Thin film surface quality detection method and system for semiconductor devices

Abstract

The invention relates to the technical field of surface quality detection, and discloses a method and system for detecting the surface quality of a thin film of a semiconductor device. The method comprises: performing multimodal scanning on the surface of a thin film of a semiconductor device to obtain an original multimodal data set; performing feature extraction and feature dimension reduction to obtain a standard multidimensional feature matrix; performing film quality assessment training by using a multi-branch convolutional neural network and a graph neural network fusion algorithm to obtain a film quality assessment model; performing edge computing real-time monitoring by using the film quality assessment model to obtain a dynamic quality assessment result and a defect prediction map; performing stress modulation Raman spectroscopy and intelligent etching analysis on the dynamic quality assessment result and the defect prediction map to obtain three-dimensional defect distribution verification data of the thin film; and optimizing thin film growth parameters of the semiconductor device based on the three-dimensional defect distribution verification data to generate a target thin film growth control parameter set. The invention improves the accuracy of thin film surface quality detection of semiconductor devices.

Description

Technical Field

[0001] The present invention relates to the technical field of surface quality detection, and in particular to a method and system for detecting the surface quality of a thin film of a semiconductor device. Background Art

[0002] With the rapid development of semiconductor technology, thin film devices play an increasingly important role in modern electronic products. However, the quality of thin films directly affects the performance and reliability of semiconductor devices, so accurate detection and evaluation of thin film surface quality becomes particularly critical. Traditional thin film quality detection methods often rely on single-mode measurement techniques, such as optical microscopy, atomic force microscopy or X-ray diffraction. Although these methods can provide information in certain specific aspects, it is difficult to fully reflect the comprehensive quality characteristics of the film.

[0003] In addition, as semiconductor devices develop towards higher integration and smaller size, the scale of film defects has also shrunk, which places higher requirements on the accuracy and resolution of detection technology. At the same time, traditional offline detection methods cannot meet the needs of real-time monitoring and rapid feedback in modern semiconductor manufacturing processes, which leads to reduced production efficiency and increased costs. Another important issue is that existing film quality assessment methods often lack the ability to comprehensively analyze multidimensional data, making it difficult to accurately capture the complex distribution and evolution of film defects. This limitation makes it difficult for researchers and engineers to deeply understand the key influencing factors in the film growth process, thereby affecting the optimization and control of film growth parameters. Summary of the invention

[0004] The invention provides a method and system for detecting the surface quality of a thin film of a semiconductor device, and the invention improves the detection accuracy of the surface quality of the thin film of the semiconductor device.

[0005] In a first aspect, the present invention provides a method for detecting the surface quality of a thin film of a semiconductor device, the method comprising:

[0006] Perform multimodal scanning on the thin film surface of the semiconductor device to obtain an original multimodal data set;

[0007] Performing feature extraction and feature dimensionality reduction on the original multimodal data set to obtain a standard multidimensional feature matrix;

[0008] Inputting the standard multi-dimensional feature matrix into a preset multi-branch convolutional neural network and graph neural network fusion algorithm for film quality assessment training to obtain a film quality assessment model;

[0009] The thin film growth process of the semiconductor device is monitored in real time by edge computing through the thin film quality assessment model to obtain a dynamic quality assessment result and a defect prediction diagram;

[0010] Perform stress modulation Raman spectroscopy and intelligent etching analysis on the dynamic quality assessment result and the defect prediction map to obtain thin film three-dimensional defect distribution verification data;

[0011] Based on the three-dimensional defect distribution verification data, thin film growth parameters of the semiconductor device are optimized to generate a target thin film growth control parameter set.

[0012] In a second aspect, the present invention provides a thin film surface quality detection system for a semiconductor device, the thin film surface quality detection system for a semiconductor device comprising:

[0013] A scanning module, used for performing multimodal scanning on the thin film surface of the semiconductor device to obtain an original multimodal data set;

[0014] An extraction module, used for performing feature extraction and feature dimension reduction on the original multimodal data set to obtain a standard multidimensional feature matrix;

[0015] A training module, used for inputting the standard multi-dimensional feature matrix into a preset multi-branch convolutional neural network and graph neural network fusion algorithm to perform film quality assessment training to obtain a film quality assessment model;

[0016] A monitoring module, used to perform edge computing real-time monitoring of the thin film growth process of the semiconductor device through the thin film quality assessment model to obtain dynamic quality assessment results and defect prediction diagrams;

[0017] An analysis module, used for performing stress modulation Raman spectroscopy and intelligent etching analysis on the dynamic quality assessment result and the defect prediction map to obtain thin film three-dimensional defect distribution verification data;

[0018] A generation module is used to optimize the thin film growth parameters of the semiconductor device based on the three-dimensional defect distribution verification data to generate a target thin film growth control parameter set.

[0019] In the technical solution provided by the present invention, the method performs multi-modal scanning on the surface of the semiconductor device film to obtain comprehensive data including mechanical, chemical and electrical properties, thereby realizing a multi-dimensional evaluation of the film quality and improving the comprehensiveness and accuracy of the detection. Advanced signal processing technologies, such as Fourier transform, wavelet packet decomposition and non-negative matrix decomposition, are used to extract features and reduce the dimension of the original multi-modal data, effectively extracting the key feature information of the film surface. A film quality assessment model is constructed using a multi-branch convolutional neural network and a graph neural network fusion algorithm to realize the intelligent analysis and evaluation of complex film surface features. Real-time monitoring of the film growth process is achieved through edge computing technology, potential quality problems are discovered and predicted in a timely manner, and the control accuracy of the production process is improved. Combined with stress-modulated Raman spectroscopy and intelligent etching analysis, accurate verification of the three-dimensional defect distribution of the film is achieved. Based on the three-dimensional defect distribution verification data, an advanced optimization algorithm is used to optimize the film growth parameters, and a target film growth control parameter set is generated, which effectively improves the quality and efficiency of film growth. From nanoscale surface scanning to macroscopic thin film growth parameter optimization, this method realizes cross-scale full-process analysis and control of thin film quality, thereby improving the accuracy of thin film surface quality detection of semiconductor devices. 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 any creative work.

[0021] Figure 1 It is a schematic diagram of the steps of a method for detecting the surface quality of a thin film of a semiconductor device according to an embodiment of the present invention;

[0022] Figure 2 Schematic diagram of the structure of a thin film surface quality detection system for semiconductor devices in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The embodiment of the present invention provides a method and system for detecting the surface quality of a thin film of a semiconductor device. 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.

[0024] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of a method for detecting the surface quality of a thin film of a semiconductor device in an embodiment of the present invention comprises:

[0025] Step S1, performing multimodal scanning on the thin film surface of the semiconductor device to obtain an original multimodal data set;

[0026] It is understandable that the execution subject of the present invention may be a semiconductor device thin film surface quality detection system, 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.

[0027] Specifically, nanoindentation scanning is performed on the film surface of the semiconductor device, and the force-displacement curve data is obtained by controlling the load and displacement applied by the indenter on the film surface. According to the force-displacement curve data, the Oliver-Pharr model is applied for analysis to obtain the elastic modulus and hardness distribution data of the film surface. The Oliver-Pharr model calculates the elastic modulus and hardness of the film material through a specific algorithm based on the indentation load and displacement curve. These data can reflect the distribution of the mechanical properties of the film. Micro-area Raman spectroscopy is performed on the film surface. Raman spectroscopy technology is used to irradiate the film surface with a laser and collect scattered spectral data to obtain Raman spectroscopy data. Raman spectroscopy data can reflect the chemical composition and molecular structure information of the film surface. By analyzing the Raman spectroscopy data, a chemical component distribution map of the film surface is obtained, showing the chemical composition and its changes in different regions. The film surface is scanned by a high-resolution electrochemical scanning probe, and a tiny probe is used to scan the film surface to obtain local electrochemical impedance spectroscopy data. According to the electrochemical impedance spectroscopy data, an equivalent circuit model is used for fitting analysis to obtain the charge transfer resistance and double-layer capacitance distribution data on the film surface. Electrochemical impedance spectroscopy data and its fitting results can reveal the electrochemical properties and local electrochemical reaction behaviors of the film. The elastic modulus and hardness distribution data, chemical component distribution map, charge transfer resistance and double layer capacitance distribution data are processed in a unified spatial coordinate system to form three-dimensional matrix format data. The data of different modes are aligned in the same coordinate system to have consistent spatial distribution and resolution. Baseline correction and noise suppression are performed on the three-dimensional matrix format data to improve the accuracy and reliability of the data. Baseline correction is to eliminate systematic errors in the data caused by instrument drift or other reasons, while noise suppression is to remove random noise in the data, enhance the significance of useful signals, and finally obtain the original multimodal data set.

[0028] Step S2, performing feature extraction and feature dimensionality reduction on the original multimodal data set to obtain a standard multidimensional feature matrix;

[0029] Specifically, the force-displacement curve data in the original multimodal data set is processed by Fourier transform, and the force-displacement signal in the time domain is converted to the frequency domain to obtain the frequency domain feature vector, which reflects the response characteristics of the film surface in different frequency ranges. Based on the frequency domain feature vector, the force-displacement curve data is decomposed by wavelet packet to analyze the multi-scale energy distribution characteristics of the signal. Wavelet packet decomposition can finely divide the energy distribution of the signal at different scales and capture the changes in the mechanical properties of the film material at different scales. The Raman spectral data in the original multimodal data set is subjected to discrete cosine transform to obtain the spectral characteristic coefficients. The spectral characteristic coefficients represent the main components of the Raman spectrum and can reveal the chemical composition characteristics of the film material. Based on the spectral characteristic coefficients, the Raman spectral data is subjected to non-negative matrix decomposition, and the spectral data is decomposed into several non-negative components to obtain the chemical component characteristic matrix, thereby realizing the fine characterization of the chemical composition of the film. For the electrochemical impedance spectroscopy data in the original multimodal data set, the Hilbert-Huang transform is used to analyze the time-frequency characteristics of the signal, and the non-stationary signal is decomposed into a series of intrinsic mode functions, and then the time-frequency characteristic diagram is constructed. The time-frequency characteristic diagram shows the changes of the signal in different time and frequency ranges. Based on the time-frequency characteristic diagram, the electrochemical impedance spectroscopy data is subjected to empirical mode decomposition to obtain the intrinsic mode function set, which can reveal the dynamic behavior of the film in the electrochemical environment. The multi-scale energy distribution characteristics, chemical component feature matrix and intrinsic mode function set are subjected to feature dimensionality reduction to reduce the data dimension and retain important information. Based on the feature set after dimensionality reduction, the kernel principal component analysis is used to perform nonlinear feature mapping of the feature space, and the complex relationship in the original feature space is mapped to a simpler nonlinear space. The feature space tensor decomposition is performed on the nonlinear feature mapping to simplify the data structure and finally obtain the standard multidimensional feature matrix.

