A Lithography Defect Detection Method and Device Based on Deep Learning

Abstract

This application belongs to the fields of artificial intelligence and optical detection technologies. Specifically, it discloses a lithography defect detection method and device based on deep learning. The method includes: scanning each defective lithography sample through the coherent Fourier scattering method to obtain a far-field differential data set, where the far-field differential data set includes defect label data and far-field differential data corresponding to each scanning point in each defective lithography sample; performing asymmetry analysis based on the far-field differential data set to obtain left-right asymmetry information, up-down asymmetry information, and differential summation information corresponding to each scanning point in each defective lithography sample; training a deep learning model using three-channel data. After collecting the diffraction information in the far field, this application uses the method of differential asymmetry analysis to obtain three-channel data as the input of the deep learning model. The deep learning model can automatically learn the characteristics of defect signals, improve the robustness of the system, and achieve lithography defect detection through a neural network.

Description

Technical Field

[0001] This application belongs to the technical fields of artificial intelligence and optical detection, and more specifically, relates to a lithography defect detection method and device based on deep learning. Background Technique

[0002] Lithography defect detection has always been a key link in the semiconductor chip manufacturing process. With the continuous progress of the integrated circuit related manufacturing industry, the feature sizes on wafers and masks are shrinking day by day, and the requirements for detection capabilities are also constantly increasing. Lithography defects usually refer to the graphic distortion or missing caused by various reasons during the lithography process. These defects will affect subsequent etching, deposition, and electrical performance, reducing the yield and reliability of devices. Coherent Fourier scatterometry (CFS) is a scanning-based scatterometry technique, initially used for high-precision measurement of the shape parameters of periodic gratings, and has been introduced in recent years for the detection of nanostructure particles in various bright-field illumination configurations. A monochromatic light source with high temporal and spatial coherence is focused on the surface of the lithography sample to form an illumination spot, which covers the entire plane in a multi-row scanning manner (there are multiple scanning points on each scanning row, and there is a certain overlap between the spots corresponding to each scanning point). In this way, the reflected light from the sample surface and the scattered light from the defects are coherently superimposed in the far field. A quadrant detector or a CCD camera is used to capture the far-field image in the Fourier plane and perform a difference with the far-field image of the defect-free sample. When a sample defect passes through the focused spot during the scanning process, it will cause the non-uniformity of the far-field intensity distribution, that is, the asymmetry of the difference signal. At the same time, it will also be observed that the intensity and positive / negative of the asymmetric signal will change relative to the change of the defect's position within the spot. According to this property, the specific position of the defect on the lithography sample can be determined.

[0003] Diffraction measurement data from defective lithography wafers and masks is a massive, high-dimensional data source. Besides the scattering signals from the defects, this data also contains interference signals such as instrument noise, background noise, and readout noise. This can lead to indistinguishable erroneous signals in far-field images, resulting in false detections. When defect sizes decrease to a certain level (subwavelength or sub-sub-wavelength scales), the scattering signals caused by the defects are easily drowned out by the background noise, making missed detections more likely. To overcome these issues, a more accurate and robust defect detection method is needed. Deep learning has been a hot research area in recent years, with widespread applications in target detection and image classification. By rationally constructing neural network models, features can be autonomously extracted from the data, which is more robust than manually extracted features and has excellent high-dimensional data processing capabilities. Trained neural networks can rapidly detect data during the inference phase, meeting the real-time detection requirements of industrial production. Implementing neural networks for lithography defect detection is a pressing technical challenge in this field. Summary of the Invention

[0004] In response to the defects of the existing technology, the purpose of this application is to realize lithography defect detection through neural networks.

[0005] To achieve the above objectives, in a first aspect, the present application provides a lithography defect detection method based on deep learning, comprising:

[0006] Scanning each defective lithography sample using a coherent Fourier scattering method to obtain a far-field differential data set, the far-field differential data set including defect label data corresponding to each defective lithography sample and far-field differential data corresponding to each scanning point in each defective lithography sample. The far-field differential data is obtained by taking the difference between the diffraction light intensity distribution of the far field corresponding to the defective lithography sample and the diffraction light intensity distribution of the far field corresponding to the defect-free lithography sample at the scanning point;

[0007] Based on the far-field differential data set, an asymmetry analysis is performed on the far-field differential data corresponding to each scanning point to obtain the left-right asymmetry information, the top-bottom asymmetry information, and the differential summation information corresponding to each scanning point in each defective lithography sample;

[0008] Based on the defect label data corresponding to each defective lithography sample and the left-right asymmetric information, top-bottom asymmetric information and differential summation information corresponding to each scanning point in each defective lithography sample, a deep learning model is trained. The deep learning model is used to detect lithography defects in lithography samples.

