Photoetching hot spot detection method and device based on deep learning

Abstract

The invention discloses a photoetching hot spot detection method and device based on deep learning, and the method comprises the steps: collecting a circuit layout image, and constructing a sample data set containing photoetching layout hot spots and having no hot spots; performing mixed data enhancement on the hotspot samples by using a generative adversarial network; training a hotspot detection network by using the improved data set; and inputting the to-be-detected layout into the trained hotspot detection network to obtain a hotspot detection result. According to the generative adversarial network, a hybrid data enhancement method is provided for generating more photoetching hot spot layout samples, the problem that the number distribution of the samples is seriously unbalanced is effectively solved, and the training stability of the GAN network and the capability of generating high-quality auxiliary samples with complete features are improved; the improved hotspot detection network reduces the calculation amount, and can solve the problem that the model is difficult to optimize due to overfitting and gradient discretization under the conditions of excessive model parameters and limited training data sets.

Description

Technical Field

[0001] The present invention relates to the field of photolithography hotspot detection, and in particular to a photolithography hotspot detection method and device based on deep learning. Background Art

[0002] In semiconductor manufacturing, photolithography is a key technology for pattern transfer, and its control accuracy and resolution directly determine the yield of semiconductors. Among the many factors that affect the photolithography process, the quality of the optical devices of the photolithography machine plays a decisive role, and the quality of imaging is directly determined by the optical devices. During the photolithography process, due to the influence of the optical proximity effect, the mask pattern will be distorted when transferred to the surface of the silicon wafer, such as: rounded corners, shortened line ends, line width deviation, and missing parts of the area. The circuit failure area generated by the mask layout after photolithography is called a photolithography hotspot. Hotspots in the photolithography layout can cause defects such as circuit short circuits, open circuits, and poor ohmic contact, which seriously affect the electrical characteristics of integrated circuits and the product yield of integrated circuit board manufacturing.

[0003] In order to improve the quality of integrated circuit manufacturing, the industry has proposed using resolution enhancement technology to repair hotspot layouts, such as sub-resolution assisted imaging technology and optical proximity correction technology. This technology can improve the quality of the circuit board after imaging, but it cannot completely eliminate the hot spots in the layout. Therefore, it is necessary to perform hot spot detection on the photolithography layout before the circuit layout is burned onto the silicon wafer. The complexity of integrated circuits is growing in a Moore's Law-like manner. Hot spot detection faces problems such as an imbalance in the ratio of hot spot samples to non-hot spot samples, complex hot spot patterns, and high false alarm rates. Hot spot detection on the photolithography layout has become a huge challenge in the field of integrated circuit manufacturing.

[0004] Lithography hotspot detection can be defined as the precise location of lithography hotspot areas in the layout within an acceptable time frame. As an important part of integrated circuit manufacturability design, the research on lithography hotspot detection has attracted widespread attention. Existing lithography hotspot detection methods can be roughly divided into three types: hotspot detection based on lithography simulation, hotspot detection based on pattern matching, and hotspot detection based on machine learning.

[0005] In order to shorten the feedback time of lithography hotspot detection on layout design, machine learning has been applied to lithography hotspot detection. Hotspot detection based on machine learning defines lithography hotspot detection as a classification problem. In the training phase, a large number of lithography hotspot and non-lithography hotspot layout samples are used as training sets and the feature vectors of each layout are extracted as the input of the machine learning model. The lithography hotspot classifier is trained through supervised learning. In the testing phase, the trained classifier can effectively determine whether the layout is a hotspot and locate the position with a high probability of belonging to the lithography hotspot area in the published drawing.

[0006] However, machine learning methods require manual extraction of image features, and the effectiveness of manual feature extraction needs further exploration and verification. As a subcategory of end-to-end, deep learning has excellent learning capabilities. Compared with machine learning methods, deep learning eliminates the tedious manual feature extraction process and can automatically extract image features using convolutional neural networks. Deep learning can effectively improve feature learning capabilities through massive data-driven methods, achieve accurate fitting of complex functional relationships, and describe rich information about data content.

