Screening method for lithography process window based on sequential convolutional neural network model

Abstract

The present invention provides a method for screening a lithography process window based on a sequential convolutional neural network model, comprising: inputting a group of lithography images and the relative position information of the images in the group of images into the sequential neural network model. During the training stage, the network model extracts the features of each image separately while fusing the features of the remaining images and the relative coordinates of the current image parameters in this group of images. And finally determines whether the current image meets the production requirements. After the training is completed, in order to make full use of a large amount of diverse unlabeled data and the data after random inspection and correction of the network results. Semi-supervised learning is used to utilize more unlabeled data for learning. Finally, the results of the artificially corrected neural network are learned through distillation learning, and the parameters are updated. The function of continuous learning during use is realized. It ensures that the generalization ability of the algorithm is continuously improved and meets the needs of lithography process window screening in actual applications.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and intelligent information processing, and specifically relates to a method for screening lithography process windows based on a sequential convolutional neural network model. Background Art

[0002] Photolithography is a key process in integrated circuit manufacturing. It utilizes photochemical reactions to transfer a pattern created on a mask onto a wafer substrate. The quality of the lithography process is ultimately determined by the interaction of the mask, exposure system, and photoresist, which is crucial to the yield rate of integrated circuit manufacturing.

[0003] Exposure focus and energy are two core parameters in the photolithography process. Due to various environmental factors, the range of fluctuation within these parameters is known as the photolithography process window. Precisely determining the appropriate photolithography process window, controlling photolithography parameters for different regions on the wafer, and monitoring and adjusting them in real time during production are crucial steps in integrated circuit production. An inappropriate process window can severely reduce product yield.

[0004] To select the appropriate lithography process window, an electron microscope typically generates corresponding lithography images for different exposure parameters, creating a focus-energy matrix (FEM). Experienced engineers then analyze the FEM matrix to determine the process window. Furthermore, during the production process, these two crucial parameters, focus and energy, can still experience "parameter drift" between their set and actual values due to interference from external factors like vibration waves, resulting in a decrease in product yield.

[0005] However, at this stage, the method of relying on engineers to sample and determine the process window can no longer meet the requirements of production testing. It is urgently necessary to rely on intelligent analysis of lithography images to effectively control and monitor more lithography process windows during the R&D process, control more exposure parameters in actual production, and monitor "parameter drift" during the production process to ensure a sufficiently high yield.

[0006] It is understood that a certain company has developed image analysis software for specific forms of measurement lines in lithography images, but it cannot process lithography images of any pattern, nor can it directly provide process window parameters. It only plays a certain auxiliary role in manual processing.

[0007] Currently, process window parameter analysis relies primarily on experienced engineers. However, these issues present the following challenges: a) Human subjectivity is high, resulting in inconsistent quality and low precision, impacting product yields, low efficiency, and relatively high labor costs; b) the high labor input and high costs, coupled with the intense workload, also contribute to an unstable engineering team.

[0008] Decision-making and monitoring of the lithography process window based on artificial intelligence technology can not only improve the production efficiency of the chip manufacturing process, but the decided lithography process window is also more stable than the results of manual screening, and the chip yield will be improved. Summary of the Invention

[0009] The purpose of the present invention is to provide a method for screening lithography process windows based on a sequential convolutional neural network model, which is used to screen the process window of the current lithography image, improve the quality inspection efficiency under the premise of providing a stable process window extraction method, and further improve the stability of production and process window screening. It also realizes regular upgrades of the model based on new data continuously obtained during the application stage, ensures the stability of the model during the application process, and does not cause significant interference to the accuracy of the model due to the increasing complexity of the image. The present invention does not have overfitting or underfitting, and has high recognition accuracy and strong network interpretability.

[0010] To achieve the above objectives, the present invention adopts the following specific technical solutions:

[0011] A method for screening a lithography process window based on a sequential convolutional neural network model, the method specifically comprising:

[0012] Pre-training phase:

[0013] First, we built a Bottleneck Mobile NetV3 sequential convolutional neural network model and trained the neural network using a manually labeled lithography process window dataset.

