Optical proximity effect correction method based on U-Net neural network

Abstract

The invention discloses an optical proximity correction method based on a U-Net neural network. The method comprises the following steps of: 1, generating a data set for training and testing by utilizing photoetching analogue simulation software; 2, preprocessing the data set image, wherein the preprocessing comprises cutting of a mask and a light intensity distribution diagram and edge extraction of the light intensity distribution diagram; 3, constructing a U-Net convolutional neural network, and training an existing data set; 4, carrying out the same preprocessing process on the light intensity distribution image under the ideal condition, taking the light intensity distribution image as input, and obtaining a target mask pattern by utilizing the network model obtained through training; and 5, simulating the target mask pattern by utilizing photoetching simulation software existing in a laboratory to obtain a light intensity distribution diagram, and comparing and observing the light intensity distribution diagram obtained by the target mask pattern and an unmodified mask. The optimized mask pattern is simulated, so that the edge placement error can be obviously reduced, and the precision and manufacturability of a photoetching pattern are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithography mask optimization, and particularly relates to an optical proximity effect correction method based on a U-Net neural network. Background Art

[0002] Integrated circuits are made by using semiconductor manufacturing technology to integrate numerous basic electronic components, such as transistors, resistors, inductors, and capacitors, on a single silicon wafer. Through a fine wiring method, these components are combined and then packaged in a casing to finally form a tiny structured device with specific circuit functions. Making integrated circuit chips smaller and smaller is one of the core development directions in the semiconductor industry, which not only means an increase in computing speed but also a reduction in cost and power consumption. Lithography is the most critical, precise, complex, and costly process in integrated circuit manufacturing (accounting for more than 30% of the integrated circuit manufacturing cost), and it is one of the core factors determining whether integrated circuit chips can be made smaller and smaller.

[0003] The lithography system is essentially a precision device that realizes the replication of micro-nano patterns through optical principles. Its core components include a light source, a mask, a projection lens system, and a silicon wafer substrate coated with photosensitive material. With the development of micro-nano manufacturing technology, the design and optimization of mask patterns have become one of the core factors restricting lithography quality and structural accuracy. Under the conditions of dense distribution of complex patterns and significant interference between adjacent structures, the optical proximity effect (OPE) will cause deviations in the topography of wafer patterns, and common problems include line width variation, corner rounding, and pattern collapse.

[0004] Currently, traditional optical proximity correction techniques highly rely on lithography process simulation and iterative verification of mask patterns. However, as a computing-intensive task, lithography simulation's computational resource consumption and time cost have become bottlenecks in process development. There is an urgent need to break through the efficiency limitations of the existing mask correction paradigm through algorithmic architecture innovation or hardware acceleration technologies. The Lin team designed a resist model construction framework based on Residual Network (ResNet), enhancing the data utilization rate of the model through transfer learning technology and significantly reducing the demand for training samples. The Yang team pioneered the introduction of Conditional Generative Adversarial Network (cGAN) into the mask optimization process of Inverse Lithography Technology (ILT). By using GAN to generate high-potential masks as initial solutions and combining traditional ILT tools for refinement, both the convergence speed is accelerated and the manufacturability of mask patterns is improved. The Chen team first fused the level set method with a deep neural network, introducing curvature constraints in the joint optimization framework to simplify the mask topology structure. At the same time, real-time mask optimization is achieved based on GPU parallel computing, greatly enhancing the printability of complex patterns. To improve pattern reproduction accuracy, traditional methods such as Rule-Based Optical Proximity Correction (OPC) and Model-Based OPC rely on lithography simulation platforms for repeated iterative calculations. Although the accuracy is high, the computational volume is huge and the efficiency is limited. Inverse Lithography Technology (ILT) attempts to directly derive the corresponding mask with the optimized target pattern as the input, but its optimization process still relies on complex simulations, and there are manufacturability problems with the generated masks.

[0005] The success of deep learning methods in the field of images has inspired research on their application in lithography optimization tasks. As an end-to-end image segmentation model with a symmetric structure, the U-Net convolutional neural network has strong pattern edge recovery and structure prediction capabilities and has been widely used in high-resolution tasks such as medical images and industrial defect recognition.

