Semiconductor wafer defect detection light source configuration method, device and equipment

Abstract

The present invention relates to the technical field of light source configuration, and discloses a method, device and equipment for configuring a light source for defect detection of a semiconductor wafer. The method comprises: acquiring a semiconductor wafer image and performing dynamic graph convolution processing and defect feature enhancement to obtain an enhanced defect feature map; inputting the enhanced defect feature map into a two-way deep neural network to perform reflected light intensity calculation and image quality scoring to obtain reflected light intensity data of a defective area and image quality scoring data; acquiring light source input current and light intensity sensor data of multiple light sources, and identifying and calculating light source angle parameters in combination with the image quality scoring data to obtain initial light source parameters; performing nonlinear control based on the reflected light intensity data and the initial light source parameters to obtain light source control parameters; calculating channel weights of multiple light sources and performing differentiation processing to obtain collaborative light source configuration parameters. The present invention improves the accuracy and reliability of defect detection and reduces the manual debugging cost of light source parameter configuration.

Description

Technical Field

[0001] The present invention relates to the technical field of light source configuration, and in particular to a method, device and equipment for configuring a light source for semiconductor wafer defect detection. Background Art

[0002] Semiconductor wafer defect detection is a key link in chip manufacturing quality control, and its detection effect directly affects the yield and performance of the chip. Traditional wafer defect detection systems usually use a fixed light source configuration scheme, which is difficult to cope with the detection needs of different types of defects, especially for complex situations such as tiny defects and compound defects. There are obvious deficiencies in detection accuracy and reliability. With the continuous evolution of semiconductor manufacturing processes, the defect characteristics on the wafer surface are becoming increasingly complex and diversified, and a single light source configuration scheme can no longer meet the requirements of high-precision detection. In addition, due to the strong nonlinearity and uncertainty of the reflection characteristics of the wafer surface, the traditional experience-based light source parameter setting method often requires a lot of manual debugging time, and it is difficult to obtain the best detection effect.

[0003] The existing light source configuration methods mainly have the following problems: First, there is a lack of effective extraction and utilization of deep semantic information of the defect area, resulting in insufficient robustness of detection features; second, the mutual influence and compensation relationship between light sources are not fully considered during the collaborative configuration of multiple light sources, affecting the overall performance of the detection system; third, the adaptive adjustment capability of light source parameters is insufficient, making it difficult to dynamically optimize lighting conditions for different defect types. Summary of the invention

[0004] The present invention provides a method, device and equipment for configuring a light source for defect detection of a semiconductor wafer, which improves the accuracy and reliability of defect detection and reduces the manual debugging cost of light source parameter configuration.

[0005] In a first aspect, the present invention provides a method for configuring a light source for defect detection of a semiconductor wafer, the method comprising:

[0006] Acquire a semiconductor wafer image and perform dynamic graph convolution processing and defect feature enhancement to obtain an enhanced defect feature map;

[0007] Inputting the enhanced defect feature map into a dual-path deep neural network to calculate reflected light intensity and image quality score, thereby obtaining reflected light intensity data and image quality score data of the defect area;

[0008] Acquire light source input current and light intensity sensor data of multiple light sources, and identify and calculate light source angle parameters in combination with the image quality score data to obtain initial light source parameters;

[0009] Based on the reflected light intensity data and the initial parameters of the light source, nonlinearly control the light source angle and the light source intensity to obtain light source control parameters;

[0010] The channel weights of the multiple light sources are calculated, and the light source control parameters are differentiated to obtain collaborative light source configuration parameters.

[0011] In a second aspect, the present invention provides a semiconductor wafer defect detection light source configuration device, the semiconductor wafer defect detection light source configuration device comprising:

[0012] An acquisition module is used to acquire a semiconductor wafer image and perform dynamic graph convolution processing and defect feature enhancement to obtain an enhanced defect feature map;

[0013] A scoring module, used for inputting the enhanced defect feature map into a dual-path deep neural network to calculate the reflected light intensity and score the image quality, so as to obtain the reflected light intensity data and image quality score data of the defect area;

[0014] A calculation module, used to obtain light source input current and light intensity sensor data of multiple light sources, and identify and calculate the light source angle parameters in combination with the image quality score data to obtain the initial parameters of the light source;

[0015] A control module, used for performing nonlinear control on the light source angle and the light source intensity based on the reflected light intensity data and the initial parameters of the light source to obtain light source control parameters;

[0016] The processing module is used to calculate the channel weights of the multiple light sources and perform differentiation processing on the light source control parameters to obtain collaborative light source configuration parameters.

[0017] A third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned semiconductor wafer defect detection light source configuration method.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium, wherein instructions are stored in the computer-readable storage medium, and when the computer-readable storage medium is run on a computer, the computer executes the above-mentioned semiconductor wafer defect detection light source configuration method.

[0019] In the technical solution provided by the present invention, by introducing a dynamic graph convolutional network and a dual attention mechanism, deep extraction and enhancement of defect features are achieved, and a dual-path deep neural network architecture is used to perform quality evaluation and reflected light intensity analysis on defect images; based on the combination of a nonlinear adaptive observer and a support vector regression model, a light source parameter identification and control system with self-learning ability is constructed; by designing a hierarchical optimization strategy and an orthogonal Pareto criterion, the coordinated configuration and dynamic adjustment of multiple light sources are achieved. This method not only significantly improves the accuracy and reliability of defect detection, but also greatly reduces the manual debugging cost of light source parameter configuration, and has a strong engineering application value. Compared with the prior art, the present invention has achieved significant improvements in detection accuracy, system adaptability, and configuration efficiency, and provides a new technical solution for the field of semiconductor wafer defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] 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 creative work.

[0021] Figure 1 A schematic diagram of the steps of a method for configuring a light source for semiconductor wafer defect detection according to an embodiment of the present invention;

[0022] Figure 2 It is a structural schematic diagram of a semiconductor wafer defect detection light source configuration device according to an embodiment of the present invention;

[0023] Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] Embodiments of the present invention provide a method, device and equipment for configuring a light source for defect detection of a semiconductor wafer. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of a semiconductor wafer defect detection light source configuration method in an embodiment of the present invention includes:

[0026] Step S1, acquiring a semiconductor wafer image and performing dynamic graph convolution processing and defect feature enhancement to obtain an enhanced defect feature map;

[0027] It is understandable that the execution subject of the present invention may be a semiconductor wafer defect detection light source configuration device, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0028] Specifically, the semiconductor wafer image is preprocessed by grayscale and normalization to obtain the preprocessed wafer image. The color image is converted into a single-channel grayscale image to reduce the complexity of image processing and make the defect features more prominent. Normalization processing is to compress the pixel value of the image within a certain range to eliminate the influence of factors such as brightness and contrast of different images on subsequent processing and ensure the consistency of the processing results. The corresponding K nearest neighbor point set is calculated for each pixel point in the preprocessed wafer image. The K nearest neighbor graph structure is constructed by the Euclidean distance measurement method to obtain a feature set containing edge points. The edge point feature set is input into the dynamic graph convolution layer. The dynamic graph convolution layer contains three graph convolution units, each of which uses the ReLU activation function and the edge feature update mechanism. Graph convolution can effectively capture local features in the image by weighting and operating the neighborhood points, and as the convolution proceeds, more abstract features can be gradually extracted to generate a dynamic feature map. The dynamic feature map is subjected to kernel point convolution processing. The similarity matrix between the center point and the neighboring points is calculated using the Gaussian kernel function, and similar feature points are clustered together to improve the representation ability of the features. The light intensity contrast information is combined for feature aggregation to enhance the contrast of the defect area, so that the defect features are more obvious in the image, and the initial defect feature map is obtained. The initial defect feature map is input into the spatial attention submodule. The spatial attention submodule contains three convolutional layers and a sigmoid activation layer. Each convolutional layer is followed by a BatchNorm layer to help accelerate training and improve the stability of the network. Through these convolution operations, the network can adaptively learn important spatial features and calculate the spatial weight matrix to obtain a spatial enhanced feature map. Channel attention processing is performed on the spatial enhanced feature map. Different channels of the feature map are weighted according to the importance of each channel to enhance the channels that are important for defect detection. Channel descriptors are generated through global average pooling and maximum pooling operations. The channel weight vector is calculated using a multi-layer perceptron, and the features of each channel are weighted to obtain a channel enhanced feature map. The spatial enhanced feature map is fused with the channel enhanced feature map. The residual connection structure and 1×1 convolution layer are used to reduce the dimension of the fused features. The introduction of residual connection effectively avoids the gradient vanishing problem and accelerates the feature fusion process. After the dimensionality reduction process, the fused feature map is obtained. The fused feature map is adaptively calibrated. The feature similarity matrix is ​​calculated, and the regional features are reconstructed. Adaptive adjustments are made according to the features of different regions to make the expression of defect features more accurate. Residual learning is performed in combination with the original features to improve the accuracy and robustness of the defect features, and finally an enhanced defect feature map is obtained.

