Semiconductor material impurity detection method using support vector machine

Abstract

The invention discloses a semiconductor material impurity detection method using a support vector machine. The method comprises the following steps: S1, obtaining an initial data set containing impurity characteristic data from a semiconductor material manufacturing process; s2, carrying out normalization and noise removal preprocessing on the data set; s3, performing feature extraction on the preprocessed data by using persistent coherence analysis, and generating a topological feature vector; s4, inputting the topological feature vector into a classifier based on a Hupfield network and a support vector machine for classification; s5, performing clustering analysis on the classification result by adopting a graph neural network to generate an impurity detection report; and S6, based on the detection report, optimizing and updating the support vector machine model by using a meta-learning technology. Through the method, different impurity types in the semiconductor material can be accurately identified and classified, and the detection accuracy and efficiency are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of semiconductor material detection, and in particular to a semiconductor material impurity detection method using a support vector machine. Background Art

[0002] In the production process of semiconductor materials, the purity of the materials and the distribution of internal impurities are crucial to the performance of the final product. The type, quantity and distribution of impurities directly affect the electrical properties, reliability and life of semiconductor devices. Therefore, accurate detection of impurities in semiconductor materials is a key link in the field of semiconductor manufacturing.

[0003] Traditional impurity detection methods mainly rely on physical testing methods, such as secondary ion mass spectrometry, transmission electron microscopy and optical microscopy. Although these methods can detect impurities in materials, they usually have the following defects:

[0004] The first is that the detection process of traditional methods is time-consuming and complicated; physical testing methods usually require multiple steps to complete impurity detection, such as sample preparation, testing instrument debugging and multiple measurements. These steps are not only time-consuming, but also easily affected by human operation, resulting in inconsistent test results.

[0005] Second, the sensitivity and resolution of existing detection methods are limited, making it difficult to accurately identify and quantify trace impurities. With the development of semiconductor technology, the size of devices is getting smaller and smaller, and the impact of trace impurities in materials has become more significant. However, traditional physical testing methods often cannot provide sufficient sensitivity and resolution when dealing with these trace impurities, resulting in missed detection or misjudgment.

[0006] Third, traditional methods also have shortcomings in the extraction and classification of impurity features. Existing methods usually rely on the experience of experts or simple rule-based classification algorithms, which perform poorly when dealing with complex impurity distributions and types. In particular, the accuracy and reliability of traditional methods are limited when dealing with impurity data with nonlinear characteristics.

[0007] Therefore, how to provide a semiconductor material impurity detection method using a support vector machine is a problem that those skilled in the art urgently need to solve. Summary of the invention

[0008] One object of the present invention is to propose a semiconductor material impurity detection method using a support vector machine. The present invention can more accurately extract, classify and analyze impurity features in semiconductor materials by utilizing persistent homology analysis, Hopfield networks, graph neural networks and meta-learning techniques. The method effectively identifies and classifies different types of impurities, especially those that are trace and difficult to detect. Compared with traditional physical detection methods, the present invention not only improves the accuracy and sensitivity of impurity detection, but also significantly improves the detection efficiency.

[0009] A semiconductor material impurity detection method using a support vector machine according to an embodiment of the present invention comprises the following steps:

[0010] S1. Acquire an initial data set containing impurity characteristic data from the manufacturing process of semiconductor materials, wherein the characteristic data includes the type, quantity and distribution information of the impurities;

[0011] S2, preprocessing the initial data set by data normalization and noise removal;

[0012] S3, using persistent homology analysis to extract features from the preprocessed data, and generating topological feature vectors by identifying persistent topological features in the data;

[0013] S4, inputting the generated topological feature vector into a classifier based on a Hopfield network combined with a support vector machine, wherein the classifier based on a Hopfield network combined with a support vector machine first uses the Hopfield network to store and retrieve feature patterns, and then classifies these topological feature vectors in combination with the support vector machine to generate a classification result;

[0014] S5. Use a clustering algorithm based on graph neural network to process and analyze the classification results, reveal the potential correlation and distribution pattern between different impurities, and generate an impurity detection report;

[0015] S6. Based on the impurity detection report, meta-learning technology is used to optimize and update the support vector machine model, and the model parameters are dynamically adjusted and optimized by sharing experience between multiple tasks and batches of materials.

[0016] Optionally, the S3 specifically includes:

[0017] S31, constructing a topological space formed by a point set for the preprocessed data set;

[0018] S32, performing persistent homology analysis on the topological space, including calculating homology groups at different scales, and obtaining the Betti number β of each dimension k , where k represents the dimension of the homology group, β k represents the number of k-dimensional homology groups formed at different scales;

[0019] S33, by analyzing the Betti number β at different scales k stability with changes in scale parameters, determining persistent topological features;

[0020] S34, converting the determined persistent topological features into a topological feature vector v topo , the topological feature vector contains the Betti number of each dimension and its corresponding persistence information.

