Photoetching hot spot detection method based on Inception block and residual network

Abstract

The invention relates to a photoetching hot spot detection method based on an Inception block and a residual network. The invention provides a data enhancement and preprocessing method, a photoetching pattern is processed firstly, a training data set is effectively expanded through the step, a data enhancement processing method is adopted, a hot spot pattern is subjected to 180-degree rotation, mirroring and 180-degree rotation and then mirroring, and the recognition capability of a model on the hot spot pattern is enhanced. On the basis, the invention provides a deep learning network adopting an Inception-Resnet neural network structure to solve the photoetching hot spot detection problem, and the neural network combines the multi-scale feature extraction capability of an Inception module and the residual connection of Resnet, so that the network can effectively solve the gradient disappearance problem while keeping the depth, and the detection accuracy is improved. And the training stability and the model learning efficiency are improved.

Description

Technical Field

[0001] The present invention relates to a lithography hotspot detection method, in particular to a lithography hotspot detection method based on an Inception block and a residual network. Background Art

[0002] With the advancement of Moore's Law, integrated circuit feature sizes have continued to shrink. Due to the limitations of photolithography technology, certain areas of the lithography process exhibit shape distortion after exposure, including bridging, pinch-offs, missing, and redundant patterns. These patterns are known as lithography hotspots. The generation of lithography hotspots is related to a variety of factors, including light radiation generated by the lithography machine during chip manufacturing and the optical proximity effect. The presence of lithography hotspots can lead to chip performance degradation or even failure, making the detection and control of lithography hotspots a crucial component of the semiconductor industry.

[0003] The photolithography hotspot detection technology was developed based on the above background. Its function is to detect the mask before manufacturing the mask and report the photolithography hotspot areas existing in the mask.

[0004] Lithography hotspot detection technology has gone through several stages of development. Lithography simulation is a traditional technology used for lithography hotspot detection. The detection steps of lithography hotspot detection based on lithography simulation are: first, import the mask and lithography model into the software, use resolution enhancement technology to optimize the mask, then set the parameters of the actual production process for the lithography model, perform lithography simulation on the optimized mask, simulate the lithography imaging under actual production conditions, and screen the layout by parameters such as the edge position error of the lithography simulation results to screen out the hotspot position. Although lithography simulation can provide an accurate calculation of lithography graphics, capture sensitive layout patterns, and has the highest detection rate for lithography hotspots, due to the gradual reduction in the feature size of current masks and the increasing complexity of the layout image, lithography simulation is very time-consuming and computationally inefficient.

[0005] To improve the efficiency of hotspot detection, a pattern-matching-based lithography hotspot detection method has emerged. This method relies on a pre-built hotspot database to identify lithography hotspots. This method first establishes a hotspot graphic library based on the shapes of known hotspots. The entire mask to be detected is then scanned, and a topological comparison is performed between all regions in the mask and the hotspots in the hotspot library. If an area in the mask has a topologically similar shape to the hotspots summarized in the hotspot library, then that area is considered a hotspot, and the graphic for that area needs to be corrected. This technology provides faster hotspot detection speeds, but its limitation is that it can only identify recorded hotspot distribution patterns and has no ability to identify unrecorded hotspot patterns.

[0006] In recent years, with the development of deep learning technology, particularly the breakthroughs achieved by deep learning, represented by convolutional neural networks, in the field of image processing, researchers have gradually turned their attention to using deep learning to detect hotspots in photolithography. Deep learning can automatically learn the characteristics of hotspots, eliminating the need for frequent manual feature extraction. Deep learning detection methods divide data into training and test sets. The training set is used to train the parameters of the network in the model. When good results are obtained, training is stopped and the final network model is retained. When testing the model, the test set is input into the final network. The model will determine whether the test set graphics are hotspots or not, and the model accuracy is measured by performance indicators. When solving practical problems, a mask template with unknown hotspots is input into the trained network model, and the model will identify the hotspot patterns within it.