[0014] The Bottleneck Mobile NetV3 sequential convolutional neural network model is used to screen a set of lithography process images for a range of lithography process parameters that meet production requirements, and the range is defined as a lithography process window.

[0015] Use phase:

[0016] A set of lithography images with the same template but different production parameter settings are input into the BottleneckMobileNetV3 sequential convolutional neural network model, deep features of each image are extracted, and relative parameters between images are analyzed. Based on this, the neural network determines which images in the current set of images meet process production requirements, and combines the processing parameters of these images into a lithography process window as output; the output result is stored as a data set of the lithography process window selected by the neural network;

[0017] Use the process to regularly upgrade the network:

[0018] After the neural network is pre-trained and put into use, the network is regularly upgraded to improve its generalization capabilities and maintain good performance when dealing with the latest lithography images. There are two training methods for upgrading the neural network:

[0019] a) Acquire the latest unlabeled lithography process images during use; Based on the semi-supervised learning method, perform two different data augmentations on a set of images with the same template but different parameters; After each data augmentation, input the two sets of images into two Bottleneck MobileNetV3s with the same structure and parameters, i.e., a pair of twin networks, and obtain the output of the two networks, i.e., the probability of judging the processing parameters of each image in the set. The probability is used as the data for training the network to obtain the new parameter θ 11 and θ 12 , and update the parameters of the two twin networks through distillation learning;

[0020] After the training is completed, a parameter is selected from the parameters θ1 and θ2 of the two twin neural networks, and the parameters of the sequential convolutional neural network model in the use stage are updated through distillation learning;

[0021] b) Spot-check the lithography process window results selected by the neural network during use and manually correct them. Regularly use the corrected results as a new dataset. Based on supervised learning, the sequential convolutional neural network model is trained using the manually corrected dataset to obtain new parameters θ3, which are then updated through distillation learning.

[0022] Distillation learning is used to prevent the network from gradually forgetting the lithography process window screening capabilities it has already mastered during the upgrade process.

[0023] The Bottleneck Mobile NetV3 sequential convolutional neural network model constructed in this paper is composed of a stack of Bottleneck Mobile NetV3 block modules and soul block modules. These two modules ensure that the model saves computational effort during the analysis process and allows the model to consider the relative position of each image in the group when analyzing a group of images.

[0024] In order to make the relationship between each layer of the neural network nonlinear, each layer of the neural network is linked using the ReLU activation function; the calculation formula of the function is:

[0025]

[0026] In order to ensure that the range of each input value is constrained between [0, 1] before being input into the ReLU function, batch normalization is used to normalize the input value size of each layer of the neural network, reducing the computational complexity of the neural network and enabling the network to be trained faster.

[0027] The Bottleneck Mobile NetV3 block module consists of a group convolution Group Conv, a bottleneck structure and a SE module;

[0028] The process of calculating a feature in a common convolutional neural network is to cover the convolution kernel on n*n*channel features and obtain a new feature through convolution operation. The convolution kernel is then covered on the next n*n*channel features for the next operation until the convolution kernel covers all input features. Where n is the width of the convolution kernel in the convolution operation, and channel is the dimension of the convolution kernel in the channel direction. For example, the channel dimension of an RGB image is 3 (red, green, blue). The channel dimension of the input in this task is 99 (99 images).

[0029] Unlike ordinary convolution, Group Conv first divides the input features into m blocks in the channel direction, and then calculates each feature block separately; finally, after a convolution operation, m features are obtained. This ensures that when the input is a group of images, the features extracted from each image are processed separately, saving computational effort.

[0030] The bottleneck structure utilizes the self-learning ability of the neural network. The output of the bottleneck structure has fewer features than the input. This method summarizes the input features and then uses the output features as new inputs for the next stage to analyze and obtain more features. This saves computational effort while preventing unimportant features from interfering with the neural network's analysis.

[0031] Because the process window is a process of selecting a parameter range, and because the image changes with the parameters, the SE module calculates the average of the features of each image, and after neural network analysis, the output is equal to the number of input features. This is then multiplied by each feature of each image, thus representing the features represented by the change relationship between each image.

