Full-layout defect rapid detection system based on pattern matching and classification

Abstract

The invention provides a full-layout defect rapid detection system based on pattern matching and classification, which relates to the technical field of semiconductor manufacturing and comprises a data acquisition and multi-source fusion module, a self-adaptive preprocessing module, a dynamic pattern matching module, a multi-source feature fusion module and a closed-loop optimization and output module, the dynamic mode matching module dynamically generates a defect template from real-time data through a clustering algorithm (such as DBSCAN) and a generative adversarial network (GAN), the speed and accuracy of defect detection are remarkably improved in combination with reinforcement learning filtering and region focusing technologies of the self-adaptive preprocessing module, a detection strategy can be adaptively adjusted according to the real-time data, and the defect detection accuracy is improved. According to the method, a high-risk area is preferentially matched, and GPU parallel computing acceleration processing is performed, so that the detection time is shortened to 50% or below of that of a traditional method, deformation or fuzzy defects can be effectively identified, the false alarm rate is reduced by about 30%, and high-quality input is provided for subsequent matching and classification.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and in particular to a full-layout defect rapid detection system based on pattern matching and classification. Background Art

[0002] According to a method and device for determining defects in scanning electron microscope images disclosed in Chinese Publication No. "CN116993669A," which relates to the field of semiconductor technology, the method includes: determining, based on a design layout, a first graphic detector and a first detection position corresponding to the first graphic detector in the design layout; determining, based on the design layout, the first graphic detector, and the first detection position corresponding to the first graphic detector, a second graphic detector and a second graphic detector corresponding to the second graphic detector in the design layout; and detecting a scanning electron microscope image corresponding to the design layout based on the first graphic detector and the corresponding first detection position, and the second graphic detector and the corresponding second detection position, to determine defects in the scanning electron microscope image. This makes the determined first and second graphic detectors more comprehensive and reliable, thereby improving the accuracy and reliability of scanning electron microscope image defect detection.

[0003] The above patent documents and prior art have the following technical problems when used:

[0004] Problem 1: In semiconductor manufacturing, the morphology of wafer defects varies greatly depending on process conditions and equipment status. Traditional defect detection systems rely on predefined static template libraries and use image matching technology to identify defects. However, as process nodes shrink, defects become smaller and more complex. Static templates cannot fully cover all possible defect types. Dynamic adjustments to process parameters can also cause defect characteristics to deviate from the template, further reducing the matching success rate.

[0005] Second, traditional defect classification methods typically use machine learning or rule matching to classify defects based on single image features. However, wafer images are often subject to noise, resulting in inaccurate feature extraction. Furthermore, unknown defect types may occur in semiconductor manufacturing, and traditional methods lack adaptability and are unable to identify them, forcing them to rely on manual analysis, which increases the workload.

[0006] Problem three: Traditional defect detection systems typically only provide test results without forming a feedback mechanism with the manufacturing process. The root causes of the defects are not effectively analyzed and corrected, resulting in the same defects recurring in subsequent production, limiting yield improvements. Engineers need to manually analyze test data and adjust processes, resulting in a delayed and inefficient process. Summary of the Invention

[0007] Technical problems solved

[0008] In response to the shortcomings of the existing technology, the present invention provides a full-layout defect rapid detection system based on pattern matching and classification, which solves the following problems:

[0009] 1. Static template matching is difficult to adapt to the diversity of defect morphology and process changes, resulting in low detection efficiency and high missed detection rate;

[0010] 2. Defect classification relies solely on image features, is easily affected by noise, has low classification accuracy, and cannot identify unknown defects;

[0011] 3. The lack of closed-loop feedback for testing and process optimization leads to repeated defects and an inability to effectively improve yield.

[0012] Technical Solution

[0013] To achieve the above objectives, the present invention is implemented through the following technical solutions: a full-layout defect rapid detection system based on pattern matching and classification, the system comprising a data acquisition and multi-source fusion module, an adaptive preprocessing module, a dynamic pattern matching module, a multi-feature fusion module, and a closed-loop optimization and output module, wherein:

[0014] The data acquisition and multi-source fusion module collects wafer full-layout image data and fuses process parameter data during the manufacturing process to generate a multi-source fusion data package containing image and process context information, wherein the process parameter data includes etching time and deposition thickness;

[0015] The adaptive preprocessing module is connected to the data acquisition and multi-source fusion module, adaptively selects filtering parameters based on the reinforcement learning algorithm to enhance defect characteristics, and predicts potential defect-prone areas through a lightweight neural network based on historical inspection data, and outputs optimized image data;

[0016] The dynamic pattern matching module is connected to the adaptive preprocessing module, dynamically generates defect pattern templates from real-time data through a clustering algorithm, and combines fast Fourier transform and generative adversarial network to perform defect matching, preferentially matching the potential defect high-incidence areas, and outputting the coordinates of suspected defects and matching confidence levels;

[0017] The multi-dimensional feature fusion classification module is connected to the dynamic pattern matching module to extract the image features, physical parameter features, and process background features of suspected defects, construct a multi-dimensional feature vector, and classify the defects using a Transformer-based cross-modal learning model. At the same time, it identifies unknown defect types through unsupervised learning and outputs defect classification labels and impact assessments.

