Semiconductor equipment wafer defect identification method based on hybrid quantum Al algorithm

Abstract

The present invention relates to the field of semiconductor chip technology, and in particular to a semiconductor equipment wafer defect recognition method based on a hybrid quantum Al algorithm. The steps are as follows: S1: Select different defect recognition and classification methods according to the context in the semiconductor process; S2: When a high-throughput process site appears in semiconductor manufacturing, select a self-proliferation and attention neural network; S3: When a low-throughput process site appears in semiconductor manufacturing, select a hybrid quantum-classical deep learning model. The semiconductor equipment wafer defect recognition method based on the hybrid quantum Al algorithm provided by the present invention can perform defect classification in a more efficient manner through improvement, can generate more feature maps with lower computational cost through a self-proliferation mechanism, and can capture the long-range correlation of feature maps through a self-attention mechanism, through a channel-mode and a spatial-mode attention mechanism, and can be applied to various defect mode classification tasks.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor chip technology, and in particular to a semiconductor equipment wafer defect identification method based on a hybrid quantum Al algorithm. Background Art

[0002] With the booming development of the Internet of Things and artificial intelligence, the demand for semiconductor chips has increased significantly. The large-scale expansion of semiconductor manufacturing and the development of new technologies will cause more defective wafers. These defective wafers will not only cause cost losses to chip manufacturers, but also lose the trust of customers. In addition, these undetected defective wafers will have an impact on our environment in subsequent invalid semiconductor processes, such as energy consumption and carbon dioxide emissions.

[0003] The semiconductor industry is highly competitive, and the primary focus of wafer manufacturing companies is to reduce manufacturing costs and improve product quality and reliability. Lower costs and higher revenues can be achieved by improving wafer yield and maximizing effective throughput. Semiconductor manufacturing is a complex and time-consuming process involving hundreds of manufacturing procedures. When variations occur in these manufacturing processes, they cause defects on the surface of the chip. These defects come from various variations in the semiconductor manufacturing process and cause huge losses to the product yield.

[0004] The single electronic components of an integrated circuit (IC) are so small that their production requires atomic-level precision. ICs are manufactured by creating circuits on chips made of very pure semiconductor materials (usually silicon) and using wire interconnect structures. In order to produce high-density ICs, the surface of the wafer must be very clean, and each layer of circuits manufactured on the wafer must be aligned with each other.

[0005] Automatic Defect Classification (ADC) uses electron microscopes to identify and classify defect patterns, providing yield engineers with important information to help them determine the root cause of wafer defects. Automated Optical Inspection (AOI) is used to visualize defect patterns and determine the root cause of chip failures. AOI can replace a large number of manual inspections with high-speed and accurate inspection technology, achieve inspection consistency, shorten inspection time, and improve product quality and competitiveness. Defects are judged from the features detected by AOI, but the ultimate goal is to determine whether the defects on the wafer are real defects or pseudo-defects. Managers need to determine the defect type and propose solutions. However, the current AOI requires manual subsequent final verification to determine the type of defect. This non-automated inspection method will cause the Manufacturing Execution System (MES) to be unable to immediately take preventive measures and reduce the losses caused by process variation.

[0006] The ability to detect defects affects the yield and productivity of products. False negatives in defect detection will cause defective wafers to be regarded as qualified wafers and shipped to customers, causing losses to customers' products and affecting the credit of wafer manufacturing companies. False positives will cause non-defective wafers to be regarded as abnormal, increase the additional review work of yield engineers, reduce production efficiency and affect the production cycle. Therefore, how to reduce false negatives and false positives is the main challenge of defect detection.

[0007] To this end, a semiconductor equipment wafer defect identification method based on a hybrid quantum Al algorithm is designed to provide another technical solution to the above technical problems. Summary of the invention

[0008] Based on this, it is necessary to provide a semiconductor equipment wafer defect identification method based on a hybrid quantum Al algorithm to solve the technical problems raised in the above background technology.

[0009] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0010] The semiconductor equipment wafer defect identification method based on the hybrid quantum Al algorithm has the following steps:

[0011] S1: Select different defect identification and classification methods according to the context in the semiconductor process;

[0012] S2: When high-throughput process sites appear in semiconductor manufacturing, self-proliferation and attention neural networks are selected;

[0013] S3: When low-throughput process sites appear in semiconductor manufacturing, choose a hybrid quantum-classical deep learning model.

[0014] As a preferred implementation of the semiconductor equipment wafer defect recognition method based on the hybrid quantum Al algorithm provided by the present invention, in the step S2, the self-propagation and attention neural network is composed of a self-propagation module and a self-attention module;

[0015] The self-propagation module generates more feature maps at a lower cost through linear transformation;

[0016] The self-attention module learns information about long-range dependencies from the features generated by the self-propagation module.

[0017] As a preferred implementation of the semiconductor equipment wafer defect identification method based on the hybrid quantum Al algorithm provided by the present invention, the self-propagation module generates the same number of features through linear transformation, and the expression is as follows:

[0018]

[0019] Among them, m′ i is the basic unit of the feature map M′; h and w are the new feature maps u i,j The height and width of ; represents a low-cost feature map generation operation.