[0030] Step S3, inputting the standard multi-dimensional feature matrix into a preset multi-branch convolutional neural network and graph neural network fusion algorithm to perform film quality assessment training to obtain a film quality assessment model;

[0031] Specifically, the standard multidimensional feature matrix is ​​feature decomposed to be divided into a first mechanical feature submatrix, a first chemical feature submatrix and a first electrical feature submatrix. Feature data of different physical properties are distinguished to facilitate subsequent independent processing and analysis. The first mechanical feature submatrix, the first chemical feature submatrix and the first electrical feature submatrix are respectively input into a preset multi-branch convolutional neural network. The multi-branch convolutional neural network consists of multiple branches, each of which focuses on processing different types of feature data. The mechanical feature branch adopts a 3D convolutional neural network (3D-CNN) structure to process mechanical feature data; the chemical feature branch adopts a 2D convolutional neural network (2D-CNN) structure to focus on chemical composition analysis; and the electrical feature branch adopts a 1D convolutional neural network (1D-CNN) structure to process electrochemical property data. In the mechanical feature branch, four 3D convolutional layers are used to extract features from the first mechanical feature submatrix. Each 3D convolutional layer is followed by a BatchNormalization and ReLU activation function, which can effectively normalize feature data and introduce nonlinear features. The last 3D convolution layer uses global average pooling to reduce the high-dimensional feature data to a fixed-length feature vector to obtain the first mechanical feature vector. At the same time, in the chemical feature branch, 5 2D convolution layers are applied to process the first chemical feature submatrix. Each 2D convolution layer is followed by BatchNormalization and LeakyReLU activation functions. LeakyReLU helps solve the "neuron death" problem that may occur in the ReLU function. The last 2D convolution layer uses global maximum pooling to ensure that the most significant feature information is extracted to obtain the first chemical feature vector. For the electrical feature branch, 6 1D convolution layers are used to process the first electrical feature submatrix. Each 1D convolution layer is followed by LayerNormalization and GELU activation functions. This setting can effectively improve the representation ability of the features. The last 1D convolution layer uses adaptive average pooling to obtain the first electrical feature vector. The first mechanical feature vector, the first chemical feature vector, and the first electrical feature vector are concatenated to form a unified feature vector to obtain the first multi-branch CNN feature vector. The feature vector is mapped through the fully connected layer to obtain the first node feature matrix. Based on the first node feature matrix, the K nearest neighbor algorithm is used to calculate the connection relationship between the nodes to obtain the first adjacency matrix. This process constructs a graph structure in which the nodes represent feature data and the edges represent the similarity or correlation between the nodes. In the graph neural network part, the first node feature matrix and the first adjacency matrix are input into the preset graph neural network for graph convolution operation. The graph neural network contains 3 layers of GraphConv layers, each of which is followed by an ELU activation function and graph batch normalization.This setting can propagate and aggregate information in the graph structure to obtain the first updated node representation. The first updated node representation is subjected to graph pooling to obtain the first graph-level representation. This graph-level representation is fused with the first multi-branch CNN feature vector, and the weights are calculated through the multi-head attention mechanism to obtain the first film quality assessment score. The attention mechanism helps to emphasize the key parts of the features and comprehensively consider the contributions of multiple feature sources. The entire model is optimized through the back-propagation algorithm to adjust the parameters of the multi-branch convolutional neural network and the graph neural network, and finally an accurate film quality assessment model is obtained.

[0032] The first mechanical feature submatrix is ​​padded with three-dimensional zeros by the mechanical feature branch to obtain a padded feature tensor. The size of the feature tensor is increased by adding zero-value elements in the boundary area to ensure that the size of the feature map remains consistent during subsequent convolution operations, thereby avoiding information loss. The padded feature tensor is subjected to a first 3D convolution operation by the first 3D convolution layer in the mechanical feature branch to obtain a first convolution feature map. This convolution operation uses 32 3x3x3 convolution kernels with a step size of 1, which means that the convolution kernel moves on the feature tensor with a step size of one unit. After the convolution operation is completed, the first convolution feature map is batch normalized to standardize the data distribution in the feature map and reduce the gradient vanishing or exploding problem in model training to obtain a first normalized feature map. The normalized feature map is subjected to a ReLU activation function operation to obtain a first activated feature map. The ReLU activation function introduces nonlinear factors, which enables the model to better learn complex feature representations. The first activated feature map is subjected to a second 3D convolution operation by the second 3D convolution layer in the mechanical feature branch to obtain a second convolution feature map. This convolution operation uses 64 3x3x3 convolution kernels with a step size of 2. Compared with the first convolution layer, the number of convolution kernels in this layer has doubled, and the step size has also doubled. This design helps to reduce the spatial size of the feature map while retaining important features, thereby reducing computational complexity. After obtaining the second convolution feature map, batch normalization is also performed to obtain the second normalized feature map, and the ReLU activation function is continued to be performed to obtain the second activated feature map. The activated feature map can more effectively capture the high-order features of the input feature tensor. The third 3D convolution operation is performed on the second activated feature map through the third 3D convolution layer to obtain the third convolution feature map. The third convolution layer uses 128 3x3x3 convolution kernels with a step size of 1, so that the network can learn more complex features. After the third convolution feature map is batch normalized, the third normalized feature map is obtained, and then the ReLU activation function is performed to obtain the third activated feature map. The fourth 3D convolution operation is performed on the third activated feature map through the fourth 3D convolution layer to obtain the fourth convolution feature map. This convolution layer uses 256 3x3x3 convolution kernels with a step size of 2. This design increases the feature extraction capability and reduces the size of the feature map through a larger step size. The fourth convolution feature map is batch normalized to obtain the fourth normalized feature map, and the ReLU activation function is performed to obtain the fourth activated feature map. The fourth activated feature map is subjected to global average pooling to obtain the first mechanical feature vector. Global average pooling generates a fixed-length vector by averaging the entire feature map. The vector integrates the information of all feature maps and represents the global features of the input data.

[0033] The first node feature matrix is ​​L2 normalized to obtain a normalized node feature matrix. The length of the node feature vector is normalized to unit length to eliminate the difference in feature magnitude. According to the normalized node feature matrix, the K nearest neighbor algorithm is used to calculate the Euclidean distance between nodes to determine the K nearest neighbor nodes of each node to obtain a K nearest neighbor relationship matrix. The application of the K nearest neighbor algorithm can effectively identify the closest node to each node in the feature space and construct a local connection relationship. According to the K nearest neighbor relationship matrix, a first adjacency matrix is ​​constructed, which represents the connection relationship between nodes. The first adjacency matrix is ​​symmetric to obtain a symmetric adjacency matrix to ensure that the edges in the graph structure are symmetric, that is, if node i is connected to node j, then node j should also be connected to node i, thereby maintaining the undirected characteristics of the graph. The symmetric adjacency matrix is ​​subjected to degree matrix normalization to obtain a normalized adjacency matrix, so that the features of each node can be equally considered in the convolution operation, avoiding feature amplification or reduction due to degree differences. The normalized node feature matrix and the normalized adjacency matrix are input into the first GraphConv layer in the graph neural network, and graph convolution operation is performed to obtain the first hidden layer features. The GraphConv layer updates the representation of each node by aggregating the feature information of neighboring nodes based on the combination of adjacency relationship and node features. The ELU activation function is performed on the first hidden layer features to obtain the first activation feature, and the first activation feature is subjected to graph batch normalization to obtain the first batch of normalized features. The application of the ELU activation function introduces nonlinear factors, which enables the model to better learn complex feature relationships, while graph batch normalization can accelerate convergence and stabilize the training process. The first batch of normalized features is input into the second GraphConv layer in the graph neural network, and graph convolution operation is performed again to obtain the second hidden layer features. This layer continues to refine the node features and updates the node representation by aggregating more neighbor information. The ELU activation function is performed on the second hidden layer features to obtain the second activation feature, and graph batch normalization is performed to obtain the second batch of normalized features. The second batch of normalized features is input into the third GraphConv layer in the graph neural network, and graph convolution is performed to obtain the third hidden layer features. After multiple layers of graph convolution operations, the model can capture deeper feature relationships and graph structure information. The ELU activation function is performed on the third hidden layer features to obtain the third activated features, which are then subjected to graph batch normalization to obtain the first updated node representation.

[0034] Step S4, using a thin film quality assessment model to perform edge computing real-time monitoring on the thin film growth process of the semiconductor device to obtain a dynamic quality assessment result and a defect prediction diagram;

[0035] Specifically, real-time multimodal data of semiconductor devices during film growth are collected, including mechanical, chemical and electrical characteristics. The edge device preprocesses the real-time multimodal data to extract useful information and generate a real-time feature matrix. The preprocessing steps may include filtering, denoising, normalization and other operations to ensure the quality and consistency of the data. The real-time feature matrix is ​​decomposed into a second mechanical feature submatrix, a second chemical feature submatrix and a second electrical feature submatrix. The three submatrices are respectively input into the multi-branch convolutional neural network of the film quality assessment model for convolution operations to obtain a second mechanical feature vector, a second chemical feature vector and a second electrical feature vector. Each feature vector represents a high-level representation of the corresponding feature data and contains rich information extracted by the convolutional network. The second mechanical feature vector, the second chemical feature vector and the second electrical feature vector are concatenated and mapped through a fully connected layer to obtain a second node feature matrix. Different types of feature data are integrated into a unified feature space for subsequent graph convolution operations. According to the second node feature matrix, a dynamic graph structure is constructed. The time-varying connection relationship between nodes is calculated using an adaptive K nearest neighbor algorithm to obtain a second adjacency matrix. The adaptive K-nearest neighbor algorithm can dynamically adjust the connection relationship according to the changes in node features and adapt to the changes in real-time data. The second node feature matrix and the second adjacency matrix are input into the graph convolutional network for message passing and aggregation operations to obtain the second updated node representation. The graph convolutional network aggregates the information of neighboring nodes so that the node representation contains not only its own information but also the contextual information of its neighbors. The second updated node representation is pooled in the time series graph, and the DiffPool method of the attention mechanism is used to calculate the time series clustering matrix to obtain the second graph-level representation. The DiffPool method selectively aggregates nodes through the attention mechanism to generate a higher-level graph representation. The second graph-level representation is multi-headedly fused with the second multi-branch CNN feature vector and feature weights are calculated to obtain a comprehensive feature representation. The comprehensive feature representation is mapped to obtain the second film quality assessment score and abnormal probability distribution. By analyzing the abnormal probability distribution, the defect area on the film surface is located using threshold segmentation and morphological operations to generate a defect prediction map. Morphological operations include corrosion and expansion, which can further remove noise and highlight defect areas. The second film quality assessment score and defect prediction graph are analyzed in time series to track the quality changes during the film growth process and obtain dynamic quality assessment results. Through time series analysis, potential problems in film growth can be identified.

