A lithography mask alignment method based on a transformer model

Abstract

The application discloses a photolithography mask alignment method based on a Transformer model, relates to the technical field of photolithography machines, and is used for solving the problem of inaccurate alignment precision of a photolithography mask; the method first collects image data of a mask and a wafer combination, and removes noise and normalizes the image. Then, the dynamic coefficients of each sub-region are calculated through region division, physical property analysis and a multi-factor fusion algorithm, the extracted feature data and the dynamic coefficient sequence are input into the Transformer model, a self-attention mechanism is processed, and the alignment deviation prediction value of the mask and the wafer is output. Finally, according to the deviation prediction value, an adaptive stepping motor control strategy is used to dynamically adjust the position and angle of the mask, and the translation and rotation errors are accurately compensated. The method can real-time and accurately adjust the position of the photolithography mask, and improve the photolithography precision and production efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photoetching machines, and more particularly to a photoetching machine mask alignment method based on a Transformer model. BACKGROUND

[0002] Photoetching technology is widely used in semiconductor manufacturing processes, especially in the production of integrated circuits (IC). The alignment accuracy of the mask directly affects the quality of the photoetching pattern and the performance of the chip. In traditional photoetching processes, the alignment accuracy of the mask and the wafer usually depends on mechanical control and image matching technology. However, as the size of integrated circuits continues to shrink, the demand for alignment accuracy is becoming increasingly high, and traditional technology is difficult to meet the dual demands of precision and efficiency. In addition, environmental factors (such as temperature and humidity fluctuations, changes in air refractive index, etc.) and the physical properties of the mask itself (such as stress distribution, uneven photoresist thickness, etc.) also affect the alignment accuracy. The existing photoetching machine alignment method is difficult to achieve real-time and accurate adjustment when facing these complex influencing factors, resulting in errors and quality fluctuations in the production process. Therefore, a new alignment method is needed that can combine image data with environmental and physical property information to perform real-time and accurate alignment compensation and improve photoetching quality and production efficiency.

[0003] To solve the above problems, the present application provides a solution. SUMMARY

[0004] To overcome the above-mentioned defects of the prior art, embodiments of the present application provide a photoetching machine mask alignment method based on a Transformer model to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purposes, the present application provides the following technical solutions:

[0006] In a preferred embodiment, it comprises:

[0007] Step 1: Collect image data of the mask and wafer combination and perform denoising, normalization and data enhancement to generate normalized distance and angle labels;

[0008] Step 2: Divide the image into multiple sub-regions, calculate dynamic coefficients and extract features in combination with physical property data, input the Transformer model for global feature regression, and output distance deviation and angle deviation;

[0009] Step 3: Adjust the mask position and angle according to the predicted distance deviation and angle deviation, compensate through a stepper motor and implement low-latency real-time inference on an FPGA.

[0010] In a preferred embodiment, in step 1, image data of different mask and wafer combinations is collected, image data noise is removed, and image data is normalized.

[0011] In a preferred embodiment, in step 1, the denoised and normalized image data is further subjected to anti-interference data enhancement processing, affine transformation matrix is used to simultaneously perform translation, rotation and scaling transformation on the image and the labeled key point coordinates, the enhanced image and the labeled coordinates are ensured to be consistent through the key point synchronous mapping interface, and local nonlinear disturbance is applied to the image combined with the elastic deformation enhancement strategy.

[0012] In a preferred embodiment, in step 2, the mask and wafer area are evenly divided into multiple sub-regions.

[0013] The internal stress distribution of the mask glass in each sub-region is analyzed to generate a stress distribution matrix; the thickness and curing uniformity of the photoresist layer in each sub-region are analyzed to generate a thickness matrix and a curing uniformity matrix, which are recorded as physical property data; temperature and humidity data in each sub-region are collected to monitor air refractive index fluctuations, and a refractive index change matrix is calculated, which is recorded as air refractive index fluctuation data.