[0018] The closed-loop optimization and output module is connected to the multi-dimensional feature fusion classification module, generates process adjustment suggestions based on the defect classification results through the inverse mapping algorithm, forms a closed-loop control of detection and process optimization, and outputs an interactive defect analysis report and self-optimization model parameters.

[0019] Preferably, the data acquisition and multi-source fusion module includes a multi-modal imaging unit, which collects multi-depth image data of the entire wafer layout through optical microscopy, infrared imaging and X-ray scanning, and aligns the multi-depth image data with the process parameter data through a time synchronization mechanism to generate a multi-source fusion data packet.

[0020] Preferably, the reinforcement learning algorithm in the adaptive preprocessing module adaptively adjusts the filtering parameters based on a reward function, wherein the reward function is defined according to the signal-to-noise ratio and edge clarity of the defect features, the lightweight neural network adopts a convolutional neural network structure, and optimizes the prediction accuracy through online updating of historical detection data.

[0021] Preferably, the clustering algorithm in the dynamic pattern matching module is the DBSCAN algorithm, which dynamically generates defect pattern templates according to the shape, size and distribution density of the defects. The generative adversarial network includes a generator and a discriminator, wherein the generator is used to simulate defect deformation and the discriminator is used to verify the authenticity of the matching results.

[0022] Preferably, the dynamic pattern matching module also includes a regional focusing unit, which adjusts the priority of the matching calculation according to the potential defect high-incidence areas predicted by the adaptive preprocessing module, and accelerates the defect matching process in the high-incidence areas through GPU parallel computing.

[0023] Preferably, the image features in the multidimensional feature fusion classification module include the shape, texture and grayscale distribution of the defect, the physical parameter features include the defect depth and material density, and the process background features include the defect occurrence process and equipment status. The multidimensional feature vector is input into the Transformer model after feature normalization.

[0024] Preferably, the unsupervised learning in the multi-dimensional feature fusion classification module adopts an autoencoder structure to detect unknown defect types through reconstruction errors, and mark the detection results as abnormal data, automatically triggering a manual confirmation process.

[0025] Preferably, the inverse mapping algorithm in the closed-loop optimization and output module is based on a Bayesian reasoning model, which deduces the cause of process deviation according to defect classification results and process background characteristics, and generates process adjustment suggestions including lithography exposure time adjustment and deposition rate optimization.

[0026] Preferably, the closed-loop optimization and output module further includes an intelligent visualization unit, which is configured to generate a three-dimensional defect distribution heat map and a dynamic process simulation view, supporting users to adjust detection parameters and view optimization effects through an interactive interface.

[0027] Preferably, the system also includes a self-evolving database, which is connected to the closed-loop optimization and output module, and is used to store detection data, classification results and process adjustment records, and update the template generation model of the dynamic pattern matching module and the Transformer model of the multidimensional feature fusion classification module through periodic training.

[0028] Beneficial effects

[0029] The present invention provides a full-layout defect rapid detection system based on pattern matching and classification. It has the following beneficial effects:

[0030] 1. The present invention uses the system's dynamic pattern matching module to dynamically generate defect templates from real-time data using clustering algorithms (such as DBSCAN) and generative adversarial networks (GANs). Combined with the reinforcement learning filtering and regional focusing technology of the adaptive preprocessing module, this significantly improves the speed and accuracy of defect detection. It can adaptively adjust the detection strategy based on real-time data, prioritize matching high-risk areas, and accelerate processing through GPU parallel computing. This not only shortens the detection time to less than 50% of traditional methods, but also effectively identifies deformed or blurred defects, reducing the false alarm rate by approximately 30%. In addition, the multi-depth image data collected by the multimodal imaging unit further enhances the visibility of defect features, providing high-quality input for subsequent matching and classification.

[0031] 2. The present invention uses a multi-dimensional feature fusion classification module. The system integrates image features (such as shape and texture), physical parameter features (such as defect depth) and process background features (such as process status) into a multi-dimensional feature vector, and uses the Transformer cross-modal learning model for classification. At the same time, it combines the autoencoder to identify unknown defects, breaking through the limitations of traditional single image analysis. The classification accuracy is improved to more than 95%, especially showing higher robustness for complex defects and minor defects. Through the introduction of process context, the cause of the defect is accurately distinguished, and unknown defects are automatically marked through unsupervised learning, reducing manual intervention by about 80%.