[0007] Although the use of GAN models to generate data samples has the advantages of simple methods, unsupervised autonomous learning, and high quality of generated samples, it also has problems such as gradient vanishing, unstable training, and mode collapse, and is not suitable for processing discrete data. Among them, mode collapse is the phenomenon that GAN generates unstable samples with poor quality during training, that is, the auxiliary samples generated by the generator are not real, but the discriminator gives a correct evaluation, and the generator continues to generate auxiliary samples similar to this sample, resulting in mutual deception between the two, resulting in the loss of features of the final generated auxiliary samples and incomplete information. Summary of the invention

[0008] The purpose of the present invention is to address the deficiencies in the prior art and to propose a lithography hotspot detection method and device based on deep learning.

[0009] The objective of the present invention is achieved through the following technical solution: a lithography hotspot detection method based on deep learning, the method comprising the following steps:

[0010] S1, collect circuit layout images and construct a sample data set containing hot spots and no hot spots in the lithography layout;

[0011] S2, use the generative adversarial network to perform hybrid data enhancement on hotspot samples;

[0012] S3, use the improved dataset to train the hotspot detection network;

[0013] S4. Input the to-be-detected layout into the trained hotspot detection network to obtain the hotspot detection result.

[0014] Furthermore, S1 specifically includes: using an optical nano-level camera to collect an integrated circuit layout image, segmenting the acquired image, and generating a sample data set including hot spots and no hot spots in the lithography layout.

[0015] Furthermore, the hybrid data enhancement includes: processing layout samples containing lithography hotspots through geometric transformation; constructing a hybrid loss based on Wassersteun loss and binary cross entropy loss, and optimizing a generative adversarial network model to enhance the generated sample data.

[0016] Further, the processing of the layout samples containing the lithography hotspots by geometric transformation includes: increasing the number of samples by using two methods, rotation and mirror transformation;

[0017] Furthermore, the hybrid loss constructed based on Wassersteun loss and binary cross entropy loss includes:

[0018] =

[0019]

[0020]

[0021] Among them, G represents the generator, D represents the discriminator, represents the real sample, represents the generator input, It means the generator generates samples, and E means mathematical expectation; is the unknown sample, N is the total number of samples, represent belongs to the true distribution probability, represent belongs to the probability of generating distribution; λ1 and λ2 are the coefficients of binary cross entropy and Wasserstein loss.

[0022] Furthermore, the hotspot detection network is a GoogLeNet model, which is composed of two 3×3 convolution blocks, two 3×3 maximum pooling blocks, two 7×7 convolution blocks, four improved Inception modules, a 7×7 maximum pooling layer, a fully connected layer, and a soft-max layer connected in sequence in the order shown in the figure.

[0023] Furthermore, the improved Inception includes four main branches, the first main branch includes only 1×1 convolution, the second main branch includes a trunk composed of 1×1 convolution and 3×3 convolution connected in sequence, and two sub-branches of 1×3 convolution and 3×1 convolution are connected after the 3×3 convolution; the third main branch includes a trunk composed of 1×1 convolution and two sub-branches connected after the 1×1 convolution, each of which includes 1×3 convolution; the fourth main branch includes 3×3 convolution and 1×1 convolution connected in sequence; the data of all channels are concatenated and combined for the next layer input.

[0024] According to another aspect of the specification, a deep learning-based lithography hotspot detection device is also provided, including a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the executable code, the deep learning-based lithography hotspot detection method is implemented.

[0025] According to another aspect of the specification, a computer-readable storage medium is also provided, on which a program is stored. When the program is executed by a processor, the method for detecting lithography hotspots based on deep learning is implemented.

[0026] Beneficial effects of the present invention:

[0027] 1. Based on the GAN network, a hybrid data enhancement method HDAM is proposed to generate more lithography hotspot layout samples, effectively solving the problem of severe imbalance in the number of samples, improving the training stability of the GAN network and the ability to generate complete features and high-quality auxiliary samples;

[0028] 2. The GoogLeNet deep learning model pre-trained with a large data set effectively solves the problem of too many model parameters and easy overfitting when the training data set is limited; it reduces the amount of calculation brought by the network structure, which is convenient for practical application; it is easier to optimize and solves the problem of difficulty in optimizing the model due to gradient discreteness. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A flow chart of a lithography hotspot detection method based on deep learning of HDAM and improved GoogleNet model provided in an embodiment of the present invention;