[0009] In one possible implementation, the far-field differential data corresponding to the scanning point is obtained by the following steps:

[0010] At the scanning point, calculate the near-field electric field distribution after the illumination probe acts on the lithography sample;

[0011] Based on the near-field electric field distribution, obtain the far-field diffraction light intensity distribution corresponding to the defective lithography sample and the far-field diffraction light intensity distribution corresponding to the defect-free lithography sample through the Fraunhofer far-field propagation model;

[0012] Based on the far-field diffraction light intensity distribution corresponding to the defective lithography sample and the far-field diffraction light intensity distribution corresponding to the defect-free lithography sample, perform a difference operation to obtain the far-field difference data corresponding to the scanning point.

[0013] In a possible implementation, at the scanning point, calculating the near-field electric field distribution after the illumination probe acts on the lithography sample includes calculating the near-field electric field distribution through the following formula:

[0014] U(x,y) = P(λ,ξ,η)S(ξ,η);

[0015] Wherein, U(x,y) represents the near-field electric field distribution, (x,y) represents the coordinates in the near-field plane coordinate system, P(λ,ξ,η) represents the input illumination matrix, (ξ,η) represents the coordinates in the sample coordinate system, λ represents the illumination wavelength, and S(ξ,η) represents the sample complex refractive index function.

[0016] For the defective lithography sample, the complex refractive index function is expressed as:

[0017] S(ξ,η) = S free (ξ, η)+D(ξ k , η k );

[0018] Where S free (ξ,η) represents the complex refractive index function of the defect-free lithography sample, D(ξ k ,η k ) represents the defect complex refractive index function, and the coordinates of the defect center position are (ξ k ,η k ).

[0019] For the defect-free lithography sample, the complex refractive index function is expressed as:

[0020] S(ξ,η) = S free (ξ, η).

[0021] In a possible implementation, the deep learning model includes a cascaded feature extraction module, a position prediction module, and a category prediction module. Correspondingly, the defect label data includes a defect position label and a defect size category label;

[0022] The feature extraction module is used to extract the defect features corresponding to the lithography sample based on the left - right asymmetry information, up - down asymmetry information, and differential summation information corresponding to the scanning points;

[0023] The position prediction module is used to predict the defect position based on the defect features corresponding to the lithography sample;

[0024] The category prediction module is used to predict the defect category based on the prediction information of the defect position and the defect features corresponding to the lithography sample.

[0025] In a possible implementation, the loss function used for training the deep - learning model is:

[0026] L = αL P +(1 - α)L c ;

[0027] where L represents the total loss, α represents the weighting coefficient, L P represents the position loss, and L c represents the category loss.

[0028] In a possible implementation, the position loss is calculated by the following formula:

[0029]

[0030] where (ξ i , η i ) represents the defect position predicted by the deep - learning model, (ξ t , η t ) represents the defect position indicated by the defect position label, and n batch represents the number of samples in a training batch. A sample includes the left - right asymmetry information, up - down asymmetry information, and differential summation information corresponding to each scanning point in a defective lithography sample.

[0031] In a possible implementation, the category loss is calculated by the following formula:

[0032]

[0033] where y k represents the category predicted by the deep - learning model, and t k represents the category indicated by the defect size category label.

[0034] In a second aspect, the present application provides a lithography defect detection device based on deep learning, including:

[0035] A dataset construction module, configured to scan each defective lithography sample by means of coherent Fourier scattering method to obtain a far-field differential dataset, where the far-field differential dataset includes defect label data corresponding to each defective lithography sample, and far-field differential data corresponding to each scanning point in each defective lithography sample. The far-field differential data is obtained by taking the difference between the diffraction intensity distribution of the far-field corresponding to the scanning point and the diffraction intensity distribution of the far-field corresponding to a defect-free lithography sample for the scanning point;

[0036] A dataset preprocessing module, configured to perform asymmetry analysis on the far-field differential data corresponding to each scanning point based on the far-field differential dataset to obtain left-right asymmetry information, up-down asymmetry information, and differential summation information corresponding to each scanning point in each defective lithography sample;

[0037] A training module, configured to train a deep learning model based on the defect label data corresponding to each defective lithography sample and the left-right asymmetry information, up-down asymmetry information, and differential summation information corresponding to each scanning point in each defective lithography sample. The deep learning model is used to detect lithography defects in lithography samples.