[0006] Therefore, an optical proximity effect correction method based on the U-Net neural network is proposed to achieve mask prediction. Taking the light intensity distribution map as the input, rapid and non-iterative prediction of the target mask pattern is realized through training. Summary of the Invention

[0007] The purpose of the present invention is to provide an optical proximity correction method based on the U-Net neural network for realizing optical proximity effect correction during the lithography simulation process. The following experimental scheme will take the optimization process of a square expected lithography pattern as an example to study the feasibility of this method to solve the technical problems mentioned in the background art.

[0008] To solve the above technical problems, the specific technical solution of the present invention is as follows:

[0009] An optical proximity correction method based on the U-Net neural network, which realizes supervised image conversion based on the U-Net network, and then completes the correction and optimization of the standard mask corresponding to the square pattern, including the following steps:

[0010] Step 1: Construct a training dataset: Use lithography simulation software to perform corner trimming on the mask pattern to obtain a new mask image, obtain the light intensity distribution maps corresponding to different masks, and form a paired dataset;

[0011] Step 2: Image preprocessing: Crop and edge-delineate the mask image and the corresponding light intensity distribution map obtained in Step 1. The unified size range is from 50×50 pixels to 100×100 pixels. Finally, a new dataset is obtained through the preprocessing process;

[0012] Step 3: Construct and train the U-Net network model: Design a U-Net network including an encoder and a decoder, and train it with the preprocessed dataset;

[0013] Step 4: Predict and optimize the mask: Input the preprocessed target light intensity distribution map in Step 2 into the trained U-Net network model in Step 3, and output the optimized mask pattern;

[0014] Step 5: Performance evaluation: Input the optimized mask pattern obtained in Step 4 into the existing lithography simulation tool in the laboratory, simulate and output the light intensity distribution map, and adopt a method for evaluating the optimization effect of the light intensity distribution map and the ideal pattern. The optimization effect is evaluated by calculating the magnitude of the edge placement error.

[0015] Further, in Step 2, the unified size is 80×80 pixels.

[0016] Further, Step 3 includes the following steps:

[0017] Step 3-1: Define the input and output size parameters according to the ideal conditions; Build the U-Net network according to the sizes of the input and output patterns defined in the preprocessing; Set the size of the output feature map to be the same as that of the input feature map, set the appropriate number of training epochs, the number of channels, and the size of the convolutional kernel; Build the U-Net network; The main body of the U-Net network includes an encoder and a decoder. After passing through the decoder output, it enters the output layer to map the obtained image. At the same time, the binary cross-entropy loss function is used to characterize the training situation:

[0018] A. Encoder

[0019] The encoder includes an input layer, 6 convolutional layers, and 3 pooling layers; The input graph X is a tensor of size m×m×1 in the real number domain, that is

[0020] X∈R m×m×1(1)

[0021] Among them, m is the number of pixels corresponding to the side length of the light intensity distribution map; the convolutional layer represents performing a convolutional operation on the previous-level image, and the mathematical expression for defining image convolution is as follows

[0022]

[0023] Among them, f is the input feature map, g is the convolutional kernel, and i and j are positive integers respectively, representing the coordinates of any point within the tensor f; (x, y) are the coordinates of the output feature map; the ReLU function is used as the activation function; the ReLU function is expressed as

[0024] ReLU(x) = max(0, x) (3)

[0025] Performing a convolutional operation on the input graph X using the convolutional kernel W1 and the bias b1 is expressed as

[0026] Conv1 = ReLU(W1 * X + b1) (4)

[0027] Conv1 is the result obtained from the first convolution; after performing two convolutional operations on the input X, Conv2 is obtained, which is a tensor of 80 * 80 * 64; the selected pooling function is the max pooling function, and the selected pooling size is 2 * 2. The max pooling function is defined as MaxPooling(x, y), as shown in Equation (5)

[0028]

[0029] Therefore, a pooling operation can be performed on Conv2, that is

[0030] Pool1 = MaxPooling(Conv2) (6)

[0031] The pooled feature map is denoted as Pool1, and its size is

[0032]