[0021] Optionally, the S33 specifically includes:

[0022] S331. Under different scale parameters ∈, calculate the Betti number β corresponding to each scale k (∈), where β k (∈) represents the number of k-dimensional homology groups when the scale parameter is ∈, where k is the dimension of the homology group;

[0023] S332. Record and plot the Betti number β k (∈) The curve β changes with the scale parameter ∈ k (∈) vs ∈;

[0024] S333, for curve β k (∈) vs∈ to analyze and calculate the persistence p of each k-dimensional homology group in various scales k , persistence p k :

[0025]

[0026] Among them, w i is the weight coefficient of different homology groups at different scales, ∈ i 出现 and ∈ i 消失 They represent the scale parameters where the homology group first appears and disappears at the i-th scale and reflect the topological changes of the impurity characteristics at different scales. n is the number of homology groups.

[0027] S334, based on the calculated persistence p k , the following formula is used to comprehensively score each persistent topological feature S k , used to screen important topological features:

[0028]

[0029] Among them, α and β are weight parameters, max(p k ) is the maximum value among all k-dimensional homology group persistence, ∈ 消失 and ∈ 出现They represent the disappearance and appearance scales of topological features in the scale space respectively;

[0030] S335, the comprehensive score S k Above the set threshold S 阈值 The topological features are used as persistent topological features and converted into topological feature vectors v persist , the topological feature vector contains the Betti number β of each persistent topological feature k , persistence k And the comprehensive score S k information.

[0031] Optionally, the S4 specifically includes:

[0032] S41, the topological feature vector v generated from step S3 topo The input is sent to the Hopfield network for processing. The Hopfield network stores and retrieves the characteristic pattern through the improved energy function E(v). The energy function expression is:

[0033]

[0034] Where v = (v 1 ,v 2 ,,v n ) is the input topological feature vector and represents the impurity features extracted from the semiconductor material, w ij is the connection weight between feature vectors and reflects the association strength between different impurity features. i is the bias term of the feature vector and is used to adjust the importance of each impurity feature in the network, γ is the regularization parameter, and n represents the number of features processed by the model;

[0035] S42, through the Hopfield network to the input topological feature vector v topo Energy minimization is performed, and the characteristic mode is iteratively updated using an optimization algorithm based on the gradient descent method to obtain a stable characteristic mode v stable , the characteristic pattern is used to store and retrieve characteristic information of different impurity types in semiconductor materials;

[0036] S43, stabilize the characteristic mode v stable Input into the support vector machine classifier, the support vector machine classifier uses the improved kernel function K(x,y) to map the input feature pattern to the feature space;

[0037] S44, classify the feature patterns mapped to the feature space based on the support vector machine classifier, use the Lagrange multiplier method to solve the classification hyperplane, and output the classification result C(v stable ), where C(v stable) are classification labels for different impurity types in semiconductor materials.

[0038] Optionally, the S43 specifically includes:

[0039] S431, stabilize the characteristic mode v stable Input to the support vector machine classifier as the initial input feature for classification;

[0040] S432, in the support vector machine classifier, a nonlinear adaptive kernel function K is used adaptive (x, y) performs high-dimensional mapping on the input features, and the nonlinear adaptive kernel function is:

[0041]

[0042] Where x and y are input feature vectors and represent the topological feature vectors extracted from different batches of semiconductor materials, σ is the width parameter of the radial basis function kernel and is used to control the similarity measure of impurity features, and λ 1 , 2 is the weight coefficient of the kernel function combination, β is the coefficient used to optimize the nonlinear relationship between the polynomial kernel and the input vector, d is the order of the polynomial kernel, and γ is the adjustment factor that controls the nonlinearity of the overall kernel function;

[0043] S433, input the mapped high-dimensional features into the optimized support vector machine classifier, and optimize the classifier by an improved objective function, wherein the objective function includes a dynamically adjusted regularization term R(α), and the optimized objective function is:

[0044]

[0045] Where maximizeW new (α) represents the optimization of the objective function W new (α) to maximize the discriminative ability of the classifier, α i is the Lagrange multiplier, which indicates the contribution of each impurity feature to the classifier, y i is the category label of the i-th sample, representing different impurity types in semiconductor materials, K adaptive (x i ,x j ) is a new adaptive kernel function used to capture the complex interactions of impurity features, μ is a dynamically adjusted regularization coefficient, and n represents the number of different impurity feature samples involved in training;

[0046] S434, by solving the optimized objective function W new (α), obtain the optimal classification hyperplane, and calculate the final classification decision function for the input feature pattern v stableClassify and output the classification result C(v stable ).

[0047] Optionally, the S5 specifically includes:

[0048] S51, sorting the classification results in step S4 to generate a classification data matrix M class , each column of the classification data matrix represents a specific impurity type, and each row is the detection result of the corresponding impurity in different batches of materials;

[0049] S52, using the clustering algorithm based on graph neural network to classify the classification data matrix M class Processing is performed to construct a graph structure G(V,E), where V represents the node set of impurity type, E represents the connection relationship between nodes, and the connection relationship is determined by the impurity feature similarity S in different batches of materials. ij OK, similarity S ij :

[0050]

[0051] in and are the stable eigenvectors of the i-th and j-th impurity types, σ s is the scale parameter for controlling the similarity calculation, and λ is the weight parameter for adjusting the influence of cross terms between feature vectors;

[0052] S53, performing spectral clustering analysis on the graph structure G(V,E), determining the potential correlation between different impurity types by analyzing the connection weights and feature vectors of the nodes in the graph, and clustering closely correlated impurity types into the same category;

[0053] S54. Generate an impurity clustering matrix M based on the results of the spectral clustering analysis cluster ;

[0054] S55, based on the impurity clustering matrix M cluster The impurity detection report includes the clustering results, quantity statistics and distribution pattern of each impurity type in the semiconductor material.