[0020] As a preferred implementation of the semiconductor equipment wafer defect identification method based on the hybrid quantum Al algorithm provided by the present invention, the self-attention module includes a channel attention mechanism and a spatial attention mechanism;

[0021] The channel attention mechanism focuses on feature patterns that are highly relevant to the output results by focusing on the input feature map and activating part of the channel;

[0022] The spatial attention mechanism enhances the features at the query position by aggregating information from other feature map positions.

[0023] As a preferred implementation of the semiconductor equipment wafer defect identification method based on the hybrid quantum AI algorithm provided by the present invention, the architecture of the channel attention mechanism is composed of the following steps:

[0024] The first is Squeeze, which flattens the 3D Feature Map into a 1D tensor through Global Average Pooling;

[0025] The second is Excitation, which captures the importance of channels through bottleneck transformation;

[0026] The third is the fusion function, which is used to propagate the element-wise multiplication of feature fusion.

[0027] As a preferred implementation of the semiconductor equipment wafer defect recognition method based on the hybrid quantum Al algorithm provided by the present invention, the architecture expression of the spatial attention mechanism is as follows:

[0028]

[0029] where h and w are the input x i The height and width, x i Represented as the feature map of the input instance; w e1 and w e2 represents the linear transformation matrix used for bottleneck transformation; w g is the weight of global attention; LN represents layer normalization.

[0030] As a preferred implementation of the semiconductor equipment wafer defect identification method based on the hybrid quantum AI algorithm provided by the present invention, the hybrid quantum-classical deep learning model includes the following three parts:

[0031] The first part is the classic convolutional neural network, which is mainly used for feature extraction;

[0032] The second part is the quantum neural network, which consists of a coding network, a parameterized quantum circuit, and a measurement layer;

[0033] The third part is a classical neural network that performs post-processing to generate predictions so that the final measurement value of the quantum neural network can be used for the classification task.

[0034] As a preferred implementation of the semiconductor equipment wafer defect recognition method based on the hybrid quantum Al algorithm provided by the present invention, a self-propagation module is used for feature extraction of a classical convolutional neural network;

[0035] The self-proliferation module consists of a compression layer, a proliferation layer and a connection layer;

[0036] The compression layer is a small number of convolutional layers, using Conv2D 1×1 convolution to reduce the output channels by half;

[0037] The proliferation layer is a linear transformation layer, which generates the other half of the feature map through linear transformation;

[0038] The connection layer is used to merge the feature maps from the compression layer and the proliferation layer.

[0039] As a preferred implementation of the semiconductor equipment wafer defect identification method based on the hybrid quantum Al algorithm provided by the present invention, the quantum neural network consists of three consecutive parts, namely:

[0040] An encoding circuit for encoding classical data into the state of a quantum bit so as to input the classical data into a quantum neural network for operation;

[0041] parameterized quantum circuits for states transitioning to their optimal positions on Hilbert space;

[0042] The measurement layer is used to prepare quantum states and perform measurements in accordance with the provisions of PQCs by the quantum computer, while allowing the measurement results to be subsequently processed by the classical computer to predict and classify wafer defects.

[0043] As a preferred implementation of the semiconductor equipment wafer defect recognition method based on the hybrid quantum Al algorithm provided by the present invention, the classical neural network consists of three parts, as follows:

[0044] First, the output state is converted from the quantum layer to the input of the classical layer;

[0045] Secondly, let the neural network choose the linear rectification function as the activation function;

[0046] Finally, the Softmax activation function is used to calculate the representative probability values ​​of various defect types.

[0047] It can be seen without a doubt that the above-mentioned technical solution of the present application can definitely solve the technical problem to be solved by the present application.

[0048] At the same time, through the above technical solutions, the present invention has at least the following beneficial effects:

[0049] The semiconductor equipment wafer defect recognition method based on the hybrid quantum Al algorithm provided by the present invention can perform defect classification in a more efficient way through improvement, can generate more feature maps with lower computational cost through a self-propagation mechanism using a series of linear transformations, and can capture the long-range correlation of feature maps through a self-attention mechanism through a channel-wise and spatial-wise attention mechanism, and can be applied to various defect pattern classification tasks. Compared with other state-of-the-art methods, it can have higher accuracy and lower computational cost in many defect detection tasks;

[0050] A hybrid quantum-classical deep learning consisting of a quantum layer and a classical layer is proposed. Feature extraction is performed on the classical layer through a combination of self-propagation and self-attention modules. The self-propagation module uses a series of linear transformations to generate more features with lower computational cost. The self-attention module then learns a lot of information about long-range dependencies from the rich features generated by the self-propagation module, thereby being able to use a model with fewer parameters to achieve more accurate defect judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0052] Figure 1 The self-value-added block architecture diagram of the present invention;

[0053] Figure 2 The channel attention block architecture diagram of the present invention;

[0054] Figure 3 This is a diagram of the self-attention block architecture of the present invention;

[0055] Figure 4 The self-value-added and self-attention block architecture diagram of the present invention;

[0056] Figure 5 A schematic diagram of the self-value-added and self-attention algorithm of the present invention;

[0057] Figure 6 This is a diagram of the hybrid quantum-classical deep learning architecture of the present invention;

[0058] Figure 7 The hybrid quantum-classical deep learning model-self-propagation block of the present invention;

[0059] Figure 8 The hybrid quantum-classical deep learning model-self-proliferation module of the present invention;

[0060] Fig. 9 The hybrid quantum-classical deep learning model of the present invention - quantum neural network;

[0061] Fig.10 The loss function and architecture diagram of the hybrid quantum-classical deep learning model of the present invention;

[0062] Fig.11 Schematic diagram of the cosine decay learning rate distribution of the present invention;

[0063] Fig.12 A schematic diagram of an image transformation case of the present invention;

[0064] Fig.13 Schematic diagram illustrating the quantum logic gate of the present invention. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0066] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings.