[0033] Continuing to perform two rounds of convolution and pooling processes on Pool1 according to Equation (4), Equation (5), and Equation (6), finally Pool3 is obtained; after pooling, the number of channel layers doubles. Let Then the finally obtained Pool3 should be a tensor of k × k × 512;

[0034] B. Decoder

[0035] The decoder is symmetric to the encoder and contains 6 convolutional layers and 3 upsampling layers; the operation of the decoder is upsampling, and the upsampling layers complete the restoration of the resolution of the feature map;

[0036] The Pool3 output from the encoder is further subjected to two convolutional operations to obtain Conv4. Conv4 is processed using the nearest neighbor interpolation method to obtain the upsampling layer Up5, satisfying

[0037]

[0038] where \(i',j'\in\{0,1,\cdots,2N - 1\}\), \(c\in\{0,1,\cdots,C - 1\}\), \(N\) is the side length of the image, and \(C\) is the number of channels; \(i',j',c\) respectively represent the plane position and channel ordinal number of any point within the tensor Up5. After interpolation, two convolutional operations are performed on Up5 to obtain Conv5. After two rounds of convolution and upsampling in the decoder, Conv7 is finally obtained, satisfying

[0039] Conv7\in\mathbb{R} m×m×64 (9)

[0040] C. Output Layer

[0041] The fully connected layer uses a 1x1 convolution to map the feature map to the number of output channels, and finally a tensor \(Y\) of size \(m\times m\times1\) is obtained as the output layer of the network. The sigmoid function needs to be applied in this convolution. Define the sigmoid function as \(\sigma(x)\), that is

[0042]

[0043] where \(e -x represents the exponential operation. Map the value of the output Conv7 to the range \((0,1)\), determine the color of the corresponding pixel in the output to be black or white, and generate a mask shape composed only of black and white pixels. Take the convolutional kernel \(W7\) and bias \(b7\), then

[0044] Y = \sigma(W7 * Conv7 + b7)\ (11)

[0045] D. Loss Function

[0046] Use the binary cross - entropy loss function as the evaluation metric. The binary cross - entropy loss function is defined as

[0047]

[0048] where \(y i is the true label, is the probability predicted by the model. Calculate the fitting degree after each round of training through the formula (12) to characterize the training situation;

[0049] Step 3 - 2: Start training the dataset; during the training process, calculate and record the cross - entropy loss respectively, and then update the model parameters through gradient descent.

[0050] Further, the method for evaluating the optimization effect of the light intensity distribution map and the ideal pattern in step 5 specifically includes the following steps:

[0051] Simulate according to the mask pattern in step 4 using the existing lithography simulation software in the laboratory to obtain the light intensity distribution map. Compare and observe the light intensity distribution map obtained with the unmodified mask, calculate the edge placement error EPE, and evaluate its optimization degree according to the magnitude of EPE. The calculation formula of EPE is shown in the following formula (13):

[0052]

[0053] Where L C is the total contour length of the target pattern, represents the edge placement error of each point on the contour, represents the loop integral operation for the curve c; the integral calculation is converted to a series calculation, that is, as shown in formula (14):

[0054]

[0055] Change the input and output image sizes, the size of the convolution kernel, and the number of training rounds respectively. Simulate the obtained mask pattern using the existing lithography simulation software in the laboratory to obtain the light intensity distribution map, calculate the EPE of the light intensity distribution map, compare the optimization degree, and screen to obtain a more optimized mask pattern.

[0056] An optical proximity correction method based on the U-Net neural network of the present invention has the following advantages:

[0057] 1. The data set required for training is easy to generate, and the training rate can be improved through appropriate preprocessing methods;

[0058] 2. Compared with the traditional optical proximity effect correction, the robustness and correction rate of the correction realized by the present invention can be greatly improved;

[0059] 3. Select the U-Net network to complete the supervised image conversion, which occupies less memory compared with the generative adversarial network and the diffusion model, and has a faster fitting rate. Brief Description of the Drawings

[0060] Figure 1 It is an operation process model diagram for realizing optical proximity effect correction based on the U-Net network;