[0007] As shown in Table 1, one existing technique uses a CNN (convolutional neural network) to extract image features. The CNN model consists of four convolution-pooling layers (CP) and two fully connected layers (FC). Each convolution-pooling layer includes maximum pooling and an activation function. The fully connected layer uses the hyperbolic tangent as the activation function. The CNN model is trained and calibrated using given hotspot and non-hotspot images. The model is trained using the stochastic gradient descent method to reduce the cross-entropy cost. During the hotspot detection phase, the trained CNN model scans the entire test layout using a sliding window method, evaluates the hotspot probability of each pixel, and generates a hotspot probability map. After obtaining the hotspot probability map, a clustering algorithm is used to determine the precise location of potential hotspot pixels. Finally, through technical optimization, such as network structure optimization, the performance of the CNN model is improved, enabling it to be used for hotspot detection.

[0008] Table 1 Convolutional neural network structure for lithography hotspot recognition

[0009]

[0010] However, the above technical solutions also have the following shortcomings:

[0011] Using multiple layers of convolutional kernels, the varying sizes of photomask patterns prevents comprehensive feature extraction. Using linear fully connected layers results in a loss of two-dimensional information due to the flattening operation. Furthermore, fully connected layers require a large number of training parameters, which can lead to overfitting. Typically, photomask pattern samples are relatively small, so high-complexity models can make it difficult to converge training results. Furthermore, the number of hotspot samples is typically smaller than the number of non-hotspot samples, creating data imbalance that can interfere with existing models. Summary of the Invention

[0012] In response to the difficulties of the above technical problems, the present invention proposes a lithography hotspot detection method based on Inception blocks and residual networks. This method improves the model performance by adding data enhancement and weighted cross-entropy loss function to the imbalanced lithography layout data set, and uses lithography hotspot classification based on Inception blocks and residual networks to improve the model rate and classification performance.

[0013] The present invention comprises the following steps:

[0014] Step 1: Collect images of lithography hotspots and non-hotspots as a total data set, and divide the total data set into a training data set and a test data set;

[0015] Step 2: performing data enhancement processing on the training data set, including rotating, mirroring, and rotating and then mirroring the hotspot image to expand the training data set;

[0016] Step 3: After preprocessing, the data-enhanced training dataset is input into a deep learning model based on the Inception block and the residual network for training. The deep learning model uses the Inception block to extract multi-scale features and uses residual connections to solve the gradient vanishing problem. During the training process, a weighted cross-entropy loss function is used for back propagation to enhance the ability to handle unbalanced datasets. At the same time, the Dropout mechanism and early stopping mechanism are introduced to prevent the model from overfitting.

[0017] Step 4: Input the test dataset into the trained model, disable the Dropout mechanism and early stopping mechanism, and obtain the lithography hotspot classification results output by the deep learning model.

[0018] Beneficial effects of the present invention:

[0019] This paper proposes a data augmentation and preprocessing method. This first processes the lithographic pattern, effectively expanding the training dataset. Using data augmentation, the hotspot pattern is rotated 180 degrees, mirrored, and then rotated 180 degrees again, enhancing the model's ability to recognize hotspot patterns. The rotation angle used here is only 180 degrees. It's important to note that rotations of 90 and 270 degrees can alter the properties of hotspots and non-hotspots in the lithographic image, leading to errors. Most data augmentation methods fail to address this issue.

[0020] On this basis, the present invention proposes a deep learning network using the Inception-Resnet neural network structure to solve the problem of lithography hotspot detection. This neural network combines the multi-scale feature extraction capability of the Inception module and the residual connection of Resnet, which enables the network to effectively solve the gradient vanishing problem while maintaining depth, thereby improving the training stability and learning efficiency of the model.

[0021] The present invention introduces the Dropout and early stopping training mechanism during the training process to reduce the probability of overfitting of the model and improve the generalization performance of the model. During each round of training, Dropout regularization is used to randomly pause the weight learning of some neurons each time to avoid overfitting. The early stopping mechanism records the training loss after each round of training. If the loss does not decrease after several rounds of training with the minimum loss function, the model is considered to have completed learning and training is stopped to prevent overtraining from causing the model to overfit the training data set.

[0022] The present invention adopts global average pooling to retain the integrity of image information to the greatest extent possible, thereby improving the accuracy of the model in determining hotspot graphics.