[0067] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features and technical solutions in the embodiments may be combined with each other.

[0068] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0069] Reference Figure 1-Figure 13 , a semiconductor equipment wafer defect identification method based on a hybrid quantum Al algorithm.

[0070] Choose different defect identification and classification methods based on the context in the semiconductor process;

[0071] When high-throughput process sites appear in semiconductor manufacturing, the Self-Proliferation-and-Attention Neural Network (SP&A-Net) is selected to meet the needs of defect inspection sites that require real-time computing and rapid response, avoiding continuous processing of defective processes and resulting in more defective products.

[0072] The Self-Proliferation and Attention Neural Network consists of two main parts: the Self-Proliferation Block and the Self-Attention Block (SA-Block). The Self-Proliferation Block uses a series of linear transformations to generate more feature maps at a lower cost. The Self-Attention Block learns a lot of information about long-range dependencies from these generated features.

[0073] Self-Proliferation Block (SP-Block), which self-proliferates enough feature maps to allow the model to fully learn the relationship between features to improve the accuracy of the classifier;

[0074] The traditional convolution process increases the depth of features through a series of convolution operations, as shown in formula (1). The self-proliferation module generates the same number of features through linear transformation, as shown in formula (2). This process is similar to the decompression and replication of DNA, which can effectively increase the number of features (such as Figure 1 The first step of self-propagation is a small number of traditional convolution operations, which can be built on the transformation The input feature map can be Mapped to a new feature map of depth n

[0075]

[0076] Where c is the number of input channels; h and w are the height and width of the input feature map; h' and w' are the height and width of the output feature map. Here the symbol * represents the convolution operation;

[0077] The second step is a low-cost feature map generation operation, represented by the notation in formula (2). This is a linear transformation to generate the other half of the feature map.

[0078]

[0079] Among them, m i ' is the basic unit of the feature map M' generated by equation (1); h and w are the new feature maps u i,j If 2n convolution kernels are used, n convolution kernels can be used for self-propagation. Therefore, the computational cost can be reduced by half (e.g. Figure 1 shown).

[0080] The Self-Attention Block includes a channel attention mechanism and a spatial attention mechanism. The channel attention mechanism focuses on the input feature map and activates a part of the channel, allowing the model to focus on feature patterns that are highly relevant to the output results. Dynamic channel features are executed through the network to recalibrate feature extraction to improve the network's recognition ability. The spatial attention mechanism helps the model understand the visual scene globally, and can model the long-distance dependency between pixels in the feature map to extract a global understanding of the visual scene.

[0081] The present invention refers to Squeeze-and-Excitation Network [Jie Hu,Li Shen,Samuel Albanie,and Gang Sun, "Squeeze-and-Excitation Networks”,2018 IEEE / CVF Conference on Computer Vision and Pattern Recognition(CVPR),pp.7132- 7141,Jun.2018] A Channel-Wise Attention block is set in each Convolution process to enhance the representation of important features. Its architecture mainly consists of three parts (such as Figure 2 shown):

[0082] (a) First, Squeeze, using Global Average Pooling to flatten the 3D Feature Map into a 1D tensor;

[0083] (b) The second is Excitation, which captures the importance of channels through the bottleneck transform. Figure 2 From the perspective of , it represents the ReLU activation function, which reduces the dimension by one factor r in the middle to improve the computing efficiency;

[0084] (c) is represented as the fusion function, which is used to propagate the element-wise multiplication of feature fusion. After activation by Sigmoid Function(), the result after Excitation is converted into a probability value between 0 and 1, which represents the importance of the original Feature Map. Both parameters W1 and W2 can be updated during the Backpropagation process.

[0085] In short, the architecture of Channel-Wise Attention not only allows the network to learn the importance of Feature Map, but also has the function of selecting Feature Map.

[0086] The spatial attention mechanism (Spatial-Wise Attention) enhances the features of the query position by aggregating information from other feature map positions. Its basic architecture is described as follows:

[0087]

[0088] is represented as the feature map of the input instance, where h and w are the input x i The height and width of e1 and w e2 represents the linear transformation matrix for bottleneck transformation (1×1 convolution); w gis the weight of global attention; LN stands for layer normalization, which is used to filter redundant information and refine the obtained context information. Specifically, the self-attention module consists of three parts (such as Figure 3 shown):

[0089] (a) Global Attention Pooling for context modeling; (b) Bottleneck Transform to capture spatial dependencies; (c) Fusion Function for propagating element-wise addition of feature fusion. This function can aggregate global context feature information into features at each query position.