[0055] Optionally, the S6 specifically includes:

[0056] S61, based on the impurity detection report generated in step S5, extracting a data set containing impurity characteristic distribution information, wherein the data set includes the type, quantity and distribution pattern of impurities in different batches of materials;

[0057] S62. Use meta-learning technology to dynamically optimize the support vector machine model, build a meta-model, analyze the impurity detection tasks of different batches of materials, extract the commonalities and differences between the tasks, and optimize and adjust the global parameters of the support vector machine model based on these analysis results;

[0058] S63, obtaining a global parameter update strategy of the support vector machine model by optimizing the meta-model, wherein the global parameter update strategy is based on a comprehensive analysis of the losses of different tasks and the changes in model parameters;

[0059] S64. After obtaining the global parameter update strategy, apply it to the support vector machine model, and further adjust the kernel function parameters and regularization coefficient of the model, so as to achieve fine adjustment of the model through iterative optimization and optimize the recognition capability of complex impurity features;

[0060] S65. Verify the optimized support vector machine model, test the model using a new batch of impurity feature data sets, evaluate the generalization ability and classification accuracy of the model in new impurity detection tasks, feed the test results back to the meta-model, further optimize and adjust the model parameters, and form a dynamic and adaptive optimization cycle.

[0061] Optionally, the S63 specifically includes:

[0062] S631, using the impurity feature data and classification results extracted from step S61 to construct a meta-learning task set T, where each task contains relevant classification results and model performance, for analyzing the commonalities and differences between different tasks;

[0063] S632. Based on the task features in the meta-learning task set T, design a global parameter update strategy, first calculate the gradient information of each task in is the loss function of the task, and then the global gradient is calculated by weighted nonlinear combination

[0064]

[0065] Among them, ω i represents the weight of the i-th task in the global parameter update, λ i is the nonlinear adjustment coefficient, which is used to control the nonlinear contribution of the task gradient, γ i is a nonlinear index used to adjust the influence of gradient modulus, α i is the amplification factor, which enhances the nonlinear effect of the gradient between tasks through the hyperbolic tangent function, ρ i is the exponential decay coefficient, used to balance the linear and nonlinear contributions of the gradient, β i is the index adjustment coefficient;

[0066] S633, using the calculated global gradient The global parameter θ of the support vector machine model is updated. In this process, the adaptive gradient descent method is used and the dynamically adjusted learning rate η is introduced. t :

[0067]

[0068] Among them, η 0 is the initial learning rate and determines the basic step size of parameter update, G t is the cumulative gradient square matrix and is used to smooth gradient fluctuations, δ is the smoothing term, τ is the time step bias term and is used to adjust the learning rate as the time step changes, ξ is the learning rate oscillation adjustment coefficient and is used to introduce slight oscillations, and T is the oscillation period;

[0069] S634. After the global parameter update is completed, further adjust the kernel function parameters and regularization coefficients of the support vector machine model to effectively cope with the changes in the distribution of impurity characteristics in different batches of materials. The adjustment strategy includes dynamically modifying the relevant parameters and regularization coefficients of the kernel function according to the complexity of the detection task, and optimizing the classification accuracy of the support vector machine model in complex detection tasks.

[0070] S635. Apply the optimized support vector machine model to the detection task of a new batch of materials, monitor the classification performance of the model in real time, and feed the test results back to the meta-learning model to dynamically adjust the global parameter update strategy to form a continuously optimized adaptive loop.

[0071] The beneficial effects of the present invention are:

[0072] (1) The present invention can more accurately extract, identify and classify complex impurity features in semiconductor materials by combining persistent homology analysis with Hopfield networks and support vector machines. Compared with traditional physical detection methods, this method can more effectively handle trace and complex impurities, thereby improving the accuracy of detection.

[0073] (2) The present invention realizes dynamic optimization and automatic adjustment of the support vector machine model by introducing meta-learning technology. Compared with the traditional manual experience-based method, this technology significantly reduces manual intervention, improves detection efficiency, and ensures the stability and consistency of the detection results.

[0074] (3) The present invention uses graph neural networks to perform cluster analysis on the classification results, which can reveal the potential correlation between impurities and effectively reduce the false alarm rate and missed detection rate in detection. This is particularly important for semiconductor material detection with high precision and high reliability requirements.