[0029] Step S2, inputting the enhanced defect feature map into a dual-path deep neural network to calculate the reflected light intensity and image quality score, and obtaining the reflected light intensity data and image quality score data of the defect area;

[0030] Specifically, the enhanced defect feature map is input into the image enhancement subnetwork and the quality assessment subnetwork respectively. The two subnetworks use different architectures and processing methods to calculate the light intensity and evaluate the quality of the image, respectively, and obtain relevant information of the defect area. The image enhancement subnetwork adopts the U-Net structure to enhance the reflected light intensity of the defect area through image enhancement; the quality assessment subnetwork uses a multi-scale feature extraction structure to evaluate the clarity, contrast and integrity of the defect in the image. The image enhancement subnetwork extracts and enhances the reflected light intensity information of the defect area from the original enhanced defect feature map through the enhanced feature extraction and decoding process. The encoder of the image enhancement subnetwork extracts features from the enhanced defect feature map. The encoder contains four downsampling units, each of which consists of two 3×3 convolutional layers and one maximum pooling layer. This process gradually reduces the spatial resolution of the image and increases the abstraction level of the features, thereby obtaining coded features with multi-scale information. The multi-scale coded features are input into the decoder of the image enhancement subnetwork. The decoder contains four upsampling units, each of which consists of one deconvolutional layer and two 3×3 convolutional layers. The decoder recovers and enhances the spatial information of the defective area by upsampling the low-resolution encoded features and fusing the encoded features of the corresponding scales through jump connections. The decoder outputs the reflected light intensity data of the defective area. At the same time, the quality assessment subnetwork extracts features through multiple parallel convolution branches, each of which uses different dilated convolution layers to extract receptive field features of different scales. Each branch of the quality assessment subnetwork contains four dilated convolution layers with different dilation rates. The use of dilated convolution can increase the range of the receptive field, allowing the network to capture a wider range of contextual information and improve the global feature extraction ability of the image. Through these parallel branches, the network can extract features from multiple angles of the image at different scales to obtain richer image information. The multi-scale receptive field features are input into the global context attention module for processing. This module can automatically identify the importance of each channel and spatial position in the image by calculating the correlation matrix between feature channels and the spatial position correlation matrix, and enhance the feature information of key areas. By calculating the correlation between features, the context attention module can adjust the weight of features globally to ensure that important features are strengthened and irrelevant features are suppressed. Adaptive feature aggregation is performed on context-enhanced features. The features of different branches are reorganized through channels through 1×1 convolutional layers to extract and fuse key information. The pyramid pooling module is used to extract multi-scale global information, enhance feature diversity and global context information, and obtain more comprehensive and detailed image features. Feature transformation is performed on the aggregated features to obtain feature vectors. These feature vectors are used to calculate the clarity score, contrast score, and defect completeness score of the image through multi-task learning. Multi-task learning optimizes multiple objectives simultaneously by sharing network weights, allowing the model to be evaluated and optimized in multiple aspects.These scores are used to evaluate the quality of the image, especially in the defect detection process, where the quality of the image directly affects the identification and location of the defect. The different scores are aggregated by weighted average to obtain the final image quality score data.

[0031] Step S3, obtaining light source input current and light intensity sensor data of multiple light sources, and identifying and calculating light source angle parameters in combination with image quality score data to obtain initial light source parameters;

[0032] Specifically, the light source input current data and light intensity sensor data of multiple light sources are collected, and the light source input current data and light intensity sensor data are normalized to eliminate the dimensional difference in the data and ensure that each data item is processed at the same scale. The current data, light intensity data and corresponding image quality score data are combined into a feature vector matrix. Each row corresponds to a light source, including the current data, light intensity data and the corresponding image quality score of the light source, forming a sample set with multidimensional features. The radial basis kernel function is used for nonlinear mapping. The feature vector is mapped from a low-dimensional space to a high-dimensional space to capture the potential nonlinear relationship in the data and obtain a high-dimensional feature mapping matrix. A quadratic programming problem is constructed for the high-dimensional feature mapping matrix, and a regression problem is solved by quadratic programming to determine the initial value of the light source angle. When constructing the quadratic programming problem, the penalty factor is set to 10 to balance the fitting accuracy and complexity of the model. At the same time, the insensitive loss parameter is set to 0.1 to control the robustness of the model and ensure that it can maintain adaptability to small fluctuations in the input data. The quadratic programming problem is solved by the sequential minimum optimization algorithm to obtain a set of support vector sets, which contains the sample data that has the greatest impact on the regression function. Based on the support vector set, the weight vector and bias term of the regression function are calculated to help predict the output value of the input data. Through regression prediction, the initial value of the light source angle of each light source is obtained. The initial value of the light source angle is input into the optical characteristic compensation module, and the correction amount of the light source angle is calculated according to the Fresnel reflection model of the wafer surface. The correction process takes into account the optical effect of the interaction between the light source and the wafer surface to ensure that the angle of the light source can match the actual optical environment. The compensated light source angle data is cross-validated. Through cross-validation, the stability and reliability of the light source angle data are evaluated to obtain the confidence of the light source angle. The results of cross-validation help to judge the generalization ability of the model, avoid overfitting or underfitting, and ensure the rationality and accuracy of the light source angle data. According to the light source angle data and angle confidence, combined with the physical constraints of the light source, parameter optimization is performed. The physical constraints of the light source include the adjustable range of the light source angle, the stability of the light source position, etc. These conditions help to limit the optimization space and ensure that the final light source parameters are not only reasonable in theory, but also feasible in practical applications. After parameter optimization, the final initial light source parameters are used for subsequent defect detection and light source control to ensure that the light source can irradiate at the optimal angle and intensity during the wafer defect detection process, thereby improving the accuracy and efficiency of detection.

[0033] Step S4: Based on the reflected light intensity data and the initial parameters of the light source, nonlinearly control the light source angle and the light source intensity to obtain the light source control parameters;

[0034] Specifically, the reflected light intensity data is preprocessed. The light intensity signal is effectively filtered and normalized. The bandpass filter is used to remove high-frequency noise and low-frequency drift in the signal, making the signal smoother and meeting the analysis requirements. The filtered light intensity signal is normalized to ensure that the data is analyzed under the same standard, eliminate the dimensional differences in the light intensity signal, and make it suitable for subsequent modeling and calculation. Combined with the initial parameters of the light source, the processed light intensity data is combined with the initial parameters to construct the state space equation and obtain the state matrix of the observation system. The state space equation describes the dynamic behavior of the light source control system, and the state matrix provides a mathematical model for the subsequent control algorithm to help the system understand the relationship between various parameters. The state matrix of the observation system is subjected to Lyapunov stability analysis. Lyapunov stability analysis is used to ensure that the system can operate stably during the control process to avoid unstable states or oscillations. By solving the Lyapunov equation, a positive definite matrix is ​​obtained, which characterizes the stability characteristics of the system state during the time evolution process. Based on the positive definite matrix, the feedback gain calculation criterion is constructed. This criterion can provide a reasonable gain constraint condition for the control system to ensure that the system will not over-respond or become unstable when adjusting the angle and intensity. The gain constraint condition is input into the pole configuration module. Pole configuration is a control optimization method that can achieve precise control of the dynamic response of the system by adjusting the pole position of the system. In the pole configuration module, the Ackerman formula is used to calculate the distribution of the closed-loop poles of the system. Through pole configuration, the dynamic characteristics of the system, such as response speed and stability, are determined. Using this information, the initial value of the state feedback gain matrix is ​​obtained through matrix operation to obtain preliminary feedback gain data. Based on the feedback gain data, the state space equation is decoupled. Since the angle control and intensity control may be coupled in the original state space equation, the two control terms are converted into independent standard forms through decoupling transformation. Through this process, the control of angle and intensity is separated, so that the subsequent control algorithm can be simplified and more efficient during optimization, and a set of independent decoupled state equations are obtained. Nonlinear compensation is performed on the decoupled state equation. By constructing a feedback linearization controller, the nonlinear terms are compensated to eliminate the nonlinear effects in the system. The design of the feedback linearization controller enables the system to be controlled in a linear manner, thereby improving the accuracy and stability of the control. The calculation of the compensation control quantity is based on the initial parameters of the light source obtained previously, combined with the feedback gain and state equation of the system, to obtain the compensated control strategy to optimize the adjustment of the light source angle and intensity. The compensation control data is input into the sliding mode controller for control quantity allocation. The sliding mode controller is a robust control method that can handle uncertainty and interference in the system. Through the sliding mode controller, saturation processing is performed according to the physical constraints of the light source angle and intensity to ensure that the control quantity does not exceed the actual capability of the light source. After saturation processing, a set of distribution control data is obtained.The distribution control data and feedback gain data are combined and optimized. According to the feedback gain and distribution control data, the final light source control parameter set is obtained, including the angle control amount and intensity control amount of each light source, and the position and intensity of the light source can be accurately adjusted in practical applications to meet the needs of defect detection.

[0035] Step S5: Calculate the channel weights of multiple light sources, and perform differentiation processing on the light source control parameters to obtain collaborative light source configuration parameters.