[0090] The self-value-added and self-attention module consists of four parts (such as Figure 4 (a) Expansion Layer: Use the self-propagation module to increase the depth of the input feature map and generate more feature maps at a lower computational cost. (b) Deep convolutional neural network is used for feature extraction, and the self-attention mechanism is used to capture the relationship between channel importance and long-range dependencies. (c) Compression Layer is used to reduce the feature dimension to the same as the input feature.

[0091] The spatial information can then be transformed at a lower computational cost. In addition, the tensor collapse problem is avoided by removing the activation of ReLU in the compression layer. [Mark Sandler,Andrew Howard,Menglong Zhu,Andrey Zhmoginov,and Liang-Chieh Chen,"MobileNetV2:Inverted Residuals an d Linear Bottlenecks ”,2018 IEEE / CVF Computer Vision and Pattern Recognition Conference(CVPR),pp.4510-4520,Jun.2018] Finally, in (d), refer to MobileNetV2 [MarkSandler ,Andrew Howard,Menglong Zhu,Andrey Zhmoginov,and Liang-Chieh Chen,"MobileNetV2:InvertedResiduals and Linear Bottlene cks”,2018 IEEE / CVF Computer Vision and Pattern Recognition Conference(CVPR),pp.4510-4520,Jun.2018] To set the inverted residuals. The residual architecture was first proposed by He et al. in ResNet [Kaiming He ,Xiangyu Zhang,Shaoqing Ren,and Jian Sun,"Deep Residual Learning for Image Recognition”,2016 IEEE Conference on Comp uter Vision and Pattern Recognition(CVPR),pp.770-778,2016.] Its design concept is to learn residuals on the network to avoid the gradient vanishing problem. The architecture of ResNet is to compress first and then expand. The architecture of MobileNet is exactly the opposite of ResNet, expanding first and then compressing. The design concept of MobileNet is to extract features in high dimensions and transmit information in low dimensions.

[0092] Self-Proliferation-and-Attention Network (SP&A-Net). In the overall network design of SP&A-Net, Howard et al. use the Auto ML [AndrewHoward,Mark Sandler,Grace Chu,Liang-Chieh Chen,Bo Chen,Mingxing Tan,Weijun Wang,Yukun Zhu,Ruoming Pang,Vijay Vasudevan,Quo c V.Le ,and Hartwig Adam,"Searching for MobileNetV3”,2019 IEEE / CVF International Conference on Computer Vision(ICCV),pp.131 4-1324,Nov.2019] The designed network architecture. As shown in Table 1, a total of 16 layers of SP&A blocks are set. Conv2D 1×1 is used to increase the number of channels to 1024. Finally, in the flattening layer, 1000 features are output to connect the categories of the output layer. In addition, this model defaults to the Self-Attention block applied to all self-propagation layers (parameter = 1);

[0093] Table 1 Self-value enhancement and self-attention network architecture

[0094]

[0095] In each iteration, SP&A-Net randomly samples n images to calculate the gradient and then updates the network parameters. It stops training after the kth dataset. All functions and hyperparameters in its algorithm can be implemented in different neural networks introduced previously. The algorithm pseudocode is written as follows Figure 5 shown.

[0096] When semiconductor manufacturing encounters low-throughput process sites, the Hybrid Quantum-Classical Deep Learning (HQCDL) model is selected. This model uses the feature extraction of traditional computers and the classification computing advantages of quantum computers to improve the accuracy of defect identification.

[0097] The hybrid quantum-classical deep learning model is composed of quantum layers and classical layers. The quantum layer is composed of quantum circuits, which operate in a more efficient way than classical neural networks. HQCDL will encode the input in the quantum state to map the data into the nonlinear characteristics of the quantum Hilbert space.

[0098] The hybrid quantum-classical deep learning model consists of the following three parts (such as Figure 6 shown).

[0099] The first part is the Classical Convolutional Neural Network (CNN), which is mainly used for feature extraction. The model architecture here can be ResNet [Kaiming He ,Xiangyu Zhang,Shaoqing Ren,and JianSun,"Deep Residual Learning for Image Recognition”,2016 IEEE Conference on Compu ter Vision and Pattern Recognition(CVPR),pp.770-778,2016.] ,MobileNet family [Mark Sandler,Andrew Howard,Menglong Zhu,Andrey Zhmoginov,and Liang-Chieh Chen,"MobileNetV2:Inverted Residuals and Linear Bottlenecks”,2018 IEEE / CVF Com puter Vision and Pattern Recognition Conference(CVPR),pp.4510-4520,Jun.2018]、[Andrew G.Howard,Menglong Zhu,Bo Chen,D mitry Kalenichenko,Weijun Wang,Tobias Weyand,Marco Andreetto,and Hartwig Adam,"MobileNets:Efficient Convolutional Ne ural Networks for Mobile Vision Applications”,arXiv:1704.04861,2017]、[Andrew Howard,Mark Sandler,Grace Chu,Liang-Chi eh Chen,Bo Chen,Mingxing Tan,Weijun Wang,Yukun Zhu,Ruoming Pang,Vijay Vasudevan,Quoc V.Le,and Hartwig Adam,"Searchin g for MobileNetV3”,2019 IEEE / CVF International Conference on Computer Vision(ICCV),pp.1314-1324,Nov.2019], or the Self-Value and Self-Attention Network (SP&A-Net).

