Training method of hotspot detection model and hotspot detection method

Abstract

The invention relates to the field of chip manufacturing, in particular to a hot spot detection model training method and a hot spot detection method, and the method comprises the steps: obtaining a sample set composed of a plurality of circuit layouts and corresponding labels; inputting the sample set into a local feature extraction network, and extracting local features; inputting the local features into a global feature extraction network, and extracting global features; inputting the global features into a decoder to obtain a prediction result of whether each circuit layout is a layout hot spot; and on the basis of the prediction result and the difference degree of the corresponding label, updating parameters of the to-be-trained hotspot detection model, and obtaining a trained hotspot detection model. According to the method, the layout is processed in the frequency domain and the time domain at the same time, and the complex layout features are automatically extracted by adopting the deep learning model, so that the prediction accuracy of the layout hotspots is remarkably improved; in addition, by using the weight of the pre-training model and a staged feature extraction strategy, the model can still have high efficiency and high accuracy when processing a large-scale layout.

Description

Technical Field

[0001] The present invention relates to the field of chip manufacturing, and in particular to a hotspot detection model training method and a hotspot detection method. Background Art

[0002] The layout of an integrated circuit is a graphic that describes how circuit components are laid out and connected on a silicon wafer. In layout design, some areas may produce defects during the photolithography process. These defects are called "layout hot spots". The appearance of layout hot spots may affect the performance of the integrated circuit, resulting in yield loss or even making the integrated circuit inoperable. Therefore, the detection of layout hot spots is very important in the design and manufacture of integrated circuits.

[0003] Existing layout hotspot detection methods mainly include lithography simulation, pattern matching-based methods, and traditional machine learning methods. Traditional machine learning methods usually use manually extracted features, such as the geometric properties and topological structure of the pattern, and then use classifiers (such as support vector machines) to predict which areas may be hotspots. However, the accuracy of this method is usually low because the manually extracted features may not fully capture the complex layout patterns and the optical effects of the lithography process. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a hotspot detection model training method and a hotspot detection method.

[0005] The first aspect of the present invention discloses a method for training a hotspot detection model, comprising:

[0006] Obtain a sample set consisting of multiple circuit layouts and labels corresponding to the multiple circuit layouts;

[0007] Input the sample set into the local feature extraction network of the hotspot detection model to be trained to extract local features;

[0008] Inputting the local features into the global feature extraction network of the hotspot detection model to be trained to extract global features;

[0009] The global features are input into the decoder of the hotspot detection model to be trained to obtain a prediction result of whether each circuit layout is a layout hotspot;

[0010] Based on the difference between the prediction result and the corresponding label, the parameters of the hotspot detection model to be trained are updated to obtain a trained hotspot detection model.

[0011] Furthermore, the step of obtaining a sample set consisting of multiple circuit layouts includes:

[0012] For each circuit layout:

[0013] Obtaining a time domain layout of a circuit layout;

[0014] Perform discrete cosine processing on the time domain map to obtain the frequency domain map;

[0015] Combining the time domain layout and the frequency domain layout into image data of different channels of the circuit layout;

[0016] The image data of multiple circuit layouts are aggregated as a sample set.

[0017] Furthermore, the step of inputting the sample set into the local feature extraction network of the hotspot detection model to be trained and extracting local features includes:

[0018] Input each circuit layout in the sample set into the block division module of the hotspot detection model to be trained for splitting to obtain multiple layout blocks of the same size;

[0019] The layout block is input into the local feature extraction network of the hotspot detection model to be trained to extract local features.

[0020] Furthermore, the step of inputting the layout block into a local feature extraction network of a hotspot detection model to be trained to extract local features includes:

[0021] Input each block into the linear embedding layer of the local feature extraction network, perform linear transformation on each pixel in the block, and embed the position information;

[0022] Input the block with embedded position information into the multi-head self-attention layer of the local feature extraction network to extract features at different positions and obtain multi-head information of the block;

[0023] The multi-head information of the layout block is input into the window-shifting self-attention layer of the local feature extraction network, the multi-head information of the layout block is divided into multiple windows, the pixels at the predetermined position of each window are removed, and self-attention calculation is performed to obtain local features.