[0038] In a third aspect, the present application provides an electronic device, including: at least one memory, configured to store a program; at least one processor, configured to execute the program stored in the memory. When the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any one of the possible implementation manners of the first aspect.

[0039] In a fourth aspect, the present application provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program. When the computer program runs on a processor, the processor is caused to execute the method described in the first aspect or any one of the possible implementation manners of the first aspect.

[0040] Generally speaking, compared with the prior art by the above technical solution conceived by the present application, the following beneficial effects are achieved:

[0041] (1) Deep learning is introduced into lithography defect detection based on the coherent Fourier scattering method (CFS). Defects with different positions and sizes are set on the lithography sample for multiple scans to obtain a sufficient amount of datasets. After collecting the diffraction information in the far field, the method of differential asymmetry analysis is used to obtain three-channel data as the input of the neural network. The neural network model will automatically learn the characteristics of the defect signal without manual extraction, improving the robustness of the detection system.

[0042] (2) The identification of lithography defects is mainly reflected in two aspects: the determination of the defect position and the identification of the defect type. By setting two types of loss functions L P and L cThe position prediction loss and the class prediction loss are punished separately, and the two loss functions are summed according to a certain weighted ratio. The neural network trained with sufficient data can process a large amount of scanning data simultaneously and has a fast processing speed to achieve more efficient detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 FIG. is one of the schematic flowcharts of the lithography defect detection method based on deep learning provided by an embodiment of the present application;

[0044] Figure 2 FIG. is another schematic flowchart of the lithography defect detection method based on deep learning provided by an embodiment of the present application;

[0045] Figure 3 FIG. is a schematic diagram of a lithography pattern and a defect provided by an embodiment of the present application;

[0046] Figure 4 FIG. is a schematic diagram of a serpentine scanning path provided by an embodiment of the present application;

[0047] Figure 5 FIG. is a schematic diagram of a lithography defect detection system based on transmission and reflection provided by an embodiment of the present application;

[0048] Figure 6 FIG. is a schematic diagram of the left-right asymmetry information of each scanning point obtained from a complete scan provided by an embodiment of the present application;

[0049] Figure 7 FIG. is a schematic diagram of the up-down asymmetry information of each scanning point obtained from a complete scan provided by an embodiment of the present application;

[0050] Figure 8 FIG. is a schematic diagram of the differential summation information of each scanning point obtained from a complete scan provided by an embodiment of the present application;

[0051] Figure 9 FIG. is a schematic diagram of the structure of a lithography defect recognition network provided by an embodiment of the present application;

[0052] Figure 10 FIG. is a schematic diagram of the structure of a lithography defect detection device based on deep learning provided by an embodiment of the present application;

[0053] Figure 11 FIG. is a schematic diagram of the structure of an electronic device provided by an embodiment of the present application.

[0054] In all the drawings, the same reference numerals are used to represent the same elements or structures, where:

[0055] 1: Helium-neon laser; 2: Attenuator; 3: Adjustable aperture; 4: Beam expander; 5: Mirror; 6: Adjustable aperture; 7: Focusing lens; 8: Lithography sample; 9: Camera; 10: Computer. Detailed implementation manners

[0056] In order to make the objectives, technical solutions and advantages of this application clearer, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application.

[0057] In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0058] In the description of the embodiments of this application, unless otherwise specified, the meaning of "a plurality" refers to two or more. For example, a plurality of processing units refers to two or more processing units, etc.; a plurality of elements refers to two or more elements, etc.

[0059] First, the technical terms involved in the embodiments of this application are introduced.

[0060] (1) Coherent Fourier Scattering (CFS) is a technique that uses a coherent light source to perform scattering experiments and analyzes the scattering signals through Fourier transform. This method is widely used in fields such as physics, chemistry, biology, and materials science to study the microscopic structure and dynamic processes of substances.

[0061] (2) The Fraunhofer diffraction model is an important theoretical model that describes the propagation and diffraction phenomena of light waves when they are far from the light source or optical elements. This model is usually used to analyze the diffraction patterns of light waves after passing through slits, optical elements, or other obstacles in the far-field region.