[0014] The dynamic coefficient D(x, y) of each sub-region is calculated, and the dynamic coefficient D(x, y) of each sub-region is compared with the dynamic coefficient threshold Yd; when the dynamic coefficient D(x, y) of the sub-region is greater than or equal to the dynamic coefficient threshold Yd, it indicates that the sub-region belongs to a high dynamic region; when the dynamic coefficient D(x, y) of the sub-region is less than the dynamic coefficient threshold Yd, it indicates that the sub-region belongs to a low dynamic region.

[0015] In a preferred embodiment, in step 2, the feature data of the image data in each sub-region is extracted and converted into a fixed-size feature vector Xi, and the feature vector Xi of each sub-region and the dynamic coefficient D(x, y) form sequence data Fi.

[0016] The sequence data Fi of each sub-region is subjected to linear transformation; the correlation matrix between the sequence data of each sub-region is calculated; the sequence data of each sub-region is weighted and fused to determine the global feature information of each sub-region through the correlation matrix, and different relationship information in the global feature information between each sub-region extracted through multiple attention heads is mapped to determine the feature mapping value of each sub-region, and the feature mapping value of each sub-region is fused to determine the alignment deviation prediction value Y of the entire mask and wafer area.

[0017] In a preferred embodiment, in step 2, after linear transformation and weighted fusion of sequence data, the sequence data is input into a regression model based on the VisionTransformer architecture, the input image is divided into image blocks and linear mapping is performed to obtain a feature vector, while a learnable position encoding is added to preserve spatial information.

[0018] The regression model based on the VisionTransformer architecture comprises a twelve-layer encoder, each layer containing multi-head self-attention and a feedforward neural network.

[0019] In a preferred embodiment, in step 2, the input data is divided into image blocks and global features are extracted, and distance deviation and angle deviation are output, the angle is encoded by sine and cosine and then calculated by arctangent, and a layered learning rate is used for different layers during training and combined with weighted mean square error.

[0020] In a preferred embodiment, in step 3, the mask image offset is calculated, and a compensation deviation scheme is designed for high dynamic regions and low dynamic regions.

[0021] In a preferred embodiment, in step 3, when implementing alignment compensation, the model is quantized and pruned, and the calculation process is scheduled in parallel pipeline.

[0022] In a preferred embodiment, the method for obtaining the dynamic coefficient threshold comprises:

[0023] The mean and standard deviation of the physical property data and the air refractive index fluctuation data in each sub-region are calculated, the mean and standard deviation of the physical property data and the air refractive index fluctuation data in each sub-region are analyzed by K-means clustering, and the dynamic coefficient threshold Yd is determined.

[0024] The technical effects and advantages of the photolithography mask alignment method based on the Transformer model are as follows:

[0025] The application introduces global context modeling and data enhancement mechanism in the traditional lithography machine alignment process, which significantly improves the alignment ability under complex working conditions. By adding affine synchronous enhancement and elastic deformation in the image preprocessing link, the data diversity and annotation consistency are improved, so that the model can adapt to the interference caused by noise, illumination and deformation. In the feature extraction and modeling link, the VisionTransformer structure is used to regress the global features, and the layered learning rate, freezing strategy and weighted loss function are combined to ensure the convergence stability of model training and the high precision of prediction. In the compensation execution link, the step motor adjustment scheme is designed for high dynamic and low dynamic regions respectively, realizing accurate compensation of translation and rotation deviation. Finally, through FPGA quantization, pruning and pipeline scheduling, the model still has real-time inference ability in the low-power hardware environment. In summary, the application has high precision, strong robustness and industrial deployment feasibility, which can significantly improve the production efficiency and product quality of the lithography machine. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 An operation flowchart of a lithography mask plate alignment method based on a Transformer model according to the application.

[0027] Figure 2 An implementation flowchart of a lithography mask plate alignment method based on a Transformer model according to the application. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0029] EMBODIMENT

[0030] The application discloses a lithography mask plate alignment method based on a Transformer model, as shown in the figure, which comprises the following steps. Figure 1

[0031] Step 1: Collect image data of different mask plates and wafer combinations and preprocess the image data.

