Semiconductor defect classification and prediction method using deep learning algorithm

Abstract

The invention discloses a semiconductor defect classification and prediction method using a deep learning algorithm, and the method comprises the following steps: S1, collecting image data and manufacturing process parameters in a semiconductor manufacturing process through a multi-modal sensor, and generating a multi-dimensional data set; s2, preprocessing the data set; s3, constructing and training a deep learning model in which a convolutional neural network and a long-short-term memory network are combined, and automatically detecting and marking a defect region; s4, classifying the defects by adopting a quantum convolutional neural network in combination with a Hash embedding algorithm; s5, performing defect prediction based on historical data and real-time process parameters through a chaotic time sequence prediction network; s6, dynamically adjusting the model structure through the self-adjusting neural network, and maintaining the detection and prediction capabilities; and S7, adopting a reinforcement learning strategy to carry out adaptive optimization on the model. The method can improve the accuracy and prediction precision of semiconductor defect detection, is suitable for different manufacturing environments, and effectively reduces the defective product rate and production loss.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and in particular to a semiconductor defect classification and prediction method using a deep learning algorithm. Background Art

[0002] In the modern semiconductor manufacturing industry, as the integration of integrated circuits continues to increase, the complexity of the manufacturing process also increases, resulting in various tiny defects in the manufacturing process. These defects not only affect the performance of semiconductor devices, but may also cause the entire batch of products to be scrapped. Therefore, how to effectively detect, classify and predict these defects has become a key link in ensuring the quality of semiconductor products.

[0003] Traditional semiconductor defect detection methods mainly rely on microscopic observation and physical inspection tools, and usually require manual operation to identify defects. These methods are not only time-consuming and inefficient, but also easily affected by the skill level and subjective judgment of the operator, resulting in insufficient accuracy and consistency of the detection results. In addition, due to the large number of process parameters involved in the semiconductor manufacturing process, such as temperature, pressure, chemical composition, etc., it is difficult for traditional detection methods to simultaneously consider the impact of these complex factors on defect formation, and it is difficult to achieve comprehensive analysis and accurate prediction of defects. With the development of artificial intelligence and deep learning technology, more and more research has begun to try to apply these advanced technologies to the field of semiconductor defect detection. Image processing methods based on deep learning can automatically extract features from large amounts of image data, thereby realizing automated detection and classification of defects. However, existing deep learning technologies still have some limitations and challenges when applied to semiconductor defect detection:

[0004] The first is the complexity of feature extraction: Although traditional convolutional neural networks perform well in image processing, when dealing with semiconductor defects, due to the wide variety of defect types and complex morphologies, traditional convolutional neural networks may find it difficult to effectively extract all key defect features. In addition, in terms of the fusion of multimodal data (such as image data and process parameters), traditional deep learning methods also show certain limitations, making it difficult to fully utilize all relevant information for comprehensive analysis.

[0005] The second is the instability of classification accuracy: Existing deep learning models usually rely on large-scale labeled data for training when classifying defects. However, some defect types in semiconductor manufacturing are relatively rare, and there is a lack of sufficient labeled data to support effective training of the model. This leads to the problem of insufficient classification accuracy of the model when facing new defect types. In addition, traditional deep learning models have poor adaptability when dealing with changes in process parameters during the production process, which easily leads to instability in classification results.

[0006] The third is the limited accuracy of defect prediction: Although some studies have tried to introduce time series analysis methods into semiconductor defect prediction, most methods can only process simple time series data and have difficulty capturing complex dynamic evolution processes. Existing defect prediction methods usually rely on linear or static prediction models, which are difficult to cope with nonlinear changes and complex dynamic environments in manufacturing processes. Therefore, there is still much room for improvement in the accuracy and adaptability of existing technologies in defect prediction.

[0007] Therefore, how to provide a semiconductor defect classification and prediction method using a deep learning algorithm is an issue that needs to be urgently addressed by those skilled in the art. Summary of the invention

[0008] One object of the present invention is to propose a semiconductor defect classification and prediction method using a deep learning algorithm. By combining convolutional neural networks, long short-term memory networks, and quantum convolutional neural networks, the present invention can more accurately extract and identify various defect features generated in the semiconductor manufacturing process, such as microcracks, surface scratches, particle contamination, etc. By introducing a chaotic time series prediction network and a similarity measurement method based on fractal dimension regulation, the present invention can make high-precision predictions of defects that may occur in the future, thereby supporting preventive adjustments to the production process. Compared with traditional defect detection and classification methods, the present invention has higher detection accuracy, classification accuracy, and prediction reliability, especially when dealing with complex manufacturing environments and dynamically changing process conditions, it shows stronger adaptability and robustness.

[0009] A semiconductor defect classification and prediction method using a deep learning algorithm according to an embodiment of the present invention includes the following steps:

[0010] S1. Collect image data from the semiconductor manufacturing process through multimodal sensors, and simultaneously collect manufacturing process parameters including temperature, pressure, chemical composition concentration and optical properties to generate a multidimensional data set containing defect characteristics and process conditions;

[0011] S2, preprocessing the image data in the generated multidimensional data set;

[0012] S3. Build and train a deep learning model that combines a convolutional neural network and a long short-term memory network to extract complex defect features from the preprocessed multidimensional data set, and automatically detect and mark defect areas in the image. The detected defect information will be passed to the classification stage;

[0013] S4. Classify the detected defect information, use quantum convolutional neural network for feature extraction, and combine hash embedding algorithm to classify the newly emerged defect types;

[0014] S5. Based on the deep learning model combining the trained convolutional neural network and the long short-term memory network, the historical defect data and the real-time collected manufacturing process parameters are used to predict defects. The model parameters are dynamically adjusted through the chaotic time series prediction network to provide the probability distribution and trend analysis of future defects and support preventive adjustments to the production process.