[0062] (3) The Softmax-with-loss layer is a commonly used layer in deep learning, usually used for multi-classification problems. It combines the softmax activation function and the cross-entropy loss function, and can calculate the output probability distribution and loss value of the model simultaneously in a single forward propagation.

[0063] The embodiments of this application are described below with reference to the accompanying drawings in the embodiments of this application.

[0064] Figure 1 is one of the schematic flowcharts of the deep learning-based lithography defect detection method provided by the embodiments of the present application. As Figure 1 shown, this method can be applied to a coherent Fourier scattering detection system, and this method includes the following steps S101 to step S103.

[0065] Step S101, scanning each defective lithography sample by the coherent Fourier scattering method to obtain a far-field differential data set. The far-field differential data set includes defect label data corresponding to each defective lithography sample and far-field differential data corresponding to each scanning point in each defective lithography sample. The far-field differential data is obtained by taking the difference between the diffraction light intensity distribution of the far field corresponding to the defective lithography sample and the diffraction light intensity distribution of the far field corresponding to the defect-free lithography sample for the scanning point;

[0066] Step S102, based on the far-field differential data set, performing an asymmetry analysis on the far-field differential data corresponding to each scanning point to obtain left-right asymmetry information, up-down asymmetry information, and differential summation information corresponding to each scanning point in each defective lithography sample;

[0067] Step S103, training a deep learning model based on the defect label data corresponding to each defective lithography sample and the left-right asymmetry information, up-down asymmetry information, and differential summation information corresponding to each scanning point in each defective lithography sample. The deep learning model is used to detect lithography defects in lithography samples.

[0068] It can be understood that the present application introduces deep learning into lithography defect detection based on the coherent Fourier scattering method (CFS). Defects with different positions and sizes are set on the lithography sample for multiple scans to obtain a sufficient amount of data sets. After collecting the diffraction information in the far field, the method of differential asymmetry analysis is used to obtain three-channel data (left-right asymmetry information, up-down asymmetry information, and differential summation information) as the input of the deep learning model. The deep learning model can automatically learn the characteristics of defect signals without manual extraction, improving the robustness of the detection system. The trained deep learning model can be used to accurately detect lithography defects in lithography samples, realizing lithography defect detection through a neural network.

[0069] Therefore, the deep learning-based lithography defect detection method provided by the present application has the advantages of no need for manual feature extraction, fast and accurate positioning and classification, effectively solving the problems of complex feature extraction and insufficient robustness of existing lithography defect detection methods, and can be applied to fields such as mask and wafer defect detection.

[0070] Figure 2 is the second schematic flowchart of the deep learning-based lithography defect detection method provided by the embodiments of the present application. AsFigure 2 As shown, it includes the following steps S201 to S205.

[0071] Step S201: Obtain a far-field differential data set by scanning the entire lithography sample completely. This step specifically includes:

[0072] S201.1 Design a lithography sample model S(ξ,η) with any pattern type, add a defect at any position on the sample pattern to form a defective lithography sample, and record the coordinates (ξ1,η1) of the defect center position on the sample pattern and the defect characteristic length D. As Figure 3 shown, Figure 3 where the value range of k in

[0073] is 1 ≤ k ≤ n;

[0074] S201.2 Construct an input illumination matrix P(λ,ξ,η), calculate the near-field electric field distribution U(x,y) after the illumination probe acts on the sample. The calculation method is as follows:

[0075] U(x,y) = P(λ,ξ,η)S(ξ,η);

[0076] where λ represents the illumination wavelength, (x,y) is the near-field plane coordinate system at a very small distance from the sample. Here, the sample surface is regarded as a function S(ξ,η), S(ξ,η) can be the sample complex refractive index function, and the near-field electric field distribution is the simple product P(λ,ξ,η)S(ξ,η) of the illumination probe and the sample complex refractive index function. and make a difference to obtain the diffraction differential light intensity distribution ΔI(x′,y′), that is, the far-field differential data. The calculation method of the Fraunhofer far-field propagation model is as follows:

[0077]

[0078] where (x′,y′) is the diffraction plane coordinate (corresponding to the detector target position), U′(x,y) is the light field amplitude distribution in the diffraction plane, k is the wave number, and z represents the diffraction propagation distance. The first two terms before the integral term affect the phase and overall amplitude of the light field. In actual detection, what the detector obtains is the intensity distribution of the light field, that is, the square of the amplitude. The phase factor does not affect the intensity distribution of the diffraction field, and mainly focuses on the relative intensity distribution. Therefore, in order to reduce the calculation amount, the first two terms before the integral can be discarded. In addition, the integral term is mathematically equivalent to performing a Fourier transform on the near-field electric field distribution U(x,y). Therefore, the final calculation formula for solving the far-field diffraction intensity distribution is as follows:

[0079]

[0080] wherein represents the Fourier transform operation.

