown, to calculate the cosine similarity between two embedding vectors, the following steps need to be performed: calculate the dot product of the two vectors. Multiply the corresponding elements of vector A and vector B, then add up all these products to get a numerical value. This numerical value is the dot product of vector A and vector B. Calculate the modulus of vector A and vector B. That is, the modulus of vector A and the modulus of vector B are the square roots of the sum of the squares of the elements in A and B, respectively. Divide the dot product of the two vectors by the product of their moduli. Divide the dot product value obtained in the first step by the product of the modulus of vector A and the modulus of vector B to get the cosine similarity:

[0055]

[0056] A i and B i represent the i-th component of vector A and vector B, respectively. θ is the angle between vector A and vector B. The cosine similarity ranges from [-1, 1], the closer the value to 1, the more similar the two vectors, the closer the value to -1, the less similar the two vectors, and the value of 0 means the two vectors are completely unrelated. Compared with traditional similarity algorithms, using the SimCLR model can better grasp the image features, thus more accurately determining whether the test image meets the requirements. When the similarity between the test image and the target image meets the requirements, it can be determined that the corresponding test image is qualified. For all qualified test images, the corresponding lithography conditions can be determined to meet the requirements of actual production. This method has high practicality and precision, which can help improve the quality and production efficiency of lithography images.

Claims

1. A contrast learning based lithography process window detection method, characterized in that, The method comprises the following steps: acquire lithography conditions in a set of lithography conditions and all lithography images corresponding to the lithography conditions, the lithography conditions including lithography energy parameters and lithography focal length parameters to be set in a lithography process flow, and the all lithography images including target lithography images and to-be-tested lithography images, and align all to-be-tested images with target images; input the aligned lithography images into a trained denoising neural network model, retain the contour details of the lithography images and remove background noise, input the denoised to-be-tested images and the denoised target images into a trained contrast learning network to calculate similarity, and determine that the corresponding to-be-tested image is qualified when the similarity between the denoised to-be-tested image and the denoised target image meets the requirements, and the lithography conditions corresponding to all qualified to-be-tested images are target lithography process windows, which meet the requirements of actual production; wherein: the step of inputting the denoised to-be-tested images and the denoised target images into the trained contrast learning network to calculate similarity specifically comprises: using a self-supervised image representation learning contrast learning model SimCLR to learn lithography image features and calculate the similarity of the lithography images, the contrast learning model SimCLR being a double-tower network, wherein each tower contains a series of convolution and pooling layers, and finally outputs a high-dimensional vector representing an image, in the training process, two towers encode two similar images input into vectors respectively and pass the vectors to a contrast loss function; the contrast loss function is the core for training the SimCLR model; the contrast loss function compares the vector encoding of the input two vectors with the vector encoding of other images, so that the vectors from the same image are closer in the embedding space, while the vectors from different images are farther away in the embedding space; the model is trained using a training data set and the performance of the model is evaluated using a validation data set; the SimCLR contrast learning model is trained, and the specific steps are as follows: ⅰ) create a data set: collect denoised lithography images to build a data set, and divide the data set into a training set and a validation set; ⅱ) data enhancement: perform random data enhancement transformation on the input image to generate a pair of similar images; the transformation includes random cropping, horizontal flipping and color jittering; the enhanced image pair is regarded as a positive sample; ⅲ) feature extraction: use a convolutional neural network as a feature extractor to map the input image to a feature space; the network is continuously updated during training to learn better feature representations; add a projection head at the top of the feature extractor, which is a simple feedforward neural network that maps the output of the feature extractor to a low-dimensional representation space; the low-dimensional representation space will be used to calculate similarity; ⅳ) calculate contrast loss: in the representation space, calculate the similarity between positive sample pairs and compare it with other negative sample pairs; negative sample pairs are generated from other images in the same batch; the goal of contrast loss is to minimize the distance between positive sample pairs while maximizing the distance between negative sample pairs; ⅴ) backpropagation and optimization: update the parameters of the feature extractor and the projection head according to the contrast loss to learn better image representations; Using the trained contrast learning model SimCLR: input the denoising test image and the denoising target image into the SimCLR model, encode them into vectors, and then calculate the cosine similarity between the vectors; the similarity between the images is represented by the cosine similarity between the encoded vectors; when the similarity between the denoising test image and the denoising target image meets the requirements, it is determined that the corresponding test image is qualified, and the lithography condition corresponding to all qualified test images is the target lithography process window, which meets the requirements of actual production.

2. The contrastive learning based lithography process window detection method of claim 1, wherein, The alignment of all test images and target images specifically includes: Selecting the center region of the target lithography image as the region of interest; performing cross-correlation operation on the regions of interest in the two images to calculate the similarity between them; determining how many pixels the test lithography image should be moved to achieve the best match according to the maximum value of the correlation measure; and translating the test lithography image to align it with the region of interest of the target lithography image.

3. The contrastive learning based photolithography process window detection method of claim 1, wherein, The input of the aligned lithography image into the trained denoising neural network model to obtain a denoised lithography image that retains the contour details and removes the background noise specifically includes: The denoising neural network model includes an encoder module, a decoder module, a refinement module, and a convolution layer; the encoder module and the decoder module are symmetrically designed, and the encoder module and the decoder module use skip connections to pass the features extracted from the lithography image; the decoder module restores the low-dimensional feature representation generated by the encoder module to a high-dimensional space; then, the refinement module is used to extract the detailed features of the lithography image; the convolution layer further processes the features through convolution operation to extract higher-level feature information; the output of the convolution layer is superimposed with the original input image to form a new feature map; the new feature map contains the combination of the original input image and the detailed features extracted by the refinement module.

4. The contrast learning based lithography process window detection method of claim 3, wherein, The encoder module and the decoder module include a plurality of Transformer modules, the number of Transformer modules in each layer of the encoder module gradually increases, the resolution of the image gradually decreases, the number of Transformer modules in each layer of the decoder module gradually decreases, and the resolution of the image gradually increases; the Transformer modules in the encoder module convert the input image into a series of feature representations to capture local and global information of the image; the Transformer modules in the decoder gradually recover the spatial information of the image using the feature representations, thereby generating a more refined output; the Transformer modules use a self-attention mechanism to handle long-range dependencies and complex structures; the self-attention mechanism assigns different weights to each element in the input sequence to focus on different parts of the image, thereby capturing local and global information; the Transformer modules use a point-wise feed-forward neural network (FFN) to enhance the expressive power of the model; the FFN performs a nonlinear transformation on the input features to capture complex image features; the Transformer modules use residual connections and layer normalization to stabilize the training process and improve the performance of the model; in the residual connection, information is directly transmitted between multiple layers to alleviate gradient vanishing; and the layer normalization accelerates the convergence of the Transformer module and improves the generalization performance of the Transformer module.

Images