[0032] When features are input into the Bottleneck Mobile NetV3 block, the output features still retain the properties of a sequence, that is, each image is always extracting its own features and drawing on the features of other images; instead of fusing all features together for analysis, the features are arranged in the order of the images.

[0033] The soul block module extracts new features while retaining shallow features, inserting the new features of each image behind its shallow features; avoiding forgetting features obtained early in the network, which leads to performance degradation; at the same time, it can still make the features have the property of sequence, that is, the order of features is arranged according to the order of each image.

[0034] The training method a) of the upgraded neural network updates the parameters of the two twin networks through distillation learning, and its formula is:

[0035]

[0036] Among them, E (T) [θ1] and E (T) [θ2] represents the updated parameters of the two twin networks after the parameters are updated in this training phase and after distillation learning;

[0037] Initial parameter E (0) [θ1]=θ,E (0) [θ2] = θ, where θ represents the initial parameters of the network before updating the parameters in this training phase;

[0038] α1 represents the weight of retaining old knowledge set by the distillation learning of the updated twin network in this training stage, and its range is (0,1];

[0039] Since the training gradually stabilizes and the performance is similar, we randomly select one parameter from the parameters θ1 and θ2 of the two twin networks and update the parameters of the sequential convolutional neural network model in the use phase through distillation learning. The formula is:

[0040] E (T) [θ]=α2*E (T-1) [θ]+(1-α2)θ1 (2)

[0041] In the formula, the parameters of the randomly selected twin network are θ1, E (T) [θ] is the updated parameter of the sequential convolutional neural network in the usage phase after the parameter update; E (T-1) [θ] is the parameter of the sequential convolutional neural network model in the usage phase before updating the parameters; α2 represents the weight of retaining old knowledge set by the distillation learning of the parameters of the sequential convolutional neural network model in the usage phase in this training phase, and its range is (0,1].

[0042] In the training method b) of the upgraded neural network, a new parameter θ3 is obtained and updated through distillation learning, and the formula is:

[0043] E (T) [θ]=α3*E (T-1) [θ]+(1-α3)θ3 (3)

[0044] Among them, E (T) [θ] is the updated parameter of the sequential convolutional neural network model in the usage phase after the parameter update; E (T -1) [θ] represents the parameters of the sequential convolutional neural network model in the usage phase before updating the parameters; θ3 represents the parameters of the network in this phase obtained based on supervised learning; α3 represents the weight of retaining old knowledge set by the distillation learning of the sequential convolutional neural network parameters in the usage phase during this training phase, and its range is (0,1].

[0045] The parameter update of the sequential convolutional neural network model described in the present invention includes the first training phase, the second training phase, and the third training phase. The first training phase is the pre-training phase before the model is put into use, and the second and third training phases are regular training and upgrades during use:

[0046] The first training includes, after the network model framework is built, using a manually labeled lithography process window dataset to train the network based on supervised learning to initially obtain a Bottleneck MobileNetV3 sequential convolutional neural network model for the use stage;

[0047] To improve the model's generalization capabilities and maintain good performance even when dealing with the latest lithography images, the model was regularly upgraded through the second and third training phases after it was put into use.

[0048] The second training involves constructing a dataset using a large number of unlabeled images collected during the photolithography production process. A set of photolithography images with the same template and different parameters are augmented twice and fed into two Bottleneck MobileNetV3 networks with the same structure and parameters, i.e., a pair of twin networks. The outputs of the two networks are obtained, i.e., the predicted values for determining whether the processing parameters of each image are qualified. The predicted values are used as training data for the network, and the network is trained based on semi-supervised learning to obtain new parameters θ. 11 and θ 12 , the parameters of the twin network are updated through distillation learning. After training, because the performance of the two networks is similar, the parameters of one network are selected from the two twin networks, and the parameters of the network in the use phase are updated through distillation learning;

[0049] The third training step involves spot-checking the lithography process windows selected by the neural network and manually correcting the results. Using this dataset of spot-checked and manually corrected lithography process windows, the network is trained using supervised learning to obtain the parameter θ3, which is then updated using distillation learning.