[0023] The present invention adopts the weighted cross entropy loss function for reverse gradient propagation. By assigning different weights to samples of different categories, the model can pay more attention to categories with fewer samples, thereby improving the model's recognition accuracy for minority classes. In practical applications, the weighted cross entropy loss function can better guide model training and improve the performance of the model on unbalanced datasets, thereby further solving the problem of unbalanced datasets. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Flow chart of the method of the present invention;

[0025] Figure 2 Inception block framework for multi-scale feature extraction;

[0026] Figure 3 Schematic diagram of ordinary network and residual network;

[0027] Figure 4 The specific framework of the model. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0029] This paper proposes a deep learning-based method for detecting hotspots in lithography. This method uses an image classification model based on Inception blocks and a residual network. The method is particularly suitable for classifying hotspots from non-hotspots based on frequency-domain image features. This method combines custom data preprocessing, the Inception network architecture, residual connections, a dropout mechanism, and an early stopping mechanism to improve the efficiency and accuracy of image classification tasks.

[0030] Furthermore, the present invention proposes a method for data enhancement and preprocessing, first processing the lithography graphics, and effectively expanding the training data set through this step. On this basis, the present invention proposes a deep learning network using the Inception-Resnet neural network structure to solve the problem of lithography hotspot detection. The neural network combines the multi-scale feature extraction capability of the Inception module and the residual connection of Resnet, which enables the network to effectively solve the gradient vanishing problem while maintaining depth, thereby improving the training stability and the accuracy of hotspot detection; the present invention adopts a weighted cross entropy loss function as the training loss function, thereby enhancing the model's processing capability for unbalanced data sets; and introduces the Dropout and early stopping training mechanism during the training process to reduce the probability of overfitting of the model and improve the generalization performance of the model; the present invention adopts global average pooling to retain the integrity of image information to the greatest extent, thereby improving the accuracy of the model in judging hotspot graphics.

[0031] During the training phase, data augmentation is used to rotate the hotspot image 180 degrees, mirror it, and then rotate it 180 degrees again to enhance the model's ability to recognize hotspot patterns. A sample of a photolithography mask image with either hotspot or non-hotspot labels is fed into the model. The image is first transformed using the DCT to obtain its spectral features. These spectral features are then fed into the first Inception block, which consists of parallel convolution kernels of varying sizes. A residual connection is introduced, and the sum of the input and output of the Inception block serves as the output of this portion of the ResNet network. This process is then passed through two Inception-ResNet networks, using convolution kernels of varying sizes to comprehensively learn image information. Residual connections propagate the input forward faster through inter-layer connections, addressing the vanishing gradient problem.

[0032] During training, the Dropout mechanism is used to randomly stop a certain proportion of neurons from training each time to prevent the model from overfitting and causing insufficient generalization ability. After each round of training, the model is tested on the test set, and the best validation set accuracy so far is recorded. When the best accuracy is not achieved for a certain number of consecutive times, the training is stopped to prevent the model from overfitting the training set data.

[0033] During the non-training phase, after performing a DCT transformation on the input photolithography mask image, the model's Dropout and early stopping mechanisms are disabled, and the trained weighted network is input. The network extracts image features from multiple angles, makes judgments, and predicts whether the input image is a photolithography hotspot.

[0034] Based on the above concept, Figure 1 As shown, the method of the present invention specifically includes:

[0035] Step 1: Collect images of lithography hotspots and non-hotspots as a total dataset, and divide the total dataset into a training dataset and a test dataset.

[0036] Step 2: The training dataset is augmented and used as the input of the Inception-Resnet network model. The data augmentation method includes rotating the hotspot image 180 degrees, mirroring it, and rotating it 180 degrees and then mirroring it.

[0037] Step 3: The augmented dataset is processed and fed into a pre-trained Inception block and residual network for training. This allows the network to learn features across multiple scales. The residual network, by introducing short-circuit connections, can mitigate the vanishing gradient problem. A loss function is used to perform gradient backpropagation to update network parameters and achieve network convergence.

[0038] Step 4: Put the test data set from step 1 into the weighted network trained in steps 2 and 3, turn off the network's Dropout and early stopping mechanisms, and obtain the network output lithography hotspot classification results.

[0039] Furthermore, in the data enhancement operation in step 2, the rotation angle is only 180 degrees. It should be noted that the rotation operations of 90 degrees and 270 degrees may cause the hotspot and non-hotspot properties of the lithography image to change, thereby causing errors.