[0036] Step S5, performing stress modulation Raman spectroscopy and intelligent etching analysis on the dynamic quality assessment results and defect prediction diagram to obtain thin film three-dimensional defect distribution verification data;

[0037] Specifically, the defect prediction map is segmented to extract obvious defect areas, generate defect area masks, and locate potential defect positions on the film surface. Based on the defect area mask, the film surface is periodically stressed to introduce stress changes in the film sample to form a stress-modulated film sample. Stress loading is usually achieved by mechanical methods, such as using a piezoelectric device or a mechanical loading system to generate a periodic stress field on the sample surface. The stress-modulated film sample is scanned by Raman spectroscopy to obtain stress-modulated Raman spectroscopy data. Raman spectroscopy data can reveal the stress state and chemical composition changes inside the material. The Raman spectroscopy data is Fourier transformed to extract the stress modulation frequency component and obtain a stress-sensitive Raman peak displacement map. The information in the Raman peak displacement map reflects the lattice distortion of the film material under stress. According to the Raman peak displacement map, the biaxial stress model is used to perform stress field inversion calculation to construct a stress distribution map on the film surface. The stress distribution map shows the difference in stress state in different regions of the film surface, so that potential stress concentration regions can be identified, which may be potential defect sites. According to the stress distribution map, threshold segmentation is performed to obtain the potential defect depth distribution map, indicating the possible defect depth range inside the film. An adaptive etching scheme is generated according to the potential defect depth distribution. The adaptive etching scheme takes into account the depth and distribution characteristics of different defects and designs a layer-by-layer peeling etching step. The film sample is subjected to intelligent etching treatment, and the film material is removed layer by layer according to the adaptive etching scheme to expose the internal defect structure. The intelligent etching treatment can accurately control the etching depth and speed to avoid damage to the sample and ensure the complete disclosure of defects. After completing multi-layer etching, the etched surface is imaged by high-resolution scanning electron microscopy to obtain the morphology and composition data of the defects. High-resolution scanning electron microscopy can provide detailed morphological characteristics of defects, including the size, shape and distribution of defects, and at the same time reveal the material composition of the defect area through composition analysis. Through the three-dimensional reconstruction algorithm, the three-dimensional defect distribution map of the film is reconstructed according to the defect morphology and composition data, stress distribution map and etching depth information.

[0038] Step S6: Based on the three-dimensional defect distribution verification data, thin film growth parameters of the semiconductor device are optimized to generate a target thin film growth control parameter set.

[0039] Specifically, a spatial statistical analysis is performed on the three-dimensional defect distribution verification data to obtain the defect distribution feature vector, which contains information such as the position, size, shape and density of the defect in space, reflecting the defect characteristics inside the film. A film growth simulator is constructed based on the defect distribution feature vector. The simulator is used to simulate the film growth process under different growth conditions to help understand how to adjust the growth parameters to reduce the formation of defects. The film growth simulator is simulated and analyzed to generate a parameter-defect mapping database. The database records the defect distribution of the film under different growth parameter combinations. Based on this database, a Gaussian process regression model is trained to construct a probabilistic relationship between growth parameters and defect distribution to form a film quality prediction model. The Gaussian process regression model can handle nonlinear relationships and provide uncertainty estimates of the impact of growth parameter changes on defect distribution. The film quality prediction model is Bayesian optimized to explore the growth parameter space and find parameter combinations that can minimize defects. Bayesian optimization combines model prediction and uncertainty estimation to efficiently search the optimal parameter space. After obtaining the candidate growth parameter set, a multi-objective optimization solution is performed based on the parameter set to determine the balance of multiple optimization objectives and obtain a Pareto optimal solution set. The Pareto optimal solution set represents the optimal parameter combination that cannot be improved between different optimization objectives. Cluster analysis is performed on the Pareto optimal solution set to identify parameter combinations with similar performance. Through cluster analysis, representative growth parameter combinations are screened out, which can cover different optimization objectives. According to the representative growth parameter combinations, an optimization experiment matrix is ​​generated for experimental verification and optimization. Response surface analysis is performed on the optimization experiment matrix, and a second-order polynomial regression model is established. The second-order polynomial regression model can capture the nonlinear relationship between growth parameters and film quality indicators and generate a parameter-quality response surface. The response surface shows the changing trend of film quality under different growth parameter combinations, which helps to understand the impact of parameters on film quality. Based on the parameter-quality response surface, the gradient descent method is used to find the global optimal point. The gradient descent method calculates the gradient of the response surface and continuously adjusts the growth parameters to find the optimal quality film growth conditions, and obtains the target film growth control parameter set, which can maximize the film quality and minimize the defect density and other unfavorable factors.

[0040] In the embodiment of the present invention, the method performs multi-modal scanning on the surface of the semiconductor device film to obtain comprehensive data including mechanical, chemical and electrical properties, thereby realizing a multi-dimensional evaluation of the film quality and improving the comprehensiveness and accuracy of the detection. Advanced signal processing techniques, such as Fourier transform, wavelet packet decomposition and non-negative matrix decomposition, are used to extract features and reduce the dimension of the original multi-modal data, effectively extracting the key feature information of the film surface. A film quality evaluation model is constructed using a multi-branch convolutional neural network and a graph neural network fusion algorithm, realizing intelligent analysis and evaluation of complex film surface features. Real-time monitoring of the film growth process is achieved through edge computing technology, potential quality problems are discovered and predicted in a timely manner, and the control accuracy of the production process is improved. Combined with stress modulation Raman spectroscopy and intelligent etching analysis, accurate verification of the three-dimensional defect distribution of the film is achieved. Based on the three-dimensional defect distribution verification data, an advanced optimization algorithm is used to optimize the film growth parameters, generate a target film growth control parameter set, and effectively improve the quality and efficiency of the film growth. From nano-scale surface scanning to macro-film growth parameter optimization, the method realizes cross-scale full-process analysis and control of film quality, thereby improving the accuracy of film surface quality detection of semiconductor devices.

[0041] In a specific embodiment, the process of executing step S1 may specifically include the following steps:

[0042] (1) Perform nanoindentation scanning on the thin film surface of the semiconductor device to obtain force-displacement curve data, and perform Oliver-Pharr model analysis on the thin film surface based on the force-displacement curve data to obtain elastic modulus and hardness distribution data;

[0043] (2) Scanning the film surface by micro-Raman spectroscopy to obtain Raman spectral data, and analyzing the film surface based on the Raman spectral data to obtain a chemical component distribution map;

[0044] (3) Scan the film surface with a high-resolution electrochemical scanning probe to obtain local electrochemical impedance spectroscopy data, and perform equivalent circuit fitting on the film surface based on the local electrochemical impedance spectroscopy data to obtain charge transfer resistance and double layer capacitance distribution data;

[0045] (4) Perform spatial coordinate unification processing on the elastic modulus and hardness distribution data, chemical composition distribution map, charge transfer resistance and double layer capacitance distribution data to obtain three-dimensional matrix format data;

[0046] (5) Perform baseline correction and noise suppression on the three-dimensional matrix format data to obtain the original multimodal dataset.

[0047] Specifically, a nanoindentation scan is performed on the surface of the film, a specific load is applied to the surface of the film using a nanoscale indenter, and the displacement of the indenter is recorded to obtain force-displacement curve data. The force-displacement curve data provides the response characteristics of the material under different loads. Using these data, the elastic modulus and hardness distribution of the film are calculated using the Oliver-Pharr model. The Oliver-Pharr model can effectively determine the mechanical properties of the material by analyzing the relationship between load and displacement, especially during the loading and unloading process of the indentation. Elastic modulus E and hardness H are key mechanical performance parameters that reflect the rigidity and surface hardness of the material, respectively, and can be calculated using the following formulas:

[0048] ;

[0049] in, is the Poisson's ratio of the material, is the stiffness of the indentation, is the maximum load, is the contact depth, and is the contact area. Perform a micro-area Raman spectrum scan on the surface of the film. Raman spectroscopy technology can provide information about the chemical composition of the material by exciting the sample with a laser and measuring the changes in the scattering spectrum. The acquired Raman spectral data can reflect the chemical composition of each region in the film. After analysis of these data, a chemical component distribution map is generated to reveal the chemical uniformity and composition changes on the surface and inside of the film. For example, different peak positions and peak intensities can indicate the presence of different chemical bonds or defect types in the film, which helps to understand the purity and chemical stability of the film. At the same time, a high-resolution electrochemical scanning probe is scanned on the surface of the film to obtain local electrochemical impedance spectroscopy data. These data help to understand the electrochemical properties of the film, such as charge transfer resistance and double layer capacitance. Through equivalent circuit fitting, the complex electrochemical behavior is simplified into several basic circuit elements to obtain the charge transfer resistance. and double layer capacitance The distribution of charge transfer resistance It reflects the ease of charge transfer from electrode to electrolyte, while the double layer capacitance It is related to the charge storage capacity at the interface. The spatial coordinates of the elastic modulus and hardness distribution data, chemical composition distribution map, charge transfer resistance and double layer capacitance distribution data are unified, and the spatial information from different data sources is unified into a common coordinate system, which is helpful for comprehensive analysis of the relationship between different physical quantities. By aligning the spatial resolution and sampling coordinates of each data set, a data structure in a three-dimensional matrix format is generated. Baseline correction and noise suppression are performed on the three-dimensional matrix format data. The purpose of baseline correction is to eliminate systematic errors caused by equipment drift or environmental factors to ensure the accuracy of the data. Noise suppression is to reduce random noise in the data and enhance the significance of useful signals through filtering or other signal processing techniques.

[0050] In a specific embodiment, the process of executing step S2 may specifically include the following steps:

[0051] (1) Perform Fourier transform processing on the force-displacement curve data in the original multimodal data set to obtain the frequency domain feature vector, and then perform wavelet packet decomposition on the force-displacement curve data based on the frequency domain feature vector to obtain the multi-scale energy distribution characteristics;

[0052] (2) Performing discrete cosine transform on the Raman spectral data in the original multimodal data set to obtain spectral characteristic coefficients, and performing non-negative matrix decomposition on the Raman spectral data based on the spectral characteristic coefficients to obtain the chemical component characteristic matrix;

[0053] (3) Performing Hilbert-Huang transform on the electrochemical impedance spectroscopy data in the original multimodal data set to obtain a time-frequency characteristic diagram, and performing empirical mode decomposition on the electrochemical impedance spectroscopy data based on the time-frequency characteristic diagram to obtain an intrinsic mode function set;

[0054] (4) Performing feature dimensionality reduction on the multi-scale energy distribution features, chemical component feature matrix, and intrinsic mode function set to obtain a feature set after dimensionality reduction;

[0055] (5) Based on the feature set after dimensionality reduction, kernel principal component analysis is performed on the feature space to obtain a nonlinear feature map, and the feature space tensor decomposition is performed on the nonlinear feature map to obtain a standard multidimensional feature matrix.