[0024] Furthermore, the step of inputting the local features into a global feature extraction network to extract global features includes:

[0025] Inputting the local features into the linear embedding and multi-head self-attention layers of the global feature extraction network to extract context information to obtain the first global feature;

[0026] Performing a linear transformation on the first global feature through a learnable weight matrix to obtain a query feature;

[0027] Mapping the query feature to the dimension of a predefined reference point, and performing nonlinear transformation and convolution processing to generate shape shift key position and value information;

[0028] Attention calculation is performed on the query feature, the shape-shifted key position and the value information to obtain a global feature.

[0029] Furthermore, the step of inputting the global features into the decoder of the hotspot detection model to be trained to obtain a prediction result of whether each circuit layout is a layout hotspot includes:

[0030] The global features are input into the inter-layer regularization layer of the decoder for regularization processing to obtain regularized features;

[0031] The regular features are input into the global pooling layer of the decoder for global average pooling to obtain the pooled features;

[0032] The pooled features are extracted through the fully connected layer of the decoder to obtain fully connected features;

[0033] The fully connected features are processed by activation functions to obtain nonlinear features;

[0034] Based on a preset threshold, the nonlinear features are binarized to obtain a prediction result of whether each circuit layout is a layout hotspot.

[0035] Furthermore, the step of updating the parameters of the hotspot detection model to be trained based on the difference between the prediction result and the corresponding label includes:

[0036] Based on the difference between the prediction results and the corresponding labels, the parameters of the hotspot detection model to be trained are updated according to a predefined loss function; wherein the loss function includes hyperparameters for judging risk, coverage loss and defining coverage weights.

[0037] A second aspect of the present invention discloses a hotspot detection method, comprising:

[0038] Obtaining a circuit layout to be tested;

[0039] The circuit layout to be detected is input into the hotspot detection model trained by the hotspot detection model training method disclosed in any one of the first aspects of the present invention to perform hotspot detection and obtain a detection result.

[0040] A third aspect of the present invention discloses a training system for a hotspot detection model, comprising:

[0041] An acquisition module, used to acquire a sample set consisting of multiple circuit layouts and labels corresponding to the multiple circuit layouts;

[0042] A first extraction module is used to input the sample set into a local feature extraction network of a hotspot detection model to be trained to extract local features;

[0043] A second extraction module is used to input the local features into the global feature extraction network of the hotspot detection model to be trained to extract the global features;

[0044] A prediction module, used for inputting global features into a decoder of a hotspot detection model to be trained to obtain a prediction result of whether each circuit layout is a layout hotspot;

[0045] The updating module is used to update the parameters of the hotspot detection model to be trained based on the difference between the prediction result and the corresponding label to obtain the trained hotspot detection model.

[0046] The fourth aspect of the present invention discloses an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the hotspot detection model training method disclosed in any one of the first aspects of the invention when executing the computer program.

[0047] The present invention processes the layout in both the frequency domain and the time domain simultaneously, and uses a deep learning model to automatically extract complex layout features, thereby significantly improving the prediction accuracy of layout hotspots; in addition, by utilizing the weights of the pre-trained model and a phased feature extraction strategy, the model can still have high efficiency and accuracy when processing large-scale layouts. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0049] Figure 1 is a network structure diagram of a hotspot detection model disclosed in an embodiment of the present invention;

[0050] Figure 2 is a flow chart of a training method for a hotspot detection model disclosed in an embodiment of the present invention;

[0051] Figure 3 is a schematic diagram of global feature extraction disclosed in an embodiment of the present invention;

[0052] Figure 4 It is a flow chart of a hotspot detection method disclosed in an embodiment of the present invention.