[0032] A high-resolution CMOS camera is used, and the movement and positioning of the mask plate and wafer combination are controlled by an automatic mechanical device. The high-resolution CMOS camera is fixed at a shooting angle parallel to the mask plate and wafer. The combination images of different mask plates and wafers are shot in a full field of view scanning mode, and the image data of different mask plate and wafer combinations are obtained. ​

[0033] Gaussian filter algorithm is used to remove image data noise, and the image data is normalized according to the formula:

[0034] ;

[0035] In the formula, is the normalized value;

[0036] After completing image acquisition and basic preprocessing, further anti-interference data enhancement is performed on the obtained image data. Specifically, affine transformation matrix is used to perform translation, rotation and scaling transformation on the image and labeled coordinates synchronously, and then Albumentations library KeypointParams interface is used to realize real-time mapping between key point coordinates and image, so that each image enhancement strictly corresponds to accurate labeled coordinates, thereby realizing consistency enhancement of image data and labeled data.

[0037] At the same time, through the elastic deformation enhancement strategy, the local thermal deformation of the mask plate that may occur in actual processing and operation is simulated, and a small nonlinear disturbance is introduced in the data enhancement process by using the random elastic distortion algorithm, so as to improve the adaptability of the model to the local deformation of the mask plate under different working conditions.

[0038] In the label generation link, the normalized distance and angle are calculated according to the synchronized key point coordinates. Specifically, the distance label is obtained by calculating the Euclidean distance between the two mark centers and dividing by the diagonal length of the image, and the angle label is obtained by calculating the angle between the line connecting the two points and the horizontal axis and mapping it to the range of zero to one hundred and eighty degrees, realizing the unified scale expression of distance and angle, and enhancing the robustness of subsequent model training.

[0039] The specific mathematical implementation is as follows:

[0040] ;

[0041] Where M is the transformation matrix, (x, y) is the original coordinate, and (x', y') is the transformed coordinate.

[0042] Label design:

[0043] Distance: calculate the Euclidean distance between the two mark centers and normalize:

[0044] ;

[0045] Where Ldiagonal is the diagonal length of the image.

[0046] Angle: calculate the angle between the line connecting the two points and the horizontal axis, and map it to the range of 0-180°:

[0047] ;

[0048] Step 2: Establish the alignment relationship between the mask and the wafer through Transformer model training;

[0049] The mask and wafer regions in the entire image data are evenly divided into multiple sub-regions using a grid-based region division method;

[0050] The internal stress distribution of the mask glass in each sub-region is analyzed using photoelasticity measurement technology to generate a stress distribution matrix S(x, y), and the thickness and curing uniformity of the photoresist layer in each sub-region are analyzed using an ellipsometric spectrometer to generate a thickness matrix T(x, y) and a curing uniformity matrix C(x, y), which are referred to as physical property data;

[0051] Temperature and humidity fluctuations in each sub-region are monitored using an array of temperature and humidity sensors to obtain temperature and humidity data in each sub-region, generating a temperature matrix H1(x, y) and a humidity matrix H2(x, y), and using interference measurement to monitor air refractive index fluctuations based on the temperature and humidity data, calculating a refractive index change matrix R(x, y), which is referred to as air refractive index fluctuation data;

[0052] Based on the physical property data and air refractive index fluctuation data in each sub-region, a multi-factor fusion algorithm is used to calculate the dynamic coefficient D(x, y) of each sub-region, specifically according to the formula: D(x, y) = W1 × S(x, y) + W2 × T(x, y) + W3 × C(x, y) + W4 × R(x, y), where W1 represents the weight coefficient of stress distribution, W2 represents the weight coefficient of photoresist thickness, W3 represents the weight coefficient of curing uniformity, and W4 represents the weight coefficient of air refractive index; The mean and standard deviation of the physical property data and air refractive index fluctuation data in each sub-region are calculated, and the K-means clustering is used to analyze the mean and standard deviation of the physical property data and air refractive index fluctuation data in each sub-region to determine the dynamic coefficient threshold Yd. The dynamic coefficient D(x, y) of each sub-region is compared with the dynamic coefficient threshold Yd. When the dynamic coefficient D(x, y) of a sub-region is greater than or equal to the dynamic coefficient threshold Yd, it indicates that the sub-region belongs to a high dynamic region; When the dynamic coefficient D(x, y) of a sub-region is less than the dynamic coefficient threshold Yd, it indicates that the sub-region belongs to a low dynamic region;