[0015] S6, through self-adjusting neural network, dynamically adjust the network structure under different production conditions to maintain defect detection, classification and prediction capabilities;

[0016] S7. By continuously collecting new data from the production process, a reinforcement learning strategy is used to adaptively optimize the deep learning model that combines convolutional neural networks and long short-term memory networks to adapt to new types of defects.

[0017] Optionally, the S3 specifically includes:

[0018] S31. Construct and train a convolutional neural network, wherein the convolutional neural network includes multiple convolutional layers, pooling layers, and fully connected layers, wherein the convolutional layers are used to extract spatial features from the preprocessed multidimensional data set, the pooling layers are used to downsample and retain important features, and the fully connected layers are used to convert the extracted features into defect feature vectors;

[0019] S32, inputting the spatial features extracted by the convolutional neural network into the long short-term memory network, and the long short-term memory network combines the spatial features with the temporal features by capturing the time series information in the multidimensional data;

[0020] S33, combining the outputs of the convolutional neural network and the long short-term memory network to form a deep learning model combining the convolutional neural network and the long short-term memory network, wherein the deep learning model combining the convolutional neural network and the long short-term memory network generates a final comprehensive defect feature vector by fusing spatial features and temporal features;

[0021] S34, performing multi-layer full connection processing on the comprehensive defect feature vector, and generating the final defect detection result through the Softmax layer;

[0022] S35, passing the generated defect feature vector and defect detection result to the classification stage.

[0023] Optionally, the S4 specifically includes:

[0024] S41, receiving a comprehensive defect feature vector and a defect detection result output by a deep learning model combining a convolutional neural network and a long short-term memory network;

[0025] S42, inputting the comprehensive defect feature vector into the quantum convolutional neural network, the quantum convolutional neural network uses the coherence and entanglement of the multi-dimensional quantum state to perform quantum state operations on the input feature vector to generate a feature representation;

[0026] S43, optimizing feature representation based on defect detection results;

[0027] S44, inputting the optimized feature representation into a hash embedding module, the hash embedding module maps the feature into a compact embedding space through an adaptive hash embedding algorithm, and the hash function is dynamically adjusted according to the complexity of the input feature;

[0028] S45. Combining the defect detection results, the generated compact feature vector is input into the evolutionary classifier. The evolutionary classifier simulates the evolutionary process, finds the optimal classification path in the feature vector space, dynamically adjusts the classification boundary, and adapts to the newly emerged defect types.

[0029] Optionally, the S42 specifically includes:

[0030] S421, input the comprehensive defect feature vector into the quantum state initialization module, and the quantum state initialization module converts the input feature vector into a quantum state representation Φ through a quantum superposition state generation algorithm. 0 , where Φ 0 It is represented as the initial representation of the eigenvector in the quantum state space, and after the quantum superposition process, a composite state is formed, which represents the superposition and entanglement combination of multiple quantum states;

[0031] S422. In the quantum state evolution module, quantum state coupling operation and adaptive evolution algorithm are applied to Φ 0 Perform multiple iterative operations to make the quantum state evolve into an optimized intermediate quantum state Φ t , where the evolution process is defined by the following quantum state update rule:

[0032] Φ t ~ε(Φ t-1 ,Q);

[0033] Among them, Φ t represents the intermediate quantum state at time t, Φ t-1 represents the quantum state at the previous time step t-1, ε represents the quantum state evolution operation, and Q is the quantum gate operation matrix used to control the dynamic parameters in the quantum state evolution process;

[0034] S423, the intermediate quantum state Φ t Input quantum convolution layer, quantum convolution layer through quantum convolution kernel Φ t Perform multi-scale quantum convolution operations to generate quantum feature representation Φ f ;

[0035] S424, through the quantum state post-processing module, Φ f The quantum measurement operation is performed. The quantum measurement process uses an optimized multi-phase measurement mechanism, and the effect of the quantum measurement is expressed by the following relationship:

[0036] v f →Optim(Measure(Φ f ));

[0037] Among them, v f Represents the optimized feature data, Optim(Measure(Φ f )) represents the optimized quantum measurement mechanism;

[0038] Optionally, the S5 specifically includes:

[0039] S51, the defect feature vector v after classification f and the feature vector v in the historical defect data h Matching is performed and v is calculated by a similarity measurement method based on fractal dimension regulation. f With v h The similarity S between fractal (v f ,v h ), the similarity measurement method based on fractal dimension regulation uses the fractal dimension change of the feature vector to adaptively adjust the similarity measurement formula:

[0040]

[0041] in, and Denotes the feature vector v f and v h The local representation in the i-th dimension, and Represents the eigenvectors and The norm of i (v f ,v h ) represents the fractal dimension difference of the eigenvector in the i-th dimension;

[0042] S52, based on the calculated fractal similarity S fractal (v f ,v h ) and the manufacturing process parameters P collected in real time t , the defect feature vector v f The data are transferred to the chaotic time-varying network for prediction, and the network structure uses time-varying coupling mapping to model the feature evolution;

[0043] S53, the predicted defect feature evolution trend T f (t) Compare and analyze the historical defect evolution data, calculate the possible defect types and their occurrence probability, and output the probability distribution P of defect occurrence d ;

[0044] S54, according to the output defect occurrence probability distribution P d , dynamically adjust the parameter set θ of the deep learning model;

[0045] S55. Use the optimized model for real-time process adjustment and preventive maintenance, continuously monitor process parameters and defect characteristics during the manufacturing process, and dynamically update the prediction model.