[0053] Figure 5 It is a structural diagram of a training system for a hotspot detection model disclosed in an embodiment of the present invention;

[0054] Figure 6 It is a schematic diagram of the structure of an electronic device disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, or product end including a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0057] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0058] The hotspot detection model referred to in the present invention is a Vision Transformer (ViT) model with a shape-shifting attention mechanism. The network structure of the hotspot detection model is as follows: Figure 1 As shown. The present invention uses the weights of a pre-trained large model as the initial weights of the hotspot detection model. The pre-trained large model refers to a deep learning model trained on a large-scale data set. In this embodiment, the weights of the ViT model trained based on the ImageNet22k data set are used as the initial weights of the hotspot detection model. Similarly, a ResNet model, or a DenseNet model, or an EfficientNet model can be used. These pre-trained large models have learned a large amount of general visual features and knowledge, which can help the hotspot detection model quickly adapt to the layout hotspot prediction task, accelerate the training process, and improve the prediction accuracy of the model.

[0059] The training method of the hotspot detection model provided by the present invention is as follows Figure 2 As shown, the invention is explained below in conjunction with a specific embodiment.

[0060] S101, obtaining a sample set consisting of multiple circuit layouts and labels corresponding to the multiple circuit layouts.

[0061] In an optional embodiment, the step of obtaining a sample set consisting of a plurality of circuit layouts includes:

[0062] For each circuit layout:

[0063] Obtaining a time domain layout of a circuit layout;

[0064] Perform discrete cosine processing on the time domain map to obtain the frequency domain map;

[0065] Combining the time domain layout and the frequency domain layout into image data of different channels of the circuit layout;

[0066] The image data of multiple circuit layouts are aggregated as a sample set.

[0067] The time domain layout is the original spatial expression form of the circuit layout, which directly presents the geometry of the circuit layout and the arrangement of circuit elements. However, the time domain layout may contain a large amount of redundant information, and may ignore some global patterns and structural characteristics that are difficult to observe intuitively. At the same time, since the layout generates a spatial domain image with hot spots through the lithography process, the time domain information of the layout is transformed into frequency domain information via Fourier transform during the lithography process, so the frequency domain information also contains key information that causes the layout to become a hot spot. In order to capture this information, the present invention performs DCT (Discrete Cosine Transform, Discrete Cosine Transform) on the time domain layout in the frequency domain, and the discrete cosine transform (Discrete Cosine Transform, DCT) obtains the frequency domain layout corresponding to the time domain layout. The time domain layout and the corresponding frequency domain layout of each circuit layout form the image data of different channels of the circuit layout, and the image data of multiple circuit layouts are used as a sample set.

[0068] S102, inputting the sample set into the local feature extraction network of the hotspot detection model to be trained to extract local features;

[0069] In an optional embodiment, the step of inputting the sample set into the local feature extraction network of the hotspot detection model to be trained and extracting the local features comprises:

[0070] Input each circuit layout in the sample set into the block division module of the hotspot detection model to be trained for splitting to obtain multiple layout blocks of the same size;

[0071] The layout block is input into the local feature extraction network of the hotspot detection model to be trained to extract local features.

[0072] In a further optional embodiment, the step of inputting the layout block into a local feature extraction network of a hotspot detection model to be trained, and extracting local features comprises:

[0073] Input each block into the linear embedding layer of the local feature extraction network, perform linear transformation on each pixel in the block, and embed the position information;

[0074] Input the block with embedded position information into the multi-head self-attention layer of the local feature extraction network to extract features at different positions and obtain multi-head information of the block;

[0075] The multi-head information of the layout block is input into the window-shifting self-attention layer of the local feature extraction network, the multi-head information of the layout block is divided into multiple windows, the pixels at the predetermined position of each window are removed, and self-attention calculation is performed to obtain local features.

[0076] like Figure 1 As shown in the figure, the two local feature extraction networks have the same network structure, which is a linear embedding layer, a multi-head self-attention layer, and a window self-attention layer. The dimensions of the two local feature extraction networks are different.

[0077] For each pixel in the circuit layout, the linear embedding layer maps it to a higher dimensional space through a linear transformation, and then adds the embedded position information. By embedding the position information, the hotspot detection model can distinguish the features of different pixel positions.

[0078] Through the multi-head self-attention layer, the hotspot detection model pays attention to the features of different positions in the block after the position information is embedded, and the multi-head information X of the block is obtained. mul .