[0053] The CNN extracts feature data from the image data in each sub-region and outputs the feature map of each sub-region to the Transforme model;

[0054] The CNN structure is designed as follows:

[0055] The first convolutional layer uses a 3x3 convolutional kernel and a ReLU activation function to extract edge, texture, and curve features from the image data in each sub-region.

[0056] The second layer of the pooling layer: using 2x2 convolution kernel, maximum pooling enhances the translational invariance;

[0057] The third layer of the convolution layer: using 5x5 convolution kernel, using ReLU activation function, extracting the pattern of image data in each sub-region, shape feature;

[0058] The fifth layer of the global connection layer: using global average pooling to convert the feature data of image data in each sub-region into a fixed size feature vector Xi;

[0059] It should be noted that the data format required by the Transformer model is serialized data, therefore, the feature vector Xi of each sub-region and the dynamic coefficient D(x, y) are formed into a sequence: Fi=[Xi, Di], wherein Di represents the dynamic coefficient D(x, y) of each sub-region, and the sequence data Fi of each sub-region is transmitted to the Transformer model;

[0060] Then, the sequence data Fi of each sub-region is linearly transformed to calculate the Q, K, and V matrices: Q=FiWq, K=FiWk, V=FiWv, wherein Wq, Wk, and Wv represent trainable parameters, Q represents the query matrix, K represents the key matrix, and V represents the value matrix;

[0061] The correlation matrix between the sequence data of each sub-region is calculated

[0062]

[0063] wherein, represents the dimension of K, represents the dot product of Q and K;

[0064] The sequence data of each sub-region is weighted and fused by the correlation matrix

[0065]

[0066] wherein, represents the global feature information of each sub-region after fusion;

[0067] Using multiple attention heads to extract different relationship information in the global feature information of each sub-region:

[0068]

[0069] wherein, h represents the number of attention heads, and Wo represents the final projection matrix;

[0070] ​​​​​The different relationship information in the global feature information output by the Transformer model between each sub-region is mapped by a two-layer fully connected network MLP in a feedforward neural network FFN, and is converted into a numerical prediction according to the formula:

[0071]

[0072] wherein, represents the feature mapping value of the i-th sub-region, Wf1 and Wf2 represent the weight matrix, and bf1 and bf2 represent the bias term.

[0073] The feature mapping values of each sub-region are fused to output the alignment deviation prediction value Y of the entire mask plate and wafer region,

[0074] According to the formula:

[0075]

[0076] wherein, represents the learnable weight, and N represents the total number of sub-regions.

[0077] Finally, the mean square error MSE is used as the loss function to optimize the model parameters.

[0078] The loss function is designed as follows:

[0079] Periodic angle loss: the sine-cosine encoding is used to avoid angle period jump:

[0080]

[0081] Weighted total loss: balance distance and angle error:

[0082]

[0083] Further, on the basis of the above feature extraction and serialization processing, the VisionTransformer architecture is further used to regress the serialized data. Specifically, the input image is divided into sixteen by sixteen pixel image blocks, and a linear mapping is performed on each image block to obtain a seven hundred and sixty-eight dimensional vector feature, and a learnable position encoding is added to retain the spatial position information of each image block.

[0084] Subsequently, the ViT structure containing twelve layers of encoders is used, each layer of encoder is composed of multi-head self-attention and feedforward neural network, and the long distance dependence between the mask plate and wafer image is described through global context modeling.