[0046] Optionally, the S52 specifically includes:

[0047] S521, the defect feature vector v f and real-time collected manufacturing process parameters P t Input to the time-varying coupling mapping module to construct the time-varying coupling matrix M based on adaptive regulation adaptive (t), based on the adaptively controlled time-varying coupling matrix, the coupling relationship is dynamically adjusted by combining characteristic complexity and time evolution:

[0048] M adaptive (t)=A·exp(γ·sin(ωt))+B·log(1+|v f |·t δ );

[0049] Among them, A and B are the basic structure of the coupling matrix controlled by the ground state matrix, γ is the adaptive control coefficient for adjusting the dynamic response of the coupling strength, ω is the frequency control parameter for adjusting the time period of the characteristic interaction, δ is the exponent for controlling the time-dependent change of the coupling strength, |v f | is the norm of the eigenvector;

[0050] S522, using the time-varying coupling matrix M adaptive (t) for the defect feature vector v f Perform dynamic coupling to generate the coupled feature representation v f ′(t), in the coupling process, the influence of feature complexity and time evolution is amplified, and the obtained feature representation v f ′(t);

[0051] S523, the feature representation after adaptive coupling v f ′(t) is input into the chaotic nonlinear mapping module, and the feature vector is further processed by combining the chaotic control mechanism and nonlinear mapping to generate the predicted defect feature evolution trend T f (t+1):

[0052] T f (t+1)=v f ′(t)·C(t)·sin(α·v f ′(t)+β)+λ·tanh(N(t)·G(V f ′,P t ));

[0053] Among them, C(t) is the time-dependent chaos control matrix, α controls the amplitude of chaos, β controls the phase of chaos, λ is the nonlinear gain coefficient to adjust the intensity of nonlinear response, tanh is the hyperbolic tangent function to increase the intensity of nonlinearity, N(t) is the nonlinear mapping weight matrix and dynamically adjusts the mapping weight between features and process parameters, G(v f ′,P t ) is a complex coupling function between the feature vector and the process parameters and reflects the nonlinear relationship between the feature evolution and the process parameters.

[0054] Optionally, the S54 specifically includes:

[0055] S541, receiving the defect occurrence probability distribution P output by the chaotic time-varying network d , and based on the defect occurrence probability distribution P d And the loss function L(θ) of the current deep learning model dynamically adjusts the model's parameter set θ;

[0056] S542. During the parameter adjustment process, a dynamic weight optimization method based on information entropy difference is used to minimize the corrected loss function L r (θ), the corrected loss function L r (θ) combines the original loss function L(θ and the probability distribution of defect occurrence P d The dynamic weight entropy difference term

[0057]

[0058] in, It represents the entropy difference between the current time step t and the previous time step t-1 and reflects the dynamic change of the probability distribution of defect occurrence. λ(t) is the dynamically adjusted regularization coefficient and is used to control the weight of the entropy difference term at different time steps.

[0059] S543, based on the corrected loss function L r (θ) Calculate the gradient of the model parameters And adjust the learning rate η when applying adaptive (t) Update the parameter set θ:

[0060]

[0061] Among them, θ t+1 represents the model parameters at time t+1 after update, η adaptive (t) represents the time-varying regulated learning rate based on information entropy fluctuations and training dynamics and is used to flexibly adjust the learning rate according to the current training stage;

[0062] S544, the optimized model parameter set θ t+1 Applied to deep learning models, the model can be based on the probability distribution P of new defects d Make more accurate classifications and predictions, and dynamically update the model’s weights and structure;

[0063] S545. Use the updated deep learning model for real-time process adjustment and preventive maintenance, and continuously monitor process parameters and defect feature data during the manufacturing process.

[0064] Optionally, the S6 specifically includes:

[0065] S61. Analyze the network structure of the deep learning model, identify the key paths and bottleneck nodes that affect the model performance under different production conditions, and use the dynamic optimization algorithm based on nonlinear complexity analysis to adjust the model's hierarchy and connection mode;

[0066] S62, introduce a dynamic architecture reorganization mechanism based on quantum state feature mapping, through the quantum state interference matrix Q adaptive (t) Adaptively control the connections between different convolutional layers and pooling layers of the model:

[0067]

[0068] Among them, Q adaptive (t) represents the quantum state interference matrix at time t and is used to adjust the connection strength between different levels in the deep learning model. U is a quantum state evolution operator and is used to describe the evolution process of the quantum state. iH(t) It is the time evolution factor of the quantum state and is used to control the time dynamics of the network structure, reflecting how the network structure is adaptively adjusted to cope with changes in the production environment at different production stages. is the Hermitian conjugate of U, V is another set of weight matrices, tanh(·) is the hyperbolic tangent function, α is the control factor, and F(t) is the eigenvector at the current time used to introduce nonlinear interference effects;

[0069] S63, the quantum state interference matrix Q adaptive (t) Applied to the network structure of the deep learning model, by adjusting the quantum connection strength between network levels, the defect detection, classification and prediction capabilities of the model can be maintained under different production conditions;

[0070] S64, Combine real-time data from the production process to dynamically adjust the quantum state interference matrix Q adaptive (t) configuration, where dynamic adjustments include fine-tuning of the model's internal weights, connection strengths, and feature maps;

[0071] S65. By introducing a continuous feedback mechanism based on entropy measurement, the prediction accuracy and classification performance of the model are monitored in real time, and the quantum state interference matrix Q is adjusted according to the feedback results. adaptive (t) Make further adjustments.

[0072] The beneficial effects of the present invention are:

[0073] (1) Improve the accuracy and reliability of defect detection. By combining convolutional neural networks and long short-term memory networks, the present invention can effectively extract and identify complex defect features in the semiconductor manufacturing process, such as tiny cracks, surface scratches, and particle contamination. Compared with traditional rule-based image processing methods, deep learning models can process multimodal data more accurately and capture complex spatial and temporal features, thereby greatly improving the accuracy and stability of defect detection.