[0079] The traditional self-attention mechanism considers the features of all positions, which will lead to excessive computation when processing large layouts. mul The image is divided into multiple windows, each of which is of size M×M. Then, self-attention calculation is performed in each window, which greatly reduces the amount of calculation. At the same time, by removing a portion of pixels in each window, the window-shifting self-attention can also allow the hotspot detection model to pay attention to the features between windows, further improving the performance of the hotspot detection model. In this embodiment, the removed pixels are Indicates rounding down.

[0080] There are L layers of window self-attention layers. Each layer normalizes the output of the previous layer and then passes through the window attention window. MSA Or move the window attention SHiftWindow MSA After processing, it is added to the output of the previous layer and processed by the multi-layer perceptron MLP as the output of this layer. The input of the first layer of the window self-attention layer is the multi-head information X of the map block mul The working principle of the window-shifting self-attention layer is as follows:

[0081]

[0082] Among them, X l represents the output of the lth layer of the window-shifting self-attention layer, LN represents the layer normalization operation, Temporary variables of the lth window-shifting self-attention layer.

[0083] The output of the last layer of window-shifting self-attention layer is used as the local feature X of the block. local .

[0084] S103, inputting the local features into the global feature extraction network of the hotspot detection model to be trained to extract global features;

[0085] In an optional embodiment, the step of inputting the local features into a global feature extraction network to extract global features includes:

[0086] Inputting the local features into the linear embedding and multi-head self-attention layers of the global feature extraction network to extract context information to obtain the first global feature;

[0087] Performing a linear transformation on the first global feature through a learnable weight matrix to obtain a query feature;

[0088] Mapping the query feature to the dimension of a predefined reference point, and performing nonlinear transformation and convolution processing to generate shape shift key position and value information;

[0089] Attention calculation is performed on the query feature, the shape-shifted key position and the value information to obtain a global feature.

[0090] like Figure 1 As shown in the figure, the network structures of the two global feature extraction networks are the same, which are linear embedding layer, multi-head self-attention layer, shape shifting network layer, among which the structure of the shape shifting network layer is downsampling layer, Gaussian error linear unit and convolution layer. The dimensions of the two global feature extraction networks are different.

[0091] The local feature X of the block local Input linear embedding layer in sequence, and multi-head self-attention layer to obtain the first global feature X gol_1 .

[0092] like Figure 3 As shown, the first global feature X gol_1 Through the learnable weight matrix W Q Perform linear changes to obtain the query feature Q:

[0093] Q=X gol_1 W Q ;

[0094] The downsampling layer maps the query feature Q to the same dimension as the reference point, where the reference point is the Z elements pre-selected from the query feature Q, and Z is a hyperparameter. The Gaussian error linear unit performs a nonlinear transformation on the transformed feature value, passes through a convolution layer with a convolution kernel of 1×1, and finally transforms it into the corresponding shape shift key position after bicubic downsampling. and quantity information

[0095]

[0096]

[0097] g(p x , r x )=max(0,1-|p x -r x |);

[0098] Among them, W K is the shape-shift bond matrix, W V is the shape displacement value information matrix, is the shape shift transformation input, p represents the first global feature X gol _1 corresponds to the position information of the reference point, and Δp represents the first global feature X gol _1 corresponds to the deformation variable of p, φ(X gol_1 ; p+Δp) represents the circuit layout position information embedding function, RP represents the reference point position information, (p x ,p y ) is the displacement position corresponding to the position information RP of the reference point, (r x ,r y ) is the first global feature X gol_1 The corresponding position information, g(p x ,r x )g(p y ,r y ) is the displacement position corresponding to RP (p x ,p y ) relative to the reference point (r x ,r y )'s redirection function.

[0099] Shift the key position Quantity information and query feature Q as the input of multi-head attention, and obtain the global feature X of the map block through multi-head attention global :

[0100]

[0101] Among them, d qRepresents the dimension of the query feature Q, and softmax is a nonlinear function.