[0032] 3. The present invention adopts a closed-loop optimization and output module to convert defect detection results into process adjustment suggestions through an inverse mapping algorithm (based on Bayesian reasoning) and an intelligent visualization unit, forming a closed-loop control of detection and optimization. The periodic training of the self-evolving database further improves the model performance, enabling the system to continuously optimize with production changes, derive process deviations (such as insufficient exposure) based on defect distribution, generate specific suggestions (such as adjusting the lithography time), and intuitively display the optimization effect through a three-dimensional heat map, shortening the process adjustment cycle by about 40%, improving the yield by 5%-10%, significantly reducing production costs and defective product rates, and realizing an intelligent transformation from detection to prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a system architecture diagram of the present invention;

[0034] Figure 2 It is a diagram of the system operation steps of the present invention;

[0035] Figure 3 A line graph comparing the detection time of the system of the present invention and the traditional method;

[0036] Figure 4 A discrete graph of defect classification distribution for the system detection of the present invention;

[0037] Figure 5 This is the yield optimization feedback table of the present invention. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. Specific embodiment one:

[0040] like Figures 1 to 5 As shown in the figure, the full-layout defect rapid detection system based on pattern matching and classification includes a data acquisition and multi-source fusion module, an adaptive preprocessing module, a dynamic pattern matching module, a multi-feature fusion module, and a closed-loop optimization and output module, among which:

[0041] The data acquisition and multi-source fusion module collects image data of the entire wafer layout and integrates the process parameter data in the manufacturing process to generate a multi-source fusion data package containing image and process context information, where the process parameter data includes etching time and deposition thickness. The data acquisition and multi-source fusion module includes a multimodal imaging unit, which collects multi-depth image data of the entire wafer layout through optical microscopy, infrared imaging and X-ray scanning, and aligns the multi-depth image data with the process parameter data through a time synchronization mechanism to generate a multi-source fusion data package. The multidimensional feature fusion classification module integrates image features (such as shape and texture), physical parameter features (such as defect depth) and process background features (such as process status) into a multidimensional feature vector, and uses the Transformer cross-modal learning model for classification. At the same time, it combines with the autoencoder to identify unknown defects, breaking through the limitations of traditional single image analysis. The classification accuracy is improved to more than 95%, especially showing higher robustness for complex defects and minor defects. By introducing process context, the cause of the defect is accurately distinguished, and unknown defects are automatically labeled through unsupervised learning, reducing manual intervention by about 80%.

[0042] The adaptive preprocessing module is connected to the data acquisition and multi-source fusion module. It adaptively selects filter parameters based on a reinforcement learning algorithm to enhance defect characteristics. It also predicts potential high-incidence areas of defects through a lightweight neural network based on historical inspection data and outputs optimized image data. The reinforcement learning algorithm in the adaptive preprocessing module adaptively adjusts the filter parameters based on a reward function defined by the signal-to-noise ratio and edge clarity of the defect characteristics. The lightweight neural network uses a convolutional neural network structure and optimizes prediction accuracy through online updates of historical inspection data.

[0043] The dynamic pattern matching module is connected to the adaptive preprocessing module. It dynamically generates defect pattern templates from real-time data through clustering algorithms, and combines fast Fourier transform and generative adversarial networks for defect matching. It prioritizes matching areas with high incidence of potential defects and outputs the coordinates of suspected defects and matching confidence. The clustering algorithm in the dynamic pattern matching module is the DBSCAN algorithm, which dynamically generates defect pattern templates according to the shape, size and distribution density of the defects. The generative adversarial network includes a generator and a discriminator. The generator is used to simulate defect deformation, and the discriminator is used to verify the authenticity of the matching results. The dynamic pattern matching module also includes a regional focusing unit, which is based on the prediction of the adaptive preprocessing module. In areas with high incidence of potential defects, the priority of matching calculations is adjusted, and the defect matching process in high-incidence areas is accelerated through GPU parallel computing. The system's dynamic pattern matching module uses clustering algorithms (such as DBSCAN) and generative adversarial networks (GAN) to dynamically generate defect templates from real-time data. Combined with the reinforcement learning filtering and regional focusing technology of the adaptive preprocessing module, the speed and accuracy of defect detection are significantly improved. It can adaptively adjust the detection strategy according to real-time data, prioritize matching high-risk areas, and accelerate processing through GPU parallel computing. Not only does it shorten the detection time to less than 50% of traditional methods, it can also effectively identify deformation or blurred defects, and reduce the false alarm rate by about 30%. In addition, the multi-depth image data collected by the multimodal imaging unit further enhances the visibility of defect features and provides high-quality input for subsequent matching and classification.

[0044] The multidimensional feature fusion classification module is connected to the dynamic pattern matching module to extract the image features, physical parameter features, and process background features of suspected defects, construct a multidimensional feature vector, and classify the defects using a Transformer-based cross-modal learning model. At the same time, it identifies unknown defect types through unsupervised learning and outputs defect classification labels and impact assessments. The image features in the multidimensional feature fusion classification module include the shape, texture, and grayscale distribution of the defect, the physical parameter features include the defect depth and material density, and the process background features include the defect occurrence process and equipment status. The multidimensional feature vector is input into the Transformer model after feature normalization. The unsupervised learning in the multidimensional feature fusion classification module adopts an autoencoder structure to detect unknown defect types through reconstruction errors, mark the detection results as abnormal data, and automatically trigger the manual confirmation process;