[0036] Specifically, the control parameters of multiple light sources are grouped by channels, and the angle and intensity of each light source are regarded as independent control units and assigned to different channels for processing. For the angle and intensity parameters of each light source channel, standardization is performed so that the parameters of different light source channels can be compared and analyzed under the same standard, eliminating the dimensional differences caused by different light sources, and obtaining standardized light source parameters. Correlation analysis is performed on the standardized light source parameters to reveal the relationship between the light source channels, obtain channel association data, and reflect the relative influence and synergy of each light source channel in the control process. The channel association data is input into the attention calculation module. A weight is assigned to each channel to reflect its importance in the overall control. By using the softmax function, the attention scores between the channels are calculated, and these scores represent the relative importance of each channel in the final control. In order to ensure the rationality of the calculation results and the feasibility of practical applications, a spatial constraint matrix is ​​constructed in combination with the physical position information of the light source. The matrix takes into account the actual physical constraints of the light source position, such as the distance between light sources, the illumination range, and other factors, and models these constraints in the attention calculation process to obtain a more accurate and practical initial channel weight. The initial channel weights are normalized and exponentially smoothed in combination with the historical weights of each light source channel. This means that a certain "memory" is given to the historical weights, making the new weight calculation more stable and avoiding control fluctuations caused by mutation data. Based on the smoothed weights, the light source control parameters are weighted and distributed, and the complementary constraints between light sources are considered. The complementary constraints consider the mutual support relationship between different light source channels when working together, that is, how the channels coordinate and cooperate to achieve the best light source configuration. Through this weighted distribution, the system can ensure the optimization of light source control parameters so that each channel can work in coordination in actual applications to achieve the best lighting effect. The optimized weight data is input into the collaborative control module to collaboratively optimize the light source control parameters to ensure that all light sources can reasonably coordinate according to their channel weights and physical constraints. The collaborative control module uses a distributed consistency protocol, which can establish consistency constraints between multiple light sources, ensure that each light source channel follows a unified rule during the optimization process, and handle the interaction between the light sources. At this stage, the light source control parameters are collaboratively adjusted through the protocol to ensure the coordination and effectiveness between the light sources. Constraint processing is performed on the collaborative control data to ensure that in the process of optimizing the light source control parameters, all results meet the actual working conditions of the light source, such as the power limit and angle range of the light source. By processing these constraints, the collaborative control data is parameter truncated, that is, the parameters that exceed the preset limits are corrected to ensure that the final result is physically feasible and meets the requirements. Through these constraint processing, the final result obtained is the constrained control data.The constraint control data is combined with the optimization weight data to obtain the final collaborative light source configuration parameters. The collaborative light source configuration parameters include the optimal angle and intensity configuration values ​​of each light source channel.

[0037] When performing constraint processing, the collaborative control data is constrained and decomposed. The light sources are divided into two groups: the first group of light sources is located directly above the wafer, and the second group of light sources is distributed around the wafer. Nonlinear constraint equations are established for the angle and intensity parameters of each group of light sources, which will constrain the working range of the light sources, including angle limits and power thresholds. The angle limit range of the first light source group is set to [-30°, 30°], and the power threshold of this group of light sources is set to 80%; while the angle range of the second light source group is [-60°, 60°], and the power threshold is set to 60%. These constraints ensure that each group of light sources will not exceed its physical limits during operation, thereby avoiding excessive or insufficient illumination. By setting these constraints, a group constraint optimization equation is constructed. The group constraint optimization equation is solved using a hierarchical quadratic programming method. According to the set angle range and power threshold, the hierarchical truncation parameters of each light source group are calculated, reflecting the specific adjustment of angle and power in different light source groups. Based on the hierarchical truncation parameters, an orthogonal multi-objective optimization model is constructed. In this model, the goal of the first light source group is to optimize the uniformity of illumination, while the second light source group mainly optimizes the contrast. In order to strike a balance between these two goals, the orthogonal Pareto criterion is introduced, and the objective function is decoupled and optimized through this criterion. Through this process, the optimization goals of the two light source groups are adjusted independently to ensure that the optimal configuration of each light source group can simultaneously meet its specific optimization goals and improve the overall lighting effect. The orthogonal optimization data is input into the genetic algorithm solver. The genetic algorithm can continuously optimize the parameters through natural selection and genetic operations to ensure the realization of the optimal solution. In this process, the dynamic adaptive crossover probability and mutation probability are set so that the algorithm can effectively explore in the search space and dynamically adjust the parameters according to the search results. The local search operator and the elite retention strategy are introduced. These strategies can improve the optimization efficiency, avoid falling into the local optimal solution, and better find the global optimal solution. After the genetic algorithm is solved, a double-layer optimization parameter sequence is obtained. The double-layer optimization parameter sequence is grouped and constrained for verification. This is done by constructing two constraint functions: the first constraint function constrains the uniformity of light intensity to ensure that the light intensity distribution within the illumination range of the light source is uniform; the second constraint function focuses on angle complementarity to ensure that the angle settings between different light source groups can complement each other and improve the overall lighting effect. Through the verification of these two constraint functions, it is ensured that the optimization results meet the physical limitations in practical applications. After the verification is completed, the parameters are corrected using the gradient projection method to ensure that a hierarchical feasible parameter set is obtained. The hierarchical feasible parameter set is orthogonally combined with the optimization weight data. The parameters are fused by constructing a combined optimization function that takes into account the phase difference of the light source. The consideration of phase difference is to ensure that the phase relationship between different light sources can be effectively coordinated, thereby avoiding the adverse effects of phase differences on the lighting effect.By combining optimization functions, the stability and coordination of the light source configuration are improved, thereby optimizing the overall lighting effect. In order to verify the stability of the obtained collaborative light source configuration, a regional stability analysis is performed. For the first light source group, the energy function method is used to verify whether the parameter configuration of this group of light sources has sufficient stability. For the second light source group, the Lyapunov direct method is used, which verifies the stability of the parameter configuration by analyzing the dynamic stability of the system. If some configurations are unstable, the system will correct them through feedback compensation to ensure that the final configuration parameters can operate stably in practical applications. Through this series of optimization, verification and correction, the collaborative light source configuration parameters that meet all constraints are finally obtained, including the optimal angle and intensity configuration values ​​for each light source channel.

[0038] In the embodiment of the present invention, by introducing a dynamic graph convolutional network and a dual attention mechanism, deep extraction and enhancement of defect features are achieved, and a dual-path deep neural network architecture is used to perform quality evaluation and reflected light intensity analysis on defect images; based on the combination of a nonlinear adaptive observer and a support vector regression model, a light source parameter identification and control system with self-learning ability is constructed; by designing a hierarchical optimization strategy and an orthogonal Pareto criterion, the coordinated configuration and dynamic adjustment of multiple light sources are achieved. This method not only significantly improves the accuracy and reliability of defect detection, but also greatly reduces the manual debugging cost of light source parameter configuration, and has a strong engineering application value. Compared with the prior art, the present invention has achieved significant improvements in detection accuracy, system adaptability, and configuration efficiency, and provides a new technical solution for the field of semiconductor wafer defect detection.

[0039] In a specific embodiment, the process of executing step S1 may specifically include the following steps:

[0040] The semiconductor wafer image is preprocessed by grayscale and normalization to obtain a preprocessed wafer image, and the corresponding K nearest neighbor point set is calculated for each pixel point in the preprocessed wafer image, and the K nearest neighbor graph structure is constructed by Euclidean distance measurement to obtain an edge point feature set;

[0041] The edge point feature set is input into the dynamic graph convolution layer. The dynamic graph convolution layer contains 3 graph convolution units. Each graph convolution unit uses the ReLU activation function and edge feature update to obtain a dynamic feature map.

[0042] Perform kernel point convolution processing on the dynamic feature map, calculate the similarity matrix between the center point and the neighborhood point through the Gaussian kernel function, and perform feature aggregation in combination with the light intensity contrast information to obtain the initial defect feature map;

[0043] The initial defect feature map is input into the spatial attention submodule, which contains three convolutional layers and one sigmoid layer. Each convolutional layer is followed by a BatchNorm layer. The spatial enhanced feature map is obtained by calculating the spatial weight matrix.

[0044] Perform channel attention processing on the spatial enhancement feature map, generate channel descriptors through global average pooling and maximum pooling, use multi-layer perceptron to calculate the channel weight vector, and obtain the channel enhancement feature map;

[0045] The spatial enhanced feature map and the channel enhanced feature map are fused, and the fused features are reduced in dimension using the residual connection structure and the 1×1 convolution layer to obtain the fused feature map. The fused feature map is then adaptively calibrated, and the regional features are reconstructed by calculating the feature similarity matrix. The original features are combined for residual learning to obtain the enhanced defect feature map.