[0102] S104, inputting the global features into the decoder of the hotspot detection model to be trained to obtain a prediction result of whether each circuit layout is a layout hotspot;

[0103] The global feature X corresponding to the circuit layout global The input is sent to a decoder composed of an inter-layer regularization layer, a global pooling layer, and a fully connected layer in sequence to obtain a prediction of whether each circuit layout in the sample set is a layout hotspot.

[0104] The global feature X corresponding to the circuit layout global The regularization layer between input layers performs regularization processing to reduce the difference between features and obtain the regularized feature X ln :

[0105]

[0106] Where μ is the global feature X global The mean of all elements, σ ​​is the global feature X global The standard deviation of all elements, γ and β are trainable parameters used to scale and translate the output after regularization.

[0107] The regular feature X ln Input the global pooling layer for global average pooling to further reduce the difference between features and make the feature expression more robust, and obtain the pooled feature X pool :

[0108]

[0109] Among them, K represents the regular feature X ln The dimension of k represents the regular feature X ln The ordinal number of the dimension.

[0110] The pooled feature X pool The fully connected features are further extracted through the fully connected layer to obtain the fully connected features, and then the fully connected features are processed by the activation function to obtain the nonlinear features X nonLinear In this embodiment, the ReLU function is selected as the activation function to introduce nonlinearity, so that the hotspot detection model can better fit the complex nonlinear relationship.

[0111] Finally, the nonlinear feature X is transformed into nonLinear Perform binarization processing to obtain the prediction f(x i ):

[0112]

[0113] Among them, x i Represents the i-th circuit layout in the sample set.

[0114] S105. Based on the difference between the prediction result and the corresponding label, update the parameters of the hotspot detection model to be trained to obtain a trained hotspot detection model.

[0115] In an optional embodiment, the step of updating the parameters of the hotspot detection model to be trained based on the difference between the prediction result and the corresponding label includes:

[0116] Based on the difference between the prediction result and the corresponding label, the parameters of the hotspot detection model to be trained are updated according to a predefined loss function; wherein the loss function includes hyperparameters for judging risk, coverage loss and defining coverage weight:

[0117] The judgment risk represents the gap between the prediction result and the corresponding label;

[0118] The coverage loss represents the proportion of the circuit board diagram coverage sample set corresponding to the prediction of the hotspot detection model to be trained.

[0119] Among them, judging the risk for:

[0120]

[0121] Among them, y i Represents the circuit layout x i The corresponding labels, N represents the number of circuit layouts in the sample set, loss() represents the calculation difference function, h(x i ) represents the circuit layout x i The size of the sample set covered, represents the label of the sample set, Coverage loss for the hotspot detection model:

[0122]

[0123] Loss function of hotspot detection model for:

[0124]

[0125] Among them, λ is a hyperparameter that defines the coverage weight, c0 is the target coverage, and Ψ is the square form of the ReLU function.

[0126] The following is an introduction to the hotspot detection method provided by the embodiment of the present invention. Figure 4 , Figure 4 A flow chart of a hotspot detection method provided by an embodiment of the present invention. The method may include:

[0127] S401, obtaining a circuit layout to be tested;

[0128] S402, input the circuit layout to be detected into the hotspot detection model trained by the training method of the above-mentioned hotspot detection model to perform hotspot detection and obtain the detection result.

[0129] The trained hotspot detection model provided by the embodiment of the present invention can automatically extract layout features including time domain information and frequency domain information of the circuit layout, thereby significantly improving the prediction accuracy of the layout hotspot. Based on the trained hotspot detection model, the circuit layout to be detected is predicted, thereby improving the accuracy of hotspot detection.

[0130] The embodiment of the present invention also provides a training system for a hotspot detection model, the structure of which is as follows: Figure 5 As shown, including:

[0131] An acquisition module 501 is used to acquire a sample set consisting of multiple circuit layouts and labels corresponding to the multiple circuit layouts;

[0132] The first extraction module 502 is used to input the sample set into the local feature extraction network of the hotspot detection model to be trained to extract local features;

[0133] The second extraction module 503 is used to input the local features into the global feature extraction network of the hotspot detection model to be trained to extract the global features;

[0134] A prediction module 504 is used to input the global features into a decoder of the hotspot detection model to be trained to obtain a prediction result of whether each circuit layout is a layout hotspot;

[0135] The updating module 505 is used to update the parameters of the hotspot detection model to be trained based on the difference between the prediction result and the corresponding label, so as to obtain a trained hotspot detection model.