[0075] (4) The present invention utilizes the flexibility of support vector machines and meta-learning techniques to adapt to different types of semiconductor materials, including silicon-based materials, compound semiconductors, and new two-dimensional materials. By adjusting model parameters, the present invention can efficiently detect impurity characteristics in various materials, expanding its application range. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0077] Figure 1 A flowchart of a semiconductor material impurity detection method using a support vector machine proposed by the present invention;

[0078] Figure 2 The present invention provides a flowchart of optimizing and updating a support vector machine model based on meta-learning in a semiconductor material impurity detection method using a support vector machine. DETAILED DESCRIPTION

[0079] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0080] refer to Figure 1-2 , a semiconductor material impurity detection method using a support vector machine, comprising the following steps:

[0081] S1. Acquire an initial data set containing impurity characteristic data from the manufacturing process of semiconductor materials, wherein the characteristic data includes the type, quantity and distribution information of the impurities;

[0082] S2, preprocessing the initial data set by data normalization and noise removal;

[0083] S3, using persistent homology analysis to extract features from the preprocessed data, and generating topological feature vectors by identifying persistent topological features in the data;

[0084] S4, inputting the generated topological feature vector into a classifier based on a Hopfield network combined with a support vector machine, wherein the classifier based on a Hopfield network combined with a support vector machine first uses the Hopfield network to store and retrieve feature patterns, and then classifies these topological feature vectors in combination with the support vector machine to generate a classification result;

[0085] S5. Use a clustering algorithm based on graph neural network to process and analyze the classification results, reveal the potential correlation and distribution pattern between different impurities, and generate an impurity detection report;

[0086] S6. Based on the impurity detection report, meta-learning technology is used to optimize and update the support vector machine model, and the model parameters are dynamically adjusted and optimized by sharing experience between multiple tasks and batches of materials.

[0087] In this implementation, S3 specifically includes:

[0088] S31, constructing a topological space formed by a point set for the preprocessed data set;

[0089] S32, performing persistent homology analysis on the topological space, including calculating homology groups at different scales, and obtaining the Betti number β of each dimension k , where k represents the dimension of the homology group, β k represents the number of k-dimensional homology groups formed at different scales;

[0090] S33, by analyzing the Betti number β at different scales k stability with changes in scale parameters, determining persistent topological features;

[0091] S34, converting the determined persistent topological features into a topological feature vector v topo , the topological feature vector contains the Betti number of each dimension and its corresponding persistence information.

[0092] In this implementation manner, the S33 specifically includes:

[0093] S331. Under different scale parameters ∈, calculate the Betti number β corresponding to each scale k (∈), where β k (∈) represents the number of k-dimensional homology groups when the scale parameter is ∈, where k is the dimension of the homology group;

[0094] S332. Record and plot the Betti number β k (∈) The curve β changes with the scale parameter ∈ k (∈) vs ∈;

[0095] S333, for curve β k (∈) vs∈ to analyze and calculate the persistence p of each k-dimensional homology group in various scales k , persistence p k :

[0096]

[0097] Among them, w i are the weight coefficients of different homology groups at different scales, and They represent the scale parameters where the homology group first appears and disappears at the i-th scale and reflect the topological changes of the impurity characteristics at different scales. n is the number of homology groups.

[0098] S334, based on the calculated persistence p k , the following formula is used to comprehensively score each persistent topological feature S k , used to screen important topological features:

[0099]

[0100] Among them, α and β are weight parameters, max(p k ) is the maximum value among all k-dimensional homology group persistence, ∈ 消失 and ∈ 出现 They represent the disappearance and appearance scales of topological features in the scale space respectively;

[0101] S335, the comprehensive score S k Above the set threshold S 阈值 The topological features are used as persistent topological features and converted into topological feature vectors v persist , the topological feature vector contains the Betti number β of each persistent topological feature k , persistence k And the comprehensive score S k information.

[0102] In this implementation manner, the S4 specifically includes:

[0103] S41, the topological feature vector v generated from step S3 topo The input is sent to the Hopfield network for processing. The Hopfield network stores and retrieves the characteristic pattern through the improved energy function E(v). The energy function expression is:

[0104]

[0105] Where v = (v 1 ,v 2 ,,v n ) is the input topological feature vector and represents the impurity features extracted from the semiconductor material, w ij is the connection weight between feature vectors and reflects the association strength between different impurity features. i is the bias term of the feature vector and is used to adjust the importance of each impurity feature in the network, γ is the regularization parameter, and n represents the number of features processed by the model;

[0106] S42, through the Hopfield network to the input topological feature vector v topoEnergy minimization is performed, and the characteristic mode is iteratively updated using an optimization algorithm based on the gradient descent method to obtain a stable characteristic mode v stable , the characteristic pattern is used to store and retrieve characteristic information of different impurity types in semiconductor materials;

[0107] S43, stabilize the characteristic mode v stable Input into the support vector machine classifier, the support vector machine classifier uses the improved kernel function K(x,y) to map the input feature pattern to the feature space;

[0108] S44, classify the feature patterns mapped to the feature space based on the support vector machine classifier, use the Lagrange multiplier method to solve the classification hyperplane, and output the classification result C(v stable ), where C(v stable ) are classification labels for different impurity types in semiconductor materials.