[0100] The second part is the Quantum Neural Network (QNN), which is mainly composed of the encoding circuit (EC), parameterized quantum circuits (PQCs), and the measurement layer.

[0101] The measurement results are post-processed by a third part - a classical neural network - to generate predictions, so that the final measurement value of the quantum neural network can be used for classification tasks.

[0102] To improve predictions, an optimization learning algorithm that updates model parameters was implemented simultaneously on both the classical and quantum computers. The entire algorithm runs in an endless loop between the classical and quantum computers, and a human-supervised process is used to optimize the network parameters for accurate defect discrimination and classification.

[0103] A self-proliferation block for classic convolutional neural network feature extraction was created, which consists of a compression layer, a proliferation layer, and a connection layer (such as Figure 7 As shown in Figure 1). The compression layer is a small number of convolutional layers that use Conv2D 1×1 convolution to reduce the number of output channels by half. The proliferation layer is a linear transformation layer that uses linear transformation to generate the other half of the feature map. The connection layer is used to merge the feature maps from the compression layer and the proliferation layer. Under this architecture, the total computational cost can be reduced by half. This architecture can obtain more features with fewer FLOPs;

[0104] =(,,) (4)

[0105] LO=depthwise(M,LK) (5)

[0106] =(,) (6)

[0107] Among them, is the Convolution Feature Map; is the output channel; is the stride that defines the step size when the kernel traverses the image, and here a step size of 3 is used for downsampling; is the convolution kernel size of Conv2D1×1; LO is the linear operation; LK is the linear operation kernel size, which is set to (3,3) here. Finally, the self-increment usage ratio (r) is set to Figure 7 For example, r is set to 50%, which means that half of the feature maps are generated by the self-propagation module.

[0108] Then, the SP-block is used to construct the Self-Proliferation Module, which is the basic module for constructing a classic convolutional neural network. It contains two traditional convolutional networks, an expansion layer, a deep convolutional layer, and a compression layer (such as Figure 8 ). The expansion layer is used to capture features in high dimensions, and the compression layer is used to transfer data in low dimensions. In deep convolution, the channel attention mechanism is used to extract relations with different semantic concepts between different channels. It can improve the model representation ability by performing dynamic channel feature extraction and recalibrating network parameters. In addition, the ReLU activation function is removed after the compression layer to avoid the tensor collapse problem mentioned by Sandler et al. [Mark Sandler,Andrew Howard,Menglong Zhu,Andrey Zhmoginov,and Liang-Chieh Chen,"MobileNetV2:Inverted Residuals an d Linear Bottlenecks ”,2018 IEEE / CVF Computer Vision and Pattern Recognition Conference(CVPR),pp.4510-4520,Jun.2018] .

[0109] Quantum Neural Network (QNN), QNN consists of three consecutive parts (by Fig. 9 As shown in Figure 1). The first is the encoding circuit, which encodes classical data into the state of quantum bits to input classical data into the quantum neural network for operation. There are four encoding circuit methods, including basis encoding, amplitude encoding, angle encoding, and dense angle encoding.

[0110] The second part is parametric quantum circuits (PQCs). The circuits transform these states to their optimal positions on the Hilbert space during the supervision process. Finally, the quantum computer prepares the quantum states and performs the measurements as specified by the PQCs. The measurement results are then processed by the classical computer to predict and classify wafer defects.

[0111] The third part, the measurement layer, is where the quantum computer prepares quantum states and performs measurements according to the PQCs. The measurement results are then processed by a classical computer to predict and classify wafer defects.

[0112] The Classical-NN Layer consists of three parts. The first part is to convert the output state from the quantum layer to the input of the classical layer.

[0113] Select the input of the classic layer {o i} as an operator The expected value of the measurement for each qubit is For the second part, most neural networks choose the Rectified Linear Unit (ReLU) as the activation function (as shown in Formula 7). For the third part, the Softmax activation function is used to calculate the representative probability values ​​of various defect types (as shown in Formula 8).

[0114] c i =ReLU(o i ) (7)

[0115] y j =softmax(∑ i w ji c i +b i ) (8)

[0116] where c i is the rectified output of the i-th layer linear unit; y j The Softmax operator is used to convert the output into a probability score to obtain the predicted probability of the jth result; w ji and b i are the weights and biases of the i-th layer.

[0117] At the same time, the quantum activation function is used to replace the traditional nonlinear activation function.

[0118] Traditional neural network representations are built on ReLU-based multi-layer perceptrons. These architectures lack the ability to represent fine details in the underlying signal, and they are generally not very good at representing the derivatives of the target signal. This is partly due to the fact that ReLU networks are piecewise linear and their second-order derivatives are zero everywhere, so they cannot model the information contained in the higher-order derivatives of natural signals. To address these limitations, quantum activation functions are utilized, which can represent the details in the signal better than ReLU-MLP.