[0030] Figure 2 A schematic diagram of a test data set sample provided by an embodiment of the present invention;

[0031] Figure 3 HDAM text flow chart provided for an embodiment of the present invention;

[0032] Figure 4 A diagram of a generator network structure of a GAN provided in an embodiment of the present invention;

[0033] Figure 5 A schematic diagram of the GoogLeNet model structure provided by an embodiment of the present invention;

[0034] Figure 6 A network structure diagram of the improved Inception module provided in an embodiment of the present invention;

[0035] Figure 7 A schematic diagram of a lithography hotspot detection device based on deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0037] like Figure 1 As shown, the present invention provides a lithography hotspot detection method based on deep learning, comprising:

[0038] S1: Collect integrated circuit layout images through an optical nano-level camera, segment the acquired images, and generate a sample data set containing hot spots and no hot spots in the lithography layout;

[0039] S2: The HDAM method based on the GAN network is used to enhance the data of the hotspot samples of the lithography pattern, reducing the distribution difference of the number of samples under different categories;

[0040] S3: Process the samples in S2 and divide them into training sample set and test sample set according to a certain ratio;

[0041] S4: Input the hotspot map sample set trained in S3 into the improved GoogLeNet model for training. The training method is to divide the training samples into a training sample set and a test sample set according to a certain ratio. The test indicators are recall rate, accuracy rate, and F1 score.

[0042] Recall rate indicates the ratio of the number of correctly predicted hotspot maps to the actual positive examples:

[0043] Recall =

[0044] The accuracy rate indicates the ratio of the actual hotspot map to the predicted hotspot map:

[0045] Precision =

[0046] However, using only accuracy or recall cannot evaluate the performance of the algorithm model well. The F1 score is needed to measure the overall performance of the model:

[0047] F1 =

[0048] TP, FP, and FN represent hotspots that are correctly identified, non-hotspots that are incorrectly identified as hotspots, and hotspots that are identified as non-hotspots. The training set and test set of the two categories are marked as and , where TR represents the training set, TE represents the test set, the input x size is 224×224, and the output is y∈[0,1]. The formula for constructing the GoogleNet mathematical model is as follows:

[0049] (1) Relationship between model input and output:

[0050]

[0051] in is the input image, X is the feature representation of the input image or the feature map between levels, This corresponds to the entire mapping operation process from input image to feature map extraction, W is the convolution kernel, b is the bias, is a hyperparameter.

[0052] (2) The classifier parameters are shown in Formula 3-9. The final output category is Y=argmax{Y(0),y(1)}, that is, the output indicates that the input image has a higher probability of belonging to the hotspot or non-hotspot.

[0053]

[0054] (3) Loss Function Given by:

[0055]

[0056] Among them, R(W) and R(θ) represent regularization terms, which sparse the parameters to prevent overfitting. represents the L2 norm, , , are all hyperparameter weights, , It represents the weight corresponding to whether the input image belongs to a hotspot or a non-hotspot. Represents the total number of input-output pairs in the training set.

[0057] S5: Input the hotspot layout sample set tested in S3 into the trained GoogLeNet model in S4 to perform hotspot detection on the lithography layout. The test sample set is the integrated circuit layout design data set of the ICCAD 2012 competition. Its benchmark has a total of 5 design layouts with hotspots and no hotspots in GDSII format. The original layout data format consists of a continuous list of vertex coordinates, so these vertex coordinates need to be encoded into a two-dimensional density image. The design sample diagram is shown in Figure 2 Given, Figure 2 (a) in the figure is the hotspot layout design. Figure 2 (b) is a layout design without hotspots. The detailed information of each design benchmark is shown in the following table;

[0058] S6: Output and analyze whether the layout contains hot spots.

[0059] Table 1: Detailed information on the design benchmarks for each training and test dataset

[0060]

[0061] Specifically, in step S2, since the layouts containing lithography hotspots only account for 5% of all layouts, the sample quantity distribution is seriously unbalanced. In order to improve the training stability of the GAN network and the ability to generate complete and high-quality auxiliary samples, a hybrid data augmentation method (HDAM) is proposed based on the GAN network to generate more lithography hotspot layout samples, and an indicator evaluation system is constructed to analyze the quality of the generated layout samples.