[0056] Specifically, the force-displacement curve data in the original multimodal data set is processed by Fourier transform to convert the time domain data into frequency domain feature vectors. Fourier transform can effectively reveal the periodic components in the signal, especially the high-frequency and low-frequency components, which can provide detailed information about the mechanical response of the material. , whose Fourier transform It is expressed as:

[0057] ;

[0058] in, represents the amplitude of the frequency domain signal, is the frequency, is a time domain signal, and i is an imaginary unit. The frequency domain eigenvector is obtained by Fourier transform, which reflects the intensity of different frequency components. According to the frequency domain eigenvector, the force-displacement curve data is decomposed by wavelet packet. Wavelet packet decomposition is a multi-scale analysis method that can capture the local characteristics of the signal at different scales. For a signal , its energy is Level and The distribution of frequency bands is expressed as:

[0059] ;

[0060] in, are the wavelet packet coefficients, is the scale of decomposition, is the index of the frequency band, is the signal length. The multi-scale energy distribution feature provides detailed energy distribution information of the signal in both time and frequency dimensions, which helps to identify the behavior of materials under different stress states. For Raman spectral data, discrete cosine transform (DCT) is performed to convert the spectral data into frequency domain characteristic coefficients. Discrete cosine transform is a transformation method that is suitable for processing signals with boundary effects and can effectively compress data and extract features. The formula for discrete cosine transform is as follows:

[0061] ;

[0062] in, It is DCT coefficients, is the input signal, is the signal length. Through DCT, the spectral characteristic coefficients are obtained, which reflect the intensity and position of different frequency components in the spectral data. According to the spectral characteristic coefficients, the Raman spectral data is subjected to non-negative matrix decomposition to obtain the chemical component characteristic matrix. Non-negative matrix decomposition is a matrix decomposition technique that can decompose the original data matrix into the product of two non-negative matrices to reveal the potential structure in the data. For Raman spectral data, non-negative matrix decomposition can separate the contributions of different chemical components and generate a chemical component characteristic matrix. The matrix represents the concentration distribution of various chemical components in different regions. The goal of non-negative matrix decomposition is to minimize the following objective function:

[0063] ;

[0064] in, is the input data matrix, is the basic component matrix, is the weight matrix, Represents the Frobenius norm. At the same time, the electrochemical impedance spectroscopy data is subjected to Hilbert-Huang transform to obtain the time-frequency characteristic diagram. Hilbert-Huang transform is a time-frequency analysis method that can handle nonlinear and non-stationary signals. Through Hilbert-Huang transform, the electrochemical impedance spectroscopy data is decomposed into several intrinsic mode functions (IMFs), and each IMF is subjected to Hilbert transform to obtain the time-frequency characteristic diagram, revealing the pattern of how different frequency components in the electrochemical signal change over time. For a signal , the Hilbert-Huang transform obtains the IMF through empirical mode decomposition (EMD):

[0065] ;

[0066] in, It is IMF, is the residual part. According to the time-frequency characteristic diagram, EMD is performed on the electrochemical impedance spectroscopy data to obtain the intrinsic mode function set. The acquisition of the intrinsic mode function set helps to understand the different physical mechanisms in the electrochemical process. Feature dimensionality reduction is performed on the multi-scale energy distribution characteristics, chemical component feature matrix and intrinsic mode function set to reduce the dimension of the features while retaining the key information of the data as much as possible. Through dimensionality reduction, a more concise feature set is obtained. Based on the reduced feature set, kernel principal component analysis (KPCA) is performed. KPCA is a nonlinear dimensionality reduction method that maps the original data to a high-dimensional feature space through kernel techniques, and then performs principal component analysis in the high-dimensional space. The central formula of KPCA is:

[0067] ;

[0068] in, It is a dataset Mapping in high-dimensional space, is the kernel matrix. Kernel function It is used to calculate the inner product of input data in high-dimensional space. It performs feature space tensor decomposition on nonlinear feature mapping to obtain a standard multi-dimensional feature matrix.

[0069] In a specific embodiment, the process of executing step S3 may specifically include the following steps:

[0070] (1) performing feature decomposition on the standard multidimensional feature matrix, dividing the standard multidimensional feature matrix into a first mechanical feature sub-matrix, a first chemical feature sub-matrix, and a first electrical feature sub-matrix;

[0071] (2) inputting the first mechanical feature sub-matrix, the first chemical feature sub-matrix and the first electrical feature sub-matrix into a preset multi-branch convolutional neural network, wherein the multi-branch convolutional neural network includes a mechanical feature branch, a chemical feature branch and an electrical feature branch, wherein the mechanical feature branch adopts a 3D-CNN structure, the chemical feature branch adopts a 2D-CNN structure, and the electrical feature branch adopts a 1D-CNN structure;

[0072] (3) Extracting features from the first mechanical feature submatrix through the mechanical feature branch to obtain the first mechanical feature vector. The mechanical feature branch includes four 3D convolutional layers. Each 3D convolutional layer is followed by BatchNormalization and ReLU activation functions. The last 3D convolutional layer uses global average pooling.

[0073] (4) Extracting features from the first chemical feature submatrix through the chemical feature branch to obtain the first chemical feature vector. The chemical feature branch includes five 2D convolutional layers. Each 2D convolutional layer is followed by a BatchNormalization and a LeakyReLU activation function. The last 2D convolutional layer uses global maximum pooling.

[0074] (5) The first electrical feature submatrix is ​​feature extracted through the electrical feature branch to obtain the first electrical feature vector. The electrical feature branch contains 6 1D convolutional layers. Each 1D convolutional layer is followed by LayerNormalization and GELU activation function. The last 1D convolutional layer uses adaptive average pooling.

[0075] (6) concatenating the first mechanical feature vector, the first chemical feature vector, and the first electrical feature vector to obtain a first multi-branch CNN feature vector, and mapping the first multi-branch CNN feature vector through a fully connected layer to obtain a first node feature matrix;

[0076] (7) Construct a graph structure based on the first node feature matrix, use the K nearest neighbor algorithm to calculate the connection relationship between nodes, obtain the first adjacency matrix, and input the first node feature matrix and the first adjacency matrix into the preset graph neural network for graph convolution operation. The graph neural network includes 3 layers of GraphConv layers, and each GraphConv layer is followed by an ELU activation function and graph batch normalization to obtain the first updated node representation;

[0077] (8) performing a graph pooling operation on the first updated node representation to obtain a first graph-level representation;

[0078] (9) The first graph-level representation is fused with the first multi-branch CNN feature vector, and the weights are calculated through a multi-head attention mechanism to obtain a first film quality assessment score. The multi-branch convolutional neural network and graph neural network are optimized through back propagation to obtain a film quality assessment model.

[0079] Specifically, the standard multi-dimensional feature matrix is ​​feature decomposed and divided into the first mechanical feature sub-matrix, the first chemical feature sub-matrix and the first electrical feature sub-matrix. Each feature sub-matrix represents detailed data on different physical properties of the film, such as mechanical properties, chemical composition and electrical properties. These sub-matrices are respectively input into a preset multi-branch convolutional neural network, which contains specialized branches for different feature types: mechanical feature branch, chemical feature branch and electrical feature branch. The mechanical feature branch adopts a 3D convolutional neural network (3D-CNN) structure to capture the three-dimensional spatial features of the mechanical property data. 3D-CNN contains 4 3D convolutional layers, each of which is followed by BatchNormalization and ReLU activation function. The role of batch normalization is to standardize each small batch of data, accelerate training convergence and stabilize the model. The ReLU activation function enhances the expressive power of the model by introducing nonlinearity. The output of each 3D convolutional layer can be expressed as:

[0080] ;

[0081] in, and Respectively The convolution kernel and bias of the layer, * represents the convolution operation, ReLU is the activation function, and BatchNorm represents the batch normalization operation. The last 3D convolution layer uses global average pooling to convert the three-dimensional feature map into a fixed-size feature vector, namely the first mechanical feature vector. The chemical feature branch adopts the 2D convolutional neural network (2D-CNN) structure, which is suitable for processing two-dimensional image data, such as the distribution map of chemical components. 2D-CNN contains 5 2D convolutional layers, each of which is followed by BatchNormalization and LeakyReLU activation function. The LeakyReLU activation function is an improved ReLU that allows negative inputs to pass through to prevent the occurrence of the "neuron death" problem. The output of each convolutional layer of the chemical feature branch can be expressed as:

[0082] ;

[0083] The last 2D convolutional layer uses global maximum pooling to extract the most significant part of the chemical features and obtain the first chemical feature vector. The electrical feature branch uses a 1D convolutional neural network (1D-CNN) structure to process one-dimensional sequence data, such as electrochemical impedance spectroscopy. This branch contains 6 1D convolutional layers, each followed by LayerNormalization and GELU activation function. LayerNormalization can stabilize the training process, while the GELU activation function is a smooth nonlinear activation function that helps retain more information. The calculation formula of GELU is:

[0084] ;

[0085] The last 1D convolutional layer uses adaptive average pooling to map the features to a fixed-size vector, namely the first electrical feature vector. The first mechanical feature vector, the first chemical feature vector, and the first electrical feature vector are concatenated to form a comprehensive first multi-branch CNN feature vector. The feature vector is mapped through the fully connected layer to obtain the first node feature matrix. The output of the fully connected layer can be expressed as:

[0086] ;

[0087] in, and are the weights and bias parameters of the fully connected layer. Based on the first node feature matrix, a graph structure is constructed. The K nearest neighbor algorithm is used to calculate the connection relationship between nodes to obtain the first adjacency matrix. The K nearest neighbor algorithm selects the nearest K neighbors by calculating the Euclidean distance between each node to construct the connection edges of the graph. The first node feature matrix and the first adjacency matrix are input into the preset graph neural network for graph convolution operation. The graph neural network contains 3 layers of GraphConv layers, each of which is followed by an ELU activation function and a graph batch normalization. The ELU activation function is a smooth nonlinear activation function that helps capture complex graph structure features, while the graph batch normalization is used to stabilize the training process. Through the operation of these layers, the first updated node representation is obtained. After obtaining the updated node representation, a graph pooling operation is performed to obtain the first graph-level representation. Graph pooling obtains the global representation of the graph by aggregating the node representation, which helps capture the global features of the entire graph. The first graph-level representation is fused with the first multi-branch CNN feature vector for attention. The attention mechanism determines the contribution of different feature vectors in the final output by calculating their weights. This mechanism helps to highlight important features and suppress irrelevant information. The weights are calculated through the multi-head attention mechanism to obtain the first film quality assessment score. The multi-head attention mechanism is a technology that enhances the learning ability of the model by focusing on different feature parts in parallel to improve the expressiveness and robustness of the model. Through the back-propagation algorithm, the multi-branch convolutional neural network and graph neural network are optimized to finally obtain an accurate film quality assessment model.

[0088] In a specific embodiment, the process of performing feature extraction on the first mechanical feature submatrix through the mechanical feature branch to obtain the first mechanical feature vector may specifically include the following steps:

[0089] (1) Performing three-dimensional zero-filling processing on the first mechanical feature submatrix through the mechanical feature branch to obtain a filled feature tensor;

[0090] (2) Perform the first 3D convolution operation on the padded feature tensor through the first 3D convolution layer in the mechanical feature branch to obtain the first convolution feature map. The first 3D convolution operation uses 32 3x3x3 convolution kernels with a step size of 1;

[0091] (3) performing batch normalization processing on the first convolution feature map to obtain a first normalized feature map, and performing a ReLU activation function operation on the first normalized feature map to obtain a first activated feature map;

[0092] (4) performing a second 3D convolution operation on the first activation feature map through the second 3D convolution layer in the mechanical feature branch to obtain a second convolution feature map. The second 3D convolution operation uses 64 3x3x3 convolution kernels with a step size of 2.

[0093] (5) performing batch normalization processing on the second convolutional feature map to obtain a second normalized feature map, and performing a ReLU activation function operation on the second normalized feature map to obtain a second activated feature map;

[0094] (6) performing a third 3D convolution operation on the second activation feature map through the third 3D convolution layer in the mechanical feature branch to obtain a third convolution feature map. The third 3D convolution operation uses 128 3x3x3 convolution kernels with a step size of 1.

[0095] (7) performing batch normalization processing on the third convolution feature map to obtain a third normalized feature map, and performing a ReLU activation function operation on the third normalized feature map to obtain a third activated feature map;

[0096] (8) performing a fourth 3D convolution operation on the third activation feature map through the fourth 3D convolution layer in the mechanical feature branch to obtain a fourth convolution feature map. The fourth 3D convolution operation uses 256 3x3x3 convolution kernels with a step size of 2;

[0097] (9) performing batch normalization processing on the fourth convolutional feature map to obtain a fourth normalized feature map, and performing a ReLU activation function operation on the fourth normalized feature map to obtain a fourth activated feature map;

[0098] (10) Perform a global average pooling operation on the fourth activated feature map to obtain a first mechanical feature vector.