[0081] S201.4 Move the illumination probe to the next scanning position, and repeat steps S201.2 - S201.3 along the serpentine scanning path until the entire sample is scanned (i.e., a complete scan is completed), while ensuring that the overlapping area of adjacent illumination spots is not less than 60%, that is, the spatial domain overlapping constraint. The scanning path is schematically shown as Figure 4 shown;

[0082] S201.5 Keep the sample pattern unchanged, adjust the coordinate position (ξ1, η1) of the defect on the sample and the characteristic length D of the defect to obtain a new defective lithography sample. Repeat steps S201.2 - S201.4 for the new defective lithography sample until the sample with the defect coordinate position (ξ n , η n ) is scanned completely, and save the far - field differential data obtained from each scan;

[0083] Figure 5 is a schematic diagram of a lithography defect detection system based on transmission and reflection provided by an embodiment of the present application. As Figure 5 shown, it includes a transmission detection system and a reflection detection system, and either one can be selected according to actual conditions during specific implementation. The lithography defect detection system is sequentially arranged as: a helium - neon laser 1, an attenuation sheet 2, an adjustable aperture 3, a beam expander 4, a mirror 5, an adjustable aperture 6, a focusing lens 7, a lithography sample 8, a camera 9, and a computer 10.

[0084] Step S202, pre - process the far - field differential data: perform asymmetry analysis on the far - field differential data corresponding to each scanning point in the dataset, and use the defect coordinates and sizes of the entire scanned sample plane each time as label data, which are jointly used as the input of the neural network.

[0085] The method for performing asymmetry analysis on the far - field differential data is as follows:

[0086]

[0087] wherein, ΔI L-R represents the left - right asymmetry information, ΔI T-B represents the up - down asymmetry information, and ΔI sum represents the differential summation information.

[0088] In the embodiments of the present application, a CMOS detector based on pixel points is used to receive the light intensity signal, and a matrix is obtained in which each pixel point corresponds to a light intensity value. The method for solving the left-right asymmetry information is to subtract the matrix corresponding to the far-field differential data ΔI(x′,y′) from the matrix corresponding to its horizontally flipped data ΔI(-x′,y′). For the matrix obtained after the subtraction, the sum of all elements in the matrix is calculated and a normalization operation is performed. The method for solving the up-down asymmetry information is to subtract the matrix corresponding to the far-field differential data ΔI(x′,y′) from the matrix corresponding to its vertically flipped data ΔI(x′,-y′). For the matrix obtained after the subtraction, the sum of all elements in the matrix is calculated and a normalization operation is performed. The method for solving the differential sum information is to calculate the sum of all elements in the matrix corresponding to the far-field differential data ΔI(x′,y′) and perform a normalization operation. After solving for all scanning points, two asymmetry matrices I L-R 、I T-B and a differential sum matrix I sum for a single complete scan can be obtained, which are represented as follows:

[0089]

[0090] wherein, the superscript of each element in the matrix represents the coordinate sequence of the center position of each scanning point in a single complete scan. For example, (i,j) represents the scanning point in the i-th row and j-th column, and 1 ≤ i ≤ m, and 1 ≤ j ≤ m. Finally, the three matrices are combined to obtain the three-channel data which is represented as follows:

[0091]

[0092] where k represents the k-th complete scan. To ensure that the network has enough data sets for training, the number of scans is not less than 800 times.