[0109] In this implementation manner, the S43 specifically includes:

[0110] S431, stabilize the characteristic mode v stable Input to the support vector machine classifier as the initial input feature for classification;

[0111] S432. In the support vector machine classifier, a nonlinear adaptive kernel function is used. K adaptive ( x,y ) performs high-dimensional mapping on the input features, and the nonlinear adaptive kernel function:

[0112]

[0113] Where x and y are input feature vectors and represent the topological feature vectors extracted from different batches of semiconductor materials, σ is the width parameter of the radial basis function kernel and is used to control the similarity measure of impurity features, and λ 1 , 2 is the weight coefficient of the kernel function combination, β is the coefficient used to optimize the nonlinear relationship between the polynomial kernel and the input vector, d is the order of the polynomial kernel, and γ is the adjustment factor that controls the nonlinearity of the overall kernel function;

[0114] S433, input the mapped high-dimensional features into the optimized support vector machine classifier, and optimize the classifier by an improved objective function, wherein the objective function includes a dynamically adjusted regularization term R(α), and the optimized objective function is:

[0115]

[0116] Where maximizeW new (α) represents the optimization of the objective function Wnew (α) to maximize the discriminative ability of the classifier, α i is the Lagrange multiplier, which indicates the contribution of each impurity feature to the classifier, y i is the category label of the i-th sample, representing different impurity types in semiconductor materials, K adaptive (x i ,x j ) is a new adaptive kernel function used to capture the complex interactions of impurity features, μ is a dynamically adjusted regularization coefficient, and n represents the number of different impurity feature samples involved in training;

[0117] S434, by solving the optimized objective function W new (α), obtain the optimal classification hyperplane, and calculate the final classification decision function for the input feature pattern v stable Classify and output the classification result C(v stable ).

[0118] In this implementation manner, S5 specifically includes:

[0119] S51, sorting the classification results in step S4 to generate a classification data matrix M class , each column of the classification data matrix represents a specific impurity type, and each row is the detection result of the corresponding impurity in different batches of materials;

[0120] S52, using the clustering algorithm based on graph neural network to classify the classification data matrix M class Processing is performed to construct a graph structure G(V,E), where V represents the node set of impurity type, E represents the connection relationship between nodes, and the connection relationship is determined by the impurity feature similarity S in different batches of materials. ij OK, similarity S ij :

[0121]

[0122] in and are the stable eigenvectors of the i-th and j-th impurity types, σ s is the scale parameter for controlling the similarity calculation, and λ is the weight parameter for adjusting the influence of cross terms between feature vectors;

[0123] S53, performing spectral clustering analysis on the graph structure G(V,E), determining the potential correlation between different impurity types by analyzing the connection weights and feature vectors of the nodes in the graph, and clustering closely correlated impurity types into the same category;

[0124] S54. Generate an impurity clustering matrix M based on the results of the spectral clustering analysis cluster ;

[0125] S55, based on the impurity clustering matrix M cluster The impurity detection report includes the clustering results, quantity statistics and distribution pattern of each impurity type in the semiconductor material.

[0126] In this implementation manner, S6 specifically includes:

[0127] S61, based on the impurity detection report generated in step S5, extracting a data set containing impurity characteristic distribution information, wherein the data set includes the type, quantity and distribution pattern of impurities in different batches of materials;

[0128] S62. Use meta-learning technology to dynamically optimize the support vector machine model, build a meta-model, analyze the impurity detection tasks of different batches of materials, extract the commonalities and differences between the tasks, and optimize and adjust the global parameters of the support vector machine model based on these analysis results;

[0129] S63, obtaining a global parameter update strategy of the support vector machine model by optimizing the meta-model, wherein the global parameter update strategy is based on a comprehensive analysis of the losses of different tasks and the changes in model parameters;

[0130] S64. After obtaining the global parameter update strategy, apply it to the support vector machine model, and further adjust the kernel function parameters and regularization coefficient of the model, so as to achieve fine adjustment of the model through iterative optimization and optimize the recognition capability of complex impurity features;

[0131] S65. Verify the optimized support vector machine model, test the model using a new batch of impurity feature data sets, evaluate the generalization ability and classification accuracy of the model in new impurity detection tasks, feed the test results back to the meta-model, further optimize and adjust the model parameters, and form a dynamic and adaptive optimization cycle.

[0132] In this implementation manner, the S63 specifically includes:

[0133] S631, using the impurity feature data and classification results extracted from step S61 to construct a meta-learning task set T, where each task contains relevant classification results and model performance, for analyzing the commonalities and differences between different tasks;

[0134] S632. Based on the task features in the meta-learning task set T, design a global parameter update strategy, first calculate the gradient information of each task in is the loss function of the task, and then the global gradient is calculated by weighted nonlinear combination

[0135]

[0136] Among them, ω i represents the weight of the i-th task in the global parameter update, λ i is the nonlinear adjustment coefficient, which is used to control the nonlinear contribution of the task gradient, γ i is a nonlinear index used to adjust the influence of gradient modulus, α i is the amplification factor, which enhances the nonlinear effect of the gradient between tasks through the hyperbolic tangent function, ρ i is the exponential decay coefficient, used to balance the linear and nonlinear contributions of the gradient, β i is the index adjustment coefficient;

[0137] S633, using the calculated global gradient The global parameter θ of the support vector machine model is updated. In this process, the adaptive gradient descent method is used and the dynamically adjusted learning rate η is introduced. t :

[0138]

[0139] Among them, η 0 is the initial learning rate and determines the basic step size of parameter update, G t is the cumulative gradient square matrix and is used to smooth gradient fluctuations, δ is the smoothing term, τ is the time step bias term and is used to adjust the learning rate as the time step changes, ξ is the learning rate oscillation adjustment coefficient and is used to introduce slight oscillations, and T is the oscillation period;

[0140] S634. After the global parameter update is completed, further adjust the kernel function parameters and regularization coefficients of the support vector machine model to effectively cope with the changes in the distribution of impurity characteristics in different batches of materials. The adjustment strategy includes dynamically modifying the relevant parameters and regularization coefficients of the kernel function according to the complexity of the detection task, and optimizing the classification accuracy of the support vector machine model in complex detection tasks.