[0046] Specifically, the semiconductor wafer image is preprocessed by grayscale and normalization, and the color image is converted into a single-channel image to facilitate subsequent feature extraction. The grayscale operation converts the RGB value of each pixel into a grayscale value according to a weighted average method. The commonly used weighted formula is:

[0047] ;

[0048] in, , and are the pixel values ​​of the red, green, and blue channels respectively, is the output grayscale value. The normalization process maps the image pixel value to the interval from 0 to 1 using the following formula:

[0049] ;

[0050] in, and are the minimum and maximum pixel values ​​of the image, respectively. Through this step, the preprocessed wafer image is obtained. The corresponding K nearest neighbor point set is calculated for each pixel in the preprocessed wafer image. For each pixel, the Euclidean distance of its surrounding pixels is calculated, and the K pixels with the smallest distance are selected as the nearest neighbor points. The Euclidean distance formula is:

[0051] ;

[0052] in, and Represent the coordinates or feature values ​​of two pixels respectively. Through Euclidean distance calculation, a K nearest neighbor graph structure is established for each pixel point. Each node of the graph represents a pixel in the image, and the weight of the edge represents the distance or similarity between pixels. In the graph structure, the edge point feature set refers to the pixel point feature set associated with the potential defect position in the image. The relative position of these points and the change in pixel value reveal the defect information on the wafer surface. The edge point feature set is input into the dynamic graph convolution layer. The dynamic graph convolution network is a neural network that can process graph structured data. The function of dynamic convolution is to locally update the features of different pixel points in the image. The dynamic graph convolution layer contains 3 graph convolution units, each of which uses the ReLU activation function and edge feature update. The formula of the ReLU activation function is:

[0053] ;

[0054] Each convolution unit updates the node features through graph convolution operation. The update formula of graph convolution is expressed as:

[0055] ;

[0056] in, Representation Node In the The characteristics of the layer, Representation Node The set of neighbor nodes of It is The weight matrix of the layer, is the activation function, Is a node and nodes The normalization coefficient between is the bias term. Through the processing of the dynamic graph convolution unit, the relationship and features between nodes can be captured in the image, so that the features of each node (pixel) are updated under the influence of the adjacent nodes, and a dynamic feature map is obtained. The dynamic feature map needs to be processed by kernel point convolution. At this stage, the Gaussian kernel function is used to calculate the similarity matrix between the center point and the neighborhood point. The formula of the Gaussian kernel function is:

[0057] ;

[0058] in, is the standard deviation of the Gaussian kernel, which determines the range of similarity. The similarity matrix of each point in the image is calculated through the Gaussian kernel function, and then the feature aggregation is performed in combination with the intensity contrast information. The intensity contrast information is measured by the difference in the values ​​of adjacent pixels in the image, and the local contrast is calculated using the following formula:

[0059] ;

[0060] in, It's a pixel The gray value of is the average grayscale value of the neighborhood area of ​​the pixel. Feature aggregation combines the similarity and contrast information of each pixel by weighted averaging to obtain the initial defect feature map. The initial defect feature map is input into the spatial attention submodule. The task of the spatial attention module is to enhance the features of different regions in the image. The module contains 3 convolutional layers and 1 sigmoid layer, each convolutional layer is followed by a BatchNorm layer. Through these convolutional layers and BatchNorm layers, the module learns the importance of different spatial regions and weights the regional features by calculating the spatial weight matrix. The calculation formula of the spatial weight matrix is:

[0061] ;

[0062] in, Indicates area The weight of Represents the spatial attention value of the area. The calculation of the weight matrix is ​​normalized by the sigmoid activation function, and the result is a weight value between 0 and 1, which indicates the importance of the area. By calculating the spatial weight matrix, the spatial enhancement feature map obtained can effectively enhance those areas that are most useful for defect detection. Channel attention processing is performed on the spatial enhancement feature map to improve the expression ability of important feature channels. Channel descriptors are generated through global average pooling and maximum pooling, and the weight vector of the channel is calculated using a multi-layer perceptron. The calculation formula for channel weight is:

[0063] ;

[0064] Among them, pooling Indicates that a global pooling operation is performed on the input feature map to obtain a channel descriptor. MLP is a multi-layer perceptron network. is a sigmoid function used to calculate the weight of each channel. The final channel enhancement feature map improves the expressiveness of important feature channels by weighting the channels. The spatial enhancement feature map and the channel enhancement feature map are fused. By using the residual connection structure and the 1×1 convolution layer, the fused feature map is reduced in dimension to obtain the fused feature map. The fused feature map is then optimized through adaptive feature calibration. The process of adaptive feature calibration is completed by calculating the feature similarity matrix, and the formula is:

[0065] ;

[0066] in, is the feature similarity matrix, and the softmax function is used to calculate the similarity between each feature. Through residual learning with the original features, an enhanced defect feature map is finally obtained, which can effectively highlight the defect area in the semiconductor wafer image.

[0067] In a specific embodiment, the process of executing step S2 may specifically include the following steps:

[0068] The enhanced defect feature map is input into the image enhancement sub-network and the quality assessment sub-network respectively. The image enhancement sub-network adopts the U-Net structure, and the quality assessment sub-network adopts the multi-scale feature extraction structure.

[0069] The enhanced defect feature map is extracted through the encoder of the image enhancement sub-network. The encoder contains four downsampling units, each of which contains two 3×3 convolutional layers and one maximum pooling layer to obtain multi-scale encoding features.

[0070] The multi-scale coded features are input into the decoder of the image enhancement sub-network. The decoder contains four upsampling units, each of which contains one deconvolution layer and two 3×3 convolution layers. The coded features of the corresponding scales are fused through jump connections to obtain the reflected light intensity data of the defect area.

[0071] The enhanced defect feature map is extracted through the three parallel convolution branches of the quality evaluation subnetwork. Each branch contains four dilated convolution layers with different dilation rates to obtain multi-scale receptive field features.

[0072] The multi-scale receptive field features are input into the global context attention module, and the context enhancement features are obtained by calculating the correlation matrix between feature channels and the spatial position correlation matrix. The context enhancement features are adaptively aggregated, and the channels of different branch features are reorganized through a 1×1 convolution layer. The pyramid pooling module is used to extract multi-scale global information to obtain the aggregated features.

[0073] The aggregated features are transformed to obtain feature vectors, and multi-task learning is performed on the feature vectors to calculate the image clarity score, contrast score, and defect completeness score respectively, and the image quality score data is obtained by weighted averaging.

[0074] Specifically, the enhanced defect feature map is input into two sub-networks respectively: the image enhancement sub-network and the quality assessment sub-network. The image enhancement sub-network adopts the U-Net structure, while the quality assessment sub-network adopts the multi-scale feature extraction structure. The two sub-networks work together to evaluate the quality of the defective image while ensuring image enhancement. The image enhancement sub-network performs deep feature extraction and image enhancement on the enhanced defect feature map. The encoder part of the network consists of 4 downsampling units, each of which contains two 3×3 convolutional layers and one maximum pooling layer. In each convolutional layer, the 3×3 convolution kernel captures the local features in the image through convolution operations, while the pooling layer can reduce the size of the feature map while retaining key information. For each downsampling unit, the function of the convolution layer is to extract the high-order features of the image, while the maximum pooling layer can further reduce the dimension and compress the spatial information in the feature map. Assume that the input feature map is , the convolution operation is expressed as:

[0075] ;

[0076] in, is the convolution kernel, is the bias term, Represents the input image feature map The convolution operation is performed and the output is the feature map after convolution. The maximum pooling operation is expressed by the following formula:

[0077] ;

[0078] Through convolution and pooling operations, the network can gradually extract low-level to high-level features in the image. After each downsampling unit, the spatial resolution of the image is reduced, but the feature representation becomes more abstract and complex. The extracted multi-scale coding features are input into the decoder part of the image enhancement subnetwork. The decoder consists of 4 upsampling units, each of which contains a deconvolution layer and two 3×3 convolution layers. The purpose of the deconvolution layer (also called the upsampling layer) is to restore the spatial size of the feature map so that the image can be gradually restored to its original resolution. Assume that the deconvolution operation in the decoder is expressed as:

[0079] ;

[0080] Among them, DeConv It means that the input feature map is deconvolved, and the output is the upsampled feature map. The deconvolution operation restores the spatial size of the feature map through interpolation or other methods. The restored feature map is further refined through two layers of 3×3 convolutional layers, so that the details of the reflected light intensity are enhanced. Through the jump connection, the features of the corresponding scales of the decoder and encoder stages are fused to ensure that the high-resolution features can be retained. The implementation formula of the jump connection is:

[0081]

[0082] in, represents the encoder output features in the skip connection, is the fused feature map. At the same time, the quality assessment subnetwork performs quality assessment on the enhanced defect feature map, including image clarity, contrast, and defect integrity. The quality assessment subnetwork contains three parallel convolution branches, each of which contains a dilated convolution layer with a different dilation rate. Dilated convolution is a method of expanding the receptive field of the convolution kernel, so that the convolution operation can capture a wider range of contextual information without increasing the amount of calculation. The formula for the dilated convolution layer is:

[0083] ;

[0084] in, is the convolution kernel, is the expansion rate, Represents the pixel point after the dilation operation in the input image. Through multiple dilated convolution branches with different dilation rates, the network can extract the multi-scale receptive field features of the image at multiple scales, thereby capturing the detailed information in the image more comprehensively. The multi-scale receptive field features are input to the global context attention module, which is responsible for calculating the correlation matrix between feature channels and the spatial position correlation matrix in order to enhance the important features in the image. The calculation of the correlation matrix is ​​expressed by the following formula:

[0085] ;

[0086] in, represents the correlation between channels, and are the feature vectors of different channels, represents the L2 norm. The spatial position correlation matrix reflects the dependency between each position in the image. By calculating these matrices, the network can obtain the contextual enhancement features of the image. Adaptive feature aggregation is performed on the contextual enhancement features. The features of different branches are reorganized through the 1×1 convolution layer to fuse the feature information of different scales. The purpose of adaptive aggregation is to enhance the key parts of the image by selecting the optimal feature combination. In order to improve the expressiveness of the features, the network uses a pyramid pooling module, which extracts the global information of the image at different scales through multiple pooling operations, thereby enriching the features. The aggregated features are input into the feature transformation module and converted into a feature vector, which contains the global information of the image and local details. Through multi-task learning, the feature vector is used to calculate the clarity score, contrast score and defect completeness score of the image. These scores reflect the quality of the image, especially the performance of the defect area. The clarity score measures the sharpness of the image, the contrast score measures the light intensity change of the image, and the defect completeness score reflects the integrity of the defect area. The three scores are weighted averaged to obtain the image quality score data, which is used as the basis for judging the surface quality of the wafer and assists in subsequent defect detection decisions.