[0045] The closed-loop optimization and output module is connected to the multi-dimensional feature fusion classification module. It generates process adjustment suggestions through the inverse mapping algorithm based on the defect classification results, forms a closed-loop control of detection and process optimization, and outputs interactive defect analysis reports and self-optimization model parameters. The inverse mapping algorithm in the closed-loop optimization and output module is based on the Bayesian reasoning model. It deduces the cause of process deviation according to the defect classification results and process background characteristics, and generates process adjustment suggestions including lithography exposure time adjustment and deposition rate optimization. The closed-loop optimization and output module also includes an intelligent visualization unit. The intelligent visualization unit is configured to generate a three-dimensional defect distribution heat map and a dynamic process simulation view, which supports users to adjust the defect through an interactive interface. Adjust detection parameters and view optimization effects. A closed-loop optimization and output module is used to convert defect detection results into process adjustment suggestions through an inverse mapping algorithm (based on Bayesian reasoning) and an intelligent visualization unit, forming a closed-loop control of detection and optimization. The periodic training of the self-evolving database further improves the model performance, enabling the system to continuously optimize with production changes, deduce process deviations (such as insufficient exposure) based on defect distribution, generate specific suggestions (such as adjusting the lithography time), and intuitively display the optimization effects through a three-dimensional heat map, shortening the process adjustment cycle by about 40%, increasing the yield by 5%-10%, significantly reducing production costs and defective product rates, and realizing an intelligent transformation from detection to prevention.

[0046] The system also includes a self-evolving database, which is connected to the closed-loop optimization and output module to store detection data, classification results and process adjustment records, and update the template generation model of the dynamic pattern matching module and the Transformer model of the multidimensional feature fusion classification module through periodic training. Specific embodiment two:

[0048] like Figures 1 to 5 As shown, based on the content in the above specific embodiments, the following contents are further disclosed:

[0049] The operation steps of the entire system are as follows:

[0050] Sp1: Data Acquisition and Multi-Source Fusion: Through the data acquisition and multi-source fusion module, multi-modal imaging units (such as optical microscopes, infrared imaging, and X-ray scanning) are used to collect multi-depth image data of the entire wafer layout. At the same time, process parameters (such as etching time and deposition thickness) are obtained from the manufacturing equipment. The image and parameters are aligned through a time synchronization mechanism to generate a multi-source fusion data package containing image and process context information, providing high-quality input for subsequent analysis;

[0051] Sp2: Adaptive image preprocessing: In this step, the adaptive preprocessing module takes over the image data and uses a reinforcement learning algorithm (based on a reward function for signal-to-noise ratio and edge clarity) to dynamically select filtering parameters (such as Gaussian kernel size), denoise, and enhance defect features. At the same time, a lightweight CNN is used to analyze historical data to predict potential high-incidence areas of defects, outputting optimized images and high-risk coordinates to improve subsequent inspection efficiency.

[0052] Sp3: Dynamic defect pattern generation: The dynamic pattern matching module uses the DBSCAN clustering algorithm to extract defect features (such as shape, size, and density) from the optimized image, dynamically generates defect pattern templates based on density distribution, removes noise points, and forms a real-time adaptive template set, providing a diverse and accurate reference standard for subsequent matching;

[0053] Sp4: Defect matching and positioning: In the dynamic pattern matching module, fast Fourier transform is used to calculate the correlation between template and image, and GAN is combined to generate deformable defect samples to enhance matching robustness. High-risk areas are matched first, and GPU parallel computing is used to accelerate processing. Finally, the coordinates of suspected defects and matching confidence are output to achieve rapid positioning.

[0054] Sp5: Multidimensional feature extraction: The multidimensional feature fusion classification module extracts image features (shape, texture), physical parameter features (depth, density), and process background features (process, equipment status) from suspected defect areas, and generates multidimensional feature vectors through standardization, providing a comprehensive data foundation for subsequent classification;

[0055] Sp6: Defect classification and unknown defect identification: The multi-dimensional feature fusion classification module inputs feature vectors into the Transformer model for classification, outputs defect labels (such as "scratch" and "particle") and impact assessment, and simultaneously detects reconstruction errors through the autoencoder, identifies unknown defects and automatically labels them, triggering manual confirmation to ensure classification accuracy and adaptability.

[0056] Sp7: Process Deviation Reasoning: The closed-loop optimization and output module is based on a Bayesian reasoning model. It combines defect classification results and process background to deduce the cause of process deviation (such as underexposure) and generate specific adjustment suggestions (such as photolithography time optimization), providing a scientific basis for process improvement.

[0057] Sp8: Result visualization and output: The closed-loop optimization and output module generates a three-dimensional defect distribution heat map and process simulation view through an intelligent visualization unit, outputs an interactive analysis report, supports user parameter adjustment, and updates the self-optimization model parameters at the same time, ensuring intuitive results and continuous improvement of system performance;

[0058] Sp9: Data archiving and model updating: Utilizes a self-evolving database to store detection data, classification results, and process records, and periodically trains and updates the template generation model and Transformer classification model of the dynamic pattern matching module to improve the system's long-term adaptability and detection accuracy.