[0099] Specifically, the first mechanical feature submatrix is ​​subjected to three-dimensional zero padding to ensure that edge information is not lost during the convolution operation. Three-dimensional zero padding is to add zero elements around each dimension of the feature tensor, thereby expanding the size of the tensor. Assume that the size of the original feature tensor is , the size of the padded feature tensor becomes ,in is the padding size. The first 3D convolution operation is performed on the padded feature tensor through the first 3D convolution layer in the mechanical feature branch. The 3D convolution operation uses 32 The convolution kernel has a step size of 1. The output of the convolution operation, that is, the first convolution feature map, can be expressed by the following formula:

[0100] ;

[0101] in, It is The convolution kernel of the layer, * indicates the convolution operation, is the input feature map of the previous layer, is the bias. The first convolution feature map is batch normalized to obtain the first normalized feature map. Batch normalization helps to alleviate the problem of internal covariate shift and accelerate model training by standardizing small batches of data. The ReLU activation function is applied to the first normalized feature map to obtain the first activated feature map. The role of the ReLU activation function is to introduce nonlinearity so that the model can learn complex features. The second 3D convolution operation is performed on the first activated feature map through the second 3D convolution layer in the mechanical feature branch. The second 3D convolution layer uses 64 The convolution kernel has a stride of 2, which means that each convolution operation will skip an element and reduce the spatial size of the feature map. Similarly, the convolution output formula is:

[0102] ;

[0103] The second convolution feature map is batch normalized to obtain the second normalized feature map, and the ReLU activation function is applied again to obtain the second activated feature map. The ReLU activation function effectively retains the non-negative part of the input and truncates the negative values ​​to zero, increasing the sparsity and computational efficiency of the model. The third 3D convolution operation is performed on the second activated feature map through the third 3D convolution layer in the mechanical feature branch. The third 3D convolution layer uses 128 The convolution kernel is 1, and the step size is 1. The output of the convolution operation is batch normalized to obtain the third normalized feature map. The ReLU activation function is applied to obtain the third activated feature map. The formula of the ReLU function is;

[0104] ;

[0105] This function helps the network learn more complex feature representations through simple nonlinear transformations. The fourth 3D convolution operation is performed on the third activation feature map through the fourth 3D convolution layer in the mechanical feature branch. The fourth 3D convolution layer uses 256 The convolution kernel with a step size of 2 is used to reduce the spatial size of the feature map and increase the abstract level of the feature. The convolved feature map is batch normalized again to obtain the fourth normalized feature map, and finally the fourth activated feature map is obtained by the ReLU activation function. The fourth activated feature map is globally average pooled, and the average value of the entire feature map is calculated. The average value of all elements of each channel is used as the final output to obtain the first mechanical feature vector.

[0106] In a specific embodiment, the execution step of constructing a graph structure according to the first node feature matrix, using the K nearest neighbor algorithm to calculate the connection relationship between nodes, obtaining a first adjacency matrix, and inputting the first node feature matrix and the first adjacency matrix into a preset graph neural network for graph convolution operation may specifically include the following steps:

[0107] (1) Perform L2 normalization on the first node feature matrix to obtain a normalized node feature matrix. Based on the normalized node feature matrix, use the K nearest neighbor algorithm to calculate the Euclidean distance between nodes, select the K nearest neighbor nodes of each node, and obtain the K nearest neighbor relationship matrix.

[0108] (2) Constructing a first adjacency matrix based on the K nearest neighbor relationship matrix, and performing symmetry processing on the first adjacency matrix to obtain a symmetric adjacency matrix;

[0109] (3) The degree matrix of the symmetric adjacency matrix is ​​normalized to obtain a normalized adjacency matrix, and the normalized node feature matrix and the normalized adjacency matrix are input into the first GraphConv layer in the graph neural network to perform graph convolution operation to obtain the first hidden layer features;

[0110] (4) Performing an ELU activation function operation on the first hidden layer feature to obtain the first activated feature, and performing batch normalization on the first activated feature to obtain the first batch of normalized features;

[0111] (5) Input the first batch of normalized features into the second GraphConv layer in the graph neural network and perform graph convolution operation to obtain the second hidden layer features;

[0112] (6) Performing an ELU activation function operation on the second hidden layer feature to obtain a second activated feature, and performing a batch normalization process on the second activated feature to obtain a second batch of normalized features;

[0113] (7) Input the second batch of normalized features into the third GraphConv layer in the graph neural network, perform graph convolution operations, and obtain the third hidden layer features;

[0114] (8) Performing an ELU activation function operation on the third hidden layer feature to obtain a third activated feature, and performing graph batch normalization on the third activated feature to obtain a first updated node representation.

[0115] Specifically, the first node feature matrix is ​​L2 normalized, and the length of each node feature vector is normalized to 1 to ensure that the scales of different features are consistent. , the formula for L2 normalization is:

[0116] ;

[0117] in, Represents the feature vector The L2 norm of is the square root of the sum of the squares of the vector components. Through the normalization operation, the normalized node feature matrix is ​​obtained. According to the normalized node feature matrix, the K nearest neighbor algorithm is used to calculate the Euclidean distance between nodes. The calculation formula of the Euclidean distance is:

[0118] ;

[0119] in, and Respectively represent and The node in The value of the feature, is the dimension of the feature. According to the calculated Euclidean distance, select K nearest neighbor nodes for each node and construct a K nearest neighbor relationship matrix. The matrix records the connection relationship between each node and its K nearest neighbors and is used to construct the structure of the graph. According to the K nearest neighbor relationship matrix, construct the first adjacency matrix. Adjacency matrix is a binary matrix where Representation Node and nodes There is a connection between them, otherwise To ensure the symmetry of the adjacency matrix, the first adjacency matrix is ​​symmetricized to obtain a symmetric adjacency matrix The formula for symmetry processing is:

[0120] ;

[0121] in, is the adjacency matrix The transposed matrix of . The symmetric adjacency matrix ensures the non-directionality of the graph, that is, the connection relationship is mutual. The symmetric adjacency matrix is ​​normalized by degree matrix to obtain the normalized adjacency matrix. Degree matrix is a diagonal matrix whose elements Representation Node The degree of the node The total number of connected nodes. The calculation formula of the normalized adjacency matrix is:

[0122] ;

[0123] in, is the inverse square root of the degree matrix. Normalization can prevent the difference in the number of connections between nodes from affecting the convolution operation, making feature aggregation more balanced. The normalized node feature matrix and the normalized adjacency matrix are input into the first GraphConv layer in the graph neural network for graph convolution operation. The function of the graph convolution layer is to aggregate the feature information of each node and its neighboring nodes to obtain a richer representation. The formula for the graph convolution operation is:

[0124] ;

[0125] in, It is The node feature matrix of the layer, It is The weight matrix of the layer, is the activation function. Through this process, the first hidden layer feature is obtained. The ELU activation function is operated on the first hidden layer feature to obtain the first activation feature. The formula of the ELU activation function is:

[0126] ;

[0127] in, Is a positive number used to control the attenuation of negative values. The ELU activation function can reduce the gradient vanishing problem while providing negative values ​​to maintain the diversity of information flow. The first activation feature is subjected to graph batch normalization to obtain the first batch of normalized features. The first batch of normalized features are input into the second GraphConv layer in the graph neural network, and similar graph convolution operations are performed to obtain the second hidden layer features. The second hidden layer features are subjected to the ELU activation function operation again to obtain the second activation features, and graph batch normalization is performed to obtain the second batch of normalized features. The second batch of normalized features are input into the third GraphConv layer in the graph neural network, and the last graph convolution operation is performed to obtain the third hidden layer features. The third hidden layer features are subjected to the ELU activation function operation to obtain the third activation features, and graph batch normalization is performed on them to finally obtain the first updated node representation.

[0128] In a specific embodiment, the process of executing step S4 may specifically include the following steps:

[0129] (1) Collect real-time multimodal data of semiconductor devices during thin film growth and pre-process the real-time multimodal data through edge devices to obtain a real-time feature matrix;

[0130] (2) decomposing the real-time characteristic matrix into a second mechanical characteristic sub-matrix, a second chemical characteristic sub-matrix, and a second electrical characteristic sub-matrix;

[0131] (3) inputting the second mechanical characteristic submatrix, the second chemical characteristic submatrix, and the second electrical characteristic submatrix into a multi-branch convolutional neural network of a film quality assessment model for convolution operation to obtain a second mechanical characteristic vector, a second chemical characteristic vector, and a second electrical characteristic vector;

[0132] (4) concatenating the second mechanical eigenvector, the second chemical eigenvector, and the second electrical eigenvector, and mapping them through a fully connected layer to obtain a second node feature matrix;

[0133] (5) Construct a dynamic graph structure based on the second node feature matrix, use the adaptive K nearest neighbor algorithm to calculate the time-varying connection relationship between nodes, and obtain the second adjacency matrix;

[0134] (6) Inputting the second node feature matrix and the second adjacency matrix into the graph convolutional network for message passing and aggregation operations to obtain the second updated node representation;

[0135] (7) Perform temporal graph pooling on the second updated node representation, use the DiffPool method of the attention mechanism, calculate the temporal clustering matrix, and obtain the second graph-level representation;

[0136] (8) Perform multi-head attention fusion and feature weight calculation on the second graph-level representation and the second multi-branch CNN feature vector to obtain a comprehensive feature representation;

[0137] (9) Mapping the comprehensive feature representation to obtain the second film quality assessment score and abnormal probability distribution, and based on the abnormal probability distribution, using threshold segmentation and morphological operations to locate the defect area on the film surface to obtain a defect prediction map;

[0138] (10) Perform a time series analysis on the second film quality assessment score and defect prediction diagram to obtain a dynamic quality assessment result.