[0136] For the specific limitations of the hotspot detection model training system, please refer to the limitations of the hotspot detection model training method above, which will not be repeated here. Each module in the above-mentioned hotspot detection model training system can be implemented in whole or in part through software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the electronic device in hardware format, or can be stored in the memory of the electronic device in software format, so that the processor can call the operations corresponding to the above modules.

[0137] It should be noted that, in order to highlight the innovative part of the present invention, the present embodiment does not introduce modules that are not closely related to solving the technical problem proposed by the present invention, but this does not mean that there are no other modules in the present embodiment.

[0138] like Figure 6 As shown, the electronic device 1 provided by the present invention may include a memory 12, a processor 13 and a bus, and may also include a computer program stored in the memory 12 and executable on the processor 13, such as a training program for a hotspot detection model.

[0139] Among them, the memory 12 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, mobile hard disk, multimedia card, card-type memory (for example: SD or DX memory, etc.), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 12 can be an internal storage unit of the electronic device 1, such as a mobile hard disk of the electronic device 1. In other embodiments, the memory 12 can also be an external storage device of the electronic device 1, such as a plug-in mobile hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the electronic device 1. Further, the memory 12 can also include both an internal storage unit of the electronic device 1 and an external storage device. The memory 12 can not only be used to store application software and various types of data installed in the electronic device 1, such as the code for training the hotspot detection model, etc., but can also be used to temporarily store data that has been output or is to be output.

[0140] In some embodiments, the processor 13 may be composed of an integrated circuit, for example, a single packaged integrated circuit, or a plurality of packaged integrated circuits with the same or different functions, including one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and a combination of various control chips. The processor 13 is the control core (Control Unit) of the electronic device 1, and uses various interfaces and lines to connect the various components of the entire electronic device 1, and executes or executes programs or modules stored in the memory 12 (such as a training program for a hotspot detection model, etc.), and calls the data stored in the memory 12 to execute various functions of the electronic device 1 and process data.

[0141] The processor 13 executes the operating system and various installed applications of the electronic device 1. The processor 13 executes the applications to implement the steps in the training method of the hotspot detection model.

[0142] Exemplarily, the computer program may be divided into one or more modules, which are stored in the memory 12 and executed by the processor 13 to complete the present application. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program in the electronic device 1. For example, the computer program may be divided into an acquisition module 501, a first extraction module 502, a second extraction module 503, a prediction module 504, and an update module 505.

[0143] The above-mentioned integrated unit implemented in the form of a software function module can be stored in a computer-readable storage medium, and the storage medium can be non-volatile or volatile. The above-mentioned software function module is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a computer device, or a network device, etc.) or a processor to perform part of the functions of the training method of the hotspot detection model described in each embodiment of the present application.

[0144] In summary, the present invention discloses a hotspot detection model training method and a hotspot detection method, which significantly improves the prediction accuracy of layout hotspots by processing the layout in the frequency domain and time domain at the same time and automatically extracting complex layout features using a deep learning model; in addition, by using the weights of the pre-trained model and the phased feature extraction strategy, the model can still have high efficiency and high accuracy when processing large-scale layouts. Therefore, the present invention effectively overcomes the various shortcomings of the prior art and has a high industrial utilization value.

[0145] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A training method for a hotspot detection model, characterized in that: include: Obtain a sample set consisting of multiple circuit layouts and labels corresponding to the multiple circuit layouts; Input the sample set into the local feature extraction network of the hotspot detection model to be trained to extract local features; Inputting the local features into the global feature extraction network of the hotspot detection model to be trained to extract global features; The global features are input into the decoder of the hotspot detection model to be trained to obtain a prediction result of whether each circuit layout is a layout hotspot; Based on the difference between the prediction result and the corresponding label, the parameters of the hotspot detection model to be trained are updated to obtain a trained hotspot detection model.