[0085] ​​​​On the basis of the above-mentioned VisionTransformer architecture, dynamic sparsification processing is performed to further reduce the computational complexity and improve the feature expression capability. Specifically, during the multi-head self-attention calculation process, the contribution of each attention head is calculated in real time according to the average attention score of each head during the training process, and through a threshold screening mechanism, only the attention heads with a contribution higher than the set threshold are retained to participate in the subsequent matrix multiplication calculation, and redundant attention heads are dynamically removed, so that the overall calculation amount is reduced by about forty-three percent on the premise of maintaining the prediction accuracy.

[0086] The specific ViT regression architecture is improved as follows:

[0087] Model structure:

[0088] Input layer: The image is divided into 16x16 pixel blocks and linearly mapped into a 768-dimensional vector.

[0089] Position encoding: Learnable position embedding is added to preserve spatial information.

[0090] Transformer Encoder: A 12-layer encoder is used, each layer containing a multi-head self-attention (Multi-HeadAttention) and an MLP module.

[0091] Regression head: The CLS token feature of ViT is input into a fully connected layer to output a two-dimensional vector (d^, θ^).

[0092] Fine-tuning strategy:

[0093] The first 6 layers of the encoder are frozen, only the last 6 layers and the regression head are trained, and the learning rate is set to 1x10−4.

[0094] AdamW optimizer is used with a weight decay coefficient λ=0.01.

[0095] At the same time, in order to enhance the feature interaction between different encoding layers, the output features of the previous layer and the input features of the current layer are concatenated in the channel dimension, then linearly transformed and integrated into a unified dimension, and then sent to the subsequent multi-head self-attention structure to realize the collaborative regression of local and global information.

[0096] In terms of training strategy, the first six layers of the encoder of the pre-trained model are frozen, only the last six layers of the encoder and the regression head are fine-tuned, AdamW optimizer is used with a weight decay coefficient of zero point one, and a hierarchical learning rate scheduling is combined, as follows:

[0097] The pre-training layer learning rate is set to one times ten raised to the power of negative four, and the regression head learning rate is set to one times ten raised to the power of negative three. During the overall training process, the learning rate is decayed to one times ten raised to the power of negative five within an initial number of training cycles through a cosine annealing scheduler, thereby accelerating the convergence of new tasks while maintaining the stability of pre-trained features.

[0098] Meanwhile, in the regression head design, the CLS token feature of ViT is input into a double-output head fully connected layer, and the outputs correspond to distance deviation and angle deviation, respectively. The angle deviation is encoded in two channels of sine and cosine, and the output value is finally restored to an angle value through an arctan2 function, effectively avoiding the periodic jump problem existing in angle prediction.

[0099] Meanwhile, to avoid numerical overflow or not meeting the constraints of the sine and cosine functions in the output space, a double-channel tanh activation function is used to normalize and limit the predicted values of the network output. Specifically, the network outputs ŷ_cos and ŷ_sin are activated through tanh, so that their values are always limited within the interval [−1, 1]. Constraint checks are performed on the output values during training and inference to ensure that ŷ_cos² + ŷ_sin² ≤ 1, thereby ensuring the stability and physical reasonableness of the angle restoration calculation.

[0100] It should be noted that in the loss function design, a weighted mean square error of distance error and angle error is needed to balance the optimization process of distance prediction and angle prediction, and finally achieve high-precision regression of global features.

[0101] Step 3: Calculate the mask image offset and adjust the position and angle of the mask in real time;

[0102] According to the alignment deviation prediction value Y output by the Transformer model in step 2 and the image data collected in step 1, the mask image offset is calculated: ; wherein, and represents the mask image offset, represents the image data collected in step 1.

[0103] For high dynamic regions, an adaptive stepper motor control strategy is used to dynamically adjust the step length of the stepper motor according to the mask image offset to compensate for the translation error; a proportional-integral-derivative controller is used to close-loop control the rotation angle of the stepper motor to compensate for the rotation error.