[0141] S635. Apply the optimized support vector machine model to the detection task of a new batch of materials, monitor the classification performance of the model in real time, and feed the test results back to the meta-learning model to dynamically adjust the global parameter update strategy to form a continuously optimized adaptive loop.

[0142] Embodiment 1:

[0143] Company B, a large semiconductor material manufacturing enterprise in Suzhou, Jiangsu Province, China, has deployed the semiconductor material impurity detection method using a support vector machine of the present invention. The company mainly produces high-purity silicon wafers for integrated circuit manufacturing, with an annual output of about 20 million wafers. Since the purity of silicon wafers directly affects the performance and reliability of integrated circuits, impurity detection has become a crucial link in the production process.

[0144] Company B originally used traditional physical detection methods, such as transmission electron microscopy and secondary ion mass spectrometry, to detect impurities. Although these methods have certain detection capabilities, they are complex, time-consuming, and have limited sensitivity, resulting in low detection efficiency and prone to false alarms and missed detections. In order to improve detection accuracy and efficiency, Company B decided to introduce the impurity detection method of the present invention and conducted a 6-month test and evaluation on the production line.

[0145] During the detection process, the method of the present invention realizes efficient detection of semiconductor material impurities through the following steps:

[0146] First, in the production line, the system automatically obtains initial data sets containing impurity feature data from the manufacturing process. These data sets include the type, quantity, and distribution information of various impurities in the silicon wafer. Then, the data sets are pre-processed such as normalization and noise removal to ensure the accuracy of subsequent analysis.

[0147] Then, the system uses persistent homology analysis to extract features from the preprocessed data and generate topological feature vectors. These topological feature vectors can accurately describe the distribution characteristics of impurities in the material, especially the characteristics of trace impurities. Subsequently, the system inputs these topological feature vectors into a classifier based on the Hopfield network combined with a support vector machine for classification and generates classification results. By processing the topological feature vectors, the classifier can effectively identify the types and distribution of different impurities.

[0148] After the classification results are generated, the system uses graph neural networks to process and analyze the classification results, revealing the potential correlation between different impurities and generating a detailed impurity detection report. Based on this detection report, the system uses meta-learning technology to dynamically optimize and update the support vector machine model, so that the model can adapt to the impurity characteristics of different batches of materials, further improving the accuracy and efficiency of detection.

[0149] During the six-month implementation, the system processed approximately 6 million silicon wafers, covering multiple production batches and different types of impurities. To verify the effectiveness of the system, we compared the test data before and after implementation, as follows:

[0150] Table 1 Impurity detection data comparison table

[0151]

[0152]

[0153] As shown in Table 1, after using the method of the present invention, the average detection time of silicon wafers was reduced from 12 minutes of the traditional method to 1.5 minutes, the total detection time was significantly reduced, and the overall efficiency was improved by 3.3 times. At the same time, the false alarm rate of the system was reduced from 3.5% to 0.7%, and the false alarm rate was reduced from 2.2% to 0.5%, which significantly improved the accuracy and reliability of the detection. In addition, the accuracy of the test results was increased from 94.3% to 99.1%, and the proportion of silicon wafers with a comprehensive quality score of "excellent" or "good" was increased from 89.5% to 97.8%.

[0154] Through the implementation of the present invention, Company B has significantly improved the efficiency and accuracy of impurity detection, reduced detection time and manual intervention, and significantly reduced the incidence of false positives and missed detections. This not only ensures the high purity and consistency of silicon wafers, but also improves the overall quality of the product, laying a solid foundation for the company to maintain its technological advantage in the fiercely competitive semiconductor market.

[0155] The advanced technical means adopted in the present invention not only effectively solves the shortcomings of traditional detection methods, but also realizes efficient and accurate detection of impurities in semiconductor materials through intelligent and automated processes, providing a strong guarantee for the production efficiency and product quality of enterprises.