[0087] In a specific embodiment, the process of executing step S3 may specifically include the following steps:

[0088] Obtain light source input current and light intensity sensor data of multiple light sources, normalize the light source input current data and light intensity sensor data, and construct a feature vector matrix in combination with the image quality score data. Each row of the feature vector matrix contains current data, light intensity data and corresponding quality score of a single light source.

[0089] The eigenvector matrix is ​​nonlinearly mapped through the radial basis kernel function to obtain a high-dimensional eigenmatrix, and a quadratic programming problem is constructed for the high-dimensional eigenmatrix. The penalty factor is set to 10, the insensitive loss parameter is set to 0.1, and the dual problem is solved through the sequential minimum optimization algorithm to obtain the support vector set.

[0090] The weight vector and bias term of the regression function are calculated based on the support vector set, and the input support vector set is regressed and predicted to obtain the initial value of the light source angle. The initial value of the light source angle is input into the optical characteristic compensation module, and the angle correction amount is calculated according to the Fresnel reflection model of the wafer surface to obtain the compensated light source angle data;

[0091] The compensated light source angle data is cross-validated to obtain the light source angle confidence, and the parameters are optimized based on the light source angle data and the light source angle confidence, combined with the physical constraints of the light source, to obtain the initial parameters of the light source.

[0092] Specifically, the light source input current and light intensity sensor data of multiple light sources are obtained. The light source input current data and light intensity sensor data are normalized to convert them into a standardized data range. Combined with the image quality score data, a feature vector matrix is ​​constructed. Assume that the image quality score data is , the feature vector of each light source includes current, light intensity and corresponding quality score, that is, the form of each row of feature vector is .for light sources, eigenvector matrix It is expressed as:

[0093] ;

[0094] The eigenvector matrix is ​​used to optimize the angle and intensity of the light source. In this process, the eigenvector matrix is ​​nonlinearly mapped through the radial basis kernel function to map the original data into a high-dimensional feature space. The form of the radial basis kernel function is:

[0095] ;

[0096] in, and are two eigenvectors, is the Euclidean distance between them, is the width parameter of the kernel function. Through this mapping, the eigenvector is projected into a high-dimensional space, making it easier to solve the optimization problem of light source angle and intensity in this space. The mapped eigenvector is input into a quadratic programming problem. The goal of the quadratic programming problem is to minimize a quadratic function, including a penalty factor and an insensitive loss parameter. The constructed quadratic programming problem is expressed as:

[0097] ;

[0098] in, is the weight vector of the regression function, is the bias term, is the slack variable, is a penalty factor used to control the balance between loss and complexity. Set the penalty factor to 10 and the insensitive loss parameter to 0.1, and solve this quadratic programming problem to obtain a set of support vectors. The support vector set contains key sample points and has an important influence on the decision boundary of the model. The weight vector and bias term of the regression function are calculated based on the support vector. The goal of the regression function is to predict the initial value of the light source angle. The regression function is expressed as:

[0099] ;

[0100] in, is the feature vector after mapping to high-dimensional space, and are the weight and bias of the regression function respectively. By performing regression prediction on the support vector set, the initial value of the light source angle is obtained. Assume that the regression prediction result is , then the initial angle data of each light source is obtained. The initial value of the light source angle is input into the optical property compensation module. This module uses the Fresnel reflection model of the wafer surface to calculate the correction of the light source angle. The Fresnel reflection model describes the law of light reflection on the surface of different media, and the relationship between the reflection angle and the incident angle is given by Snell's law. Through this model, the angle correction is calculated based on the initial light source angle and the physical properties of the wafer surface (such as refractive index). Assume that the correction is , then the compensated light source angle It is expressed as:

[0101] ;

[0102] In order to improve the accuracy of the model, the compensated light source angle data is cross-validated. Cross-validation is a technique that effectively evaluates the performance of the model by dividing the data set into multiple subsets for training and validation. Through cross-validation, the confidence of the light source angle is obtained, that is, the reliability of the angle prediction. The result of cross-validation generates a confidence value for each light source angle prediction. , indicating the reliability of the prediction results. According to the light source angle data and confidence data, combined with the physical constraints of the light source (such as angle range and light intensity power limit), parameter optimization is performed. In the process of optimizing the light source angle, the angle range and power limit are included in the optimization problem as constraints. Assume that the physical constraints of the light source angle are and power constraints By solving the optimization problem, the final initial parameters of the light source are obtained, including the optimal angle and intensity configuration of each light source.

[0103] In a specific embodiment, the process of executing step S4 may specifically include the following steps:

[0104] The reflected light intensity data is preprocessed, the light intensity signal is bandpass filtered and normalized, and the state space equation is constructed in combination with the initial parameters of the light source to obtain the state matrix of the observation system;

[0105] Perform Lyapunov stability analysis on the state matrix of the observation system, obtain the positive definite matrix by solving the Lyapunov equation, and construct the feedback gain calculation criterion based on the positive definite matrix to obtain the gain constraint condition;

[0106] The gain constraint condition is input into the pole placement module, the closed-loop pole distribution is calculated using the Ackerman formula, and the initial value of the state feedback gain matrix is ​​obtained through matrix operation to obtain the feedback gain data;

[0107] Decoupling transformation is performed on the state space equation according to the feedback gain data, and the coupled angle state and intensity state are converted into the standard form of independent control to obtain the decoupled state equation;

[0108] The nonlinear compensation of the decoupled state equation is performed, the nonlinear term of the system is eliminated by constructing a feedback linearization controller, and the compensation control quantity is calculated in combination with the initial parameters of the light source to obtain the compensation control data;

[0109] Input the compensation control data into the sliding mode controller to distribute the control quantity, perform saturation processing according to the physical constraints of the light source angle and intensity, and solve the distribution control data;

[0110] The distribution control data and the feedback gain data are combined and optimized to obtain light source control parameters, which include the angle control amount and intensity control amount of each light source.

[0111] Specifically, the reflected light intensity data is preprocessed to ensure the quality and availability of the data. The light intensity signal is bandpass filtered to remove noise with too low or too high frequency and retain the effective components of the signal. The transfer function of the bandpass filter is expressed as:

[0112] ;

[0113] in, and are the low-frequency and high-frequency cutoff points of the bandpass filter, respectively. is the frequency of the signal. The filtered signal can effectively remove high-frequency noise and low-frequency drift, retaining the main characteristics of the light intensity signal. Normalization is performed to make the signal amplitude in a uniform range. Combined with the initial parameters of the light source (such as angle, intensity, and light source position, etc.), a state space equation is constructed to describe the dynamic behavior of the system. Assume that the system state is , the input is , the output is , the state space equation is expressed as:

[0114] ;

[0115] ;

[0116] in, is the state transition matrix of the system, is the input matrix, is the output matrix, is a direct transfer matrix. By setting these matrices, the dynamic response behavior of the system and how the reflected light intensity data affects the control target of the system are described. The constructed observation system state matrix is ​​subjected to Lyapunov stability analysis. Lyapunov stability analysis determines whether the system can return to equilibrium under disturbance by analyzing the state space equations of the system. A common form of the Lyapunov method is to construct a Lyapunov function , and verify that it satisfies the Lyapunov equation:

[0117] ;

[0118] If this condition is met, the system is stable. In this process, the construction of the Lyapunov function needs to select an appropriate function form according to the specific form of the state space equation. Usually, the quadratic function is selected. ,in is a positive definite matrix. Based on the positive definite matrix Construct a feedback gain calculation criterion. The goal of this criterion is to design a feedback controller that makes the system stable and achieves the desired performance indicators. The feedback gain matrix The solution is obtained through the pole placement method. The pole placement method optimizes the response speed and stability of the system by adjusting the closed-loop poles of the system. The position of the closed-loop poles is calculated using the Ackerman formula, and the initial value of the state feedback gain matrix is ​​obtained through matrix operations. The form of the Ackerman formula is:

[0119] ;

[0120] Among them, place , desired poles) means configuring the closed-loop poles of the system to the desired position, and obtaining the feedback gain matrix . Through the feedback gain matrix , decoupling transformation is performed on the state space equation. Since the original state space equation usually involves the coupling of angle state and light intensity state, the two coupled states are converted into a standard form of independent control through decoupling transformation. to complete, so that the new state variable satisfy:

[0121] ;

[0122] in, It is a state transformation matrix that can map the coupled angle and intensity states to independent standard states. After the decoupling transformation, the state equation is nonlinearly compensated. The light source control system is affected by nonlinear factors, such as the nonlinear relationship between light intensity and angle. These nonlinear terms are eliminated by constructing a feedback linearization controller. By designing appropriate control input , so that the output behavior of the system becomes linear. Assume that the nonlinear control term of the system is and , then by choosing a suitable control law , the nonlinear term and Compensate it to make the system behavior linear. The input is sent to the sliding mode controller for control quantity distribution. The sliding mode controller can ensure that the control quantity of the light source angle and intensity is within the physical constraint range by saturating the control input. For example, assuming that the angle range of the light source is , the intensity range is , the sliding mode controller performs saturation processing on the control variable to ensure that the angle and intensity are always within the acceptable range. The saturation processing is performed using the following formula:

[0123] ;

[0124] ;

[0125] in, and are the control values ​​of angle and intensity respectively, and is the control quantity after saturation processing. The distribution control data and feedback gain data are combined and optimized to obtain the final light source control parameters. The light source control parameters include the angle control quantity and intensity control quantity of each light source. The purpose is to achieve optimal control of the semiconductor wafer defect detection system under the premise of meeting physical constraints.