[0093] Figures 6 - 8 The following is a schematic diagram of the three-channel data obtained from a certain complete scan: Figure 6 、 Figure 7 、 Figure 8 are respectively the left-right asymmetry information, up-down asymmetry information, and differential sum information of each scanning point in a single scan. Here, the units of the coordinate axes in Figures 6 - 8 are explained. The coordinates in the two horizontal directions represent the positions of the scanning points corresponding to a single complete scan, and the vertical coordinate represents the normalized signal intensity corresponding to each scanning point. From Figure 6 、 Figure 7 it can be seen that when scanning to the defect position, the asymmetry information will show positive and negative peaks along the corresponding direction. When the scanning line is offset from the defect position, the peak height will decrease; according to Figure 8From the differential summation information, it can be seen that the closer the scanning point is to the defect position, the higher the result of the differential summation; the larger the defect characteristic size, the wider the width of the generated signal; all of these have very obvious characteristics and are easier to extract using a neural network.

[0094] Furthermore, at the end of each scan, it is necessary to record the defect position and size as label data T k ,

[0095] which is expressed as follows:

[0096] T k =[P(ξ k ,η k ),D k ;

[0097] where P represents the position of the defect, (ξ k ,η k ) represents the coordinate position of the defect center on the sample, and D k represents the defect size.

[0098] Finally, the preprocessed training data and label data are expressed as follows:

[0099]

[0100] T = [T1, T2 … T n ;

[0101] Step S203, establish a prediction network for the defect position and size. It includes a feature extraction part, a position prediction part, and a category prediction part, and the output is the defect position coordinates and the defect category. The network structure includes, but is not limited to, networks such as LeNet, AlexNet, and VGGNet, including convolutional layers, pooling layers, batch normalization layers, fully connected layers, etc. The feature extraction module needs to be determined according to the data scale. In the embodiment of the present application, the feature extraction module is composed of a convolutional layer, a pooling layer, and a ReLU activation function layer (considering that the number of layers can be continuously increased according to the input information dimension); the position prediction module is composed of a BatchNorm (batch normalization) layer and a fully connected layer, and outputs the predicted position coordinates of the defect on the sample; the category prediction module is composed of a fully connected layer and a Softmax-with-loss layer, and outputs the predicted defect category. The network structure is as Figure 9 shown.

[0102] The classification method of the defect is divided according to the size range of the defect characteristic length D, and the endpoint values of each interval need to be set in advance. For example, if the interval endpoints are set as Ω = [40, 60, 80, 100, 120], with the unit of nm, then the defect is divided into 4 categories, which are respectively in the four intervals corresponding to every two endpoints.

[0103] Step S204, set the network loss function and parameter update method: The loss functions used in the network include a location loss function \(L\) P and a class loss function \(L\) C , and the two loss functions are expressed as follows:

[0104]

[0105] The loss function \(L\) P for the defect location is selected as the root mean square error, representing the geometric distance between the predicted defect location \((\xi\) i , \(\eta\) i ) and the actual location \((\xi\) t , \(\eta\) t ). The defect type is represented using one - hot, with the correct solution label being 1 and the rest being 0. For example, in the set interval endpoints \(\Omega=[40,60,80,100,120]\) above, the one - hot representation of the correct solution label for the defect size in the range \([60,80)\) is \(t\) k = [0,1,0,0]. The defect type loss function is the cross - entropy error, \(y\) k is the output of the neural network, and \(t\) k is the label data. This network uses mini - batch learning, that is, each time of learning will randomly select a batch of \(n\) batch (for example, \(n\) batch = 64) data sets for learning, and will also output \(n\) batch prediction results. Therefore, it is necessary to take the average of the loss function to obtain the "average loss function" of a single data.

[0106] Finally, the total loss function is the weighted sum of the location loss function \(L\) P and the class loss function \(L\) C , expressed as follows:

[0107] \(L=\alpha L\) P +(1 - \(\alpha\))\(L\) c ;

[0108] Select the Adam optimizer to update the parameters of the network, and set the moment estimation decay exponents \(\beta_1\) and \(\beta_2\) to 0.9 and 0.999.

[0109] Step S205, train the network and adjust the system hyperparameters: The number of neurons inside the neural network needs to be determined according to the dimension of the input data. When determining the initial weights of each neuron, in order to prevent the problem of gradient disappearance, "He Initialization" is selected. When the number of neuron nodes in the current layer is \(n\) neuron , a Gaussian distribution with a standard deviation of is used.

[0110] By performing multiple trainings to find the optimal learning rate, considering that a complete training may take a long time, the range of the learning rate is specified on a logarithmic scale, and random sampling is performed within this range. For example, the range of the learning rate is set to 10 -3 ~10 3 , and the learning rate for each complete training is set to That is, the parameter n lr Random sampling is performed within the range of [-3, 3] each time, and finally a relatively appropriate result is found.

