A lithography hotspot detection method and system based on deep learning LHD model

Abstract

The present invention discloses a lithography hotspot detection method and system based on a deep learning LHD model. The present invention proposes a lightweight hotspot detection model LHD, which reduces the complexity of the model structure and the number of parameters, uses multi-scale convolution kernels to process different types of hotspots, and introduces appropriate pooling layers and batch normalization layers to accelerate training and improve the robustness of the model. The model LHD of the present invention has high adaptability and accuracy, and can effectively solve the problems of long detection time of lithography simulation and the inability of pattern matching to identify unknown hotspots. At the same time, because the deep learning model can automatically extract image features, the process of manual extraction is omitted, the training process is simplified, and the detection effect is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical proximity correction of microelectronic layout data, and relates to a lithography hotspot detection method and system based on a deep learning LHD model. Background Art

[0002] As chip size continues to shrink and graphic structures become denser, photolithography hotspots have become one of the manufacturability issues that cannot be ignored in chip production. The presence of photolithography hotspots can cause problems such as short circuits and open circuits in chips, seriously threatening the stability and safety of chips, thereby reducing the yield of chips. Therefore, how to effectively detect photolithography hotspots has become a key issue for the industry. Hotspot detection generally uses photolithography simulation to write OPC (optical proximity correction) verification scripts to define rules to capture the graphic features of the hotspot area, thereby locating the hotspot and completing the detection.

[0003] Traditional hotspot detection methods based on lithography simulation require a lot of time and computing resources, and there are widespread false positives and missed negatives. With the development of integrated circuits, lithography simulation models and layouts are becoming more and more complex, and traditional methods can no longer meet the needs. In contrast, traditional machine learning and pattern matching-based methods have certain advantages in lithography hotspot detection, but their performance will also decline as the transistor layout pattern becomes more complex and the sample size increases. Therefore, it is necessary to explore new methods to solve the problem of lithography hotspot detection.

[0004] To this end, the present invention provides a new lithography hotspot detection method based on a deep learning model to improve the hotspot detection performance. Summary of the invention

[0005] The purpose of the present invention is to provide a lithography hotspot detection method based on a deep learning LHD model in view of the deficiencies in the prior art.

[0006] In a first aspect, the present invention provides a lithography hotspot detection method based on a deep learning LHD model, comprising:

[0007] Step (1), obtaining a labeled layout lithography image and constructing a data set; dividing the data set into a training set and a test set according to a ratio;

[0008] Extract various features of the pattern from the layout lithography image using the SONR tool from at least one of layout size, pattern density, DFM (manufacturability) attributes, inter-layer features, contour size, image intensity measurement and OPC mask measurement; the features extracted from the pattern density angle include density features of Tophat kernels and Gaussian kernels with kernel sizes ranging from nanometers to millimeters; the features extracted from the inter-layer feature angle include with or without directional information, layer attributes; the features extracted from the layout size angle include geometric features of a single or multiple different layers in the layout, where different layers include target layers, re-target layers, and OPC layers; the features extracted from the contour size angle include process features of the lithography model; the features extracted from the image intensity measurement and OPC mask measurement angles include user-defined features of png images;

[0009] Labeling the layout lithography image according to all the above characteristics, wherein the label is a hot spot or a non-hot spot; the hot spots include a standard bridge type, a variable bridge type, a standard shrinkage type, a variable shrinkage type, and a line end shrinkage type;

[0010] The ratio of data labeled as non-hotspot and hotspot in the dataset is greater than 1;

[0011] The ratio of data with the labels of bridge standard type, bridge variation type, contraction standard type, contraction variation type, and line end contraction type in the data set is (3-5):1:(11-14):(11-14):(11-14), preferably 3.9:1:12.6:12.5:12.5;

[0012] Step (2), building a lightweight lithography hotspot detection model LHD, and using the data set and test set for training and testing;

[0013] The lightweight lithography hotspot detection model LHD is composed of a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a third maximum pooling layer, a fourth convolutional layer, a fifth convolutional layer, a fourth maximum pooling layer, a sixth convolutional layer, a seventh convolutional layer, an average pooling layer, a first fully connected layer, and a second fully connected layer in sequence, and the number of neurons in the convolutional layer is reduced, and a dropout algorithm is used to randomly discard neurons to reduce the false alarm rate and reduce overfitting;

[0014] Step (3), building a balanced classifier, wherein the input of the balanced classifier is the output of the lightweight lithography hotspot detection model LHD;

[0015] The balancing classifier includes first to n-th classifiers, each of which corresponds to a label, and n is the total number of label types;

[0016] The initial weight of each classifier is the proportion of the classifier's label images in the dataset;

[0017] Step (4), using the trained and verified lightweight lithography hotspot detection model LHD and the balanced classifier to realize hotspot detection of the layout lithography image.