[0074] (2) Enhance the adaptability and accuracy of defect classification. The present invention optimizes the classification ability of newly emerging defect types by introducing quantum convolutional neural networks and hash embedding algorithms. Quantum convolutional neural networks use the coherence and entanglement of multidimensional quantum states to deeply extract input defect features, and map the features to a compact embedding space through a hash embedding algorithm, thereby achieving accurate classification of rare or newly emerging defects. This method enhances the adaptability of the model in dealing with different production conditions and ensures high accuracy and consistency of classification results.

[0075] (3) Improve the accuracy and practicality of defect prediction. The present invention can effectively capture the complex dynamic changes in the manufacturing process and make high-precision predictions of defects that may occur in the future through the innovative design based on the chaotic time series prediction network. Combined with the similarity measurement method of fractal dimension regulation, the present invention can perform more accurate similarity analysis on defect characteristics, thereby providing a scientific basis for preventive adjustment of production processes, reducing defective product rates, and improving the efficiency and stability of production lines. 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 defect classification and prediction method using a deep learning algorithm proposed by the present invention;

[0078] Figure 2 A semiconductor defect classification flow chart based on quantum convolutional neural network in a semiconductor defect classification and prediction method using a deep learning algorithm proposed by the present invention; 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 defect classification and prediction method using a deep learning algorithm, comprising the following steps:

[0081] S1. Collect image data from the semiconductor manufacturing process through multimodal sensors, and simultaneously collect manufacturing process parameters including temperature, pressure, chemical composition concentration and optical properties to generate a multidimensional data set containing defect characteristics and process conditions;

[0082] S2, preprocessing the image data in the generated multidimensional data set;

[0083] S3. Build and train a deep learning model that combines a convolutional neural network and a long short-term memory network to extract complex defect features from the preprocessed multidimensional data set, and automatically detect and mark defect areas in the image. The detected defect information will be passed to the classification stage;

[0084] S4. Classify the detected defect information, use quantum convolutional neural network for feature extraction, and combine hash embedding algorithm to classify the newly emerged defect types;

[0085] S5. Based on the deep learning model combining the trained convolutional neural network and the long short-term memory network, the historical defect data and the real-time collected manufacturing process parameters are used to predict defects. The model parameters are dynamically adjusted through the chaotic time series prediction network to provide the probability distribution and trend analysis of future defects and support preventive adjustments to the production process.

[0086] S6, through self-adjusting neural network, dynamically adjust the network structure under different production conditions to maintain defect detection, classification and prediction capabilities;

[0087] S7. By continuously collecting new data from the production process, a reinforcement learning strategy is used to adaptively optimize the deep learning model that combines convolutional neural networks and long short-term memory networks to adapt to new types of defects.

[0088] In this implementation, S3 specifically includes:

[0089] S31. Construct and train a convolutional neural network, wherein the convolutional neural network includes multiple convolutional layers, pooling layers, and fully connected layers, wherein the convolutional layers are used to extract spatial features from the preprocessed multidimensional data set, the pooling layers are used to downsample and retain important features, and the fully connected layers are used to convert the extracted features into defect feature vectors;

[0090] S32, inputting the spatial features extracted by the convolutional neural network into the long short-term memory network, and the long short-term memory network combines the spatial features with the temporal features by capturing the time series information in the multidimensional data;

[0091] S33, combining the outputs of the convolutional neural network and the long short-term memory network to form a deep learning model combining the convolutional neural network and the long short-term memory network, wherein the deep learning model combining the convolutional neural network and the long short-term memory network generates a final comprehensive defect feature vector by fusing spatial features and temporal features;

[0092] S34, performing multi-layer full connection processing on the comprehensive defect feature vector, and generating the final defect detection result through the Softmax layer;

[0093] S35, passing the generated defect feature vector and defect detection result to the classification stage.

[0094] In this implementation, S4 specifically includes:

[0095] S41, receiving a comprehensive defect feature vector and a defect detection result output by a deep learning model combining a convolutional neural network and a long short-term memory network;

[0096] S42, inputting the comprehensive defect feature vector into the quantum convolutional neural network, the quantum convolutional neural network uses the coherence and entanglement of the multi-dimensional quantum state to perform quantum state operations on the input feature vector to generate a feature representation;

[0097] S43, optimizing feature representation based on defect detection results;

[0098] S44, inputting the optimized feature representation into a hash embedding module, the hash embedding module maps the feature into a compact embedding space through an adaptive hash embedding algorithm, and the hash function is dynamically adjusted according to the complexity of the input feature;

[0099] S45. Combining the defect detection results, the generated compact feature vector is input into the evolutionary classifier. The evolutionary classifier simulates the evolutionary process, finds the optimal classification path in the feature vector space, dynamically adjusts the classification boundary, and adapts to the newly emerged defect types.

[0100] In this implementation manner, the S42 specifically includes:

[0101] S421, input the comprehensive defect feature vector into the quantum state initialization module, and the quantum state initialization module converts the input feature vector into a quantum state representation Φ through a quantum superposition state generation algorithm. 0 , where Φ 0 It is represented as the initial representation of the eigenvector in the quantum state space, and after the quantum superposition process, a composite state is formed, which represents the superposition and entanglement combination of multiple quantum states;

[0102] S422. In the quantum state evolution module, quantum state coupling operation and adaptive evolution algorithm are applied to Φ 0 Perform multiple iterative operations to make the quantum state evolve into an optimized intermediate quantum state Φ t , where the evolution process is defined by the following quantum state update rule:

[0103] Φ t ~ε(Φ t-1 ,Q);

[0104] Among them, Φ t represents the intermediate quantum state at time t, Φ t-1 represents the quantum state at the previous time step t-1, ε represents the quantum state evolution operation, and Q is the quantum gate operation matrix used to control the dynamic parameters in the quantum state evolution process;