[0119] By choosing a positive solution R(z) for ReLU ReLU And the negative solution R(z) of Leaky ReLU (LReLU) LReLU To apply the multi-step quantum activation function method [Luca Parisi,Daniel Neagu,Renfei Ma,and Felician Campean,"QReLU and m-QReLU:Two Novel Quantum Activation Function s toAid Medical Diagnostics”,ExpertSystemswithApplications,Vol.187,Jan2022] By applying the principle of quantum entanglement, H ReLU and H LReLU The tensor product of the two candidate Hilbert state spaces is:

[0120]

[0121] Quantum entanglement allows overcoming the limitation of ReLU that it runs into vanishing gradients for negative inputs. The resulting state in the hybrid system is described as:

[0122]

[0123] In the entangled or inseparable state, the product state formula of Quantum ReLU (QReLU) can be summarized as:

[0124]

[0125] The positive output is kept in QReLU, but the novelty of entanglement solution for negative values ​​is added. This is applicable to complex signals such as natural 2D images such as texture, color, and their derivatives.

[0126] Optimizing training

[0127] Loss Function, after getting all the features from PQCs, feed them to the classic neural network layer. Using the Softmax activation function, the output of this layer will be a probability distribution. The i-th element of the output is the probability that this data point belongs to the i-th category, and the data point is predicted to belong to the category with the highest probability. In order to predict the actual ground truth, the cumulative distance between the predicted label and the actual label is calculated as the optimized loss function:

[0128]

[0129] where ω 1 ,b 1 is the parameter of the classical layer; θ is the parameter of the quantum layer; ω 2 ,b 2 is the parameter of the fully connected layer. Perform the above process enough times to obtain the function with the minimum loss The hybrid quantum-classical deep learning model partially differentiates the loss function and simultaneously calculates all the gradients of the classical convolutional neural network, quantum neural network, and classical neural network, and then uses this result to update all parameters ( 1 , 1 ,, 2 , 2 ), and iterate until the end (such as Fig.10 shown).

[0130] Learning Rate

[0131] When training deep learning models, one of the most difficult parameters to set is the learning rate. If this value is large, the model's weights will begin to oscillate and prevent the model from adjusting the error. If the learning rate is too small, the model will have a high learning cost and may get stuck in a local minimum. When training deep networks, it is often useful to anneal the learning rate over time. [Ilya w Loshchilovand FrankHutter ,"SGDR:Stochastic Gradient Descent with Warm Restarts”,International Conference on Learning Representations(ICLR),Ma y 2017] . Use cosine decay with preheating as the learning rate function, the formula is as follows:

[0132] ηt=0.5×(1+cos(π×T Cur / T i )) (13)

[0133] decayed=(1-α)×ηt+α (14)

[0134] decayed_learning_rate=learning_rate×decayed (15)

[0135] Where ηt is the decay rate of the cosine decay learning rate in the tth period; T Cur is the current training stage; T i is the total number of training epochs for the i-th cosine decay learning rate; α is a hyperparameter that controls the degree of decay; learning_rate is the base learning rate. Fig.11 As shown in Figure 1, this case is a schematic diagram of the distribution of cosine decay learning rate for 4 periods (t=4). In the first generation, the number of training generations is 100 (T0=100). As the number of generations increases, the learning rate decays gradually, avoiding jumping out of the optimal solution interval at the end of training without sacrificing training time.

[0136] Research and Experimentation

[0137] In deep learning, a certain amount of data with a balanced number of categories is often required to ensure that overfitting does not occur during training. However, in today's digital age, data is the foundation of driving artificial intelligence. Most valuable data is controlled by well-funded companies or institutions in related fields. It is difficult for individual developers or ordinary companies to own or collect the data required for a complete model. Therefore, data enhancement is generally used to compensate for the overfitting problem caused by insufficient data.

[0138] Data augmentation has been shown to solve the problem of insufficient data and improve the accuracy of system training. It is the process of generating samples by transforming the training data. The purpose is to improve the accuracy and robustness of the model.

[0139] In the present invention, more images are added through data augmentation, including random cropping, rotation, and resizing of images (e.g. Fig.12 After this work, the eye can still recognize it as the same picture, but it is a completely different new image for the machine. The present invention creates more images for the four defect patterns (Ball, Hump, Fallon, Color mark) with relatively small sample numbers by modifying the existing images in the data set to make up for the lack of data. The image transformation method and transformation range are set as follows:

[0140] Random cropping (±32pix)

[0141] Random flip (0°, 90°, 180°, 270°)

[0142] Randomly adjust brightness (±16, limited to 0-255)

[0143] Randomly adjust saturation (limited to 45% to 125%)

[0144] Randomly adjust the hue (±0.1)

[0145] Randomly adjust contrast (limited to 45% to 125%).

[0146] Through the above data value-added method, the number of defect types with less than 500 samples is increased to 500 data to balance the data; for the wafer defect map, the number of Scratch, Centripetal Stripes, Full Map, and Semicircle Map is increased respectively.