[0126] In a specific embodiment, the process of executing step S5 may specifically include the following steps:

[0127] The light source control parameters are grouped by channel, the angle and intensity parameters of each light source channel are standardized to obtain standardized light source parameters, and the standardized light source parameters are subjected to correlation analysis to obtain channel correlation data;

[0128] The channel association data is input into the attention calculation module, the attention scores between channels are calculated through the softmax function, and the spatial constraint matrix is ​​constructed in combination with the physical position information of the light source to obtain the initial channel weights;

[0129] The initial channel weights are normalized and exponentially smoothed in combination with the historical weights of the light source channels to obtain smoothed channel weights. The light source control parameters are weighted and allocated according to the smoothed channel weights. The complementary constraints between light sources are considered to obtain optimized weight data.

[0130] The optimized weight data is input into the collaborative control module, and the light source control parameters are collaboratively optimized by building a distributed consistency protocol to obtain collaborative control data;

[0131] The collaborative control data is constrained and parameter truncation is performed according to the power limit and angle range of the light source. The optimal solution that satisfies the constraints is solved to obtain the constrained control data. The constrained control data is combined with the optimized weight data to obtain the collaborative light source configuration parameters. The collaborative light source configuration parameters include the optimal angle and intensity configuration values ​​of each light source channel.

[0132] Specifically, the control parameters of multiple light sources are grouped by channels, and the angle and intensity parameters of each light source channel are standardized to eliminate the impact of different dimensions between different channels, so that the parameters are on a unified scale and the standardized light source parameters are obtained. The standardized light source parameters are subjected to correlation analysis, and the channel correlation data is obtained by calculating the correlation between different light source channels. The purpose of correlation analysis is to evaluate the linear relationship between different light source channels in order to identify which light source channels have strong synergy. Correlation analysis is achieved by calculating the Pearson correlation coefficient. The standardized parameters of the two channels are set to be and , and its correlation coefficient is calculated by the following formula:

[0133] ;

[0134] in, It is a channel and The correlation coefficient between and Channel and The mean of and is the specific data point of the channel. A channel correlation matrix is ​​constructed based on the results of the correlation analysis, in which each element reflects the strength of the relationship between different light source channels. The channel correlation data is input into the attention calculation module, which calculates the attention score between each channel by using the softmax function. The softmax function is used to convert a set of real numbers into a probability distribution, and its formula is:

[0135] ;

[0136] in, is the correlation measure for each channel, It is a channel The attention score, is the total number of channels. The attention score reflects the relative importance of each light source channel in the entire system. A higher score means that the channel needs more attention during the optimization process. A spatial constraint matrix is ​​constructed in combination with the physical position information of the light source. This matrix is ​​used to describe the spatial relationship between different light source channels, such as their position distribution on the wafer surface. The physical position information is described by parameters such as distance and angle, which affect the working efficiency and mutual interference of the light source. The relationship between the physical position of the light source and its operating performance is introduced into the optimization process through the spatial constraint matrix. Based on the preliminary attention calculation results, the initial channel weights are obtained. The initial weights are normalized and exponentially smoothed in combination with the historical weights of the light source channels to obtain smoothed channel weights, thereby improving the adaptability of the light source control system to long-term changes. Exponential smoothing is performed using the following formula:

[0137] ;

[0138] in, is the smoothed weight, is the historical weight, is the smoothing factor, usually taken between In this way, the weights gradually tend to be stable under the influence of historical data, and can better cope with the long-term changes of light sources. According to the smooth channel weights, the light source control parameters are weighted and distributed, and the complementary constraints between light sources, that is, the synergy between different light source channels, are considered to ensure that the working effect of the light source will not lead to overall performance degradation due to the excessive enhancement of a single channel. The complementary constraint is represented by designing an objective function, which comprehensively considers the collaborative working effect between multiple light source channels. The optimized weight data is input into the collaborative control module, and the distributed consistency protocol in the module is used to collaboratively optimize the light source control parameters. The goal of the distributed consistency protocol is to ensure that different light source channels can maintain consistency under different control conditions and achieve optimal control in a global scope. Through communication and information exchange, the protocol enables each light source channel to be adjusted in a local range to achieve the global optimal configuration. The collaborative control data is constrained to ensure that the optimized light source control parameters meet the actual physical constraints. For example, the angle range and power limit of the light source are fixed, and these constraints are included in the optimization process. Constraint processing is achieved by parameter truncation. If a parameter exceeds the allowable range, it is adjusted back to a reasonable range. Assume that the range of the light source angle is The intensity range is , then parameter truncation is performed using the following formula:

[0139] ;

[0140] ;

[0141] Constraint processing ensures that the control parameters of each light source channel meet physical constraints. The control data after constraint processing is combined with the optimized weight data to obtain the final collaborative light source configuration parameters, including the optimal angle and intensity configuration values ​​of each light source channel. Through these configurations, the system can achieve the optimal semiconductor wafer defect detection effect while meeting various physical and operational constraints.

[0142] In a specific embodiment, the execution step constrains the collaborative control data, performs parameter truncation according to the power limit and angle range of the light source, solves the optimal solution that satisfies the constraints, obtains the constrained control data, and combines the constrained control data with the optimized weight data to obtain the collaborative light source configuration parameters, and the collaborative light source configuration parameters include the optimal angle and intensity configuration values ​​of each light source channel. The process can specifically include the following steps:

[0143] The collaborative control data is constrained and decomposed, and nonlinear constraint equations are established for the angle parameters of the first light source group and the intensity parameters of the second light source group, respectively, where the first light source group is located directly above the wafer and the second light source group is located around the wafer, to construct a group constraint optimization equation;

[0144] The hierarchical quadratic programming is performed to solve the group constraint optimization equation, the angle range is set to [-30°, 30°] and the power threshold is set to 80% for the first light source group, the angle range is set to [-60°, 60°] and the power threshold is set to 60% for the second light source group, and the hierarchical truncation parameters are calculated;

[0145] A multi-objective optimization model with orthogonality is constructed based on hierarchical truncation parameters. The illumination uniformity of the first light source group and the contrast of the second light source group are set as the first optimization objective and the second optimization objective respectively. The orthogonal Pareto criterion is introduced to decouple and optimize the objective function to obtain orthogonal optimization data.

[0146] The orthogonal optimization data is input into the genetic algorithm solver, the dynamic adaptive crossover probability and mutation probability are set, and the parameter optimization calculation is performed by introducing the local search operator and the elite retention strategy to obtain the double-layer optimization parameter sequence;

[0147] The double-layer optimization parameter sequence is grouped and constrained to verify. The first constraint function is constructed to constrain the light intensity uniformity, and the second constraint function is constructed to constrain the angle complementarity. The parameters are corrected by the gradient projection method to obtain a hierarchical feasible parameter set.

[0148] The hierarchical feasible parameter set is orthogonally combined with the optimized weight data, and the parameters are fused by constructing a combined optimization function that takes into account the phase difference of the light source to obtain the orthogonal combined parameter data;

[0149] The orthogonal combination parameter data are subjected to regional stability analysis. The energy function method is used to verify the stability of the parameter configuration of the first light source group. The Lyapunov direct method is used to verify the stability of the parameter configuration of the second light source group. The unstable configuration is corrected through feedback compensation to obtain the collaborative light source configuration parameters, which include the optimal angle and intensity configuration values ​​of each light source channel.