[0050] The sequential convolutional neural network model constructed by the present invention is different from the convolutional neural network for a single image. In the present invention, it is necessary to consider the photolithography process effect of the same template under different parameters, and to analyze whether the current image is qualified based on its relative position in a group of images. Based on this, a sequential convolutional neural network model is designed that can independently analyze each image, continuously integrate features of other images, and always retain its sequential properties. The network model is a stack of Bottleneck MobileNetV3 Block modules and soul block modules; wherein the Soul Block module can still retain shallow features after extracting new features. This avoids the gradual dilution of effective shallow features during the decoupling process.

[0051] The present invention can achieve the following beneficial technical effects:

[0052] 1. This invention overcomes the overfitting problem of ordinary convolutional neural networks and has strong robustness.

[0053] 2. The present invention can self-upgrade during the application process and has good adaptability.

[0054] 3. In this invention, by designing Bottleneck MobileNetV3 Block and Soul Block, the model can consider the characteristics of associated images and their parameters while performing single image feature analysis, and save as much computation as possible.

[0055] 4. The present invention uses the distillation learning method to ensure that the model can continuously upgrade itself during engineering application while avoiding forgetting the learned features.

[0056] 5. The present invention can simultaneously process the sampling of more photolithography process windows, thereby improving the yield and stability of the photolithography process flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a flow chart of the present invention;

[0058] Figure 2 It is a schematic diagram of updating parameters of network training based on semi-supervised learning in the present invention;

[0059] Figure 3 This is the module structure diagram of Bottleneck MobileNetV3 Block of the present invention;

[0060] Figure 4 It is a structural diagram of the Soul Block of the model of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention and the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0062] Example

[0063] like Figure 1-4 As shown, the present invention provides a method for screening a lithography process window based on a sequential convolutional neural network model, comprising the following steps:

[0064] A Bottleneck Mobile NetV3 sequential convolutional neural network model is constructed, and the Bottleneck Mobile NetV3 sequential convolutional neural network model is used to determine and screen a process window that meets production requirements in a set of lithography process images.

[0065] 1) Extract features from each image in a set of lithography process images using the Bottleneck MobileNetV3 sequential convolutional neural network model, and simultaneously analyze the position and relative features of each image within the set. Based on this, determine the process window within the current set of images that meets process production requirements.

[0066] like Figure 1 As shown in the figure, during the usage phase, a set of 99 images (112 pixels by 112 pixels) are arranged into a matrix (112, 112, 99) and fed into the sequential neural network model. After analysis, the model generates a 99x3 vector representing the probability of each image being blank, unqualified, or qualified. The maximum probability for each image is taken as the output, or softmax. This value is saved as the lithography process window parameter range that meets production process requirements.

[0067] During the use phase, a large number of unlabeled lithography process images are collected. Every time 1,000 new images are collected, the neural network is regularly updated based on semi-supervised learning. The process is as follows: Figure 2 As shown in Figure 1, a set of data is augmented twice. The data is input into the twin network to obtain the predicted value, and the predicted value is used as supervision for each other to update the parameters through distillation learning.

[0068] The process window selected by the neural network is sampled and corrected. The corrected results are regularly fed into the network as new data for training, and parameters are updated through distillation learning. This prevents the extracted effective features from being forgotten when learning new data.

[0069] 2) In order to make the parameters of the network model meet the requirements, the present invention updates the parameters of the network model through training, and the training process includes first training, second training and third training:

[0070] Before training a neural network, a loss function is designed to calculate the difference between the model's predicted results and the actual results. Then, the gradient descent method and the loss function are used to update the model's network parameters.

[0071] The first training includes updating the parameters of the network through supervised learning using labeled manually labeled data.

[0072] Among them, the loss function of the first training stage is expressed as:

[0073]

[0074] in:

[0075] M is the number of categories; N is the number of image groups in the training set at the current training stage;

[0076] yic is a sign function that takes 1 if the true category of sample i is equal to c, otherwise it takes 0;

[0077] p ic is the predicted probability that the observed sample i belongs to category c;

[0078] The parameters of the neural network are updated based on the gradient descent method, and the formula is:

[0079]

[0080] Among them, η is the learning rate, the size is 0.0001, This is the derivative of the loss function for the first training phase. Completing one operation means the network has completed one training run. The first training phase involves 200 training runs.