[0139] Specifically, real-time multimodal data, including mechanical, chemical and electrical properties, are collected, and the real-time multimodal data is preprocessed through edge devices to reduce noise and redundant information to obtain a high-quality real-time feature matrix. The preprocessing steps may include signal filtering, normalization, and feature extraction processes to ensure the quality and consistency of the data. The real-time feature matrix is ​​decomposed into a second mechanical feature sub-matrix, a second chemical feature sub-matrix, and a second electrical feature sub-matrix. Each sub-matrix contains characteristic information of the film under specific physical properties. For example, the mechanical feature sub-matrix may contain information on stress and elastic modulus, the chemical feature sub-matrix may contain the concentration and distribution of elements, and the electrical feature sub-matrix contains changes in resistance and capacitance. The sub-matrix is ​​input into a preset multi-branch convolutional neural network for convolution operation. In the multi-branch convolutional neural network structure, each branch processes a specific type of feature data. For example, the second mechanical feature submatrix is ​​processed by the 3D convolution layer of the mechanical feature branch to extract the second mechanical feature vector; the second chemical feature submatrix is ​​processed by the 2D convolution layer of the chemical feature branch to extract the second chemical feature vector; the second electrical feature submatrix is ​​processed by the 1D convolution layer of the electrical feature branch to obtain the second electrical feature vector. The purpose of the convolution operation is to extract local features in the data through the convolution kernel, and the formula is:

[0140] ;

[0141] in, is the activation function, It is The convolution kernel of the layer, is the bias. The second mechanical eigenvector, the second chemical eigenvector, and the second electrical eigenvector are concatenated and mapped through a fully connected layer to generate a second node feature matrix. Features of different types are integrated into a common feature space. According to the second node feature matrix, a dynamic graph structure is constructed. Due to the time-varying characteristics of real-time data, an adaptive K-nearest neighbor algorithm is used to calculate the time-varying connection relationship between nodes to obtain the second adjacency matrix. The adaptive K-nearest neighbor algorithm dynamically adjusts the number of neighbors and the distance threshold to ensure that new adjacency relationships are reflected in a timely manner when the data changes. The second node feature matrix and the second adjacency matrix are input into the graph convolutional network for message passing and aggregation operations to obtain the second updated node representation. The graph convolutional network updates node features by aggregating information from neighboring nodes. The formula is:

[0142] ;

[0143] in, It is The node feature matrix of the layer, is the normalized adjacency matrix, is the weight matrix, is an activation function. This process makes the features of each node contain not only its own information but also the information of its neighboring nodes. The second updated node representation is subjected to time-series graph pooling, and the DiffPool method of the attention mechanism is used to calculate the time-series clustering matrix to obtain the second graph-level representation. The DiffPool method dynamically aggregates nodes through the attention mechanism to form a higher-level graph representation and capture the global pattern in the data. The second graph-level representation is multi-headed fused with the second multi-branch CNN feature vector and feature weights are calculated to obtain a comprehensive feature representation. The comprehensive feature representation is mapped to obtain the second film quality evaluation score and abnormal probability distribution. The evaluation score reflects the overall status of the film quality, while the abnormal probability distribution points out the areas where there may be problems. According to the abnormal probability distribution, the defect area on the film surface is located using threshold segmentation and morphological operations to generate a defect prediction map. Threshold segmentation can accurately identify high-risk areas, while morphological operations (such as dilation and corrosion) further correct and refine the defect boundaries. The second film quality evaluation score and defect prediction map are subjected to time-series analysis to obtain dynamic quality evaluation results. By tracking the quality changes during the film growth process, the process parameters can be adjusted in real time to improve the uniformity and stability of the film quality. For example, when monitoring the growth of a certain semiconductor film, by identifying and analyzing the defect information in the real-time data, the parameters of the chemical vapor deposition (CVD) process, such as gas flow and reaction temperature, can be adjusted in time to avoid the expansion and formation of defects.

[0144] In a specific embodiment, the process of executing step S5 may specifically include the following steps:

[0145] (1) Performing image segmentation on the defect prediction image to obtain a defect area mask, and performing periodic stress loading on the film surface based on the defect area mask to obtain a stress-modulated film sample;

[0146] (2) Performing Raman spectroscopy scanning on the stress-modulated thin film sample to obtain stress-modulated Raman spectroscopy data, and performing Fourier transform on the stress-modulated Raman spectroscopy data to extract the stress modulation frequency component and obtain a stress-sensitive Raman peak displacement map;

[0147] (3) Based on the Raman peak displacement map, the biaxial stress model is used to perform stress field inversion calculation to obtain the stress distribution map on the film surface, and the stress distribution map on the film surface is threshold segmented to obtain the potential defect depth distribution;

[0148] (4) Generate an adaptive etching scheme based on the potential defect depth distribution, and perform intelligent etching on the thin film sample, peeling off layer by layer according to the adaptive etching scheme to obtain a multi-layer etched surface;

[0149] (5) Perform high-resolution scanning electron microscopy imaging on the multi-layer etched surface to obtain defect morphology and composition data. Then, through a three-dimensional reconstruction algorithm, obtain three-dimensional defect distribution verification data of the thin film based on the defect morphology and composition data, stress distribution map and etching depth information.

[0150] Specifically, the defect prediction map is segmented to distinguish different areas in the image. Through segmentation algorithms, such as threshold segmentation, edge detection or watershed algorithms, the defect area is accurately identified, and a defect area mask is generated to mark the location of defects on the film surface. According to the defect area mask, the film surface is periodically stressed. Periodic stress loading can be applied to the film mechanically, with the purpose of inducing and amplifying stress concentration areas that may exist in the film, making these areas more significant in subsequent spectral analysis, and obtaining stress-modulated film samples. Raman spectroscopy is performed on the stress-modulated film sample. Raman spectroscopy technology can sensitively detect stress changes in materials, and stress-modulated Raman spectroscopy data can be obtained by analyzing changes in the scattering spectrum. These data contain information on Raman peak displacement caused by stress changes. Fourier transform the Raman spectroscopy data to extract the frequency components of stress modulation. The formula for Fourier transform is:

[0151] ;

[0152] in, represents the amplitude of the frequency domain signal, is the frequency, is the time domain signal, It's time. is an imaginary unit. Through Fourier transform, the spectral data in the time domain is converted into frequency domain data, so that the periodic changes of stress can be observed more clearly. The obtained frequency components are used to construct a stress-sensitive Raman peak displacement map, which shows the displacement of the Raman peak under the influence of stress. According to the Raman peak displacement map, the biaxial stress model is used to perform stress field inversion calculation to obtain the stress distribution map on the film surface. The biaxial stress model assumes that the film is subjected to stress in two directions, and the basic equation of the model is:

[0153] ;

[0154] in, and are the stresses in the x-axis and y-axis directions, and is the initial stress, and is the stress sensitivity coefficient, is the Raman peak shift. These parameters are obtained by fitting the Raman peak shift data, thereby inferring the surface stress distribution. Threshold segmentation is performed on the surface stress distribution map of the film to identify the areas of stress concentration and mark the depth distribution of potential defects. Threshold segmentation can divide the image into different regions, where the areas with greater stress correspond to the potential defect areas. According to the potential defect depth distribution, an adaptive etching scheme is generated. The adaptive etching scheme dynamically adjusts the etching parameters, such as etching depth and etching speed, according to the stress and defect characteristics of different regions to ensure effective stripping and analysis of the defect area. The film sample is subjected to intelligent etching treatment, and the layer-by-layer stripping is performed according to the adaptive etching scheme to obtain a multi-layer etched surface. High-resolution scanning electron microscopy imaging is performed on the multi-layer etched surface to obtain the morphology and composition data of the defects. High-resolution scanning electron microscopy imaging can provide high-resolution surface morphology images, allowing the system to observe the defect details exposed after etching, such as cracks, holes or particles. The chemical composition data of the defect area is obtained by energy spectrum analysis. Through the three-dimensional reconstruction algorithm, a three-dimensional defect distribution map of the film is constructed based on the defect morphology and composition data, stress distribution map and etching depth information.

[0155] In a specific embodiment, the process of executing step S6 may specifically include the following steps:

[0156] (1) Perform spatial statistical analysis on the three-dimensional defect distribution verification data to obtain the defect distribution feature vector, and construct a thin film growth simulator based on the defect distribution feature vector;

[0157] (2) Perform simulation analysis on the thin film growth simulator to generate a parameter-defect mapping database. Based on the parameter-defect mapping database, a Gaussian process regression model is trained to construct a probabilistic relationship between growth parameters and defect distribution, thereby obtaining a thin film quality prediction model.

[0158] (3) Perform Bayesian optimization on the film quality prediction model to obtain a candidate growth parameter set, and perform multi-objective optimization based on the candidate growth parameter set to obtain a Pareto optimal solution set;

[0159] (4) Perform cluster analysis on the Pareto optimal solution set to obtain representative growth parameter combinations, and generate an optimized experimental matrix based on the representative growth parameter combinations;

[0160] (5) A response surface analysis was performed on the optimization experimental matrix, and a second-order polynomial regression model was established to obtain the parameter-quality response surface. Based on the parameter-quality response surface, the gradient descent method was used to find the global optimal point and obtain the target film growth control parameter set.

[0161] Specifically, spatial statistical analysis is performed to obtain defect distribution feature vectors. Spatial statistical analysis describes the overall picture of defects by quantifying the distribution characteristics of defects on the surface of the film, such as density, size, shape and position. Feature vectors can accurately capture the spatial distribution pattern of defects in the film. For example, by calculating the number of defects per unit area, the defect density distribution is obtained, or by analyzing the geometric shape of the defects, the shape characteristics of the defects are obtained. A film growth simulator is constructed based on the defect distribution feature vector. A film growth simulator is a numerical model that simulates the interaction of various physical and chemical processes during film growth and how these processes affect the formation and distribution of defects. By inputting different growth parameters (such as temperature, pressure, gas flow rate, etc.), the simulator predicts the impact of these parameter changes on the defect distribution. The output results of the simulator form a parameter-defect mapping database, which records the changes in defect distribution under different parameter settings. The parameter-defect mapping database is used to train the Gaussian process regression model. The Gaussian process regression model is a non-parametric Bayesian regression method that describes the properties of the potential function by constructing a prior distribution and updates it to the posterior distribution through training data. The Gaussian process is used to construct the probabilistic relationship between growth parameters and defect distribution. The Gaussian process regression model is described by the following formula:

[0162] ;

[0163] in, is the predicted defect distribution, is the growth parameter vector, is the mean function, usually set to zero, is a kernel function that describes the similarity between points in the input space. Through training data, the Gaussian process regression model learns the effect of growth parameters on defect distribution. The film quality prediction model is optimized by Bayesian optimization to determine the best growth parameter set. Bayesian optimization is a sequential optimization strategy that uses previous experimental results to guide the selection of the next experiment to find the optimal solution in the least number of experiments. In this optimization, the goal is to find a set of growth parameters that minimize defects. The objective function of Bayesian optimization usually uses methods such as expected improvement or upper confidence limit to measure the usefulness of potential experimental points. Through Bayesian optimization, a set of candidate growth parameter sets is obtained. Multi-objective optimization is performed based on the candidate growth parameter set to obtain the Pareto optimal solution set. The Pareto optimal solution set refers to a solution set in which no solution can be further optimized without compromising other objectives among all objectives. Multi-objective optimization is usually achieved through evolutionary algorithms or other heuristic methods. Each point on the Pareto front represents an equilibrium state in which the optimality of each objective is the best in some sense. After obtaining the Pareto optimal solution set, cluster analysis is performed on it to obtain representative growth parameter combinations. Cluster analysis can cluster similar parameter combinations together to identify several key growth patterns or strategies. Representative growth parameter combinations provide a variety of process route options for further experimental verification. Based on the representative growth parameter combinations, an optimized experimental matrix is ​​generated. The experimental matrix refers to a system used to arrange and control experiments in experimental design, which can effectively evaluate the impact of various factors on the results. Response surface analysis is performed on the optimized experimental matrix, and a second-order polynomial regression model is established to obtain the parameter-quality response surface. The form of the response surface model is:

[0164] ;

[0165] in, It is a quality indicator. is the growth parameter, , , and are the model coefficients. These coefficients are obtained by least squares fitting. The gradient descent method is used to find the global optimal point on the parameter-quality response surface to obtain the target film growth control parameter set. The gradient descent method is an iterative optimization algorithm that gradually approaches the optimal solution by calculating the gradient of the objective function and moving along the gradient direction.