[0062] like Figure 3 As shown, the steps of the hybrid data enhancement method in the S2 process are as follows:

[0063] S21: Processing layout samples containing lithography hotspots through geometric transformations such as rotation and translation;

[0064] S22: A hybrid loss is constructed based on Wassersteun loss and binary cross entropy loss to optimize the original GAN ​​model to enhance the generated sample data.

[0065] Its specific structure is as follows:

[0066] (1) Geometric transformation

[0067] In step S21, considering that the auxiliary samples generated by a single data enhancement method are repetitive, it is necessary to use a hybrid data enhancement method for data enhancement. In the field of image processing, commonly used methods include rotation, brightness adjustment, adding random noise, etc. In order to ensure that the image generated after processing maintains the original features, this paper only uses two methods, rotation and mirror transformation, to increase the number of samples. The general expression is:

[0068]

[0069] Among them, the coefficient matrix A is the rotation matrix coefficient, θ represents the rotation angle, and its value is {π / 2, π, 3π / 2} to prevent edge feature loss. The coefficient matrix B is the mirror coefficient matrix, which rotates about the mirror line ax+by+c=0. In the formula, RN represents random noise, P represents the original sample, and P' represents the generated sample.

[0070] (2) Generator network and discriminator network structure

[0071] The generator structure is as follows Figure 4As shown in the figure, assuming that the goal of training the generator G is to generate images of size 224×224×3, the input of the generator is 100-dimensional random noise that follows a Gaussian distribution, with a standard deviation of 1 and a mean of 0. First, the random noise is projected and reshaped into a feature map of size 7×7×512. These feature blocks are then converted into a convolutional representation of size 224×224×3 through 5 deconvolution blocks. Each convolutional block includes a deconvolution layer, a batch normalization layer (denoted as BN in the figure), and a PReLU activation function layer.

[0072] The discriminator extracts features through convolution operations, reduces the dimension through pooling operations, and then randomly discards a certain proportion of neurons (set to 0.1 in this invention) through the dropout layer to prevent the model from overfitting. Finally, the fully connected layer obtains the predicted probability value by mapping the output to the sample label space.

[0073] (3) Loss Function

[0074] In step S22, when the discriminator is optimal, the loss of the generator in the generative adversarial network can be equivalently converted to minimizing the Jensen-Shannon (JS) divergence between the real sample and the generated auxiliary sample. As an improvement of the KL (Kullback–Leibler divergence) divergence, the JS divergence can overcome the limitations of the asymmetric KL divergence and has the ability to characterize the distance. Its calculation formula is as follows:

[0075]

[0076] Among them, P r Represents the true sample probability distribution, P g represents the probability of generating samples, D KL (∙) represents the KL divergence between the two, separated by the || symbol. Usually, P r and P g The overlap between them is negligible, so the JS divergence is approximately a constant, which will lead to the problem of gradient vanishing when training the generator model. The Wasserstein distance has approximate differentiability and good smoothness, which can better characterize the difference between the two distributions. It can be expressed as P g Converges to the true distribution P r The minimum "cost":

[0077]

[0078] Among them, inf represents the lower limit of the threshold, γ represents the joint distribution of the true sample probability and the generated sample probability, S(,) represents all possible joint distributions of the two, and γ(x, y) represents the mapping from x to y so that P g and P r The loss required to obey the same distribution. The above formula can be transformed into a solvable Wasserstein loss according to the Kantorovich-Rubinstein criterion:

[0079]

[0080]

[0081] Among them, G represents the generator, D represents the discriminator, represents the real sample, represents the generator input, Represents the sample generated by the generator, and E represents the mathematical expectation. L represents the set of 1-Lipschitz functions. Since Wasserstein loss is a critical function, the generator is easier to optimize and can effectively avoid the gradient vanishing phenomenon. Therefore, in order to make the training process of the constructed model more stable and easy to converge, the present invention proposes a binary cross entropy loss A hybrid optimization objective combined with Wasserstein loss, namely:

[0082]

[0083] =

[0084] in, is the unknown sample, N is the total number of samples, represent belongs to the true distribution probability, represent Belongs to the probability of generating distribution. λ1 and λ2 are the coefficients of binary cross entropy and Wasserstein loss, which are set to 1 and 0.01 respectively. That is, in the early stage of model training, the former plays a leading role and guides the model gradient update; when the gradient disappears, the latter plays an auxiliary correction role.