[0059] Sp10: Closed-loop feedback execution: The closed-loop optimization and output module feeds back process adjustment suggestions to the production line, adjusts parameters (such as deposition rate) in real time, monitors subsequent defect changes and records feedback data to the database, forming a closed-loop control of detection and optimization to improve yield and production efficiency. Specific embodiment three:

[0061] like Figures 1 to 5 As shown, based on the content in the above specific embodiments, the following contents are further disclosed:

[0062] The specific formulas of each algorithm in the above system are as follows:

[0063] The reinforcement learning algorithm formula of the adaptive preprocessing module is as follows:

[0064] Q(s,a)←Q(s,a)+α[R(s,a)+γmax a′ Q(s ′ , a ′ )-Q(s,a)]

[0065] Reward function:

[0066] R(s,a)=w1·SNR+w2·EdgeClarity

[0067] Where: Q(s, a) is the Q value of executing action a in state s, which represents the long-term reward of the filter parameter selection;

[0068] s: current state, including defect characteristics of the input image (such as noise level, edge strength);

[0069] a: action, i.e. the selected filtering parameters (e.g. Gaussian kernel size σ);

[0070] α: learning rate, controls the update step size, and its value range is [0, 1], such as 0.1;

[0071] R(s, a): instant reward, composed of the weighted signal-to-noise ratio (SNR) and edge clarity (EdgeClarity);

[0072] γ: discount factor, which measures the importance of future returns and has a value range of [0, 1], such as 0.9;

[0073] s ′ : The new state after executing action a;

[0074] a ′ : Possible actions in the new state;

[0075] w1, w2: Weight coefficients, which adjust the relative importance of SNR and edge clarity, such as w1 = 0.6, w2 = 0.4.

[0076] The reinforcement learning algorithm sets the initial Q table, all Q(s, a) = 0, defines the state space (image features) and action space (filter parameters), inputs the image, calculates the current SNR and edge strength, determines the state s, selects the filter parameter a (such as σ = 1.5) based on the ∈-greedy strategy, applies Gaussian filtering, generates a new image, and calculates the new state s. ′ The system dynamically adjusts the filtering parameters through trial and error learning, exploring multiple σ values ​​in the initial stage. As training accumulates, it tends to choose the action that maximizes the reward. The reward function ensures that the filter can both denoise (high SNR) and retain defect edges (high Edge Clarity), achieving adaptive optimization. Compared with fixed parameter filtering, reinforcement learning improves the defect feature enhancement effect by about 20%, reduces the missed detection rate by 10%, and significantly improves the ability to adapt to different image noise and defect types.

[0077] The DBSCAN algorithm of the dynamic pattern matching module includes the following:

[0078] Core point determination:

[0079] |N ∈ (p)|≥MinPts

[0080] Distance function:

[0081]

[0082] Where: N ∈ (p): the set of points in the ∈-neighborhood of point p;

[0083] |N ∈ (p)|: the number of points in the neighborhood;

[0084] MinPts: minimum point count threshold, the standard for judging core points, such as 5. ·∈: neighborhood radius, defining the clustering range, such as 0.1;

[0085] d(p,q): Euclidean distance between points p and q;

[0086] p i ,q i : The characteristic values ​​of points p and q in the i-th dimension (such as shape, grayscale).

[0087] The DBSCAN algorithm extracts real-time defect features (such as shape, size, density), forms a feature point set, initializes ∈ and MinPts, and adjusts the defect distribution density. For each point p, calculate N ∈ (p), if |N ∈ (p)|≥MinPts is marked as a core point. Starting from this core point, all density-reachable points are connected to form a cluster. 5. Template Generation: Each cluster generates a defect pattern template. DBSCAN clusters based on the local density of defect features, eliminating the need to pre-set the number of clusters and automatically separating noise points from abnormal defects. Core points drive cluster expansion, generating dynamic templates that adapt to real-time changes in defect morphology. Compared to K-means, DBSCAN eliminates the need to specify the number of clusters, increases template generation speed by 30%, and improves coverage of irregular defects by 25%, significantly enhancing dynamic adaptability.

[0088] The Generative Adversarial Network (GAN) of the dynamic pattern matching module includes the following:

[0089] Objective function:

[0090]

[0091] Where: G: generator, generates simulated defect images;

[0092] D: Discriminator, distinguishes between real defects and generated defects;

[0093] V(D, G): adversarial loss function;

[0094] x: real defect image, obeying data distribution p data (x);

[0095] z: random noise input, obeying distribution p z (z), such as normal distribution;

[0096] D(x): discriminator output, probability of real image;

[0097] G(z): generator output, simulated defect image;

[0098] E[·]: Expected operation.

[0099] The neural network parameters of G and D are randomly initialized, and the real defect image x and the generated image G(z) are input. D is optimized to maximize V(D, G). Noise z is input, and G is optimized to minimize log(1-D(G(z))). D and G are trained alternately until the generated image is realistic. The defects generated by G are used to match the deformed samples. The discriminator learns to distinguish between real and fake defects. The generator improves its simulation ability through adversarial means, and ultimately generates realistic defect deformed samples, enhancing matching robustness. GAN generates diverse defect samples, which improves matching accuracy by 15% and recognition rate of blurred or deformed defects by 20%, making up for the singleness of traditional templates.