[0105] S423, the intermediate quantum state Φ t Input quantum convolution layer, quantum convolution layer through quantum convolution kernel Φ t Perform multi-scale quantum convolution operations to generate quantum feature representation Φ f ;

[0106] S424, through the quantum state post-processing module, Φ f The quantum measurement operation is performed. The quantum measurement process uses an optimized multi-phase measurement mechanism, and the effect of the quantum measurement is expressed by the following relationship:

[0107] v f →Optim(Measure(Φ f ));

[0108] Among them, v f Represents the optimized feature data, Optim(Measure(Φ f )) represents the optimized quantum measurement mechanism;

[0109] In this implementation manner, S5 specifically includes:

[0110] S51, the defect feature vector v after classification f and the feature vector v in the historical defect data hMatching is performed and v is calculated by a similarity measurement method based on fractal dimension regulation. f With v h The similarity S between fractal (v f ,v h ), the similarity measurement method based on fractal dimension regulation uses the fractal dimension change of the feature vector to adaptively adjust the similarity measurement formula:

[0111]

[0112] in, and Denotes the feature vector v f and v h The local representation in the i-th dimension, and Represents the eigenvectors and The norm of i (v f ,v h ) represents the fractal dimension difference of the eigenvector in the i-th dimension;

[0113] S52, based on the calculated fractal similarity S fractal (v f ,v h ) and the manufacturing process parameters P collected in real time t , the defect feature vector v f The data are transferred to the chaotic time-varying network for prediction, and the network structure uses time-varying coupling mapping to model the feature evolution;

[0114] S53, the predicted defect feature evolution trend T f (t) Compare and analyze the historical defect evolution data, calculate the possible defect types and their occurrence probability, and output the probability distribution P of defect occurrence d ;

[0115] S54, according to the output defect occurrence probability distribution P d , dynamically adjust the parameter set θ of the deep learning model;

[0116] S55. Use the optimized model for real-time process adjustment and preventive maintenance, continuously monitor process parameters and defect characteristics during the manufacturing process, and dynamically update the prediction model.

[0117] In this implementation manner, the S52 specifically includes:

[0118] S521, the defect feature vector v f and real-time collected manufacturing process parameters P tInput to the time-varying coupling mapping module to construct the time-varying coupling matrix M based on adaptive regulation adaptive (t), based on the adaptively controlled time-varying coupling matrix, the coupling relationship is dynamically adjusted by combining characteristic complexity and time evolution:

[0119] M adaptive (t)=A·exp(γ·sin(ωt))+B·log(1+|v f |·t δ );

[0120] Among them, A and B are the basic structure of the coupling matrix controlled by the ground state matrix, γ is the adaptive control coefficient for adjusting the dynamic response of the coupling strength, ω is the frequency control parameter for adjusting the time period of the characteristic interaction, δ is the exponent for controlling the time-dependent change of the coupling strength, |v f | is the norm of the eigenvector;

[0121] S522, using the time-varying coupling matrix M adaptive (t) for the defect feature vector v f Perform dynamic coupling to generate the coupled feature representation v f '(t), in the coupling process, the influence of feature complexity and time evolution is amplified, and the obtained feature representation v f '(t);

[0122] S523, the feature representation after adaptive coupling v f '(t) is input into the chaotic nonlinear mapping module, and the feature vector is further processed by combining the chaotic control mechanism and nonlinear mapping to generate the predicted defect feature evolution trend T f (t+1):

[0123] T f (t+1)=v f ′(t)·C(t)·sin(α·v f ′(t)+β)+λ·tanh(N(t)·G(v f ′,P t ));

[0124] Among them, C(t) is the time-dependent chaos control matrix, α controls the amplitude of chaos, β controls the phase of chaos, λ is the nonlinear gain coefficient to adjust the intensity of nonlinear response, tanh is the hyperbolic tangent function to increase the intensity of nonlinearity, N(t) is the nonlinear mapping weight matrix and dynamically adjusts the mapping weight between features and process parameters, G(v f ',P t ) is a complex coupling function between the feature vector and the process parameters and reflects the nonlinear relationship between the feature evolution and the process parameters.

[0125] In this implementation manner, the S54 specifically includes:

[0126] S541, receiving the defect occurrence probability distribution P output by the chaotic time-varying network d , and based on the defect occurrence probability distribution P d And the loss function L(θ) of the current deep learning model dynamically adjusts the model's parameter set θ;

[0127] S542. During the parameter adjustment process, a dynamic weight optimization method based on information entropy difference is used to minimize the corrected loss function L r (θ), the corrected loss function L r (θ) combines the original loss function L(θ and the probability distribution of defect occurrence P d The dynamic weight entropy difference term

[0128]

[0129] in, It represents the entropy difference between the current time step t and the previous time step t-1 and reflects the dynamic change of the probability distribution of defect occurrence. λ(t) is the dynamically adjusted regularization coefficient and is used to control the weight of the entropy difference term at different time steps.

[0130] S543, based on the corrected loss function L r (θ) Calculate the gradient of the model parameters And adjust the learning rate η when applying adaptive (t) Update the parameter set θ:

[0131]

[0132] Among them, θ t+1 represents the model parameters at time t+1 after update, η adaptive (t) represents the time-varying regulated learning rate based on information entropy fluctuations and training dynamics and is used to flexibly adjust the learning rate according to the current training stage;

[0133] S544, the optimized model parameter set θ t+1 Applied to deep learning models, the model can be based on the probability distribution P of new defects d Make more accurate classifications and predictions, and dynamically update the model’s weights and structure;

[0134] S545. Use the updated deep learning model for real-time process adjustment and preventive maintenance, and continuously monitor process parameters and defect feature data during the manufacturing process.