[0085] Specifically, in step S3, since the size of the sample set is relatively large, i.e., 1200×1200×1, the image is compressed to 224×224×1 by a linear interpolation method, and then the single channel is copied to a sample set of 224×224×3.

[0086] Specifically, in step S4, the pre-trained GoogLeNet model is used to detect lithography hotspots. The structure is as follows: Figure 5 As shown in the figure, two 3×3 convolution blocks, two 3×3 maximum pooling blocks, two 7×7 convolution blocks, four Inception modules, a 7×7 maximum pooling layer, a fully connected layer, and a soft-max layer are connected in sequence in the figure. Each convolution block contains a convolution layer, a Relu activation function layer, and a batch normalization layer. The core of GoogLeNet is the Inception module, which is used to extract rich and delicate image features. Multiple Inception modules are superimposed to form a densely connected network, the purpose is to extract as many rich features as possible at different levels and improve the feature extraction effect. The features extracted by the Inception module are further compressed by the 7×7 maximum pooling layer, while retaining the main information. The fully connected layer extracts the pooled compressed results to determine whether it is a hotspot (output dual-channel data). The soft-max layer then approximates the output of the fully connected layer with an s-shaped function to normalize it to prevent instability. Finally, the hotspot judgment result of the map is output according to the normalized value. The original Inception module uses four branches, each of which is composed of 1×1 convolution, 3×3 convolution, 5×5 convolution and 3×3 maximum pooling layers, which significantly expands the width of the network layer domain and increases the number of neuron units at each level. At the same time, the small-size convolution kernel can overcome the problem of increased algorithm complexity caused by the deepening of the traditional neural network. At the same time, the step size is used to ensure that the feature dimensions of the four branches are the same, and then superimposed to achieve the purpose of multi-scale feature extraction. However, 3×3 convolution and 5×5 convolution still bring a large amount of calculation, so 1×1 convolution is used to reduce the dimension of the feature map, that is, a 1×1 convolution is added before the 3×3 convolution and 5×5 convolution to reduce the calculation bottleneck. At the same time, the transformation increases the number of network layers and improves the network's expressiveness. In order to further improve the detection accuracy and compress the calculation time, with the help of the idea of ​​the VGG network, the calculation efficiency is improved by reducing the large-size convolution kernel, and the 5×5 convolution is replaced by two 3×3 convolutions connected in sequence to reduce the number of parameters and alleviate the overfitting phenomenon. On the basis of replacing 5×5 with 3×3, 1×3 convolution and 3×1 convolution are used to replace 3×3 convolution, and finally the data of all channels are spliced ​​and merged for the next layer input.

[0087] Improved Inception module such as Figure 6As shown in the figure, the improved Inception includes four main branches, the first main branch includes only 1×1 convolution, the second main branch includes a trunk composed of 1×1 convolution and 3×3 convolution connected in sequence, and two sub-branches of 1×3 convolution and 3×1 convolution are connected after the 3×3 convolution; the third main branch includes a trunk composed of 1×1 convolution and two sub-branches connected after the 1×1 convolution, each of which has 1×3 convolution; the fourth main branch includes 3×3 convolution and 1×1 convolution connected in sequence; the data of all channels are concatenated and used for the next layer input.

[0088] This asymmetric convolution structure splitting has a more obvious effect than splitting it into several identical convolution kernels. It can process more and richer image features and increase feature diversity. At the same time, it also adds a layer of nonlinear mapping to make the feature information more discriminative.

[0089] Corresponding to the aforementioned embodiment of a lithography hotspot detection method based on deep learning, the present invention also provides an embodiment of a lithography hotspot detection device based on deep learning.