[0150] Specifically, the collaborative control data is constrained and decomposed. In order to meet specific control requirements, the light sources are divided into two groups: the first light source group is located directly above the wafer, and the second light source group is located around the wafer. Each light source group has different control requirements and physical constraints, and different nonlinear constraint equations are established for each group of light sources. The control parameters of the first light source group involve angle restrictions, while the control parameters of the second light source group mainly involve intensity restrictions. A nonlinear constraint equation about the angle is established for the first light source group, and its mathematical form is: ,in The second light source group establishes an intensity constraint equation based on the power limit of the light source, and its mathematical form is:

[0151] ;

[0152] These constraints ensure that the angle and intensity of the light source meet the actual working requirements. By constructing a group constraint optimization equation, the constraints of each group of light sources are introduced into the optimization model. When solving the optimization equation, a hierarchical quadratic programming method is used. For the first light source group, the angle is restricted to , and set its power threshold to 80%; for the second light source group, its angle limit is , the power threshold is 60%. This means that the angle and intensity of the light source need to be adjusted within these physical limitations to ensure the overall efficiency of the system. To this end, a hierarchical truncation parameter calculation method is designed to constrain the control parameters of each light source and ensure that it meets specific hard requirements. Hierarchical truncation gradually adjusts and restricts parameter values ​​according to the constraints of different light source groups to obtain parameter configurations that meet physical constraints. Based on the truncated parameters, an orthogonal multi-objective optimization model is constructed, in which the objective function includes two parts: the first objective is illumination uniformity, and the second objective is contrast. Illumination uniformity and contrast are two key factors affecting the quality of semiconductor wafer inspection, and these two indicators are optimized separately. In order to effectively resolve the conflict between the two optimization objectives, the orthogonal Pareto criterion is introduced. This criterion can decouple the optimization of the two objectives so that the optimization process can meet the requirements of the two objectives at the same time. Pareto optimization gradually adjusts the parameters so that no objective can be further improved while keeping other objectives unchanged. It is expressed by the following formula:

[0153] , and Cannot be improved at the same time;

[0154] in, and are the optimization objective functions of illumination uniformity and contrast, respectively. is the control parameter. In the orthogonal optimization process, the optimization data is input into the genetic algorithm solver, and the optimal solution is sought through the genetic algorithm. Genetic algorithm is an optimization method that simulates the natural selection process. The algorithm's search process is adjusted by setting the dynamic adaptive crossover probability and mutation probability. The local search operator and elite retention strategy in the genetic algorithm can help the algorithm converge to the optimal solution more quickly during the search process and avoid falling into the local optimal solution. The local search operator explores better solutions through local fine-tuning, while the elite retention strategy ensures that the best individuals in each generation can be passed to the next generation to maintain the quality of the solution. Through the calculation of the genetic algorithm, a double-layer optimization parameter sequence is obtained. In order to verify whether these optimization results meet the constraints, the optimization parameters are grouped and constrained. A first constraint function is constructed to constrain the uniformity of light intensity to ensure uniform distribution of light intensity between each light source; secondly, a second constraint function is constructed to ensure that the angles between different light source channels have certain complementarity to avoid overlap and interference of light sources. When correcting the constraints, the gradient projection method is used, which can effectively adjust the parameters that do not meet the constraints to the feasible solution area. The gradient projection method calculates the gradient of the objective function and projects it onto the constraint set to obtain the optimal solution that meets the constraints. The gradient projection formula is as follows:

[0155] ;

[0156] in, is the parameter vector, is the gradient of the objective function, is the learning rate, is a constraint function. After completing these steps, a hierarchical feasible parameter set is obtained, in which each parameter satisfies the corresponding constraint and can achieve the best performance under the given optimization goal. The hierarchical feasible parameter set is orthogonally combined with the optimization weight data to form a new optimization data set. In the combination process, the phase difference between the light sources is considered. By constructing a combined optimization function, the phase difference is comprehensively considered and introduced into the parameter optimization process to improve the overall performance of the system. The combined optimization function is expressed in the following form:

[0157] ;

[0158] in, is the regularization factor, is the phase difference between the light sources, and They are the optimization targets of illumination uniformity and contrast, respectively. The combined orthogonal parameter data is subjected to regional stability analysis, and two different methods are used to verify the stability of the parameter configuration. For the first light source group, the energy function method is used to analyze its stability; and for the second light source group, the Lyapunov direct method is used to verify the stability. These two methods can analyze the stability of the light source configuration from the two aspects of energy balance and system dynamics stability, respectively. If an unstable configuration is found in the analysis, it will be corrected through a feedback compensation mechanism to ensure that the final light source configuration is stable. After this series of optimization and verification steps, the collaborative light source configuration parameters are obtained, including the optimal angle and intensity configuration values ​​of each light source channel, and the best performance is achieved while meeting all constraints.

[0159] The above describes the semiconductor wafer defect detection light source configuration method in the embodiment of the present invention. The following describes the semiconductor wafer defect detection light source configuration device in the embodiment of the present invention. Figure 2 , an embodiment of a semiconductor wafer defect detection light source configuration device in an embodiment of the present invention includes:

[0160] An acquisition module is used to acquire a semiconductor wafer image and perform dynamic graph convolution processing and defect feature enhancement to obtain an enhanced defect feature map;

[0161] A scoring module is used to input the enhanced defect feature map into a dual-channel deep neural network to calculate the reflected light intensity and image quality score, so as to obtain the reflected light intensity data of the defect area and the image quality score data;

[0162] A calculation module is used to obtain the light source input current and light intensity sensor data of multiple light sources, and identify and calculate the light source angle parameters in combination with the image quality score data to obtain the initial parameters of the light source;

[0163] A control module, used to perform nonlinear control on the light source angle and light source intensity based on the reflected light intensity data and the initial parameters of the light source to obtain the light source control parameters;

[0164] The processing module is used to calculate the channel weights of multiple light sources and perform differentiated processing on the light source control parameters to obtain the collaborative light source configuration parameters.

[0165] Through the collaborative cooperation of the above-mentioned components, by introducing dynamic graph convolutional networks and dual attention mechanisms, deep extraction and enhancement of defect features are achieved, and a dual-path deep neural network architecture is used to evaluate the quality of defect images and analyze the reflected light intensity; based on the combination of nonlinear adaptive observers and support vector regression models, a light source parameter identification and control system with self-learning ability is constructed; by designing hierarchical optimization strategies and orthogonal Pareto criteria, collaborative configuration and dynamic adjustment of multiple light sources are achieved. This method not only significantly improves the accuracy and reliability of defect detection, but also greatly reduces the manual debugging cost of light source parameter configuration, and has strong engineering application value. Compared with the prior art, the present invention has achieved significant improvements in detection accuracy, system adaptability, and configuration efficiency, and provides a new technical solution for the field of semiconductor wafer defect detection.

[0166] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0167] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0168] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0169] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0170] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0171] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0172] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for configuring a light source for semiconductor wafer defect detection, characterized in that: The method comprises: Acquire a semiconductor wafer image and perform dynamic graph convolution processing and defect feature enhancement to obtain an enhanced defect feature map; Inputting the enhanced defect feature map into a dual-path deep neural network to calculate reflected light intensity and image quality score, thereby obtaining reflected light intensity data and image quality score data of the defect area; Acquire light source input current data and light intensity sensor data of multiple light sources, and identify and calculate light source angle parameters in combination with the image quality score data to obtain initial light source parameters; Based on the reflected light intensity data and the initial parameters of the light source, nonlinearly control the light source angle and the light source intensity to obtain light source control parameters; The channel weights of the multiple light sources are calculated, and the light source control parameters are differentiated to obtain collaborative light source configuration parameters.

2. The semiconductor wafer defect detection light source configuration method according to claim 1, characterized in that: The method of acquiring a semiconductor wafer image and performing dynamic graph convolution processing and defect feature enhancement to obtain an enhanced defect feature map includes: The semiconductor wafer image is preprocessed by grayscale and normalization to obtain a preprocessed wafer image, and a corresponding K nearest neighbor point set is calculated for each pixel point in the preprocessed wafer image, a K nearest neighbor graph structure is constructed by Euclidean distance metric, and an edge point feature set is obtained; Inputting the edge point feature set into a dynamic graph convolution layer, wherein the dynamic graph convolution layer comprises three graph convolution units, each of which uses a ReLU activation function and an edge feature update to obtain a dynamic feature graph; Performing kernel point convolution processing on the dynamic feature map, calculating the similarity matrix between the center point and the neighborhood points through the Gaussian kernel function, and performing feature aggregation in combination with the light intensity contrast information to obtain an initial defect feature map; The initial defect feature map is input into a spatial attention submodule, wherein the spatial attention submodule comprises three convolutional layers and one sigmoid layer, each convolutional layer is followed by a BatchNorm layer, and a spatial enhancement feature map is obtained by calculating a spatial weight matrix; Performing channel attention processing on the spatial enhancement feature map, generating a channel descriptor through global average pooling and maximum pooling, and calculating a channel weight vector using a multi-layer perceptron to obtain a channel enhancement feature map; The spatial enhancement feature map and the channel enhancement feature map are feature fused, and the fused features are reduced in dimension using a residual connection structure and a 1×1 convolutional layer to obtain a fused feature map. The fused feature map is adaptively calibrated, and regional features are reconstructed by calculating a feature similarity matrix. Residual learning is performed in combination with the original features to obtain an enhanced defect feature map.