[0135] In this implementation manner, S6 specifically includes:

[0136] S61. Analyze the network structure of the deep learning model, identify the key paths and bottleneck nodes that affect the model performance under different production conditions, and use the dynamic optimization algorithm based on nonlinear complexity analysis to adjust the model's hierarchy and connection mode;

[0137] S62, introduce a dynamic architecture reorganization mechanism based on quantum state feature mapping, through the quantum state interference matrix Q adaptive (t) Adaptively control the connections between different convolutional layers and pooling layers of the model:

[0138]

[0139] Among them, Q adaptive (t) represents the quantum state interference matrix at time t and is used to adjust the connection strength between different levels in the deep learning model. U is a quantum state evolution operator and is used to describe the evolution process of the quantum state. iH(t) It is the time evolution factor of the quantum state and is used to control the time dynamics of the network structure, reflecting how the network structure is adaptively adjusted to cope with changes in the production environment at different production stages. is the Hermitian conjugate of U, V is another set of weight matrices, tanh(·) is the hyperbolic tangent function, α is the control factor, and F(t) is the eigenvector at the current time used to introduce nonlinear interference effects;

[0140] S63, the quantum state interference matrix Q adaptive (t) Applied to the network structure of the deep learning model, by adjusting the quantum connection strength between network levels, the defect detection, classification and prediction capabilities of the model can be maintained under different production conditions;

[0141] S64, Combine real-time data from the production process to dynamically adjust the quantum state interference matrix Q adaptive (t) configuration, where dynamic adjustments include fine-tuning of the model's internal weights, connection strengths, and feature maps;

[0142] S65. By introducing a continuous feedback mechanism based on entropy measurement, the prediction accuracy and classification performance of the model are monitored in real time, and the quantum state interference matrix Q is adjusted according to the feedback results. adaptive (t) Make further adjustments.

[0143] Embodiment 1:

[0144] The semiconductor defect classification and prediction system using deep learning algorithm of the present invention is deployed in Company B, a world-leading semiconductor manufacturing company located in Wuxi City, Jiangsu Province, China. Company B focuses on the production of high-end semiconductor devices, with an annual output of more than 5 million wafers, which are mainly used in 5G communications, artificial intelligence, high-performance computing and other fields. Since these application scenarios have extremely high quality requirements for semiconductor devices, any minor manufacturing defects may cause device failure, resulting in significant losses.

[0145] Traditionally, Company B has used a combination of microscopic observation and physical measurement to perform defect detection. These methods are not only time-consuming and cumbersome to operate, but are also easily affected by human factors, resulting in instability in the detection results. In order to solve these problems and improve detection efficiency and accuracy, Company B decided to introduce the deep learning algorithm system of the present invention.

[0146] During the implementation of the system, Company B integrated multimodal sensors into the production line to collect image data and key process parameters (such as temperature, pressure, optical properties, etc.) during wafer manufacturing. These data are transmitted in real time to the deep learning algorithm platform, which processes the data using a deep learning model that combines convolutional neural networks and long short-term memory networks. Through this model, the system can automatically identify and mark various defects on the surface of the wafer, and pass this defect information to the quantum convolutional neural network for classification.

[0147] In actual application, the system significantly improved the detection accuracy and classification accuracy of wafer defects at Company B. In order to verify the effectiveness of the system, Company B conducted comprehensive quality inspections on different types of wafers produced on the production line within 6 months. The data showed that during this period, about 3 million wafers were processed, involving more than 100 process parameter changes, and the defect types covered scratches, particles, pits and other forms. In order to evaluate the performance of the system, we conducted a comparative analysis with traditional methods, and the results are shown in the table below.

[0148] Table 1: Wafer defect detection and classification data comparison table

[0149]

[0150] As shown in Table 1 above, the deep learning system of the present invention significantly surpasses traditional detection methods in many aspects. Specifically, the average detection time was shortened from 5 minutes to 45 seconds, the total detection time was reduced from 1.5 million minutes to 225,000 minutes, and the detection efficiency was improved by about 6.67 times. The system's false alarm rate was reduced from 3.5% to 0.5%, and the false alarm rate was reduced from 4.8% to 1.2%, which shows that the system performs well in accurately identifying and classifying defects. In addition, the proportion of wafers with a quality rating of "excellent" increased from 82% to 93.5%, which directly reflects the effectiveness of the system in improving product quality.

[0151] The system of the present invention also has efficient defect prediction capabilities, with the prediction accuracy increased from 65% to 91%. This means that the system can not only detect current defects in a timely manner, but also predict potential defect trends in advance, thereby helping Company B to make preventive adjustments during the production process and reduce the defective product rate.

[0152] In addition, the introduction of the system has greatly reduced process adjustment time and production line downtime. Through real-time monitoring and prediction systems, Company B is able to respond more quickly to abnormal situations in production, reducing process adjustment time by 30% and production line downtime by 25%, significantly improving equipment utilization and the overall efficiency of the production line.

[0153] In summary, the semiconductor defect classification and prediction system using deep learning algorithms in the implementation process of Company B has greatly improved the detection efficiency and accuracy, reduced manual intervention, and improved the response speed and product quality of the production line. Through the application of the system, Company B not only reduced operating costs, but also provided strong support for its competitiveness in the global market.