[0061] Figure 2 It is a schematic diagram of generating a data set by modifying a mask using rule-based OPC;

[0062] Figure 3(a) is the operation of extracting the edge features of the light intensity distribution map, which is histogram equalization, and Figure 3(b) reflects the effect diagram of edge extraction;

[0063] Figure 4 It describes the U-Net-based neural network graph constructed in this processing procedure;

[0064] Figure 5 It represents the function relationship graph between cross-entropy loss and the number of training epochs;

[0065] Figure 6(a) represents the expected light intensity edge extraction graph; Figure 6(b) represents the mask corner graph obtained by bringing in the training results of the U-Net network. Specific implementation manners

[0066] To better understand the purpose, structure and function of the present invention, the following further describes in detail an optical proximity correction method based on a U-Net neural network of the present invention in conjunction with the accompanying drawings.

[0067] Figure 1 It describes the overall process of implementing optical proximity effect correction based on the U-Net network. The dataset generated based on RBOPC is preprocessed, then a U-Net network is designed to complete supervised image conversion, realizing the training of the dataset. Then, the expected light intensity distribution map is used to bring in the model obtained by training, and the expected mask is converted, and then the conversion effect is evaluated. Figure 1 In it, Input represents the input tensor, Output represents the output tensor, Encoder and Decoder respectively represent the encoder and the decoder, and classifier represents the output layer with classification function.

[0068] Step 1: Generate the dataset for training and testing. Based on the existing lithography simulation software in the laboratory, it can output the simulated light intensity distribution map during the lithography process by inputting the pattern of the mask. Using the above method, more than 700 pairs of datasets can be obtained for training by modifying the edges of the mask corresponding to the ideal pattern with different degrees and different shapes.

[0069] Specific modifications can be realized by using figures with different sizes, different rotation angles and different shapes at the corners. Figure 1 It provides an idea for modification: add modified parts of rhombus, square and triangle respectively to the corners of the original figure. The size and placement direction of each figure can be changed. A total of 720 masks are generated in this way, and through the lithography simulation technology, the respective light intensity distribution maps are obtained. Thus, the generation of the dataset is completed.

[0070] Step 2: Preprocess the dataset images. To reduce unnecessary operations, the generated dataset is preprocessed, including a series of operations such as cropping and edge carving, so as to improve the processing rate of the image.

[0071] Step 2-1: Cropping of the mask and the light intensity distribution map. The four corners of the square are symmetric. Only extracting the mask and the light intensity distribution map at one corner will greatly improve the training rate. Here, we crop them into feature maps of size 80×80.

[0072] Step 2-2: Edge extraction of the light intensity distribution map. The light intensity distribution map obtained by the lithography simulation software is a grayscale image. To characterize the features of the light intensity distribution of each dataset and facilitate convolution calculation, edge extraction is performed on the grayscale image of the light intensity distribution to obtain Figure 2 the corresponding pattern.

[0073] The algorithm principle used for edge extraction is as follows: Since the corner feature extracted is the upper right corner, the gray value is detected pixel by pixel from top to bottom and from right to left. When it is greater than the threshold, the point is retained and the next line is moved to. For the selection of the threshold, first draw the gray histogram of the diffraction image to ensure that the threshold selection can achieve the segmentation of the two peaks of the histogram. Based on the gray histogram, the gray threshold is selected as 64, and the threshold realizes the segmentation of the two-peak feature of the gray histogram. The specific process of edge extraction can be specifically referred to Figure 3(a) and Figure 3(b), which are the related schematic diagrams for extracting features based on the light intensity distribution map. Figure 3(a) shows the operation of histogram equalization, and Figure 3(b) reflects the effect diagram of edge extraction.

[0074] Step 3: Construct a U-Net convolutional neural network to train the existing dataset.