[0018] In a second aspect, the present invention provides a lithography hotspot detection system for implementing the method, comprising:

[0019] A data acquisition module, which acquires the layout lithography image;

[0020] The detection module uses the trained and verified lightweight lithography hotspot detection model LHD and the balanced classifier to realize hotspot detection of the layout lithography image.

[0021] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method described.

[0022] In a fourth aspect, the present invention provides a computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the described method is implemented.

[0023] The present invention proposes a lightweight hotspot detection model LHD, which reduces the complexity of the model structure and the number of parameters, uses multi-scale convolution kernels to process different types of hotspots, and introduces appropriate pooling layers and batch normalization layers to accelerate training and improve the robustness of the model. The model is trained using the stochastic gradient descent method, which first transforms the two-dimensional image and then performs a sliding window scan on the two-dimensional image to obtain a hotspot probability map. The judgment probability is continuously increased to 1 through the back propagation algorithm and the optimization method of the loss function, thereby improving the model detection accuracy and bringing excellent hotspot detection performance.

[0024] The present invention proposes a balanced classifier design of six classifiers, reduces the number of neurons in the convolutional layer, and uses a dropout algorithm to randomly discard neurons to reduce the false alarm rate, reduce overfitting, and enhance the generalization ability of the model to improve the accuracy of the test set. The weight parameters of the neural network model are optimized through training data, and the weight update is achieved through the gradient descent method, so that it can better fit the data and achieve the prediction task. Therefore, the present invention has high adaptability and accuracy, and can effectively solve the problems of long detection time of lithography simulation and the inability of pattern matching to identify unknown hotspots. At the same time, because the deep learning model can automatically extract image features, the process of manual extraction is omitted, the training process is simplified, and the detection effect is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 To adjust the weight classifier process;

[0026] Figure 2This is the LHD model network architecture flow chart;

[0027] Figure 3(a) is a schematic diagram of the LHD hotspot detection process;

[0028] Figure 3(b) is a schematic diagram of the LHD model network structure;

[0029] Figure 4 It is a dataset of hot spots and non-hot spots of CMOS poly layers;

[0030] Figure 5 is the specific category proportion of the data set;

[0031] Figure 6 is the accuracy confusion matrix of the LHD model;

[0032] Figure 7(a) shows the accuracy comparison;

[0033] Figure 7(b) shows the comparison of parameter weight. DETAILED DESCRIPTION

[0034] The present invention is further analyzed below in conjunction with specific embodiments.

[0035] The input of the lithography hotspot detection model is the lithography image, which may contain hotspot areas. Figure 2 The convolutional neural network shown in the figure extracts features from the image and predicts the hotspot area through a series of convolutional layers, pooling layers, and fully connected layers. The convolutional layer uses a learnable convolution kernel to extract local features from the input image and capture the spatial relationship of the hotspot area. The pooling layer is used to reduce the dimension of the feature map and retain the main feature information. Finally, the fully connected layer maps the extracted features to the output layer, and the probability prediction of the hotspot is obtained through the sigmoid activation function. The training of the model depends on the labeled lithography image dataset. The error between the predicted result and the true label is calculated and optimized using the loss function. After training, the model can predict the hotspot area of ​​the new lithography image. The threshold can be used to determine whether it is a hotspot area. The deep learning model for lithography hotspot detection is based on the characteristics of convolutional neural networks and realizes the detection of hotspot areas in lithography images through feature extraction and prediction. The model is trained on a large-scale dataset and provides efficient and accurate hotspot detection capabilities.

[0036] Although complex network structures such as convnext have deep layers and can be applied to a wider range of data sets, the excessive number of parameters will cause overfitting problems, and the actual detection accuracy is not high. The lithography hotspot dataset is different from natural object recognition. The model needs to have strong depth and generalization capabilities. Efficiency and accuracy are the goals of the hotspot detection task of the present invention. Therefore, a lightweight lithography hotspot detection model LHD (lithography hotspot detection) is proposed to complete the specific classification detection of hotspot sets and non-hotspot sets. At the same time, the design of a balancing weight classifier reduces the overfitting problem, reduces the number of parameters, and achieves very good accuracy results. At the same time, the training time of the model is relatively short. Overall, it is very suitable for lithography hotspot detection tasks.

[0037] The basic implementation steps can be summarized as follows:

[0038] 1) The graphics of the polysilicon layer poly target layer are cropped and converted into images as a data set, which are marked into 5 types of hot spots and 1 type of non-hot spots according to the simulation results.