[0040] Furthermore, the dataset in step 3 is discrete cosine transformed and then input into the Inception block and residual network. The discrete cosine transform formula is as follows:

[0041]

[0042] Where N1 is the number of pixels in the image width direction, N2 is the number of pixels in the image height direction, k1, k2 and n1, n2 are the index numbers of the frequency domain and the local area of the original image, respectively. After compressing the image, a discrete cosine transform is performed to obtain its spectrum. Each expansion term is calculated by taking the first several expansion coefficients to form the input matrix. and It is the row and column scaling factor normalization factor (or scaling factor), which can make DCT and inverse DCT maintain orthogonality and unit energy.

[0043] Furthermore, during training, step three, the input data is extracted through an Inception block and a residual connection network. Multi-scale feature extraction is achieved by stacking two Inception blocks. An improved Inception structure is employed, combined with residual connections to enhance performance. Each module extracts features from different receptive fields through a multi-branch design (including 1×1 convolution, 1×3 and 3×1 convolution, and max pooling). The outputs are then integrated through channel concatenation. Residual connections are introduced to directly transmit inputs to mitigate the vanishing gradient problem. From the shallow Inception1 layer to the deep Inception2 layer, the network gradually increases the number of feature channels, improving feature expression capabilities and achieving efficient feature extraction and network optimization.

[0044] Furthermore, during training in step 3, Dropout and early stopping are used. Dropout regularization randomly pauses the learning of some neuron weights during each training round to prevent overfitting. The early stopping mechanism records the training loss after each round of training. If the loss does not decrease after several rounds of training that minimize the loss function, the model is considered to have learned enough and training is stopped. This prevents overtraining and overfitting of the model to the training dataset.

[0045] Furthermore, in step 3, a weighted cross entropy loss function is used for back propagation of gradients. This loss function increases the penalty for misjudgment of fewer categories, thereby solving the problem of unbalanced data sets. The weighted cross entropy loss function is defined as follows:

[0046]

[0047] Where ln is a one-dimensional vector, x is the input, y is the target, c is the number of categories, and w c is the weight.

[0048] Furthermore, the model also includes evaluation metrics, using a confusion matrix as a criterion for lithography hotspot detection: True Positives (TP): the number of positive samples predicted as positive; True Negatives (TN): the number of negative samples predicted as negative; False Positives (FP): the number of negative samples predicted as positive; and False Negatives (FN): the number of positive samples predicted as negative. In lithography hotspot detection, positive samples are hotspots.

[0049] Based on the above confusion matrix concept, the accuracy is the ratio of the number of hotspot samples correctly predicted by the model to the number of actual hotspot samples, which represents the precision rate of the lithography hotspot detection model in detecting lithography hotspots.

[0050]

[0051] The number of false positives is the number of samples that the model identifies as hotspots but are actually non-hotspots, that is, the number of FPs.

[0052] The recall rate R represents the completeness of the lithography hotspot detection model in detecting lithography hotspots. The higher the recall rate, the fewer lithography hotspots missed by the model.

[0053]

[0054] The F1 value represents the harmonic mean of the precision and recall of the lithography hotspot detection model. Generally, precision and recall are conflicting metrics for lithography hotspot detection models; increasing one directly decreases the other. A higher F1 value indicates that the hotspot detection model strikes a better balance between false positives and missed detections.

[0055]

[0056] Test Time (TT) indicates the time it takes to detect a lithography hotspot. The unit is seconds. The shorter the test time, the more efficient the detection model.

[0057] Example:

[0058] This embodiment includes the following steps:

[0059] Step 1: Using the ICCAD 2012 dataset as the master dataset, we collected images of lithography hotspots and non-hotspots as the master dataset, and divided the master dataset into a training dataset and a test dataset. Table 2 shows the specific number of hotspots and non-hotspots.

[0060] Table 2 Dataset statistics

[0061]

[0062] Step 2: The training dataset is subjected to data augmentation processing and used as the input of the model. The data augmentation method includes three operations: rotating the hotspot image by 180 degrees, mirroring it, and rotating it by 180 degrees and then mirroring it.

[0063] Step 3: The data set after data augmentation is preprocessed and compressed through discrete cosine transform to achieve data simplification and improve training efficiency. The processed data is input into the Inception block and residual network for information learning.