[0075] Step 3-1: Define the input and output size parameters according to the ideal conditions; build the U-Net network according to the sizes of the input and output patterns defined in the preprocessing; set the size of the output feature map to be the same as that of the input feature map, set appropriate training epochs, number of channels, and the size of the convolutional kernel; build the U-Net network; note the relevant parameters of this U-Net network. In the embodiment, m is taken as 80 for easy explanation. The network includes two parts: an encoder and a decoder. After the output of the decoder, it enters the output layer to map the obtained image. At the same time, the binary cross-entropy loss function is used to characterize the training situation:

[0076] A. Encoder

[0077] The encoder includes an input layer, 6 convolutional layers and 3 pooling layers; the input graph X is a tensor of 80×80×1 in the real number domain, that is

[0078] X∈R 80×80×1 (1)

[0079] The convolutional layer represents performing a convolution operation on the previous-level image. The mathematical expression for image convolution is defined as follows

[0080]

[0081] Among them, f is the input feature map, g is the convolution kernel, and i and j are positive integers respectively, representing the coordinates of any point in the tensor f; (x, y) are the coordinates of the output feature map; the ReLU function is used as the activation function; the ReLU function is expressed as

[0082] ReLU(x) = max(0, x) (3)

[0083] Performing a convolution operation on the input graph X using the convolution kernel W1 and the bias b1 is expressed as

[0084] Conv1 = ReLU(W1 * X + b1) (4)

[0085] Conv1 is the result obtained from the first convolution; after performing two convolution operations on the input X, Conv2 is obtained, which is a tensor of 80×80×64; the selected pooling function is the max pooling function, the selected pooling size is 2×2, and the max pooling function is defined as MaxPooling(x, y), as shown in Equation (5)

[0086]

[0087] Therefore, a pooling operation can be performed on Conv2, that is

[0088] Pool1 = MaxPooling(Conv2) (6)

[0089] The pooled feature map is denoted as Pool1, and its size is

[0090] Pool1 ∈ R 40×40×128 (7)

[0091] Continuing to perform two rounds of convolution and pooling processes on Pool1 according to Equation (4), Equation (5), and Equation (6), finally Pool3 is obtained; after pooling, the number of channel layers doubles, and the finally obtained Pool3 should be a tensor of 10×10×512;

[0092] B. Decoder

[0093] The decoder is symmetric to the encoder and contains 6 convolutional layers and 3 upsampling layers; the operation of the decoder is upsampling, and the upsampling layers complete the restoration of the resolution of the feature map;

[0094] Continuing to perform two convolution operations on the Pool3 output by the encoder to obtain Conv4, and processing Conv4 using the nearest neighbor interpolation method to obtain the upsampling layer Up5, satisfying

[0095]

[0096] where \(i', j' \in \{0, 1, \ldots, 2N - 1\}\), \(c \in \{0, 1, \ldots, C - 1\}\), \(N\) is the side length of the image, and \(C\) is the number of channels; \(i'\), \(j'\), and \(c\) represent the plane position and channel ordinal number of any point in the tensor Up5 respectively; after interpolation, two convolutions are performed on Up5 to obtain Conv5; after two rounds of convolution and up-connection in the decoder, Conv7 is finally obtained, satisfying

[0097] Conv7 \(\in \mathbb{R}\) 80×80×64 (9)

[0098] C. Output Layer

[0099] The fully connected layer uses a 1x1 convolution to map the feature map to the number of output channels, and finally a tensor \(Y\) of \(80\times80\times1\) is obtained as the output layer of the network; the sigmoid function needs to be used in this convolution. Define the sigmoid function as \(\sigma(x)\), that is

[0100]

[0101] where \(e\) -x represents the exponential operation. Map the value of the output Conv7 to the range within \((0, 1)\), determine the color of the corresponding pixel in the output to be black or white, and generate a mask shape composed only of black and white pixels; take the convolution kernel \(W7\) and the bias \(b7\), then

[0102] \(Y = \sigma(W7 * Conv7 + b7)\) (11)

[0103] D. Loss Function

[0104] Use the binary cross-entropy loss function as the evaluation metric. The binary cross-entropy loss function is defined as

[0105]

[0106] where \(y\) i is the true label, is the probability predicted by the model; calculate the fitting degree after each round of training through the calculation of formula (12), so as to characterize the training situation;

[0107] The finally obtained U-Net network can be seen in Figure 4 as shown, where Encoder Path and Decoder Path represent the encoding path and the decoding path respectively.