[0166] The above is a description of the thin film surface quality detection method of the semiconductor device in the embodiment of the present invention. The following is a description of the thin film surface quality detection system of the semiconductor device in the embodiment of the present invention. Figure 2 , an embodiment of a semiconductor device thin film surface quality detection system in an embodiment of the present invention includes:

[0167] A scanning module, used for performing multimodal scanning on the thin film surface of the semiconductor device to obtain an original multimodal data set;

[0168] The extraction module is used to perform feature extraction and feature dimension reduction on the original multimodal data set to obtain a standard multidimensional feature matrix;

[0169] A training module is used to input a standard multi-dimensional feature matrix into a preset multi-branch convolutional neural network and graph neural network fusion algorithm to perform film quality assessment training to obtain a film quality assessment model;

[0170] A monitoring module is used to perform edge computing real-time monitoring of the thin film growth process of semiconductor devices through a thin film quality assessment model to obtain dynamic quality assessment results and defect prediction diagrams;

[0171] The analysis module is used to perform stress modulation Raman spectroscopy and intelligent etching analysis on the dynamic quality assessment results and defect prediction diagrams to obtain verification data on the three-dimensional defect distribution of the thin film;

[0172] The generation module is used to optimize the thin film growth parameters of the semiconductor device based on the three-dimensional defect distribution verification data and generate a target thin film growth control parameter set.

[0173] Through the collaboration of the above components, the surface of the semiconductor device film is scanned multimodally to obtain comprehensive data including mechanical, chemical and electrical properties, thus realizing a multi-dimensional evaluation of the film quality and improving the comprehensiveness and accuracy of the detection. Advanced signal processing technologies such as Fourier transform, wavelet packet decomposition and non-negative matrix decomposition are used to extract features and reduce the dimension of the original multimodal data, effectively extracting the key feature information of the film surface. The film quality evaluation model is constructed using a multi-branch convolutional neural network and a graph neural network fusion algorithm, realizing the intelligent analysis and evaluation of complex film surface features. Real-time monitoring of the film growth process is achieved through edge computing technology, and potential quality problems are discovered and predicted in time, improving the control accuracy of the production process. Combined with stress-modulated Raman spectroscopy and intelligent etching analysis, the three-dimensional defect distribution of the film is accurately verified. Based on the three-dimensional defect distribution verification data, the film growth parameters are optimized using advanced optimization algorithms to generate the target film growth control parameter set, effectively improving the quality and efficiency of the film growth. From nano-scale surface scanning to macro-scale film growth parameter optimization, the cross-scale full-process analysis and control of film quality is realized, thereby improving the accuracy of film surface quality detection of semiconductor devices.

[0174] 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.

[0175] 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.

[0176] 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 detecting the surface quality of a thin film of a semiconductor device, characterized in that: The method comprises: Perform multimodal scanning on the thin film surface of the semiconductor device to obtain an original multimodal data set; Performing feature extraction and feature dimensionality reduction on the original multimodal data set to obtain a standard multidimensional feature matrix; The standard multi-dimensional feature matrix is ​​input into a preset multi-branch convolutional neural network and a graph neural network fusion algorithm for film quality assessment training to obtain a film quality assessment model; specifically comprising: performing feature decomposition on the standard multi-dimensional feature matrix, dividing the standard multi-dimensional feature matrix into a first mechanical feature sub-matrix, a first chemical feature sub-matrix and a first electrical feature sub-matrix; the first mechanical feature sub-matrix, the first chemical feature sub-matrix and the first electrical feature sub-matrix are input into a preset multi-branch convolutional neural network, the multi-branch convolutional neural network includes a mechanical feature branch, a chemical feature branch and an electrical feature branch, the mechanical feature branch adopts a 3D-CNN structure, and the chemical feature branch adopts a 2D-CN N structure, the electrical feature branch adopts a 1D-CNN structure; the first mechanical feature submatrix is ​​feature extracted by the mechanical feature branch to obtain a first mechanical feature vector, the mechanical feature branch includes 4 3D convolutional layers, each 3D convolutional layer is followed by BatchNormalization and ReLU activation function, and the last 3D convolutional layer uses global average pooling; the first chemical feature submatrix is ​​feature extracted by the chemical feature branch to obtain a first chemical feature vector, the chemical feature branch includes 5 2D convolutional layers, each 2D convolutional layer is followed by BatchNormalization and LeakyReLU activation function, and the last 2D convolutional layer uses global average pooling. The first electrical feature submatrix is ​​subjected to feature extraction through the electrical feature branch to obtain a first electrical feature vector, wherein the electrical feature branch comprises 6 1D convolutional layers, each 1D convolutional layer is followed by LayerNormalization and GELU activation function, and the last 1D convolutional layer uses adaptive average pooling; the first mechanical feature vector, the first chemical feature vector and the first electrical feature vector are concatenated to obtain a first multi-branch CNN feature vector, and the first multi-branch CNN feature vector is mapped through a fully connected layer to obtain a first node feature matrix; a graph structure is constructed according to the first node feature matrix, and the K nearest neighbor algorithm is used to calculate the connections between nodes. The first node feature matrix and the first adjacency matrix are input into a preset graph neural network for graph convolution operation, wherein the graph neural network comprises three GraphConv layers, each GraphConv layer is followed by an ELU activation function and a graph batch normalization, and a first updated node representation is obtained; a graph pooling operation is performed on the first updated node representation to obtain a first graph-level representation; the first graph-level representation is fused with the first multi-branch CNN feature vector, and the weights are calculated through a multi-head attention mechanism to obtain a first film quality assessment score; the multi-branch convolutional neural network and the graph neural network are optimized through back propagation to obtain a film quality assessment model; The thin film growth process of the semiconductor device is monitored in real time by edge computing through the thin film quality assessment model to obtain dynamic quality assessment results and defect prediction diagrams; specifically, the method includes: collecting real-time multimodal data of the semiconductor device during the thin film growth process, and preprocessing the real-time multimodal data through edge devices to obtain a real-time feature matrix; decomposing the real-time feature matrix into a second mechanical feature sub-matrix, a second chemical feature sub-matrix, and a second electrical feature sub-matrix; inputting the second mechanical feature sub-matrix, the second chemical feature sub-matrix, and the second electrical feature sub-matrix into a multi-branch convolutional neural network of a thin film quality assessment model for convolution operation to obtain a second mechanical feature vector, a second chemical feature vector, and a second electrical feature vector; concatenating the second mechanical feature vector, the second chemical feature vector, and the second electrical feature vector, and mapping them through a fully connected layer to obtain a second node feature matrix; according to the second node feature The matrix constructs a dynamic graph structure, uses an adaptive K-nearest neighbor algorithm to calculate the time-varying connection relationship between nodes, and obtains a second adjacency matrix; the second node feature matrix and the second adjacency matrix are input into the graph convolutional network for message passing and aggregation operations to obtain a second updated node representation; the second updated node representation is pooled in a time series graph, and the DiffPool method of the attention mechanism is used to calculate the time series clustering matrix to obtain a second graph-level representation; the second graph-level representation is multi-headedly fused with the second multi-branch CNN feature vector and feature weights are calculated to obtain a comprehensive feature representation; the comprehensive feature representation is mapped to obtain a second film quality assessment score and anomaly probability distribution, and according to the anomaly probability distribution, the defect area on the film surface is located using threshold segmentation and morphological operations to obtain a defect prediction map; the second film quality assessment score and the defect prediction map are subjected to time series analysis to obtain a dynamic quality assessment result; The dynamic quality assessment result and the defect prediction map are subjected to stress modulation Raman spectroscopy and intelligent etching analysis to obtain three-dimensional defect distribution verification data of the thin film; specifically comprising: performing image segmentation on the defect prediction map to obtain a defect area mask, and periodically stress loading the surface of the thin film according to the defect area mask to obtain a stress-modulated thin film sample; performing Raman spectroscopy scanning on the stress-modulated thin film sample to obtain stress-modulated Raman spectroscopy data, and performing Fourier transform on the stress-modulated Raman spectroscopy data to extract the stress modulation frequency component to obtain a stress-sensitive Raman peak displacement map; according to the Raman peak displacement map , using a biaxial stress model to perform stress field inversion calculation to obtain a stress distribution map on the film surface, and performing threshold segmentation on the stress distribution map on the film surface to obtain a potential defect depth distribution; generating an adaptive etching scheme based on the potential defect depth distribution, and performing intelligent etching processing on the film sample, peeling off layer by layer according to the adaptive etching scheme to obtain a multi-layer etched surface; performing high-resolution scanning electron microscopy imaging on the multi-layer etched surface to obtain defect morphology and composition data, and using a three-dimensional reconstruction algorithm, according to the defect morphology and composition data, the stress distribution map and the etching depth information, obtaining three-dimensional defect distribution verification data of the film; Based on the three-dimensional defect distribution verification data, thin film growth parameters of the semiconductor device are optimized to generate a target thin film growth control parameter set.

2. The method for detecting the surface quality of a thin film of a semiconductor device according to claim 1, characterized in that: The multimodal scanning of the thin film surface of the semiconductor device is performed to obtain an original multimodal data set, including: Performing nanoindentation scanning on the film surface of the semiconductor device to obtain force-displacement curve data, and performing Oliver-Pharr model analysis on the film surface based on the force-displacement curve data to obtain elastic modulus and hardness distribution data; Performing micro-area Raman spectroscopy scanning on the surface of the film to obtain Raman spectroscopy data, and analyzing the surface of the film based on the Raman spectroscopy data to obtain a chemical component distribution map; Scanning the surface of the film with a high-resolution electrochemical scanning probe to obtain local electrochemical impedance spectroscopy data, and performing equivalent circuit fitting on the surface of the film based on the local electrochemical impedance spectroscopy data to obtain charge transfer resistance and double layer capacitance distribution data; Performing spatial coordinate unification processing on the elastic modulus and hardness distribution data, the chemical component distribution map, and the charge transfer resistance and double-layer capacitance distribution data to obtain three-dimensional matrix format data; Baseline correction and noise suppression are performed on the three-dimensional matrix format data to obtain an original multimodal data set.

3. The method for detecting the surface quality of a thin film of a semiconductor device according to claim 2, characterized in that: The feature extraction and feature dimensionality reduction are performed on the original multimodal data set to obtain a standard multidimensional feature matrix, including: Performing Fourier transform processing on the force-displacement curve data in the original multimodal data set to obtain frequency domain feature vectors, and performing wavelet packet decomposition on the force-displacement curve data according to the frequency domain feature vectors to obtain multi-scale energy distribution characteristics; Performing discrete cosine transformation on the Raman spectral data in the original multimodal data set to obtain spectral characteristic coefficients, and performing non-negative matrix decomposition on the Raman spectral data according to the spectral characteristic coefficients to obtain a chemical component characteristic matrix; Performing Hilbert-Huang transform on the electrochemical impedance spectroscopy data in the original multimodal data set to obtain a time-frequency characteristic diagram, and performing empirical mode decomposition on the electrochemical impedance spectroscopy data according to the time-frequency characteristic diagram to obtain an intrinsic mode function set; Performing feature dimensionality reduction on the multi-scale energy distribution features, the chemical component feature matrix, and the intrinsic mode function set to obtain a feature set after dimensionality reduction; According to the feature set after dimension reduction, a kernel principal component analysis is performed on the feature space to obtain a nonlinear feature map, and a feature space tensor decomposition is performed on the nonlinear feature map to obtain a standard multidimensional feature matrix.