[0111] The lithography defect detection device based on deep learning provided by the present application will be described below. The lithography defect detection device based on deep learning described below can be referred to in correspondence with the lithography defect detection method described above.

[0112] Figure 10 is a schematic structural diagram of the lithography defect detection device based on deep learning provided by an embodiment of the present application. As Figure 10 shown, the device includes: a dataset construction module 20, a dataset preprocessing module 30, and a training module 40. Among them:

[0113] The dataset construction module 20 is configured to scan each defective lithography sample by the coherent Fourier scattering method to obtain a far-field difference dataset. The far-field difference dataset includes defect label data corresponding to each defective lithography sample, and far-field difference data corresponding to each scanning point in each defective lithography sample. The far-field difference data is obtained by taking the difference between the diffraction light intensity distribution of the far field corresponding to the defective lithography sample and the diffraction light intensity distribution of the far field corresponding to the defect-free lithography sample for the scanning point;

[0114] The dataset preprocessing module 30 is configured to perform asymmetry analysis on the far-field difference data corresponding to each scanning point based on the far-field difference dataset to obtain left-right asymmetry information, up-down asymmetry information, and difference summation information corresponding to each scanning point in each defective lithography sample;

[0115] The training module 40 trains a deep learning model based on the defect label data corresponding to each defective lithography sample and the left-right asymmetry information, up-down asymmetry information, and difference summation information corresponding to each scanning point in each defective lithography sample. The deep learning model is used to detect lithography defects in the lithography sample.

[0116] It can be understood that the detailed function implementation of the above-mentioned each unit / module can be referred to the introduction in the foregoing method embodiment, and will not be elaborated here.

[0117] It should be understood that the above device is used to execute the method in the above embodiment. For the corresponding program modules in the device, their implementation principles and technical effects are similar to those described in the above method. The working process of the device can refer to the corresponding process in the above method and will not be elaborated here.

[0118] Based on the method in the above embodiment, an embodiment of the present application provides an electronic device. Figure 11 It is a schematic structural diagram of the electronic device provided by the embodiment of the present application. As Figure 11 shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communication bus 840. Among them, the processor 810, the communications interface 820, and the memory 830 complete mutual communication through the communication bus 840. The processor 810 can call the logical instructions in the memory 830 to execute the method in the above embodiment.

[0119] In addition, when the logical instructions in the above-mentioned memory 830 are implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application.

[0120] Based on the method in the above embodiment, an embodiment of the present application provides a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program runs on the processor, it causes the processor to execute the method in the above embodiment.

[0121] Based on the method in the above embodiment, an embodiment of the present application provides a computer program product. When the computer program product runs on the processor, it causes the processor to execute the method in the above embodiment.

[0122] It can be understood that the processor in the embodiments of the present application may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0123] The method steps in the embodiments of the present application may be implemented in a hardware manner or by a processor executing software instructions. The software instructions may be composed of corresponding software modules, and the software modules may be stored in a random access memory (RAM), flash memory, read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), registers, hard disks, removable hard disks, CD-ROMs, or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium may also be a component of the processor. The processor and the storage medium may be located in an ASIC.

[0124] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired manner (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or a wireless manner (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0125] It can be understood that the various numerical numbers involved in the embodiments of the present application are only for the convenience of description and are not used to limit the scope of the embodiments of the present application.

[0126] Those skilled in the art can easily understand that the above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A lithography defect detection method based on deep learning, characterized in that Including: Scanning each defective lithography sample by coherent Fourier scattering method to obtain a far-field differential data set, which includes defect label data corresponding to each defective lithography sample and far-field differential data corresponding to each scanning point in each defective lithography sample. The far-field differential data is obtained by taking the difference between the diffraction intensity distribution of the far-field corresponding to the defective lithography sample and the diffraction intensity distribution of the far-field corresponding to the defect-free lithography sample for the scanning point; Based on the far-field differential data set, performing an asymmetry analysis on the far-field differential data corresponding to each scanning point to obtain the left-right asymmetry information, up-down asymmetry information, and differential summation information corresponding to each scanning point in each defective lithography sample; Based on the defect label data corresponding to each defective lithography sample and the left-right asymmetry information, up-down asymmetry information, and differential summation information corresponding to each scanning point in each defective lithography sample, training a deep learning model, which is used to detect lithography defects in the lithography sample.