[0104] For low dynamic regions, a fixed step control strategy is used to adjust the step length of the stepper motor to compensate for the translation error; a simple proportional control is used to compensate for the rotation error.

[0105] Meanwhile, in order to meet the real-time and low-power requirements of the semiconductor production line, the above-mentioned VisionTransformer regression model is deployed to the FPGA platform for hardware acceleration. Specifically, the model is quantized by the VitisAI tool chain, the attention layer uses INT8 fixed-point representation, the fully connected layer uses BF16 precision representation, and the output layer uses INT16 representation, which significantly reduces the computational complexity and power consumption of the model.

[0106] Meanwhile, in the model compression process, the attention heads in the ViT whose average attention scores are lower than the preset threshold are removed by the pruning strategy, so that the number of model parameters is reduced, and the storage and calculation pressure of the FPGA is effectively reduced.

[0107] In terms of operation scheduling, the multi-head attention calculation, patch division and matrix multiplication operation are divided into parallel sub-modules, and the clock cycle utilization rate of each calculation module is improved by using pipeline scheduling, so that the real-time inference delay of a single frame of image is reduced on the Xilinx DPU platform, and finally the real-time control requirement of the mask plate alignment of the lithography machine is met.

[0108] As shown in Figure 2 The implementation process of the mask plate alignment method of the present application is composed of a complete execution closed loop according to the above steps. First, the mask plate and wafer combined image is obtained by image acquisition and denoising, normalization and data enhancement operations are performed, forming a standardized input; then, the image is divided into several sub-regions and the stress, thickness, curing, temperature and humidity and other physical property data are extracted, and the dynamic coefficient D(x, y) is fused and constructed, and the high dynamic and low dynamic regions are divided; then, the sequence Fi is constructed combined with the image features and the dynamic coefficient, and input into the model based on the VisionTransformer structure, the global feature modeling and alignment deviation prediction are completed; finally, according to the deviation result, the differential compensation strategy is executed in different regions, and the low-delay hardware inference and control closed loop are realized on the FPGA platform through model pruning, quantization and parallel scheduling. Figure 2 The flowchart fully reflects the whole chain alignment mechanism of the present application from perception, modeling, prediction to control, which ensures the accuracy, stability and real-time performance of the method in industrial environment.

[0109] The FPGA deployment optimization

[0110] Dynamic quantization: use the Vitis AI tool chain to quantize the model to INT8, and the calibration set contains 1000 representative images.

[0111] Model pruning: remove the attention heads in the ViT whose attention scores are lower than the threshold (a < 0.1), compress the model volume by 40%.

[0112] Hardware acceleration: configure parallel computing units through Xilinx DPU, optimize matrix multiplication and Softmax operation, inference delay <10ms.

[0113] The above formulas are dimensionless values calculated, and the formulas are obtained by collecting a large amount of data to simulate the most recent real situation, and the preset parameters in the formula are set by the person skilled in the art according to the actual situation.

[0114] The above embodiments can be realized wholly or partially by software, hardware, firmware or any other combination. When realized by software, the above embodiments can be realized wholly or partially in the form of a computer program product.

[0115] Those skilled in the art can realize that the modules and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and the constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0116] In addition, each functional module in each embodiment of the present application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module.

[0117] The above is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0118] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application, should be included in the protection scope of the present application.

Claims

1. A lithography mask alignment method based on a Transformer model, The application is characterized in that it comprises: Step 1: Collecting image data of the mask and wafer combination and performing denoising, normalization and data enhancement to generate normalized distance and angle labels; Step 2: Dividing the image into multiple sub-regions, calculating dynamic coefficients and extracting features in combination with physical property data, inputting the Transformer model for global feature regression, and outputting distance deviation and angle deviation; Step 3: Adjusting the mask position and angle according to the predicted distance deviation and angle deviation, compensating through a stepping motor and implementing low-latency real-time inference on an FPGA.