[0100] The Transformer of the multi-dimensional feature fusion classification module further exposes the following:

[0101] Attention Mechanism:

[0102]

[0103] Where: Q: query matrix, query representation of feature vector;

[0104] K: key matrix, key representation of eigenvector;

[0105] V: value matrix, the value representation of the eigenvector;

[0106] d k : The dimension of the key vector, used for scaling to prevent excessive values, such as 64;

[0107] QK T : The dot product of the query and the key to calculate the similarity;

[0108] Softmax: Normalization function, generating attention weights.

[0109] The image, physical parameters and process background features are standardized into vectors to form Q, K, and V. The attention weights between each feature are calculated using a formula. Multiple attention heads are used in parallel to fuse the results. The classification probability is output through a feedforward network and residual connection. Cross-entropy loss is used and backpropagation is used to update the parameters. The Transformer captures the association between multi-dimensional features (such as the relationship between defect shape and process) through the attention mechanism. The multi-head structure enhances classification robustness and ultimately outputs accurate labels. The classification accuracy is increased to 95%, 10% higher than the traditional CNN. Cross-modal fusion increases the recognition rate of complex defects by 15%, significantly enhancing robustness.

[0110] The autoencoder of the multi-dimensional feature fusion classification module further includes the following:

[0111] Loss function:

[0112]

[0113] Reconstruction error:

[0114] Where: x i : Input defect feature vector;

[0115] The output vector after autoencoder reconstruction;

[0116] L: mean square error loss, measuring the quality of reconstruction;

[0117] E i : The reconstruction error of the i-th sample, used for anomaly detection;

[0118] n: sample size.

[0119] An encoder (compressed features) and a decoder (reconstructed features) are designed. Known defect data is input, L is minimized, and E is calculated for new data. If the threshold is exceeded (such as the mean + 2 times the standard deviation), it is marked as an unknown defect. Abnormal data is automatically recorded and triggered for confirmation. The autoencoder learns the characteristic distribution of normal defects. High-error samples are identified as abnormal and automatically marked as unknown defects. The unknown defect detection rate reaches 90%, and the workload of manual confirmation is reduced by 80%, which makes up for the limitations of supervised learning.

[0120] The Bayesian reasoning of the closed-loop optimization and output module further includes the following:

[0121] Posterior probability:

[0122]

[0123] Where: P(C|D): the posterior probability of process cause C given defect D;

[0124] P(D|C): conditional probability, the likelihood that process cause C leads to defect D;

[0125] P(C): prior probability of process cause;

[0126] P(D): marginal probability of defect, normalization factor.

[0127] Historical defects D and process parameters C are recorded, and P(C) and P(D|C) are calculated based on historical data. P(D|C) is updated based on the current defect characteristics, and P(C|D) is calculated to identify the most likely process deviation. Parameters (such as exposure time) are adjusted based on C. The Bayesian model combines historical and real-time data to infer the cause of the defect and generate optimization suggestions. The accuracy of process deviation identification reaches 85%, the adjustment cycle is shortened by 40%, and the yield is improved by 5%-10%.

[0128] like Figure 3As shown in the figure, the detection time of the traditional method (fixed 35 seconds) and the system (about 28 seconds) is compared, which verifies the efficiency improvement of dynamic pattern matching and parallel computing. The traditional time is a constant, and the time of this system is the sum of the time taken for each step plus random noise. The line graph shows the time trend. The console output efficiency is improved by about 20%. Dynamic pattern matching and parallel computing reduce the detection time from 35 seconds to 28 seconds, an improvement of about 20%.

[0129] like Figure 4 As shown, the defect coordinates and classification results (scratches, particles, short circuits, unknowns) are displayed to verify the multi-dimensional feature fusion classification capability of Transformer. 5,000 defect coordinates are randomly generated. The classification results are based on the true distribution plus a 5% error to simulate 95% accuracy. The discrete graph shows the spatial distribution of defects. The console outputs the proportion of each type (such as scratches, about 40%). The Transformer classification accuracy is about 95%, and the multi-dimensional feature fusion clearly distinguishes the defect types.

[0130] like Figure 5 As shown, the traditional yield (92%) is compared with the optimized yield (about 95.2%), verifying the closed-loop optimization effect of Bayesian reasoning, simulating the distribution of short-circuit defects, and improving the yield after adjusting the exposure time. The table contains wafer ID, yield, number of defects and parameters. The table shows the data of each wafer. The console outputs an average yield increase of 3.2% and a short circuit reduction. The closed-loop optimization increases the yield from 92% to 95.2%, and the short-circuit defects are significantly reduced. Specific embodiment four:

[0132] like Figures 1 to 5 As shown, based on the content in the above specific embodiments, the following contents are further disclosed:

[0133] The entire system is used in actual application scenarios, combined with the above system steps, to further disclose the following content:

[0134] The scenario is set as real-time defect detection and optimization of a 7nm wafer production line in a semiconductor factory. Combining the above system steps (Sp1-Sp10) and core technologies, the application process is described in detail and its actual benefits are analyzed.