[0154] 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 defect classification and prediction method using a deep learning algorithm, characterized in that: The following steps are involved: S1. Collect image data from the semiconductor manufacturing process through multimodal sensors, and simultaneously collect manufacturing process parameters including temperature, pressure, chemical composition concentration and optical properties to generate a multidimensional data set containing defect characteristics and process conditions; S2, preprocessing the image data in the generated multidimensional data set; S3. Build and train a deep learning model that combines a convolutional neural network and a long short-term memory network to extract complex defect features from the preprocessed multidimensional data set, and automatically detect and mark defect areas in the image. The detected defect information will be passed to the classification stage; S4. Classify the detected defect information, use quantum convolutional neural network for feature extraction, and combine hash embedding algorithm to classify the newly emerged defect types; S5. Based on the deep learning model combining the trained convolutional neural network and the long short-term memory network, the historical defect data and the real-time collected manufacturing process parameters are used to predict defects. The model parameters are dynamically adjusted through the chaotic time series prediction network to provide the probability distribution and trend analysis of future defects and support preventive adjustments to the production process. S6, through self-adjusting neural network, dynamically adjust the network structure under different production conditions to maintain defect detection, classification and prediction capabilities; S7. By continuously collecting new data from the production process, a reinforcement learning strategy is used to adaptively optimize the deep learning model that combines convolutional neural networks and long short-term memory networks to adapt to new types of defects.

2. The semiconductor defect classification and prediction method using a deep learning algorithm according to claim 1, characterized in that: The S3 specifically includes: S31. Construct and train a convolutional neural network, wherein the convolutional neural network includes multiple convolutional layers, pooling layers, and fully connected layers, wherein the convolutional layers are used to extract spatial features from the preprocessed multidimensional data set, the pooling layers are used to downsample and retain important features, and the fully connected layers are used to convert the extracted features into defect feature vectors; S32, inputting the spatial features extracted by the convolutional neural network into the long short-term memory network, and the long short-term memory network combines the spatial features with the temporal features by capturing the time series information in the multidimensional data; S33, combining the outputs of the convolutional neural network and the long short-term memory network to form a deep learning model combining the convolutional neural network and the long short-term memory network, wherein the deep learning model combining the convolutional neural network and the long short-term memory network generates a final comprehensive defect feature vector by fusing spatial features and temporal features; S34, performing multi-layer full connection processing on the comprehensive defect feature vector, and generating the final defect detection result through the Softmax layer; S35, passing the generated defect feature vector and defect detection result to the classification stage.

3. The semiconductor defect classification and prediction method using a deep learning algorithm according to claim 1, characterized in that: The S4 specifically includes: S41, receiving a comprehensive defect feature vector and a defect detection result output by a deep learning model combining a convolutional neural network and a long short-term memory network; S42, inputting the comprehensive defect feature vector into the quantum convolutional neural network, the quantum convolutional neural network uses the coherence and entanglement of the multi-dimensional quantum state to perform quantum state operations on the input feature vector to generate a feature representation; S43, optimizing feature representation based on defect detection results; S44, inputting the optimized feature representation into a hash embedding module, the hash embedding module maps the feature into a compact embedding space through an adaptive hash embedding algorithm, and the hash function is dynamically adjusted according to the complexity of the input feature; S45. Combined with the defect detection results, the generated compact feature vector is input into the evolutionary classifier. The evolutionary classifier simulates the evolutionary process, finds the optimal classification path in the feature vector space, dynamically adjusts the classification boundary, and adapts to the newly emerging defect types.

4. The semiconductor defect classification and prediction method using a deep learning algorithm according to claim 3, characterized in that: The S42 specifically includes: S421, inputting the comprehensive defect feature vector into the quantum state initialization module, the quantum state initialization module converts the input feature vector into a quantum state representation Φ0 through a quantum superposition state generation algorithm, wherein Φ0 represents the initial representation of the feature vector in the quantum state space, and forms a composite state after the quantum superposition process, and the composite state represents the superposition and entanglement combination of multiple quantum states; S422, in the quantum state evolution module, apply quantum state coupling operation and adaptive evolution algorithm to perform multiple iterations on Φ0 to make the quantum state evolve into an optimized intermediate quantum state Φ t , where the evolution process is defined by the following quantum state update rule: F t ~ε(Φ t-1 ,Q); Among them, Φ t represents the intermediate quantum state at time t, Φ t-1 represents the quantum state at the previous time step t-1, ε represents the quantum state evolution operation, and Q is the quantum gate operation matrix used to control the dynamic parameters in the quantum state evolution process; S423, the intermediate quantum state Φ t Input quantum convolution layer, quantum convolution layer through quantum convolution kernel Φ t Perform multi-scale quantum convolution operations to generate quantum feature representation Φ f ; S424, through the quantum state post-processing module, Φ f The quantum measurement operation is performed. The quantum measurement process uses an optimized multi-phase measurement mechanism, and the effect of the quantum measurement is expressed by the following relationship: v f →Optim(Measure(Φ f )); Among them, v f Represents the optimized feature data, Optim(Measure(Φ f )) represents the optimized quantum measurement mechanism.

5. The semiconductor defect classification and prediction method using a deep learning algorithm according to claim 1, characterized in that: The S5 specifically includes: S51, the defect feature vector v after classification f and the feature vector v in the historical defect data h Matching is performed and v is calculated by a similarity measurement method based on fractal dimension regulation. f With v h The similarity S between fractal (v f ,v h ), the similarity measurement method based on fractal dimension regulation uses the fractal dimension change of the feature vector to adaptively adjust the similarity measurement formula: in, and Denotes the feature vector v f and v h The local representation in the i-th dimension, and Represents the eigenvectors and The norm of i (v f ,v h ) represents the fractal dimension difference of the eigenvector in the i-th dimension; S52, based on the calculated fractal similarity S fractal (v f ,v h ) and the manufacturing process parameters P collected in real time t , the defect feature vector v f The data are transferred to the chaotic time-varying network for prediction, and the network structure uses time-varying coupling mapping to model the feature evolution; S53, the predicted defect feature evolution trend T f (t) Compare and analyze the historical defect evolution data, calculate the possible defect types and their occurrence probability, and output the probability distribution P of the defect occurrence d ; S54, according to the output defect occurrence probability distribution P d , dynamically adjust the parameter set θ of the deep learning model; S55. Use the optimized model for real-time process adjustment and preventive maintenance, continuously monitor process parameters and defect characteristics during the manufacturing process, and dynamically update the prediction model.