[0064] Furthermore, in this embodiment, the input image size is 1200×1200, and after discrete cosine transform, the coefficients of the first 12 harmonics are taken, and its data format becomes 12×12×32; further, in this embodiment, the training data set is randomly shuffled during training, and in one round of training, 128 images are extracted from all training data sets according to the training batch as input data, and are input into the training network one by one for training.

[0065] In step three, the processed data is further input into the pre-trained Inception block and residual network for training, so that the network can learn features in multiple scale ranges. Multi-scale feature extraction is achieved through the Inception block.

[0066] In this embodiment, the Inception block consists of four parallel convolution-pooling branches, combining different convolutional layers in parallel. The resulting matrices from the different convolutional layers are concatenated along the depth dimension to form a deeper matrix. The advantage of the Inception module is that it can initially reduce the dimensionality of larger matrices while simultaneously aggregating visual information at different scales, facilitating feature extraction at different scales.

[0067] Among them: the first convolution-pooling branch first performs maximum pooling, and then convolves the pooling result with a convolution kernel size of 1×1; the second convolution-pooling branch first performs convolution with a convolution kernel size of 1×1, and then performs maximum pooling on the convolution result; the third convolution-pooling branch first performs three convolutions with convolution kernel sizes of 1×1, 1×3, and 3×1, and then performs maximum pooling on the convolution result; the third convolution-pooling branch first performs three convolutions with convolution kernel sizes of 1×1, 1×5, and 5×1, and then performs maximum pooling on the convolution result; finally, the results of the four convolution branches are superimposed in the channel dimension as the output of the Inception block to improve the feature expression ability, thereby achieving efficient feature extraction and network optimization. The framework of the Inception block is as follows Figure 2 shown.

[0068] In deep learning, it is generally believed that deeper networks are generally more expressive and can learn more complex features. However, there are hidden training problems and possible network degradation in deep networks, so network optimization is reasonable. The introduction of residuals can enable the network to maintain its depth while having the advantages of shallow networks and can effectively avoid network degradation. Figure 3 Schematic diagram of ordinary network and residual network.

[0069] Introducing residual connections into the network, the input of the Inception block is convolved with a convolution kernel size of 1×1, and then directly added to the output of the Inception block, so that the data is directly transmitted in the network channel, reducing the number of parameters, improving the training speed, avoiding gradient explosion and model convergence problems. The specific framework of the deep learning model based on the Inception block and the residual network is as follows: Figure 4 .

[0070] Furthermore, a loss function is used to backpropagate the gradients of the network parameters to achieve network convergence. A weighted cross-entropy loss function is used for backpropagation of the gradients. This loss function increases the penalty for misclassification of a small number of categories, thereby solving the problem of unbalanced datasets.

[0071] Furthermore, during training, the batch size is set to 128, and the initial setting is 100 rounds of training in total. The Adam optimizer is used, and the initial learning rate is set to 0.001. An early stopping mechanism is adopted to record the training loss after each round of model training. If the loss does not decrease after several rounds of training with the minimum loss function, the model is considered to have completed learning and training is stopped. In this embodiment, a value of 10 is selected. If the loss does not decrease after 10 rounds of training, the model is considered to have completed learning. In practice, if higher accuracy is sought, the value can be set to a larger value, but a larger value will bring about the problem of slow training. The early stopping mechanism records the training loss after each round of training to prevent over-training from causing the model to overfit the training data set.

[0072] Furthermore, Dropout regularization is used to randomly pause the weight learning of some neurons during each round of training to avoid overfitting.

[0073] Step 4: Apply the test dataset from Step 1 to the weighted network trained in Steps 2 and 3. Disable the network's dropout and early stopping mechanisms to obtain the network's output for lithography hotspot classification. Furthermore, after training and testing, analyze the results using performance metrics to evaluate network performance.