[0135] Practical application scenario: Real-time defect detection and optimization of 7nm wafer production lines

[0136] Scenario Background: A semiconductor fab produces logic chips using a 7nm process node. The wafers are 300mm in diameter and each batch processes 100 wafers. Due to the complex manufacturing process and the diverse defect types (such as tiny scratches, photolithography shorts, and particle deposits), traditional manual inspection and static mask methods are inefficient, with a missed detection rate as high as 20% and yield rates fluctuating between 90% and 92%. The fab introduced this system to improve inspection speed, classification accuracy, and yield.

[0137] Practical application steps;

[0138] Sp1: Data acquisition and multi-source fusion: After the wafer lithography process, a multimodal imaging unit (optical microscope resolution 0.1μm, infrared imaging, X-ray scanning) is used to collect full-plate images of each wafer (resolution 4096×4096 pixels), and process parameters (lithography exposure time 20ms, etching time 30s, deposition thickness 50nm) are extracted from the lithography machine and etching equipment. Through time synchronization, a multi-source fusion data packet is generated. Multimodal imaging captures deep defects (such as internal shorts) that traditional optics cannot detect. After the data packet is fused with process parameters, the accuracy of defect cause analysis is improved by 30%. Data acquisition for each wafer takes 10 seconds and generates approximately 500MB of data packets.

[0139] Sp2: Adaptive image preprocessing: A reinforcement learning algorithm is applied to the image (learning rate α = 0.1, discount factor γ = 0.9, initial attempt Gaussian kernel σ = 1.0-2.0, optimized to σ = 1.5 based on the reward function (SNR weight 0.6, edge clarity 0.4). A lightweight CNN analyzes the past 100 wafer data and predicts that the lithography edge area is a high-incidence area. Adaptive filtering improves the defect edge contrast by 25%, the prediction accuracy of high-risk areas reaches 85%, and the full image scanning time is reduced by about 40%. The preprocessing takes 5 seconds per wafer, and the optimized image and high-risk coordinates are output;

[0140] Sp3: Dynamic defect pattern generation: Extract features (shape, grayscale, density) from the optimized image and use DBSCAN (∈=0.1, MinPts=5) to generate templates. Five types of defect patterns, such as scratches and particles, are discovered, and approximately 10% of noise points are eliminated. The dynamic template has strong adaptability and a coverage rate 20% higher than the static template. The generation time is only 2 seconds per piece. Five templates are generated and stored in the dynamic template library.

[0141] Sp4: Defect matching and positioning: Using FFT to calculate the correlation between template and image (threshold 0.8), GAN generates deformable samples (training for 1000 iterations), prioritizes matching of high-risk areas, and GPU-accelerated processing (NVIDIA A100). GAN enhances the fuzzy defect matching rate by 15%. Regional focusing and parallel computing reduce the matching time from 30 seconds to 8 seconds, achieving a positioning accuracy of 90%. 50 suspected defects are located on each wafer with a coordinate error of <0.5μm.

[0142] Sp5: Multidimensional feature extraction: Extracts image features of defects (aspect ratio 1.5, grayscale mean 120), physical parameters (depth 0.2μm, density 2.5g / cm 3), process background (photolithography process, equipment temperature 50°C), after standardization, a 128-dimensional feature vector is generated. The multi-dimensional feature provides comprehensive information. Compared with single image features, the quality of classification preparation data is improved by 30%. It takes 3 seconds to generate 50 feature vectors for 50 defects;

[0143] Sp6: Defect classification and unknown defect identification: The feature vector is input into the Transformer (8-head attention, dimension 64) and classified as "scratches" (20), "particles" (25), and "short circuits" (3). The autoencoder (reconstruction error threshold 0.05) marks two unknown defects. The Transformer classification accuracy is 95%, the autoencoder's unknown defect recognition rate is 90%, and manual confirmation is reduced by 80%. The classification takes 5 seconds and outputs a label and two anomaly tags.

[0144] Sp7: Process deviation reasoning: The Bayesian model analyzes the relationship between the "short circuit" defect and exposure time (prior probability 0.3, likelihood 0.7), derives underexposure, and recommends increasing the exposure time to 22ms. The reasoning accuracy is 85%, the suggestion is highly targeted, and the trial and error cost is reduced by 50%. It takes 2 seconds to generate the adjustment suggestion.

[0145] Sp8: Result visualization and output: Generates a 3D heat map to display defect distribution. The simulation view predicts that short circuits will be reduced by 50% after adjustment. Outputs a report and updates Transformer parameters. Visualization and intuitiveness improve engineering decision-making efficiency by 30%. Model updates improve subsequent accuracy by 2%. Report generation takes 3 seconds, and the user confirms the adjustment plan.

[0146] Sp9: Data Archiving and Model Update: Stores 50 defect data and adjustment records, trains and updates DBSCAN and Transformer models every 10 batches (1000 pieces), and uses a self-evolving database to enhance system adaptability over time, improving model accuracy by 5%. Archiving takes 1 second, and the update cycle is 1 hour.