[0156] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A semiconductor material impurity detection method using a support vector machine, characterized in that: The steps include: S1. Acquire an initial data set containing impurity characteristic data from the manufacturing process of semiconductor materials, wherein the characteristic data includes the type, quantity and distribution information of the impurities; S2, preprocessing the initial data set by data normalization and noise removal; S3, using persistent homology analysis to extract features from the preprocessed data, and generating topological feature vectors by identifying persistent topological features in the data; S4, inputting the generated topological feature vector into a classifier based on a Hopfield network combined with a support vector machine, wherein the classifier based on a Hopfield network combined with a support vector machine first uses the Hopfield network to store and retrieve feature patterns, and then classifies these topological feature vectors in combination with the support vector machine to generate a classification result; S5. Use a clustering algorithm based on graph neural network to process and analyze the classification results, reveal the potential correlation and distribution pattern between different impurities, and generate an impurity detection report; S6. Based on the impurity detection report, meta-learning technology is used to optimize and update the support vector machine model, and the model parameters are dynamically adjusted and optimized by sharing experience between multiple tasks and batches of materials.

2. A semiconductor material impurity detection method using a support vector machine according to claim 1, characterized in that: The S3 specifically includes: S31, constructing a topological space formed by a point set for the preprocessed data set; S32, performing persistent homology analysis on the topological space, including calculating homology groups at different scales, and obtaining the Betti number β of each dimension k , where k represents the dimension of the homology group, β k represents the number of k-dimensional homology groups formed at different scales; S33, by analyzing the Betti number β at different scales k stability with changes in scale parameters, determining persistent topological features; S34, converting the determined persistent topological features into a topological feature vector v topo , the topological feature vector contains the Betti number of each dimension and its corresponding persistence information.

3. A semiconductor material impurity detection method using a support vector machine according to claim 2, characterized in that: The S33 specifically includes: S331. Under different scale parameters ∈, calculate the Betti number β corresponding to each scale k (∈), where β k (∈) represents the number of k-dimensional homology groups when the scale parameter is ∈, where k is the dimension of the homology group; S332. Record and plot the Betti number β k (∈) The curve β changes with the scale parameter ∈ k (∈) vs ∈; S333, for curve β k (∈) vs∈ to analyze and calculate the persistence p of each k-dimensional homology group in various scales k , persistence p k : Among them, w i are the weight coefficients of different homology groups at different scales, and They represent the scale parameters where the homology group first appears and disappears at the i-th scale and reflect the topological changes of the impurity characteristics at different scales. n is the number of homology groups. S334, based on the calculated persistence p k , the following formula is used to comprehensively score each persistent topological feature S k , used to screen important topological features: Among them, α and β are weight parameters, max(p k ) is the maximum value among all k-dimensional homology group persistence, ∈ 消失 and ∈ 出现 They represent the disappearance and appearance scales of topological features in the scale space respectively; S335, the comprehensive score S k Above the set threshold S 阈值 The topological features are used as persistent topological features and converted into topological feature vectors v persist , the topological feature vector contains the Betti number β of each persistent topological feature k , persistence k And the comprehensive score S k information.

4. The semiconductor material impurity detection method using a support vector machine according to claim 1, characterized in that: The S4 specifically includes: S41, the topological feature vector v generated from step S3 topo The input is sent to the Hopfield network for processing. The Hopfield network stores and retrieves the characteristic pattern through the improved energy function E(v). The energy function expression is: Where v = (v1, v2,, v n ) is the input topological feature vector and represents the impurity features extracted from the semiconductor material, w ij is the connection weight between feature vectors and reflects the association strength between different impurity features. i is the bias term of the feature vector and is used to adjust the importance of each impurity feature in the network, γ is the regularization parameter, and n represents the number of features processed by the model; S42, through the Hopfield network to the input topological feature vector v topo Energy minimization is performed, and the characteristic mode is iteratively updated using an optimization algorithm based on the gradient descent method to obtain a stable characteristic mode v stable , the characteristic pattern is used to store and retrieve characteristic information of different impurity types in semiconductor materials; S43, stabilize the characteristic mode v stable Input into the support vector machine classifier, the support vector machine classifier uses the improved kernel function K(x,y) to map the input feature pattern to the feature space; S44, classify the feature patterns mapped to the feature space based on the support vector machine classifier, use the Lagrange multiplier method to solve the classification hyperplane, and output the classification result C(v stable ), where C(v stable ) are classification labels for different impurity types in semiconductor materials.

5. The semiconductor material impurity detection method using a support vector machine according to claim 4, characterized in that: The S43 specifically includes: S431, stabilize the characteristic mode v stable Input to the support vector machine classifier as the initial input feature for classification; S432, in the support vector machine classifier, a nonlinear adaptive kernel function K is used adaptive (x, y) performs high-dimensional mapping on the input features, and the nonlinear adaptive kernel function is: Among them, x and y are input feature vectors and represent the topological feature vectors extracted from different batches of semiconductor materials, σ is the width parameter of the radial basis function kernel and is used to control the similarity measure of impurity characteristics, λ1 and λ2 are the weight coefficients of the kernel function combination, β is the coefficient used to optimize the nonlinear relationship between the polynomial kernel and the input vector, d is the order of the polynomial kernel, and γ is the adjustment factor that controls the degree of nonlinearity of the overall kernel function; S433, input the mapped high-dimensional features into the optimized support vector machine classifier, and optimize the classifier by an improved objective function, wherein the objective function includes a dynamically adjusted regularization term R(α), and the optimized objective function is: Where, maximizeW new (α) represents the optimization of the objective function W new (α) to maximize the discriminative ability of the classifier, α i is the Lagrange multiplier, which indicates the contribution of each impurity feature to the classifier, y i is the category label of the i-th sample, representing different impurity types in semiconductor materials, K adaptive (x i ,x j ) is a new adaptive kernel function used to capture the complex interactions of impurity features, μ is a dynamically adjusted regularization coefficient, and n represents the number of different impurity feature samples involved in training; S434, by solving the optimized objective function W new (α), obtain the optimal classification hyperplane, and calculate the final classification decision function for the input feature pattern v stable Classify and output the classification result C(v stable ).