[0039] 2) A lightweight hotspot detection model (LHD model) is proposed, which takes 80% of the data as input to train the model and uses 20% of the data as test to predict the hotspot type.

[0040] 3) Improve the hotspot detection effect through data optimization and model optimization. The LHD network structure constructed by the present invention is as follows Figure 2 As shown, the process of hotspot detection method based on deep learning is shown in Figure 3(a).

[0041] 1) Hotspot detection method

[0042] The convolutional neural network takes the image as input, realizes image feature extraction through the convolution kernel of the convolution layer, and obtains the corresponding output after combining the extracted features. The layout data can be represented as graphic data, and the convolutional neural network is suitable for hotspot detection. The principle of the hotspot detection method based on the convolutional neural network is shown in Figure 3 (a), including a model training stage and a hotspot detection stage. In the model training stage, we use a labeled layout graphic data set to train the model. Through the back propagation algorithm and optimization method, the model can automatically learn the feature extraction ability suitable for hotspot detection. This is achieved by adjusting the convolution kernel parameters in the convolution layer, so that the model can effectively capture the features in the image. In the hotspot detection stage, we use the trained model to perform hotspot detection on the layout graphics to be tested. By inputting the image into the model, after a series of convolution, pooling and fully connected layer operations, an output result is finally obtained. According to the threshold judgment of the output result, we can determine whether there is a hotspot in the image. The present invention uses a convolutional neural network to detect and classify the graphic data into hotspots and non-hotspots, and then reclassifies the hotspot data to obtain 5 types of hotspots.

[0043] 2) Balancing classifier design

[0044] The present invention sets weights according to the proportion of each data set, and classifies each hotspot set and non-hotspot set into single-class images to obtain classifiers at each level. Figure 1 , reduce the false alarm rate and overfitting at the same time, and achieve the best classification effect. For the first type of bridging standard hotspot set, there are 64 hotspot images. The weight of its proportion in all image sets is set to 0.02 for single-class image training to obtain the first type of classifier. Similarly, the weight of the second type of bridging change hotspot set is set to 0.005, the weight of the third type of shrinkage standard hotspot set is set to 0.065, the weight of the fourth type of shrinkage change hotspot set is set to 0.065, the weight of the fifth type of line end shrinkage hotspot set is set to 0.065, and the weight of the sixth type of non-hotspot standard set is set to 0.78. The present invention solves the problem of multi-hotspot classification by adjusting the weights.

[0045] The weight parameters of the neural network model are optimized through training data so that it can better fit the data and achieve the prediction task. The weight update method is mainly achieved through the gradient descent method: Forward propagation: The input data is forward propagated through the neural network model. Each layer performs a series of linear and nonlinear calculation operations to generate the final output result. Calculate the loss function: Compare the output result of the model with the true label to calculate the loss function value of the model under the current parameter setting. Common loss functions include mean square error (MSE), cross entropy, etc. Backward propagation: According to the loss function, the gradient of each parameter to the loss function is calculated layer by layer through the chain rule. The back propagation process starts from the output layer of the model and reversely calculates the gradient layer by layer until it reaches the input layer. Weight update: According to the gradient calculated by back propagation, the network parameters are updated using an optimization algorithm (such as the gradient descent method). The basic idea of ​​the gradient descent method is to update the parameters in the opposite direction of the gradient to minimize the loss function.

[0046] 3) Network structure

[0047] Although complex network structures such as convnext have deep layers and can be applied to a wider range of data sets, they have problems with the number of parameters and overfitting, and the actual detection accuracy is not high. Therefore, for the hotspot detection data set of the present invention, a new hotspot detection network model is proposed: the LHD model. As shown in Figure 3(b), the LHD model consists of a 10-layer network, including 7 convolutional layers and 3 fully connected layers, and the number of neurons in the convolutional layer is reduced. The dropout algorithm is used to randomly discard neurons to reduce the false alarm rate and reduce overfitting.

[0048] The lightweight lithography hotspot detection model LHD is composed of a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a third maximum pooling layer, a fourth convolutional layer, a fifth convolutional layer, a fourth maximum pooling layer, a sixth convolutional layer, a seventh convolutional layer, an average pooling layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer in sequence, and the number of neurons in the convolutional layer is reduced, and a dropout algorithm is used to randomly discard neurons to reduce the false alarm rate and reduce overfitting;

[0049] The convolution kernels of the first to seventh convolutional layers are 64, 64, 128, 256, 256, 512, 512 respectively. Between every two layers, there is a pooling layer, which uses a 2×2 pooling matrix (maximum pooling) and a shift of 2 steps. The average pooling layer is used to flatten the feature map. After convolution block 5 (conv3-512), the feature map is flattened into a 1×1×512 feature vector through the average pooling layer, and finally mapped to a 1×1×6 output layer through three fully connected layers. The number of convolutional layers in the convolutional block is 1, 1, 1, 2, 2, for a total of 7 convolutional layers. Repeated experiments have shown that this structure has the best detection effect while maintaining stable convergence.