2. The method for training a hotspot detection model according to claim 1, characterized in that: The step of obtaining a sample set consisting of a plurality of circuit layouts comprises: For each circuit layout: Obtaining a time domain layout of a circuit layout; Perform discrete cosine processing on the time domain map to obtain the frequency domain map; Combining the time domain layout and the frequency domain layout into image data of different channels of the circuit layout; The image data of multiple circuit layouts are aggregated as a sample set.

3. The training method of a hotspot detection model according to claim 1, characterized in that: The step of inputting the sample set into the local feature extraction network of the hotspot detection model to be trained and extracting local features comprises: Input each circuit layout in the sample set into the block division module of the hotspot detection model to be trained for splitting to obtain multiple layout blocks of the same size; The layout block is input into the local feature extraction network of the hotspot detection model to be trained to extract local features.

4. The method for training a hotspot detection model according to claim 3, characterized in that: The step of inputting the layout block into a local feature extraction network of a hotspot detection model to be trained, and extracting local features comprises: Input each block into the linear embedding layer of the local feature extraction network, perform linear transformation on each pixel in the block, and embed the position information; Input the block with embedded position information into the multi-head self-attention layer of the local feature extraction network to extract features at different positions and obtain multi-head information of the block; The multi-head information of the layout block is input into the window-shifting self-attention layer of the local feature extraction network, the multi-head information of the layout block is divided into multiple windows, the pixels at the predetermined position of each window are removed, and self-attention calculation is performed to obtain local features.

5. The method for training a hotspot detection model according to claim 1, characterized in that: The step of inputting the local features into a global feature extraction network to extract global features comprises: Inputting the local features into the linear embedding and multi-head self-attention layers of the global feature extraction network to extract context information to obtain the first global feature; Performing a linear transformation on the first global feature through a learnable weight matrix to obtain a query feature; Mapping the query feature to the dimension of a predefined reference point, and performing nonlinear transformation and convolution processing to generate shape shift key position and value information; Attention calculation is performed on the query feature, the shape-shifted key position and the value information to obtain a global feature.

6. The method for training a hotspot detection model according to claim 1, characterized in that: The step of inputting the global features into the decoder of the hotspot detection model to be trained to obtain a prediction result of whether each circuit layout is a layout hotspot comprises: The global features are input into the inter-layer regularization layer of the decoder for regularization processing to obtain regularized features; The regular features are input into the global pooling layer of the decoder for global average pooling to obtain the pooled features; The pooled features are extracted through the fully connected layer of the decoder to obtain fully connected features; The fully connected features are processed by activation functions to obtain nonlinear features; Based on a preset threshold, the nonlinear features are binarized to obtain a prediction result of whether each circuit layout is a layout hotspot.

7. The method for training a hotspot detection model according to claim 1, characterized in that: The step of updating the parameters of the hotspot detection model to be trained based on the difference between the prediction result and the corresponding label includes: Based on the difference between the prediction results and the corresponding labels, the parameters of the hotspot detection model to be trained are updated according to a predefined loss function; wherein the loss function includes hyperparameters for judging risk, coverage loss and defining coverage weights.

8. A hotspot detection method, characterized in that: include: Obtaining a circuit layout to be tested; The circuit layout to be detected is input into the trained hotspot detection model obtained by the training method of any one of claims 1 to 7 to perform hotspot detection and obtain a detection result.

9. A training system for a hotspot detection model, characterized in that: include: An acquisition module, used to acquire a sample set consisting of multiple circuit layouts and labels corresponding to the multiple circuit layouts; A first extraction module is used to input the sample set into a local feature extraction network of a hotspot detection model to be trained to extract local features; A second extraction module is used to input the local features into the global feature extraction network of the hotspot detection model to be trained to extract the global features; A prediction module, used for inputting global features into a decoder of a hotspot detection model to be trained to obtain a prediction result of whether each circuit layout is a layout hotspot; The updating module is used to update the parameters of the hotspot detection model to be trained based on the difference between the prediction result and the corresponding label to obtain the trained hotspot detection model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the training method of the hotspot detection model as described in any one of claims 1 to 7 are implemented.

Images