6. The semiconductor material impurity detection method using a support vector machine according to claim 1, characterized in that: The S5 specifically includes: S51, sorting the classification results in step S4 to generate a classification data matrix M class , each column of the classification data matrix represents a specific impurity type, and each row is the detection result of the corresponding impurity in different batches of materials; S52, using the clustering algorithm based on graph neural network to classify the classification data matrix M class Processing is performed to construct a graph structure G(V,E), where V represents the node set of impurity type, E represents the connection relationship between nodes, and the connection relationship is determined by the impurity feature similarity S in different batches of materials. ij OK, similarity S ij : in and are the stable eigenvectors of the i-th and j-th impurity types, σ s is the scale parameter for controlling the similarity calculation, and λ is the weight parameter for adjusting the influence of cross terms between feature vectors; S53, performing spectral clustering analysis on the graph structure G(V,E), determining the potential correlation between different impurity types by analyzing the connection weights and feature vectors of the nodes in the graph, and clustering closely correlated impurity types into the same category; S54. Generate an impurity clustering matrix M based on the results of the spectral clustering analysis cluster ; S55, based on the impurity clustering matrix M cluster The impurity detection report includes the clustering results, quantity statistics and distribution pattern of each impurity type in the semiconductor material.

7. The semiconductor material impurity detection method using a support vector machine according to claim 1, characterized in that: The S6 specifically includes: S61, based on the impurity detection report generated in step S5, extracting a data set containing impurity characteristic distribution information, wherein the data set includes the impurity type, quantity and distribution pattern in different batches of materials; S62. Use meta-learning technology to dynamically optimize the support vector machine model, build a meta-model, analyze the impurity detection tasks of different batches of materials, extract the commonalities and differences between the tasks, and optimize and adjust the global parameters of the support vector machine model based on these analysis results; S63, obtaining a global parameter update strategy of the support vector machine model by optimizing the meta-model, wherein the global parameter update strategy is based on a comprehensive analysis of the losses of different tasks and the changes in model parameters; S64. After obtaining the global parameter update strategy, apply it to the support vector machine model, and further adjust the kernel function parameters and regularization coefficient of the model, so as to achieve fine adjustment of the model through iterative optimization and optimize the recognition capability of complex impurity features; S65. Verify the optimized support vector machine model, test the model using a new batch of impurity feature data sets, evaluate the generalization ability and classification accuracy of the model in new impurity detection tasks, feed the test results back to the meta-model, further optimize and adjust the model parameters, and form a dynamic and adaptive optimization cycle.

8. The semiconductor material impurity detection method using a support vector machine according to claim 7, characterized in that: The S63 specifically includes: S631, using the impurity feature data and classification results extracted from step S61 to construct a meta-learning task set T, where each task contains relevant classification results and model performance, for analyzing the commonalities and differences between different tasks; S632. Based on the task features in the meta-learning task set T, design a global parameter update strategy, first calculate the gradient information of each task in is the loss function of the task, and then the global gradient is calculated by weighted nonlinear combination Among them, ω i represents the weight of the i-th task in the global parameter update, λ i is the nonlinear adjustment coefficient, which is used to control the nonlinear contribution of the task gradient, γ i is a nonlinear index used to adjust the influence of gradient modulus, α i is the amplification factor, which enhances the nonlinear effect of the gradient between tasks through the hyperbolic tangent function, ρ i is the exponential decay coefficient, used to balance the linear and nonlinear contributions of the gradient, β i is the index adjustment coefficient; S633, using the calculated global gradient The global parameter θ of the support vector machine model is updated. In this process, the adaptive gradient descent method is used and the dynamically adjusted learning rate η is introduced. t : Among them, η0 is the initial learning rate and determines the basic step size of parameter update, G t is the cumulative gradient square matrix and is used to smooth gradient fluctuations, δ is the smoothing term, τ is the time step bias term and is used to adjust the learning rate as the time step changes, ξ is the learning rate oscillation adjustment coefficient and is used to introduce slight oscillations, and T is the oscillation period; S634. After the global parameter update is completed, further adjust the kernel function parameters and regularization coefficients of the support vector machine model to effectively cope with the changes in the distribution of impurity characteristics in different batches of materials. The adjustment strategy includes dynamically modifying the relevant parameters and regularization coefficients of the kernel function according to the complexity of the detection task, and optimizing the classification accuracy of the support vector machine model in complex detection tasks. S635. Apply the optimized support vector machine model to the detection task of a new batch of materials, monitor the classification performance of the model in real time, and feed the test results back to the meta-learning model to dynamically adjust the global parameter update strategy to form a continuously optimized adaptive loop.

Images