[0050] 4) Evaluation function

[0051] In deep learning classification models, commonly used evaluation indicators include accuracy, recall, precision and F1 score. The present invention uses these four evaluation functions to evaluate the performance of the lhd model in hotspot detection and conducts comparative experiments with the traditional classification model convnext.

[0052] The results of hotspot detection are divided into the following cases: True positive (TP): a hotspot image is identified as a hotspot image. True negative (TN): a non-hotspot image is identified as a non-hotspot image. False positive (FP): a non-hotspot image is identified as a hotspot image. False negative (FN): a hotspot image is identified as a non-hotspot image. The accuracy, recall, precision and F1 score are defined as follows:

[0053] accuracy=(TP+TN) / (TP+TN+FP+FN) (1)

[0054] recall=TP / (TP+FN) (2)

[0055] precision=TP / (TP+FP) (3)

[0056] F1=2×(precision×recall) / (precision+recall) (4)

[0057] Among them, the accuracy rate represents the ratio between the number of hotspot areas correctly predicted by the model and the total number of samples. The recall rate represents the ratio between the number of hotspot areas correctly predicted by the model and the actual number of hotspot areas. The precision rate represents the ratio between the number of hotspot areas correctly predicted by the model and the total number of hotspot areas predicted by the model. The F1 score is the harmonic mean of the recall rate and the precision rate. The values ​​of the accuracy rate, recall rate, precision rate and F1 score are all between 0 and 1. The larger the value, the better the model performance.

[0058] 5) Dataset preparation

[0059] In the chip manufacturing process, photolithography hotspot detection is crucial to ensure the quality of the chip, especially in the hotspot detection of the silicon gate layer, which has a significant impact on the yield of the chip. In order to train and evaluate the model, we collected 3064 silicon gate layer layout data from a 12-inch 55nm wafer factory as the data set for this experiment. Figure 4 As shown, the resolution of each image is 224×224, and we use a ratio of 8:2 between the training set and the test set. The data collection object is the 55nm CMOS polysilicon gate layer layout image. The mentor tool SNOR is used to sample the hotspot and non-hotspot graphics and prepare the input data set. The SONR tool extracts various features of the pattern from layout size, pattern density, DFM (manufacturability) attributes, inter-layer features, contour size, image intensity measurement, and OPC mask measurement, thereby achieving the sampling of hotspot patterns and non-hotspot patterns. In order to ensure the accuracy of the test results, we ensure that there is no overlap between the test set and the training set, so that the generalization ability of the model can be better evaluated.

[0060] Among these data, there are 680 hotspot maps and 2384 non-hotspot maps. The hotspot set includes the following five categories: standard bridge type, bridge variation type, standard contraction type, contraction variation type, line end contraction type and non-hotspot set, with a number of more than 200,000. Figure 5 As shown in the figure, the proportion of the above data sets is 63:16:201:200:200:2384, which is completely in accordance with the hot spot distribution ratio of the actual layout to reflect the real situation. Among them, the standard bridge type and the standard shrinkage type respectively represent the bridge problem caused by the adhesion between the two lines under the normal process conditions of the production line and the pinch-off problem caused by the shrinkage of the middle section or the end of the line; the variable bridge type and the variable shrinkage type respectively represent the bridge and shrinkage and pinch-off problems under the process conditions of 50nm defocus and 5% dose; the line end shrinkage type is the shrinkage problem at the end of the line under the normal conditions of the production line.

[0061] The non-hotspot set is the layout image without any circuit defects.

[0062] Step 6: Experimental Results

[0063] This experiment uses NVIDIA GeForce RTX4090, Windows 11 operating system, the network model is coded in Python3.8, and Pytorch is used as our deep learning framework. The network uses the SGD optimizer for 200 iterations, with an initial learning rate of 0.001, a weight decay of 0.0001, and a momentum of 0.9.