2. The deep learning-based lithography defect detection method according to claim 1, wherein The far-field differential data corresponding to the scanning point is obtained through the following steps: At the scanning point, calculating the near-field electric field distribution after the illumination probe acts on the lithography sample; Based on the near-field electric field distribution, through the Fraunhofer far-field propagation model, obtaining the diffraction intensity distribution of the far-field corresponding to the defective lithography sample and the diffraction intensity distribution of the far-field corresponding to the defect-free lithography sample; Based on the diffraction intensity distribution of the far-field corresponding to the defective lithography sample and the diffraction intensity distribution of the far-field corresponding to the defect-free lithography sample, taking the difference to obtain the far-field differential data corresponding to the scanning point.

3. The method for detecting lithography defects based on deep learning according to claim 2, wherein The calculating the near-field electric field distribution after the illumination probe acts on the lithography sample at the scanning point includes calculating the near-field electric field distribution through the following formula: U(x,y) = P(λ,ξ,η)S(ξ,η); Where, U(x,y) represents the near-field electric field distribution, (x,y) represents the coordinates in the near-field plane coordinate system, P(λ,ξ,η) represents the input illumination matrix, (ξ,η) represents the coordinates in the sample coordinate system, λ represents the illumination wavelength, and S(ξ,η) represents the sample complex refractive index function.

4. The deep learning-based lithography defect detection method according to any one of claims 1-3, characterized in that, The deep learning model includes a cascaded feature extraction module, a position prediction module, and a category prediction module. Correspondingly, the defect label data includes a defect position label and a defect size category label; The feature extraction module is used to extract the defect features corresponding to the lithography sample based on the left-right asymmetry information, up-down asymmetry information, and differential summation information corresponding to the scanning point; The position prediction module is used to predict the defect position based on the defect features corresponding to the lithography sample; The category prediction module is used to predict the defect category based on the prediction information of the defect position and the defect features corresponding to the lithography sample.

5. The method for detecting lithography defects based on deep learning according to claim 4, wherein The loss function used for training the deep learning model is: L = αL P +(1 - α)L c ; Among them, L represents the total loss, α represents the weighting coefficient, L P represents the position loss, and L c represents the class loss.

6. The method for detecting lithography defects based on deep learning according to claim 5, wherein The position loss is calculated through the following formula: Among them, (ξ i , η i ) represents the defect position predicted by the deep learning model, (ξ t , η t ) represents the defect position indicated by the defect position label, n batch represents the number of samples in a training batch, and a sample includes the left-right asymmetry information, up-down asymmetry information, and differential summation information corresponding to each scan point in a defective lithography sample.

7. The method for detecting lithography defects based on deep learning according to claim 5, wherein The category loss is calculated through the following formula: Among them, y k represents the category predicted by the deep learning model, and t k represents the category indicated by the defect size category label.

8. A lithography defect detection device based on deep learning, characterized in that, Including: A dataset construction module, configured to scan each defective lithography sample by coherent Fourier scattering method to obtain a far-field differential dataset, where the far-field differential dataset includes defect label data corresponding to each defective lithography sample and far-field differential data corresponding to each scanning point in each defective lithography sample, and the far-field differential data is obtained by taking the difference between the diffraction light intensity distribution of the far-field corresponding to the scanning point and the diffraction light intensity distribution of the far-field corresponding to a defect-free lithography sample for the scanning point; A dataset preprocessing module, configured to perform asymmetry analysis on the far-field differential data corresponding to each scanning point based on the far-field differential dataset to obtain left-right asymmetry information, up-down asymmetry information, and differential summation information corresponding to each scanning point in each defective lithography sample; A training module, configured to train a deep learning model based on the defect label data corresponding to each defective lithography sample and the left-right asymmetry information, up-down asymmetry information, and differential summation information corresponding to each scanning point in each defective lithography sample, where the deep learning model is used to detect lithography defects in lithography samples.

9. An electronic device, characterized in that, Comprising: At least one memory, configured to store a computer program; At least one processor, configured to execute the program stored in the memory, and when the program stored in the memory is executed, the processor is configured to execute the method according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program runs on the processor, the processor is caused to execute the method according to any one of claims 1-7.

Images