3. The semiconductor wafer defect detection light source configuration method according to claim 1, characterized in that: The enhanced defect feature map is input into a dual-path deep neural network to calculate the reflected light intensity and image quality score, and the reflected light intensity data and image quality score data of the defect area are obtained, including: The enhanced defect feature map is input into an image enhancement sub-network and a quality assessment sub-network respectively, wherein the image enhancement sub-network adopts a U-Net structure and the quality assessment sub-network adopts a multi-scale feature extraction structure; Extracting features from the enhanced defect feature map through an encoder of an image enhancement subnetwork, wherein the encoder includes four downsampling units, each of which includes two 3×3 convolutional layers and one maximum pooling layer, to obtain multi-scale coding features; The multi-scale coding features are input into the decoder of the image enhancement sub-network, the decoder includes 4 upsampling units, each of which includes a deconvolution layer and two 3×3 convolution layers, and the coding features of the corresponding scales are fused through jump connections to obtain the reflected light intensity data of the defect area; The enhanced defect feature map is subjected to feature extraction through three parallel convolution branches of the quality evaluation subnetwork, each branch comprising four dilated convolution layers with different dilation rates, to obtain multi-scale receptive field features; Input the multi-scale receptive field features into the global context attention module, obtain context enhancement features by calculating the correlation matrix between feature channels and the spatial position correlation matrix, perform adaptive feature aggregation on the context enhancement features, reorganize channels of different branch features through a 1×1 convolution layer, and use a pyramid pooling module to extract multi-scale global information to obtain aggregated features; The aggregated features are transformed to obtain feature vectors, and multi-task learning is performed on the feature vectors to calculate image clarity scores, contrast scores, and defect completeness scores, respectively, and image quality score data is obtained by weighted averaging.

4. The semiconductor wafer defect detection light source configuration method according to claim 1, characterized in that: The acquiring of light source input current data and light intensity sensor data of the plurality of light sources, and identifying and calculating the light source angle parameters in combination with the image quality score data to obtain the initial parameters of the light source includes: Acquire light source input current data and light intensity sensor data of multiple light sources, normalize the light source input current data and the light intensity sensor data, and construct a feature vector matrix in combination with the image quality score data, wherein each row of the feature vector matrix contains current data, light intensity data and corresponding quality score of a single light source; The eigenvector matrix is ​​nonlinearly mapped by a radial basis kernel function to obtain a high-dimensional eigenmatrix, and a quadratic programming problem is constructed for the high-dimensional eigenmatrix, the penalty factor is set to 10, the insensitive loss parameter is set to 0.1, and the dual problem is solved by a sequential minimum optimization algorithm to obtain a support vector set; Calculating the weight vector and bias term of the regression function based on the support vector set, performing regression prediction on the input support vector set to obtain an initial value of the light source angle, and inputting the initial value of the light source angle into an optical characteristic compensation module, calculating the angle correction amount according to the Fresnel reflection model of the wafer surface, and obtaining compensated light source angle data; The compensated light source angle data is cross-validated to obtain light source angle confidence, and parameter optimization is performed based on the light source angle data and the light source angle confidence in combination with physical constraints of the light source to obtain initial light source parameters.

5. The semiconductor wafer defect detection light source configuration method according to claim 1, characterized in that: The nonlinear control of the light source angle and the light source intensity based on the reflected light intensity data and the initial parameters of the light source to obtain the light source control parameters includes: The reflected light intensity data is subjected to data preprocessing, the light intensity signal is subjected to bandpass filtering and normalization processing, and a state space equation is constructed in combination with the initial parameters of the light source to obtain an observation system state matrix; Performing Lyapunov stability analysis on the state matrix of the observation system, obtaining a positive definite matrix by solving the Lyapunov equation, constructing a feedback gain calculation criterion based on the positive definite matrix, and obtaining a gain constraint condition; The gain constraint condition is input into the pole placement module, the closed-loop pole distribution is calculated using the Ackerman formula, and the initial value of the state feedback gain matrix is ​​obtained through matrix operation to obtain feedback gain data; Decoupling transformation is performed on the state space equation according to the feedback gain data, and the coupled angle state and intensity state are converted into a standard form of independent control to obtain a decoupled state equation; Performing nonlinear compensation on the decoupled state equation, eliminating the nonlinear term of the system by constructing a feedback linearization controller, and calculating the compensation control amount in combination with the initial parameters of the light source to obtain compensation control data; Input the compensation control data into the sliding mode controller to distribute the control quantity, perform saturation processing according to the physical constraints of the light source angle and intensity, and solve the distribution control data; The distribution control data and the feedback gain data are combined and optimized to obtain light source control parameters, where the light source control parameters include an angle control amount and an intensity control amount for each light source.

6. The semiconductor wafer defect detection light source configuration method according to claim 1, characterized in that: The calculating the channel weights of the multiple light sources and performing differentiation processing on the light source control parameters to obtain collaborative light source configuration parameters includes: The light source control parameters are grouped by channel, the angle and intensity parameters of each light source channel are standardized to obtain standardized light source parameters, and the standardized light source parameters are subjected to correlation analysis to obtain channel correlation data; The channel association data is input into the attention calculation module, the attention scores between channels are calculated by the softmax function, and the spatial constraint matrix is ​​constructed in combination with the physical position information of the light source to obtain the initial channel weights; The initial channel weight is normalized, and exponentially smoothed in combination with the historical weight of the light source channel to obtain a smoothed channel weight, and the light source control parameters are weightedly allocated according to the smoothed channel weight, and the complementarity constraints between light sources are considered to obtain optimized weight data; The optimized weight data is input into a collaborative control module, and the light source control parameters are collaboratively optimized by constructing a distributed consistency protocol to obtain collaborative control data; The collaborative control data is constrained, parameters are truncate according to the power limit and angle range of the light source, the optimal solution that satisfies the constraints is solved, the constrained control data is obtained, and the constrained control data is combined with the optimized weight data to obtain collaborative light source configuration parameters, which include the optimal angle and intensity configuration values ​​of each light source channel.

7. The semiconductor wafer defect detection light source configuration method according to claim 6, characterized in that: The collaborative control data is subjected to constraint processing, parameter truncation is performed according to the power limit and angle range of the light source, an optimal solution satisfying the constraint is solved, the constraint control data is obtained, and the constraint control data is combined with the optimization weight data to obtain collaborative light source configuration parameters, wherein the collaborative light source configuration parameters include the optimal angle and intensity configuration values ​​of each light source channel, including: The collaborative control data is subjected to constraint decomposition, and nonlinear constraint equations are established for angle parameters of the first light source group and intensity parameters of the second light source group, respectively, wherein the first light source group is located directly above the wafer and the second light source group is located around the wafer, to construct a group constraint optimization equation; Performing hierarchical quadratic programming to solve the grouping constraint optimization equation, setting the angle range of the first light source group to [-30°, 30°] and the power threshold to 80%, setting the angle range of the second light source group to [-60°, 60°] and the power threshold to 60%, and calculating the hierarchical truncation parameters; Based on the hierarchical truncation parameters, a multi-objective optimization model with orthogonality is constructed, the illumination uniformity of the first light source group and the contrast of the second light source group are respectively set as the first optimization objective and the second optimization objective, and the objective function is decoupled and optimized by introducing the orthogonal Pareto criterion to obtain orthogonal optimization data; Inputting the orthogonal optimization data into a genetic algorithm solver, setting a dynamic adaptive crossover probability and a mutation probability, and performing parameter optimization calculation by introducing a local search operator and an elite retention strategy to obtain a double-layer optimization parameter sequence; Performing group constraint verification on the double-layer optimization parameter sequence, constructing a first constraint function to constrain light intensity uniformity, constructing a second constraint function to constrain angle complementarity, and correcting the parameters by a gradient projection method to obtain a hierarchical feasible parameter set; orthogonally combining the hierarchical feasible parameter set with the optimization weight data, fusing the parameters by constructing a combined optimization function that takes into account the phase difference of the light source, and obtaining orthogonal combination parameter data; A regional stability analysis is performed on the orthogonal combination parameter data, an energy function method is used to verify the stability of the parameter configuration of the first light source group, and a Lyapunov direct method is used to verify the stability of the parameter configuration of the second light source group. The unstable configuration is corrected through feedback compensation to obtain collaborative light source configuration parameters, which include optimal angle and intensity configuration values ​​for each light source channel.

8. A semiconductor wafer defect detection light source configuration device, characterized in that: Used to execute the semiconductor wafer defect detection light source configuration method according to any one of claims 1 to 7, the semiconductor wafer defect detection light source configuration device comprising: An acquisition module is used to acquire a semiconductor wafer image and perform dynamic graph convolution processing and defect feature enhancement to obtain an enhanced defect feature map; A scoring module, used for inputting the enhanced defect feature map into a dual-path deep neural network to calculate the reflected light intensity and score the image quality, so as to obtain the reflected light intensity data and image quality score data of the defect area; A calculation module, used to obtain light source input current data and light intensity sensor data of multiple light sources, and identify and calculate light source angle parameters in combination with the image quality score data to obtain initial light source parameters; A control module, used for performing nonlinear control on the light source angle and the light source intensity based on the reflected light intensity data and the initial parameters of the light source to obtain light source control parameters; The processing module is used to calculate the channel weights of the multiple light sources and perform differentiation processing on the light source control parameters to obtain collaborative light source configuration parameters.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, the semiconductor wafer defect detection light source configuration method described in any one of claims 1 to 7 is implemented. 10 . A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor is enabled to execute the method for configuring a light source for defect detection of a semiconductor wafer according to any one of claims 1 to 7 .

Images