6. The semiconductor defect classification and prediction method using a deep learning algorithm according to claim 5, characterized in that: The S52 specifically includes: S521, the defect feature vector v f and real-time collected manufacturing process parameters P t Input to the time-varying coupling mapping module to construct the time-varying coupling matrix M based on adaptive regulation adaptive (t), based on the adaptively controlled time-varying coupling matrix, the coupling relationship is dynamically adjusted by combining characteristic complexity and time evolution: M adaptive (t)=A·exp(γ·sin(ωt))+B·log(1+|v f |·t δ ); Among them, A and B are the basic structure of the base state matrix controlling the coupling matrix, γ is the adaptive control coefficient for adjusting the dynamic response of the coupling strength, ω is the frequency control parameter for adjusting the time period of the characteristic interaction, 6 is the exponent for controlling the time-dependent change of the coupling strength, |v f | is the norm of the eigenvector; S522, using the time-varying coupling matrix M adaptive (t) for the defect feature vector v f Perform dynamic coupling to generate the coupled feature representation v f ′(t), in the coupling process, the influence of feature complexity and time evolution is amplified, and the obtained feature representation v f ′(t); S523, the feature representation after adaptive coupling v f ′(t) is input into the chaotic nonlinear mapping module, and the feature vector is further processed by combining the chaotic control mechanism and nonlinear mapping to generate the predicted defect feature evolution trend Tf(t+1): T f (t+1)=V f ′(t)·C(t)·sin(α·v f ′(t)+β)+λ·tanh(N(t)·G(v f ′,P t )); Among them, C(t) is the time-dependent chaos control matrix, α controls the amplitude of chaos, β controls the phase of chaos, λ is the nonlinear gain coefficient to adjust the intensity of nonlinear response, tanh is the hyperbolic tangent function to increase the intensity of nonlinearity, N(t) is the nonlinear mapping weight matrix and dynamically adjusts the mapping weight between features and process parameters, G(v f ′,P t ) is a complex coupling function between the feature vector and the process parameters and reflects the nonlinear relationship between the feature evolution and the process parameters.

7. The semiconductor defect classification and prediction method using a deep learning algorithm according to claim 5, characterized in that: The S54 specifically includes: S541, receiving the defect occurrence probability distribution P output by the chaotic time-varying network d , and based on the defect occurrence probability distribution P d And the loss function L(θ) of the current deep learning model dynamically adjusts the model's parameter set θ; S542. During the parameter adjustment process, a dynamic weight optimization method based on information entropy difference is used to minimize the corrected loss function L r (θ), the modified loss function L r (θ) combines the original loss function L(θ) and the defect probability distribution P d The dynamic weight entropy difference term in, It represents the entropy difference between the current time step t and the previous time step t-1 and reflects the dynamic change of the probability distribution of defect occurrence. λ(t) is the dynamically adjusted regularization coefficient and is used to control the weight of the entropy difference term at different time steps. S543, based on the corrected loss function L r (θ) Calculate the gradient of the model parameters And adjust the learning rate η when applying adaptive (t) Update the parameter set θ: Among them, θ t+1 represents the model parameters at time t+1 after update, η adaptive (t) represents the time-varying regulated learning rate based on information entropy fluctuations and training dynamics and is used to flexibly adjust the learning rate according to the current training stage; S544, the optimized model parameter set θ t+1 Applied to deep learning models, the model can be based on the probability distribution P of new defects d Make more accurate classifications and predictions, and dynamically update the model’s weights and structure; S545. Use the updated deep learning model for real-time process adjustment and preventive maintenance, and continuously monitor process parameters and defect feature data during the manufacturing process.

8. The semiconductor defect classification and prediction method using a deep learning algorithm according to claim 1, characterized in that: The S6 specifically includes: S61. Analyze the network structure of the deep learning model, identify the key paths and bottleneck nodes that affect the model performance under different production conditions, and use the dynamic optimization algorithm based on nonlinear complexity analysis to adjust the model's hierarchy and connection mode; S62, introduce a dynamic architecture reorganization mechanism based on quantum state feature mapping, through the quantum state interference matrix Q adaptive (t) Adaptively control the connections between different convolutional layers and pooling layers of the model: Among them, Q adaptive (t) represents the quantum state interference matrix at time t and is used to adjust the connection strength between different levels in the deep learning model. U is a quantum state evolution operator and is used to describe the evolution process of the quantum state. iH(t) It is the time evolution factor of the quantum state and is used to control the time dynamics of the network structure, reflecting how the network structure is adaptively adjusted to cope with changes in the production environment at different production stages. is the Hermitian conjugate of U, V is another set of weight matrices, tanh(·) is the hyperbolic tangent function, α is the control factor, and F(t) is the eigenvector at the current time used to introduce nonlinear interference effects; S63, the quantum state interference matrix Q adaptive (t) Applied to the network structure of the deep learning model, by adjusting the quantum connection strength between network levels, the defect detection, classification and prediction capabilities of the model can be maintained under different production conditions; S64, Combine real-time data from the production process to dynamically adjust the quantum state interference matrix Q adaptive (t) configuration, where dynamic adjustments include fine-tuning of the model’s internal weights, connection strengths, and feature maps; S65. By introducing a continuous feedback mechanism based on entropy measurement, the prediction accuracy and classification performance of the model are monitored in real time, and the quantum state interference matrix Q is adjusted according to the feedback results. adaptive (t) Make further adjustments.

Images