[0081] The second training involves using a large number of unlabeled images collected during the use of the lithography process, performing two data enhancements on the images through semi-supervised learning, and using the results as labels for network learning, and updating parameters through distillation learning. The training process is as follows: Figure 2 shown.

[0082] The third training includes training using the manually corrected process window, whose loss function is the same as the loss function of the first stage, and updating the network parameters through distillation learning.

[0083] 3) The parameter update according to the second training described in step 2) is divided into two stages: semi-supervised learning and distillation learning.

[0084] First, design the loss function for semi-supervised learning in the second training phase:

[0085]

[0086]

[0087] Among them, C j (F(x i |E (T)) [θ j ])), j∈{11,12}. Represents the classifier of the two twin networks, whose output is a vector of [1,99*3], representing the probability of each image being empty (blank), unqualified, and qualified;

[0088] x′ i and x i Represents two data enhancements of the i-th sample in the target domain;

[0089] L sid (θ 11|θ 12 ) represents the loss function based on the Siamese network model Net2;

[0090] L sid (θ 12 |θ 11 ) represents the loss function based on the twin network model Net1;

[0091] Then the parameters θ are adjusted based on the gradient descent method and two loss functions. 11 and θ 12 To update:

[0092]

[0093]

[0094] Then the parameters θ1 and θ2 of the twin network model are updated by the distillation learning method, and the formula is:

[0095] E (T) [θ1]=α1*E (T-1) [θ1]+(1-α1)θ 11

[0096] E (T) [θ2]=α1*E (T-1) [θ2]+(1-α1)θ 12

[0097] Among them, E (T) [θ1] and E (T) [θ2] represents the updated parameters of the Siamese network model after the parameters are updated in this training phase through distillation learning;

[0098] Initial parameter E (0) [θ1]=θ,E (0) [θ2] = θ, where θ represents the initial parameters of the network before updating the parameters in this training phase;

[0099] α1 represents the weight of retaining old knowledge set by the distillation learning of the updated twin network in this training stage, and its range is (0,1]. The semi-supervised learning in the second training stage is trained for 200 times;

[0100] Since the training gradually stabilizes and the performance is similar, we randomly select one parameter from the parameters θ1 and θ2 of the two twin networks and update the parameters of the sequential convolutional neural network model in the use phase through distillation learning. The formula is:

[0101] E (T) [θ]=α2*E (T-1) [θ]+(1-α2)θ1

[0102] In the formula, the parameters of the randomly selected twin network are θ1, E (T) [θ] is the updated parameter of the sequential convolutional neural network model in the usage phase after the parameter update; E (T-1) [θ] is the parameter of the sequential convolutional neural network model in the usage phase before updating the parameters; α2 represents the weight of retaining old knowledge set by the distillation learning of the sequential convolutional neural network model parameters in the usage phase during this training phase, and its range is (0,1].

[0103] 4) The loss function and gradient descent formula of the third training stage are the same as those of the first training stage. 200 training cycles are performed to obtain a new parameter θ3, which is then updated through distillation learning. The formula is:

[0104] E (T) [θ]=α3*E (T-1) [θ]+(1-α3)θ3

[0105] Among them, E (T) [θ] is the updated parameter of the sequential convolutional neural network model in the usage phase after the parameter update; E (T -1) [θ] represents the parameters of the sequential convolutional neural network model in the usage phase before updating the parameters; θ3 represents the parameters of the network in this phase obtained based on supervised learning; α3 represents the weight of retaining old knowledge set by the distillation learning of the sequential convolutional neural network model parameters in the usage phase during this training phase, and its range is (0,1].

[0106] 5) The Bottleneck MobileNetV3 described in step 1) consists of a Bottleneck MobileNetV3 block and a soul block. This is used to ensure that the model saves computational effort during the decoupling process and constantly considers the relative position of each image within the set of images.

[0107] The Bottleneck MobileNetV3 block uses group convolution, bottleneck structure and SE module to filter out redundant features while saving computation and re-summarize new features based on the refined features. Its module structure is as follows Figure 3 As shown in the figure:

[0108] ReLU represents the activation function, which makes the model space nonlinear instead of linear.