[0108] Step 3-2: Start the training of the dataset. During the training process, calculate and record the cross-entropy loss respectively, and then update the model parameters through gradient descent. After 50 rounds of training, it can be seen that the cross-entropy loss continues to decrease and stabilizes, as Figure 5 shown. Among them,Figure 5 The Training and Validation Loss over Epochs represents the cross-entropy loss of the training set and the test set during the training process.

[0109] Step 4: After performing the same preprocessing process on the light intensity distribution image under ideal conditions as the input, use the network model trained in Step 3 to obtain the target mask pattern.

[0110] Figure 6(a) is the input image for edge extraction, and Figure 6(b) is the corresponding mask prediction output image (only the corners, not processed as a mask).

[0111] Step 5: Simulate the mask pattern generated according to Step 4 using the existing lithography simulation software in the laboratory to obtain the light intensity distribution map. Compare and observe the light intensity distribution map obtained with the unmodified mask, calculate the Edge Placement Error (EPE), and evaluate its optimization degree according to the magnitude of the EPE. The calculation formula for EPE is shown as follows in Equation (13)

[0112]

[0113] where L C is the total contour length of the target pattern, represents the edge placement error at each point on the contour, represents the loop integral operation for the curve c; the integral calculation is converted to a series calculation, as shown in Equation (14):

[0114]

[0115] Change the sizes of the input and output images, the size of the convolutional kernel, and the number of training epochs respectively. Simulate the obtained mask patterns using the existing lithography simulation software in the laboratory to obtain the light intensity distribution maps, calculate the EPE of the light intensity distribution maps, compare the optimization degrees, and screen to obtain more optimized mask patterns.

[0116] It can be understood that the present invention is described through some embodiments. Those skilled in the art know that without departing from the spirit and scope of the present invention, various changes or equivalent replacements can be made to these features and embodiments. Additionally, under the teaching of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the scope protected by the present invention.

Claims

1. An optical proximity effect correction method based on a U-Net neural network, characterized in that, It includes the following steps: Step 1: Construct a training dataset: Using lithography simulation software, perform corner rounding on the mask pattern to obtain a new mask pattern, acquire the light intensity distribution maps corresponding to different masks, and form a paired dataset. Step 2: Image preprocessing: Crop and edge-delineate the mask pattern and the corresponding light intensity distribution map obtained in Step 1. The unified size range is from 50×50 pixels to 100×100 pixels. Finally, a new dataset is obtained through the preprocessing process. Step 3: Construct and train a U-Net network model: Design a U-Net network including an encoder and a decoder, and train it with the preprocessed dataset. Step 4: Predict and optimize the mask: Input the preprocessed target light intensity distribution map in Step 2 into the trained U-Net network model in Step 3, and output the optimized mask pattern. Step 5: Performance evaluation: Input the optimized mask pattern obtained in Step 4 into the existing lithography simulation tool in the laboratory, simulate and output the light intensity distribution map, and adopt a method for evaluating the optimization effect of the light intensity distribution map and the ideal pattern. Evaluate the optimization effect by calculating the magnitude of the edge placement error.

2. The optical proximity effect correction method based on the U-Net neural network according to claim 1, characterized in that In Step 2, the unified size is 80×80 pixels.