[0064] In order to intuitively evaluate the detection effect of the LHD network on various types of hotspot and non-hotspot graphics, this experiment produced a confusion matrix in tabular form. The values ​​in the table represent the probability that the model identifies the true type as the predicted type. From the elements on the diagonal, it can be seen that the LHD model has achieved a high accuracy rate on the experimental dataset. Figure 6 The confusion matrix results show that the LHD model performs excellently, with accuracy rates of 98.84%, 100%, 94.19%, 98.81%, 82.71% and 73.59% on the bridging standard set, bridging variation set, shrinkage standard set, shrinkage variation set, line end shrinkage set and standard non-hotspot set, respectively.

[0065] The present invention compares the performance of the LHD model and the classic classification model convnext as shown in Figure 7(a). The results show that the accuracy of the convnext network used in the hotspot detection scenario begins to converge after 133 generations, and the converged accuracy is 90.18%. Compared with convnext, the accuracy of the LHD network is 5.07% higher, reflecting higher performance. In addition, the parameter comparison in Figure 7(b) shows that the LHD model has excellent lightness. The parameters of the two models are calculated separately, and the parameters of the LHD model are 22131014, and the parameters of the convnext model are 197734120. The LHD model is more streamlined, and the parameters are relatively reduced by 88.8%.

Claims

1. A lithography hotspot detection method based on deep learning LHD model, characterized in that include: Step (1), obtaining a labeled layout lithography image and constructing a data set; dividing the data set into a training set and a test set according to a ratio; Step (2), building a lightweight lithography hotspot detection model LHD, and using the data set and test set for training and testing; The lightweight lithography hotspot detection model LHD is composed of a first convolutional layer, a first maximum pooling layer, a second convolutional layer, a second maximum pooling layer, a third convolutional layer, a third maximum pooling layer, a fourth convolutional layer, a fifth convolutional layer, a fourth maximum pooling layer, a sixth convolutional layer, a seventh convolutional layer, an average pooling layer, a first fully connected layer, and a second fully connected layer in sequence, and the number of neurons in the convolutional layer is reduced, and a dropout algorithm is used to randomly discard neurons to reduce the false alarm rate and reduce overfitting; Step (3), building a balanced classifier, wherein the input of the balanced classifier is the output of the lightweight lithography hotspot detection model LHD; The balancing classifier includes first to n-th classifiers, each of which corresponds to a label, and n is the total number of label types; The initial weight of each classifier is the proportion of the classifier's label images in the dataset; Step (4), using the trained and verified lightweight lithography hotspot detection model LHD and the balanced classifier to realize hotspot detection of the layout lithography image.

2. The method according to claim 1, characterized in that: The construction process of the tag: Extract various characteristics of the pattern from the layout lithography image using the SONR tool from at least one of layout size, pattern density, DFM attribute, inter-layer feature, contour size, image intensity measurement, and OPC mask measurement; The layout lithography image is labeled according to all the features, and the labels are hot spots or non-hot spots; the hot spots include standard bridge type, variable bridge type, standard shrinkage type, variable shrinkage type, and line end shrinkage type.

3. The method according to claim 2, characterized in that: The features extracted from the pattern density angle include density features of Tophat kernels and Gaussian kernels with kernel sizes ranging from nanometers to millimeters; The features extracted from the inter-layer feature angle include the features with or without directional information, layer attributes; Features extracted from the layout dimension angle include geometric features of a single or multiple different layers in the layout, where the different layers include a target layer, a retarget layer, and an OPC layer; Features extracted from the profile dimension angle include process features of the lithography model; Image intensity measurement and OPC mask measurement angle extracted features including user defined features for png images.

4. The method according to claim 2, characterized in that: The ratio of data labeled as non-hotspot and hotspot in the dataset is greater than 1.

5. The method according to claim 2, characterized in that: The proportion of data in the dataset labeled as bridge standard type, bridge variation type, contraction standard type, contraction variation type, and line end contraction type is (3-5):1:(11-14):(11-14):(11-14).

6. The method according to claim 5, characterized in that: The data ratios of the labels bridge standard type, bridge variable type, contraction standard type, contraction variable type, and line end contraction type are 3.9:1:12.6:12.5:12.

5.

7. The method according to claim 1, characterized in that: The convolution kernels of the first to seventh convolutional layers are 64, 64, 128, 256, 256, 512, and 512 respectively; The first to fourth maximum pooling layers all use a 2×2 pooling matrix and a shift stride of 2.

8. A lithography hotspot detection system implementing the method according to any one of claims 1 to 7, characterized in that include: A data acquisition module, which acquires the layout lithography image; The detection module uses the trained and verified lightweight lithography hotspot detection model LHD and the balanced classifier to realize hotspot detection of the layout lithography image.

9. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 7.

10. A computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method according to any one of claims 1 to 7 is implemented.

Images