[0109] Bn stands for batch normalization. Before the feature is input to the ReLU function, it is normalized so that the range of each input value is constrained between (0,1]

[0110] Group Conv stands for group convolution. Compared to ordinary convolution, group convolution performs convolution operations on every n layers of feature channels to extract features.

[0111] Through grouped convolution, the model always maintains the characteristics of the sequence. That is, although each image will continuously integrate the characteristics of other images, it always summarizes its own characteristics under the premise of considering the characteristics of other images.

[0112] The soul block module can extract new features while retaining shallow features, thus avoiding excessive decoupling of the model.

[0113] Its network structure is as follows Figure 4 shown.

[0114] Based on the soul block module, Bottleneck MobileNetV3 block module, ReLU, the sequential neural network model built by the present invention has a structure as shown in Table 1. A group of 99 images with a length and width of 112 pixels are arranged into a matrix of (112, 112, 99) as input, and pass through various modules of the sequential convolutional neural network model in sequence. New features are analyzed from one module each time as the input of the next module. And finally a 99*3 vector is output, with each three values as a group, representing the probability of a picture being blank, unqualified, and qualified. The maximum probability of each picture is taken as the output result, that is, softmax. It is used as the output result. And the processing parameters of all qualified pictures are used as the lithography process window of the template represented by the current group of images, that is, the parameter range that can be processed qualifiedly.

[0115] Table 1 Sequential neural network model structure constructed by the present invention

[0116]

[0117]

[0118] The present invention is carried out in the following environment:

[0119] We used the NVIDIA Jetson NX as the edge algorithm deployment platform and the NVIDIA Jetson TX2 as the cloud deployment platform to achieve cloud-edge AI algorithm collaboration. The edge algorithm deployment platform used the Ubuntu 18.04.5 operating platform and the Linux kernel version 4.9.20-tegra. The bottleneckmobilenetV3 algorithm was built based on OpenCV 4.5.1, CUDA Toolkit 11.6, PyTorch, and other frameworks, and was developed in Python 3.7.

[0120] Compared with the algorithm based on MSSIM, PSNR, the consistency of the algorithm of the present invention is 90.6, and the overlap rate of the manual label is 86.2%.

[0121] The present invention adopts a collaborative deployment of edge and cloud AI, which can achieve real-time diagnosis and accumulate data for correction and upgrade, avoiding the considerable time of uploading videos and downloading results when using cloud servers as computing power platforms.

[0122] The present invention can complete all diagnostic work locally, avoiding risks such as information leakage and interference caused by network transmission.

Claims

1. A method for screening lithography process windows based on a sequential convolutional neural network model, characterized in that: The method specifically includes: Pre-training phase: First, we built a Bottleneck MobileNetV3 sequential convolutional neural network model and trained the neural network using a manually labeled lithography process window dataset. The Bottleneck MobileNetV3 sequential convolutional neural network model is used to screen a set of lithography process images for a parameter range of the lithography process that meets production requirements, where the parameter range is defined as a lithography process window. Use phase: A set of lithography images with the same template but different processing parameter settings are input into the Bottleneck MobileNetV3 sequential convolutional neural network model, deep features of each image are extracted, and relative parameters between images are analyzed. Based on this, the sequential convolutional neural network model determines which images in the current set of images have processing parameters that meet process production requirements, and combines the processing parameters of these images into a lithography process window as output; the output result is stored as a data set of the selected lithography process window; Use the process to regularly upgrade the network: After being put into use, the sequential convolutional neural network model in the use phase is upgraded every time 1,000 new sets of images are collected. The training of the upgraded neural network specifically includes: a) Acquire the latest unlabeled lithography process images during use; Based on the semi-supervised learning method, perform two different data augmentations on a set of images with the same template but different parameters; After the data augmentation, the two sets of images are input into two Bottleneck Mobile NetV3 sequential convolutional neural network models with the same structure and parameters, i.e., a pair of twin network models, and obtain the output of the two network models, i.e., the probability of judging the processing parameters of each image in the set of images to be qualified. The probability is used as the data for training the network model to obtain the new parameter θ 11 and θ 12 , and update the parameters of the two twin network models through distillation learning; After the training is completed, a parameter is selected from the parameters θ1 and θ2 of the two twin network models, and the parameters of the sequential convolutional neural network model in the use stage are updated through distillation learning; b) Spot-check the lithography process window results selected by the sequential convolutional neural network model after the use phase and manually correct them; regularly use the corrected results as a new dataset, and based on supervised learning, use the manually corrected dataset to train the sequential convolutional neural network model to obtain new parameters θ3, and update the parameters of the sequential convolutional neural network model in the use phase through distillation learning; Through distillation learning, the sequential convolutional neural network model can maintain good performance when facing new images in the future, while not causing performance degradation when processing old data due to parameter updates; The Bottleneck Mobile NetV3 sequential convolutional neural network model is constructed by stacking a Bottleneck Mobile NetV3 block module and a soul block module. These two modules ensure that the model saves computational effort during the analysis process and allows the model to consider the relative position of each image in the group when analyzing a group of images. In order to make the relationship between each layer of the sequential convolutional neural network model nonlinear, each layer of the model is linked using the ReLU activation function; the calculation formula of the function is: (4)； In order to ensure that the input of each layer is constrained between [0, 1], batch normalization is used to normalize the input before inputting it into the ReLU function, and each input value is constrained between [0, 1] to obtain the input x of the ReLU function.