[0147] In model design, MobileNetV3 is used for traditional convolutional neural networks. [Andrew Howard ,Mark Sandler,Grace Chu,Liang-Chieh Chen,Bo Chen,Mingxing Tan,Wei jun Wang,Yukun Zhu,Ruoming Pang,Vijay Vasudevan,Qu oc V.Le ,and Hartwig Adam,"Searching for MobileNetV3”,2019 IEEE / CVF International Conference on Computer Vision(ICCV),pp.131 4-1324,Nov.2019] Used as a basic model for comparison with the baseline. This model is the architecture designed by Google in 2019 using Auto ML-Neural Architecture Search, replacing the traditional network designed by humans.

[0148] The basic module for building MobileNetV3 (called Bottleneck Block) consists of three parts. The first is the expansion layer (Expansion Layer) composed of Pointwise Convolution, which is normalized with BatchNormalization and uses ReLU as the activation function at the end; the second layer is the Linear Bottleneck Layer composed of DepthwiseConvolution as the feature extraction layer, which is normalized with BatchNormalization and uses ReLU as the activation function at the end; the last layer is the compression layer (Compression Layer) composed of PointwiseConvolution, which is normalized with Batch Normalization. The purpose of this layer is to compress the dimension to the feature map with the same dimension as the input, and connect the output and input with the inverted residuals (Inverted Residuals). The main concept of this design is to encourage the deep learning network to learn the residuals rather than the actual values, so as to avoid the gradient disappearance problem caused by the network being too deep. It is worth noting that no activation function (such as ReLU) is used in the compression layer. The reason is that performing nonlinear transformations such as ReLU on lower-dimensional tensor representations will result in a large information loss, which will cause the tensor collapse problem during training.

[0149] It is also worth noting that residual learning was first proposed in ResNet

[12] , but ResNet first compresses and then expands. MobileNetV3, on the other hand, is the opposite, expanding first and then compressing. The design concept of MobileNetV3 is "extracting features in high dimensions and transmitting information in low dimensions." This design can effectively reduce the amount of model calculations to make it suitable for MobileDevice. MobileNetV3 uses a total of 15 layers of Bottleneck Block (bneck), and Conv2D 1×1 is used at the end to increase the dimension to 960 and 1280. Finally, in the flattening layer, 1000 features are output to connect to the neurons for defect classification, and finally Softmax is used to convert them into probability values ​​represented by each defect type for loss function calculation.

[0150] Four indicators are used to evaluate the performance of the model: Accuracy, Precision, Recall, and F1-Score. The formula is as follows:

[0151]

[0152]

[0153]

[0154]

[0155] Among them, TP is true positive (True Positive); TN is true negative (True Negative); FP is false positive (False Positive); FN is false negative (False Negative).

[0156] Accuracy is a comprehensive indicator, but for data sets with uneven number of categories, this indicator cannot reflect the performance of the model, so it is necessary to combine precision and recall for additional judgment.

[0157] Precision refers to the model's ability to judge positives. A model with too low precision will cause non-defective wafers to be misjudged as defective, resulting in additional manpower impact for defect review and loss of production capacity of the measurement equipment.

[0158] A model with too low a recall rate may cause defective wafers to be misjudged as non-defective, resulting in invalid subsequent processing, affecting the yield and causing credit losses to the company.

[0159] The F1-Score is the harmonic mean of precision and recall. It takes both precision and recall into account, taking into account both production capacity loss and wafer yield. Therefore, the F1-Score is used as the main indicator for measuring the model.

[0160] In the classic neural network part, we use Howard et al. in Auto ML [AndrewHoward,Mark Sandler,Grace Chu, Liang-Chieh Chen,Bo Chen,Mingxing Tan,Weijun Wang,Yukun Zhu,Ruoming Pang,Vijay Vasudevan,Quoc V.Le,and HartwigAdam, "Searching for MobileNetV3”,2019IEEE / CVF International Conference on Computer Vision(ICCV),pp.1314-1324,Nov.2019] The designed network architecture has 16 layers of SP&A blocks. Conv2D 1×1 is used to increase the number of channels to 1280. Finally, in the flattening layer, 1000 features are output to connect the categories of the output layer.

[0161] In the loss function part, in addition to the traditional L1 and L2 Loss, we also consider Sun et al. [Yifan Sun,Changmao Cheng,Yuhan Zhang,Chi Zhang,Liang Zheng,Zhongdao Wang,and Yichen Wei,"Circle Loss:A Unified Perspective of Pair Sim ilarity Optimization”,2020 IEEE / CVF Conference on Computer Vision and Pattern Recognition(CVPR),pp.6397-6406,2020]The proposed Cycle Loss. This method is different from the traditional loss function in that it adopts a symmetrical optimization parameter mode, hoping to simultaneously reduce the intra-class similarity (Between-Class Similarity) and increase the extra-class similarity (Within-ClassSimilarity). This method will lead to the problem of lack of optimization flexibility. Cycle Loss takes the difference between the intra-class similarity and the extra-class similarity (s n -s p ) can be summarized as (α n s n -α n s p ), where α n and α p are independent weighting factors, allowing s n and p Using different gradients to learn, different similarities have different penalty strengths. The CircleLoss formula is designed as follows:

[0162]

[0163] Where K and L are the number of groups within and outside the category respectively; γ is a hyperparameter used to control the strength of the similarity difference. n and p They represent the similarity within the category and the similarity outside the category respectively.