[0090] See also Figure 7 A deep learning-based lithography hotspot detection device provided in an embodiment of the present invention includes a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the executable code, it is used to implement a deep learning-based lithography hotspot detection method in the above embodiment.

[0091] An embodiment of a lithography hotspot detection device based on deep learning provided by the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the non-volatile memory into the internal memory for execution. From a hardware perspective, if Figure 7 As shown, it is a hardware structure diagram of any device with data processing capability where a photolithography hotspot detection device based on deep learning provided by the present invention is located, except Figure 7 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiments is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0092] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0093] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0094] An embodiment of the present invention also provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, a deep learning-based lithography hotspot detection method in the above embodiment is implemented.

[0095] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be an external storage device of any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capability. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0096] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the deep learning-based lithography hotspot detection method.

[0097] Those skilled in the art will readily appreciate other embodiments of the present application after considering the description and practicing the contents disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that are not disclosed in the present application. The description and examples are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the claims.

[0098] It should be understood that the above general description and the detailed description below are only exemplary and explanatory and cannot limit the present application. The present application is not limited to the precise structure described above and shown in the drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present application is limited only by the attached claims.

Claims

1. A lithography hotspot detection method based on deep learning, characterized in that: The method comprises the following steps: S1, collect circuit layout images and construct a sample data set containing hot spots and no hot spots in the lithography layout; S2, use the generative adversarial network to perform hybrid data enhancement on hotspot samples; S3, use the improved dataset to train the hotspot detection network; S4. Input the to-be-detected layout into the trained hotspot detection network to obtain the hotspot detection result.

2. The method for detecting lithography hotspots based on deep learning according to claim 1, characterized in that: The S1 specifically includes: using an optical nano-level camera to collect an integrated circuit layout image, segmenting the acquired image, and generating a sample data set including hot spots and no hot spots in the lithography layout.

3. The method for detecting lithography hotspots based on deep learning according to claim 1, characterized in that: The hybrid data enhancement includes: processing a layout sample containing lithography hotspots through geometric transformation; constructing a hybrid loss based on Wassersteun loss and binary cross entropy loss, and optimizing a generative adversarial network model to enhance the generated sample data.

4. The method for detecting lithography hotspots based on deep learning according to claim 3, characterized in that: The processing of layout samples containing photolithography hot spots by geometric transformation includes: increasing the number of samples by using two methods: rotation and mirror transformation.

5. The method for detecting lithography hotspots based on deep learning according to claim 3, characterized in that: The hybrid loss constructed based on Wassersteun loss and binary cross entropy loss includes: ; = ； ； ； Among them, G represents the generator, D represents the discriminator, represents the real sample, represents the generator input, It means the generator generates samples, and E means mathematical expectation; is the unknown sample, N is the total number of samples, represent belongs to the true distribution probability, represent belongs to the probability of generating distribution; λ1 and λ2 are the coefficients of binary cross entropy and Wasserstein loss.

6. The method for detecting lithography hotspots based on deep learning according to claim 1, characterized in that: The hotspot detection network is a GoogLeNet model, which is composed of two 3×3 convolution blocks, two 3×3 maximum pooling blocks, two 7×7 convolution blocks, four improved Inception modules, a 7×7 maximum pooling layer, a fully connected layer, and a soft-max layer connected in sequence in the figure.

7. The method for detecting lithography hotspots based on deep learning according to claim 6, characterized in that: The improved Inception includes four main branches. The first main branch includes only 1×1 convolution. The second main branch includes a trunk composed of 1×1 convolution and 3×3 convolution connected in sequence, and two sub-branches of 1×3 convolution and 3×1 convolution are connected after the 3×3 convolution. The third main branch includes a trunk composed of 1×1 convolution and two sub-branches connected after the 1×1 convolution, each of which includes 1×3 convolution. The fourth main branch includes 3×3 convolution and 1×1 convolution connected in sequence. The data of all channels are concatenated and combined for the next layer input.

8. A lithography hotspot detection device based on deep learning, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the processor executes the executable code, a lithography hotspot detection method based on deep learning as described in any one of claims 1 to 7 is implemented.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, a lithography hotspot detection method based on deep learning as described in any one of claims 1 to 7 is implemented.

Images