2. The screening method according to claim 1, wherein The Bottleneck MobileNetV3 block module consists of a Group Conv, a bottleneck structure, and an SE module. The Group Conv first divides the input features into m blocks along the channel dimension, and then calculates each feature block separately. Finally, after a convolution operation, m features are obtained. When the input is a group of images, the features extracted from each image are processed separately. The bottleneck structure uses the self-learning ability of the neural network to make the output have fewer features than the input, thereby summarizing the input features; then the output features are used as the new input for the next stage, and more features are analyzed to obtain new outputs; The SE module averages the features of each image and uses a neural network to analyze the output to obtain the same number of features as the input. The average is then multiplied by each feature of each image to represent the features represented by the change relationship between each image. When features are input into the Bottleneck Mobile NetV3 block module, the output features still retain the properties of a sequence, that is, each image is always extracting its own features, and the features are arranged in the order of the images.

3. The screening method according to claim 1, wherein The soul block module extracts new features while retaining shallow features, inserting the new features of each image behind its shallow features; at the same time, it can still make the features have sequential properties, that is, the order of features is arranged according to the order of each image.

4. The screening method according to claim 1, wherein The training method a) of the upgraded neural network updates the parameters of the two twin network models through distillation learning, and its formula is: （1）； in, and They represent the parameters updated in this training phase and the parameters updated in the two twin network models after distillation learning; Initial parameters , , θ represents the initial parameters of the network before updating the parameters in this training phase and in the usage phase; Represents the weight of retaining old knowledge set by the distillation learning of updating the Siamese network model in this training stage, and its range is (0,1]; Select any one of the parameters θ1 and θ2 of the two twin network models and update the parameters of the sequential convolutional neural network model in the use phase through distillation learning. The formula is: (2)； In the formula, the parameter of the randomly selected twin network is θ1, The updated parameters of the sequential convolutional neural network model in the usage phase after the parameters are updated; Parameters of the sequential convolutional neural network model in the usage phase before updating the parameters Represents the weight of retaining old knowledge set by the distillation learning of the sequential convolutional neural network model parameters in the update use phase in this training phase, and its range is (0,1].

5. The screening method according to claim 1, wherein In the training method b) of the upgraded neural network, a new parameter θ3 is obtained, and the parameters of the sequential convolutional neural network model in the use phase are updated through distillation learning, and the formula is: (3)； in, The updated parameters of the sequential convolutional neural network model in the usage phase after the parameters are updated; is the parameter of the sequential convolutional neural network model in the use phase before updating the parameters; θ3 represents the parameter of the network model in this phase obtained based on supervised learning Represents the weight of retaining old knowledge set by the distillation learning of the sequential convolutional neural network model parameters in the update use phase in this training phase, and its range is (0,1].

Images