[0164] In the learning rate part, linear decay (decay rate is fixed to 0.001), Exponential Decay, Triangular Decay, and Cosine Annealing are considered respectively.

[0165] In the encoding circuit part, basic encoding, amplitude encoding, angle encoding and dense angle encoding are used. In the parameterized quantum circuit part, refer to Sim et al. [Sukin Sim,Peter D.Johnson,and Alan Aspuru-Guzik,"Expressibilityand Entangling Capability of Parameterized Quantum Circuits for Hybrid Quantum-Classical Algorithms”,arXiv:1905.10876,2019] The four circuits proposed in the published paper (Circuit 5, Circuit 6, Circuit 13, Circuit 14). Sim et al. believe that the design of these four circuits has better expression and entanglement capabilities, and they are verified by real experiments in their paper. These four circuits are used as the basic architecture of parameterized quantum circuits.

[0166] Parameterized quantum circuits are mainly composed of four types of quantum logic gates (such as Fig.13). Quantum logic gates are used in quantum computing and are essentially transformations of one or more qubits. A description of the types of quantum gates is given below.

[0167] The CNOT quantum logic gate (Controlled-NOT Gate) refers to a quantum logic gate that acts on a pair of qubits through a linear image. The linear image is given on a pair of qubits in this way. The first quantum is usually called the control qubit and the second qubit is the target qubit. The CNOT quantum logic gate operation is expressed as follows in the ground state:

[0168] (1) Keep the control qubit unchanged and perform a Pauli-X Gate (Pauli matrix quantum logic gate) on the target qubit when the control qubit is in state |1〉.

[0169] (2) When the control qubit is in state |0>, the target qubit remains unchanged.

[0170] The rotation gate generates a rotation operator that rotates the Bloch sphere vector by a given angle θ around the x, y, and z axes.

[0171] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the present invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A semiconductor equipment wafer defect identification method based on a hybrid quantum Al algorithm, characterized in that: Here are the steps: S1: Select different defect identification and classification methods according to the context in the semiconductor process; S2: When high-throughput process sites appear in semiconductor manufacturing, self-proliferation and attention neural networks are selected; S3: When semiconductor manufacturing encounters low-throughput process sites, choose a hybrid quantum-classical deep learning model; In the step S2, the self-proliferation and attention neural network consists of a self-proliferation module and a self-attention module; the self-proliferation module generates more feature maps at a lower cost through linear transformation; the self-attention module learns information about long-range dependencies from the features generated by the self-proliferation module; the self-proliferation module generates the same number of features through linear transformation, and the expression is as follows: Among them, m' i is the basic unit of the feature map M′; h and w are the new feature maps u i，j The height and width of ; ρ represents a low-cost feature map generation operation; the self-attention module includes a channel attention mechanism and a spatial attention mechanism; the channel attention mechanism focuses on feature patterns that are highly correlated to the output results by focusing on the input feature map and activating part of the channel; the spatial attention mechanism enhances the features of the query position by aggregating information from other feature map positions; the architecture of the channel attention mechanism consists of the following steps: the first is Squeeze, which flattens the 3D Feature Map into a 1D tensor through Global AveragePooling; the second is Excitation, which captures the importance of the channel through bottleneck transformation; the third is the fusion function, which is used to propagate the element-wise multiplication of feature fusion; the architecture expression of the spatial attention mechanism is as follows: Where h and w are input x i The height and width, x i Represented as the feature map of the input instance; w e1 and w e2 represents the linear transformation matrix used for bottleneck transformation; w g is the weight of global attention; LN represents layer normalization; the hybrid quantum-classical deep learning model includes the following three parts: the first part is the classical convolutional neural network, which is mainly used for feature extraction; the second part is the quantum neural network, which is composed of an encoding network, a parameterized quantum circuit, and a measurement layer; the third part is the classical neural network, which is post-processed to generate predictions so that the final measurement value of the quantum neural network can be used for classification tasks; the self-propagation module is used for feature extraction of the classical convolutional neural network; the self-propagation module consists of a compression layer, a proliferation layer and a connection layer; the compression layer is a small number of convolutional layers, which uses Conv2D1×1 convolution to reduce the output channels by half; the proliferation layer is a linear transformation layer, which generates the other half of the feature map through linear transformation; the connection layer is used for and feature maps from the compression layer and the proliferation layer; the quantum neural network consists of three consecutive parts, namely: an encoding circuit, which is used to encode classical data into the state of quantum bits so as to input the classical data into the quantum neural network for calculation; a parameterized quantum circuit, which is used to convert the states to their optimal positions in the Hilbert space; a measurement layer, which is used to prepare quantum states and perform measurements according to the provisions of PQCs by the quantum computer, and at the same time, let the measurement results be subsequently processed by the classical computer to predict and classify wafer defects; the classical neural network consists of three parts, as follows: first, convert the output state from the quantum layer to the input of the classical layer; second, let the neural network select the linear rectification function as the activation function; finally, use the Softmax activation function to calculate the representative probability values ​​of various defect types.

Images