4. The method for detecting the surface quality of a thin film of a semiconductor device according to claim 1, characterized in that: The extracting features from the first mechanical feature submatrix through the mechanical feature branch to obtain a first mechanical feature vector includes: Performing three-dimensional zero-filling processing on the first mechanical feature submatrix through the mechanical feature branch to obtain a filled feature tensor; Performing a first 3D convolution operation on the padded feature tensor through a first 3D convolution layer in the mechanical feature branch to obtain a first convolution feature map, wherein the first 3D convolution operation uses 32 3x3x3 convolution kernels with a step size of 1; Performing batch normalization processing on the first convolution feature map to obtain a first normalized feature map, and performing a ReLU activation function operation on the first normalized feature map to obtain a first activated feature map; Performing a second 3D convolution operation on the first activation feature map through a second 3D convolution layer in the mechanical feature branch to obtain a second convolution feature map, wherein the second 3D convolution operation uses 64 3x3x3 convolution kernels with a step size of 2; Performing batch normalization processing on the second convolution feature map to obtain a second normalized feature map, and performing a ReLU activation function operation on the second normalized feature map to obtain a second activated feature map; Performing a third 3D convolution operation on the second activation feature map through a third 3D convolution layer in the mechanical feature branch to obtain a third convolution feature map, wherein the third 3D convolution operation uses 128 3x3x3 convolution kernels with a step size of 1; Performing batch normalization on the third convolutional feature map to obtain a third normalized feature map, and performing a ReLU activation function operation on the third normalized feature map to obtain a third activated feature map; Performing a fourth 3D convolution operation on the third activation feature map through a fourth 3D convolution layer in the mechanical feature branch to obtain a fourth convolution feature map, wherein the fourth 3D convolution operation uses 256 3x3x3 convolution kernels with a step size of 2; Performing batch normalization on the fourth convolutional feature map to obtain a fourth normalized feature map, and performing a ReLU activation function operation on the fourth normalized feature map to obtain a fourth activated feature map; A global average pooling operation is performed on the fourth activated feature map to obtain a first mechanical feature vector.

5. The method for detecting the surface quality of a semiconductor device thin film according to claim 1, characterized in that: The step of constructing a graph structure according to the first node feature matrix, calculating the connection relationship between nodes using a K-nearest neighbor algorithm to obtain a first adjacency matrix, and inputting the first node feature matrix and the first adjacency matrix into a preset graph neural network for a graph convolution operation includes: Performing L2 normalization processing on the first node feature matrix to obtain a normalized node feature matrix, and using a K nearest neighbor algorithm to calculate the Euclidean distance between nodes based on the normalized node feature matrix, selecting K nearest neighbor nodes of each node, and obtaining a K nearest neighbor relationship matrix; Constructing a first adjacency matrix according to the K nearest neighbor relationship matrix, and performing symmetry processing on the first adjacency matrix to obtain a symmetric adjacency matrix; Performing degree matrix normalization on the symmetric adjacency matrix to obtain a normalized adjacency matrix, and inputting the normalized node feature matrix and the normalized adjacency matrix into the first GraphConv layer in the graph neural network to perform a graph convolution operation to obtain the first hidden layer features; Performing an ELU activation function operation on the first hidden layer feature to obtain a first activated feature, and performing a batch normalization process on the first activated feature to obtain a first batch of normalized features; Input the first batch of normalized features into the second GraphConv layer in the graph neural network, perform graph convolution operation, and obtain the second hidden layer features; Performing an ELU activation function operation on the second hidden layer feature to obtain a second activated feature, and performing a batch normalization process on the second activated feature to obtain a second batch of normalized features; Input the second batch of normalized features into the third GraphConv layer in the graph neural network, perform graph convolution operation, and obtain the third hidden layer features; An ELU activation function operation is performed on the third hidden layer feature to obtain a third activated feature, and a graph batch normalization process is performed on the third activated feature to obtain a first updated node representation.

6. The method for detecting the surface quality of a thin film of a semiconductor device according to claim 1, characterized in that: The step of optimizing thin film growth parameters of the semiconductor device based on the three-dimensional defect distribution verification data to generate a target thin film growth control parameter set includes: Performing spatial statistical analysis on the three-dimensional defect distribution verification data to obtain a defect distribution feature vector, and constructing a thin film growth simulator based on the defect distribution feature vector; Performing simulation analysis on the thin film growth simulator to generate a parameter-defect mapping database, and training a Gaussian process regression model based on the parameter-defect mapping database to construct a probability relationship between growth parameters and defect distribution, thereby obtaining a thin film quality prediction model; Performing Bayesian optimization on the film quality prediction model to obtain a candidate growth parameter set, and performing multi-objective optimization based on the candidate growth parameter set to obtain a Pareto optimal solution set; Performing cluster analysis on the Pareto optimal solution set to obtain representative growth parameter combinations, and generating an optimization experiment matrix based on the representative growth parameter combinations; A response surface analysis is performed on the optimization experiment matrix, a second-order polynomial regression model is established, a parameter-quality response surface is obtained, and based on the parameter-quality response surface, a gradient descent method is used to find the global optimal point to obtain a target film growth control parameter set.

7. A semiconductor device thin film surface quality detection system, characterized in that: A system for performing a thin film surface quality detection method for a semiconductor device according to any one of claims 1 to 6, the system comprising: A scanning module, used for performing multimodal scanning on the thin film surface of the semiconductor device to obtain an original multimodal data set; An extraction module, used for performing feature extraction and feature dimension reduction on the original multimodal data set to obtain a standard multidimensional feature matrix; The training module is used to input the standard multi-dimensional feature matrix into a preset multi-branch convolutional neural network and a graph neural network fusion algorithm for film quality assessment training to obtain a film quality assessment model; specifically comprising: performing feature decomposition on the standard multi-dimensional feature matrix, dividing the standard multi-dimensional feature matrix into a first mechanical feature sub-matrix, a first chemical feature sub-matrix and a first electrical feature sub-matrix; inputting the first mechanical feature sub-matrix, the first chemical feature sub-matrix and the first electrical feature sub-matrix into a preset multi-branch convolutional neural network, the multi-branch convolutional neural network including a mechanical feature branch, a chemical feature branch and an electrical feature branch, the mechanical feature branch adopts a 3D-CNN structure, and the chemical feature branch adopts a 2D-CNN structure, the electrical feature branch adopts a 1D-CNN structure; the first mechanical feature submatrix is ​​feature extracted by the mechanical feature branch to obtain a first mechanical feature vector, the mechanical feature branch includes 4 3D convolution layers, each 3D convolution layer is followed by BatchNormalization and ReLU activation function, and the last 3D convolution layer uses global average pooling; the first chemical feature submatrix is ​​feature extracted by the chemical feature branch to obtain a first chemical feature vector, the chemical feature branch includes 5 2D convolution layers, each 2D convolution layer is followed by BatchNormalization and LeakyReLU activation function, and the last 2 The D convolution layer uses global maximum pooling; the first electrical feature submatrix is ​​feature extracted through the electrical feature branch to obtain a first electrical feature vector, the electrical feature branch includes 6 1D convolution layers, each 1D convolution layer is followed by LayerNormalization and GELU activation function, and the last 1D convolution layer uses adaptive average pooling; the first mechanical feature vector, the first chemical feature vector and the first electrical feature vector are concatenated to obtain a first multi-branch CNN feature vector, and the first multi-branch CNN feature vector is mapped through a fully connected layer to obtain a first node feature matrix; a graph structure is constructed according to the first node feature matrix, and the node is calculated using the K nearest neighbor algorithm. The first node feature matrix and the first adjacency matrix are input into a preset graph neural network for graph convolution operation, wherein the graph neural network comprises three GraphConv layers, each GraphConv layer is followed by an ELU activation function and a graph batch normalization, and a first updated node representation is obtained; a graph pooling operation is performed on the first updated node representation to obtain a first graph-level representation; the first graph-level representation is fused with the first multi-branch CNN feature vector, and the weights are calculated through a multi-head attention mechanism to obtain a first film quality assessment score; the multi-branch convolution neural network and the graph neural network are optimized through back propagation to obtain a film quality assessment model; A monitoring module is used to perform edge computing real-time monitoring of the thin film growth process of the semiconductor device through the thin film quality assessment model to obtain dynamic quality assessment results and defect prediction diagrams; specifically comprising: collecting real-time multimodal data of the semiconductor device during the thin film growth process, and preprocessing the real-time multimodal data through an edge device to obtain a real-time feature matrix; decomposing the real-time feature matrix into a second mechanical feature submatrix, a second chemical feature submatrix, and a second electrical feature submatrix; inputting the second mechanical feature submatrix, the second chemical feature submatrix, and the second electrical feature submatrix into a multi-branch convolutional neural network of a thin film quality assessment model for convolution operation to obtain a second mechanical feature vector, a second chemical feature vector, and a second electrical feature vector; splicing the second mechanical feature vector, the second chemical feature vector, and the second electrical feature vector, and mapping through a fully connected layer to obtain a second node feature matrix; according to the second The node feature matrix constructs a dynamic graph structure, and uses an adaptive K-nearest neighbor algorithm to calculate the time-varying connection relationship between nodes to obtain a second adjacency matrix; the second node feature matrix and the second adjacency matrix are input into the graph convolutional network for message passing and aggregation operations to obtain a second updated node representation; the second updated node representation is subjected to time-series graph pooling, and the DiffPool method of the attention mechanism is used to calculate the time-series clustering matrix to obtain a second graph-level representation; the second graph-level representation is subjected to multi-head attention fusion and feature weight calculation with the second multi-branch CNN feature vector to obtain a comprehensive feature representation; the comprehensive feature representation is mapped to obtain a second film quality assessment score and anomaly probability distribution, and according to the anomaly probability distribution, the defect area on the film surface is located using threshold segmentation and morphological operations to obtain a defect prediction map; the second film quality assessment score and the defect prediction map are subjected to time-series analysis to obtain a dynamic quality assessment result; The analysis module is used to perform stress modulation Raman spectroscopy and intelligent etching analysis on the dynamic quality assessment result and the defect prediction map to obtain three-dimensional defect distribution verification data of the thin film; specifically comprising: performing image segmentation on the defect prediction map to obtain a defect area mask, and according to the defect area mask, periodically stress loading the surface of the thin film to obtain a stress modulated thin film sample; performing Raman spectroscopy scanning on the stress modulated thin film sample to obtain stress modulation Raman spectroscopy data, and performing Fourier transform on the stress modulation Raman spectroscopy data to extract the stress modulation frequency component to obtain a stress-sensitive Raman peak displacement map; according to the Raman peak displacement map, using a biaxial stress model to perform stress field inversion calculation to obtain a stress distribution map on the film surface, and performing threshold segmentation on the stress distribution map on the film surface to obtain a potential defect depth distribution; generating an adaptive etching scheme based on the potential defect depth distribution, and performing intelligent etching processing on the film sample, peeling off layer by layer according to the adaptive etching scheme to obtain a multi-layer etched surface; performing high-resolution scanning electron microscopy imaging on the multi-layer etched surface to obtain defect morphology and composition data, and using a three-dimensional reconstruction algorithm, according to the defect morphology and composition data, the stress distribution map and the etching depth information, obtaining three-dimensional defect distribution verification data of the film; A generation module is used to optimize the thin film growth parameters of the semiconductor device based on the three-dimensional defect distribution verification data to generate a target thin film growth control parameter set.

Images