2. The lithography mask alignment method based on a Transformer model according to claim 1, wherein: In step 1, image data of different mask and wafer combinations is collected, image data noise is removed, and image data is normalized.

3. The lithography mask alignment method based on a Transformer model according to claim 2, wherein: In step 1, further anti-interference data enhancement processing is performed on the denoised and normalized image data, affine transformation matrix is used to simultaneously translate, rotate and scale transform the image and the labeled key point coordinates, the key point synchronous mapping interface is used to ensure that the enhanced image and the labeled coordinates are consistent, and the elastic deformation enhancement strategy is used to apply local nonlinear disturbance to the image.

4. The lithography mask alignment method based on a Transformer model according to claim 3, wherein: In step 2, the mask and wafer area is evenly divided into multiple sub-regions; Analyzing the internal stress distribution of the mask glass in each sub-region to generate a stress distribution matrix; analyzing the thickness and curing uniformity of the photoresist layer in each sub-region to generate a thickness matrix and a curing uniformity matrix, denoted as physical property data; collecting temperature and humidity data in each sub-region to monitor air refractive index fluctuations and calculate a refractive index change matrix, denoted as air refractive index fluctuation data; Calculate the dynamic coefficient D(x,y) of each sub-region, and compare the dynamic coefficient D(x,y) of each sub-region with the dynamic coefficient threshold Yd. When the dynamic coefficient D(x,y) of a sub-region is greater than or equal to the dynamic coefficient threshold Yd, it indicates that the sub-region belongs to a high dynamic region; when the dynamic coefficient D(x,y) of a sub-region is less than the dynamic coefficient threshold Yd, it indicates that the sub-region belongs to a low dynamic region.

5. The lithography mask alignment method based on a Transformer model according to claim 4, wherein; In step 2, the feature data of the image data in each sub-region is extracted and converted into a fixed-size feature vector Xi, and the feature vector Xi of each sub-region and the dynamic coefficient D(x,y) form a sequence data Fi; Performing linear transformation on the sequence data Fi of each sub-region; calculating the correlation matrix between the sequence data of each sub-region; determining the global feature information of each sub-region by weighted fusion of the sequence data of each sub-region through the correlation matrix, and mapping the different relationship information in the global feature information between each sub-region extracted through multiple attention heads to determine the feature mapping value of each sub-region, and fusing the feature mapping values of each sub-region to determine the alignment deviation prediction value Y of the entire mask and wafer area.

6. The lithography mask alignment method based on a Transformer model according to claim 5, wherein: In step 2, after linear transformation and weighted fusion of the sequence data, the sequence data is input into a regression model based on the VisionTransformer architecture, the input image is divided into image blocks and linear mapping is performed to obtain feature vectors, and learnable position encoding is added to preserve spatial information; The regression model based on the VisionTransformer architecture includes twelve layers of encoders, each layer containing multi-head self-attention and a feedforward neural network.

7. The lithography mask alignment method based on a Transformer model according to claim 6, wherein: In step 2, the input data is divided into image blocks and global features are extracted, and the distance deviation and angle deviation are output, the angle is encoded by sine cosine and then calculated by arctangent, and different layers are trained with layered learning rate and combined with weighted mean square error.

8. The lithography mask alignment method based on a Transformer model according to claim 7, wherein: In step 3, the mask image offset is calculated, and a compensation deviation scheme is designed for high dynamic and low dynamic regions.

9. The lithography mask alignment method based on a Transformer model according to claim 8, wherein: In step 3, when implementing alignment compensation, the model is quantized and pruned, and the calculation process is scheduled in parallel pipeline.

10. The lithography mask alignment method based on a Transformer model according to claim 9, wherein: The method for obtaining the dynamic coefficient threshold value comprises the following steps: The mean and standard deviation of the physical property data and the air refractive index fluctuation data in each sub-region are calculated, and the mean and standard deviation of the physical property data and the air refractive index fluctuation data in each sub-region are analyzed by K-means clustering to determine the dynamic coefficient threshold value Yd.

Images