3. The optical proximity effect correction method based on the U-Net neural network according to claim 1, wherein Step 3 includes the following steps: Step 3-1: Define the input and output size parameters according to the ideal conditions; build the U-Net network according to the sizes of the input and output patterns defined in the preprocessing; set the size of the output feature map to be the same as that of the input feature map, set appropriate training epochs, number of channels, and the size of the convolutional kernel; build the U-Net network. The main body of the U-Net network includes two parts: an encoder and a decoder. After passing through the decoder output, it enters the output layer to map the obtained image. At the same time, use the binary cross-entropy loss function to characterize the training situation: A. Encoder The encoder includes an input layer, 6 convolutional layers, and 3 pooling layers; the input graph X is a tensor of size m×m×1 in the real number domain, that is X ∈ R m×m×1 (1) where m is the number of pixels corresponding to the side length of the light intensity distribution map; the convolutional layer represents performing a convolution operation on the previous-level image. The mathematical expression for image convolution is defined as follows where f is the input feature map, g is the convolutional kernel, i and j are positive integers representing the coordinates of any point within the tensor f; (x, y) are the coordinates of the output feature map; use the ReLU function as the activation function; the ReLU function is expressed as ReLU(x) = max(0, x) (3) Perform a convolution operation on the input graph X using the convolutional kernel W1 and the bias b1, which is expressed as Conv1 = ReLU(W1 * X + b1) (4) Conv1 is the result of the first convolution; after performing two convolution operations on the input X, Conv2 is obtained, which is a tensor of 80*80*64; the selected pooling function is the max pooling function, and the pooling size is 2*2. The max pooling function is defined as MaxPooling(x, y), as shown in Equation (5) Thus, a pooling operation can be performed on Conv2, that is Pool1 = MaxPooling(Conv2) (6) The feature map after pooling is denoted as Pool1, and its size is: Perform two more rounds of convolution and pooling processes on Pool1 according to Formula (4), Formula (5), and Formula (6), and finally obtain Pool3; after pooling, the number of channel layers doubles, and let then the finally obtained Pool3 should be a tensor of k×k×512; B. Decoder The decoder is symmetrical to the encoder and contains 6 convolutional layers and 3 upsampling layers; the operation of the decoder is upsampling, and the upsampling layers complete the restoration of the resolution of the feature map. Perform two more convolutional operations on the output Pool3 of the encoder to obtain Conv4, and process Conv4 using the nearest neighbor interpolation method to obtain the upsampling layer Up5, satisfying where i', j' ∈ {0, 1,..., 2N - 1}, c ∈ {0, 1,..., C - 1}, N is the side length of the image, and C is the number of channels; i', j', and c respectively represent the plane position and channel ordinal number of any point within the tensor Up5; after interpolation, perform two more convolutions on Up5 to obtain Conv5; after two rounds of convolution and upsampling by the decoder, finally obtain Conv7, satisfying Conv7 ∈ R m×m×64 (9) C. Output layer Use a 1x1 convolution on the fully connected layer to map the feature map to the number of output channels, and finally obtain a tensor Y of m×m×1 as the output layer of the network; this convolution needs to use the sigmoid function, and define the sigmoid function as σ(x), that is Among them, e -x represents the exponential operation of e. Map the value of the output Conv7 within the range of (0, 1), determine the color of the corresponding pixel in the output as black or white, and generate a mask shape composed only of black and white pixels; take the convolutional kernel W7 and the bias b7, then D. Loss function Use the binary cross-entropy loss function as the evaluation metric, and the binary cross-entropy loss function is defined as where y i is the true label, is the probability predicted by the model; the fitting degree after each round of training is calculated by formula (12) to characterize the training situation; Step 3-2: Start the training of the dataset; during the training process, calculate and record the cross-entropy loss respectively, and then update the model parameters through gradient descent.

4. The optical proximity effect correction method based on the U-Net neural network according to claim 1, characterized in that The method for evaluating the optimization effect of the light intensity distribution map and the ideal pattern in Step 5 specifically includes the following steps: Simulate according to the mask pattern in Step 4 using the existing lithography simulation software in the laboratory to obtain the light intensity distribution map, compare and observe it with the light intensity distribution map obtained from the unmodified mask, calculate the edge placement error EPE, and evaluate its optimization degree according to the size of EPE. The calculation formula of EPE is as shown in (13) below where L C is the total length of the contour of the target figure, S EPEC represents the edge placement error of each point on the contour, ∮ c S EPEC dc represents performing a loop integral operation on the curve c; the integral calculation is converted to a series calculation, as shown in formula (14): Respectively change the sizes of the input and output images, the size of the convolutional kernel, and the number of training epochs. Simulate the obtained mask pattern using the existing lithography simulation software in the laboratory to obtain the light intensity distribution map, calculate the EPE of the light intensity distribution map, compare the optimization degrees, and screen to obtain a more optimized mask pattern.

Images