[0074] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0075] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A lithography hotspot detection method based on Inception blocks and residual networks, characterized in that: The following steps are involved: Step 1: Collect images of lithography hotspots and non-hotspots as a total data set, and divide the total data set into a training data set and a test data set; Step 2: performing data enhancement processing on the training data set, including rotating, mirroring, and rotating and then mirroring the hotspot image to expand the training data set; Step 3: After preprocessing, the data-enhanced training dataset is input into a deep learning model based on the Inception block and the residual network for training. The deep learning model uses the Inception block to extract multi-scale features and uses residual connections to solve the gradient vanishing problem. During the training process, a weighted cross-entropy loss function is used for back propagation to enhance the ability to handle unbalanced datasets. At the same time, the Dropout mechanism and early stopping mechanism are introduced to prevent the model from overfitting. Step 4: Input the test dataset into the trained model, disable the Dropout mechanism and early stopping mechanism, and obtain the lithography hotspot classification results output by the deep learning model.

2. The lithography hotspot detection method based on Inception blocks and residual networks according to claim 1, characterized in that: In the data enhancement process, the rotation angle of the hotspot image is 180 degrees to avoid changes in the properties of the lithography image caused by rotations of 90 degrees and 270 degrees.

3. The lithography hotspot detection method based on Inception blocks and residual networks according to claim 1 or 2, characterized in that: The preprocessing includes performing discrete cosine transform on the image to extract the frequency spectrum features of the image.

4. The lithography hotspot detection method based on Inception blocks and residual networks according to claim 3 is characterized in that: The Inception block consists of four parallel convolution-pooling branches. The first convolution-pooling branch first performs maximum pooling and then convolves the pooling result with a convolution kernel size of 1×1. The second convolution-pooling branch first performs convolution with a convolution kernel size of 1×1 and then performs maximum pooling on the convolution result. The third convolution-pooling branch first performs three convolutions with convolution kernel sizes of 1×1, 1×3, and 3×1 and then performs maximum pooling on the convolution result. The third convolution-pooling branch first performs three convolutions with convolution kernel sizes of 1×1, 1×5, and 5×1 and then performs maximum pooling on the convolution result. Finally, the results of the four convolution branches are superimposed in the channel dimension as the output of the Inception block. The output is integrated by channel splicing and a residual connection is introduced. The input of the Inception block is convolved with a convolution kernel size of 1×1 and then directly matrix-joined with the output of the Inception block to alleviate the gradient vanishing problem.

5. The lithography hotspot detection method based on Inception block and residual network according to claim 1 is characterized in that: The weighted cross entropy loss function is used for back propagation of gradients. This loss function increases the penalty for misjudgment of fewer categories and alleviates the problem of imbalanced data sets. The weighted cross entropy loss function is defined as: where l n is a one-dimensional vector, x is the input, y is the target, c is the number of types, w c is the weight.

6. The lithography hotspot detection method based on Inception blocks and residual networks according to claim 1 or 5, characterized in that: During the training process, global average pooling is used to retain the integrity of image information to the greatest extent possible, thereby improving the accuracy of the model in determining hotspot graphics.

7. The lithography hotspot detection method based on Inception blocks and residual networks according to claim 6, characterized in that: In the Dropout mechanism, the weight learning of some neurons is randomly suspended each time, and the suspension ratio is 10% to 50%.

8. The lithography hotspot detection method based on Inception blocks and residual networks according to claim 6, characterized in that: In the early stopping mechanism, the early stopping mechanism records the training loss after each round of training. If the loss function does not decrease after several consecutive rounds of training, the training is stopped to prevent the model from overfitting.

9. The lithography hotspot detection method based on Inception blocks and residual networks according to claim 1, characterized in that: The model was trained using the Adam optimizer with an initial learning rate set to 0.001, and the neural network weights were iteratively updated based on the training data.

10. The lithography hotspot detection method based on Inception block and residual network according to claim 1, characterized in that: After the test data set is input into the model, the network outputs the lithography hotspot classification results, and the network's excellence is evaluated by benchmarking with defined performance indicators; The defined performance indicators include accuracy, false alarms, recall, F1 value and detection time; the accuracy is the ratio of the number of hotspot samples correctly predicted by the model to the number of actual hotspot samples, which represents the precision rate of the lithography hotspot detection model in detecting lithography hotspots; the false alarms are the number of samples identified as hotspots by the model but are actually non-hotspots; the recall rate represents the recall rate of the lithography hotspot detection model in detecting lithography hotspots; the F1 value represents the harmonic mean of the precision and recall rates of the lithography hotspot detection model; and the detection time represents the detection time of the lithography hotspot detection model.

Images