[0147] Sp10: Closed-loop feedback execution: The exposure time was adjusted to 22ms, and the number of short-circuit defects in the next batch of detection was reduced to 1. The yield rate increased from 92% to 94.5%. The feedback data was archived. The closed-loop control shortened the optimization cycle by 40%, and the yield rate was significantly improved. The adjustment execution took 5 minutes, and the feedback verification took 10 pieces.

[0148] Application results summary: Inspection speed is shortened from the traditional 30 seconds per piece to 28 seconds per piece (including the entire process); classification accuracy is improved from 80% to 95%; yield rate is increased from 92% to 94.5%, and approximately 200 defective pieces are reduced per batch; manual intervention is reduced from 10 hours per batch to 2 hours;

[0149] Dynamic pattern matching (DBSCAN+GAN): Adapts to complex defects and reduces missed detection rate from 20% to 5%;

[0150] Multi-dimensional feature fusion (Transformer + autoencoder): Improves classification robustness by 15%, filling a technical gap in unknown defect identification;

[0151] Closed-loop optimization (Bayesian reasoning): Automated process adjustments increased yield by 2.5% and achieved significant cost savings.

[0152] Actual benefits: Batch inspection time was reduced from 50 hours to 30 hours, and production capacity increased by 33%; each batch saved approximately 100,000 yuan in defective product costs (calculated at 500 yuan per piece); the system ran 100 batches with no performance degradation and continued to improve its adaptability.

[0153] Conclusion: Through application on a 7nm wafer production line, this system demonstrated its efficiency, accuracy, and intelligence in real-time scenarios. The core technology was fully validated in experiments, addressing the low efficiency, poor adaptability, and lack of feedback inherent in traditional methods and providing an advanced solution for semiconductor manufacturing.

[0154] Extended Scenario 2: 5nm process post-lithography defect detection:

[0155] Scenario Background: A factory produces high-performance computing chips using a 5nm process, with each batch containing 50 wafers and a target yield of 96%. Post-lithography defects (such as residues and micro-shorts) are extremely small (<0.1μm). Traditional methods have a 25% miss rate and require 40 seconds per wafer for inspection. This system is deployed at the post-lithography inspection station.

[0156] Practical application steps and specific data:

[0157] Sp1: Acquired a 5nm wafer image (6144×6144 pixels, 600MB) with an exposure time of 15ms (±1ms) and a deposition thickness of 30nm (±3nm). The acquisition took 12 seconds and identified 3,000 suspected defects (60 / wafer).

[0158] Sp2: Reinforcement learning optimization: σ = 1.2 (SNR increased to 28dB), CNN predicts edge and center high-incidence areas (covering 90% of defects), taking 6 seconds;

[0159] Sp3: DBSCAN generates 7 types of templates (residues account for 30%), takes 3 seconds, and has a coverage rate of 94%;

[0160] Sp4: FFT+GAN matching, localizing 2700 defects (missed detection rate 10%), taking 10 seconds, with an error of 0.2μm;

[0161] Sp5: Feature extraction (depth 0.1 μm, grayscale 110), takes 4 seconds and generates 2700 vectors;

[0162] Sp6: Transformer classification (1200 residuals, 200 short circuits), autoencoder labels 30 unknown defects, takes 6 seconds, and has an accuracy of 97%;

[0163] Sp7: Bayesian recommendation exposure time increased to 16ms, took 2 seconds, and had an accuracy of 90%;

[0164] Sp8: The heat map shows the center of the residue concentration, which takes 4 seconds and the user confirms the adjustment;

[0165] Sp9: archived 2700 data (1.5GB), and updated the model in 1 hour;

[0166] Sp10: After adjustment, the residues dropped to 300 and the yield increased to 96.5%, taking 5 minutes.

[0167] Quantitative results: Inspection speed of 31 seconds per piece (traditional 40 seconds); missed detection rate of 4% (traditional 25%); yield rate of 96.5% (target 96%), defective products reduced from 4 pieces to 1.75 pieces; saving of 1,125 yuan per batch (500 yuan per piece);

[0168] Analysis of core technology contributions:

[0169] Dynamic pattern matching: The missed detection rate in 7nm scenarios is reduced to 5%, and in 5nm to 4%, and the ability to adapt to small defects is improved by 20%;

[0170] Multi-dimensional feature fusion: Classification accuracy exceeds 95%, the recognition rate of 5nm complex defects is increased by 17%, and manual intervention is reduced by 80%;

[0171] Closed-loop optimization: 7nm yield increased by 3.2%, 5nm yield increased by 0.5% (baseline higher), and optimization cycle shortened by 40%-45%;

[0172] Summary: Through the application of 7nm and 5nm scenarios, this system demonstrates its high efficiency and adaptability under different processes.

[0173] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further restrictions, an element defined by the statement "comprising a reference structure" does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0174] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A rapid full-layout defect detection system based on pattern matching and classification, featuring: The system includes a data acquisition and multi-source fusion module, an adaptive preprocessing module, a dynamic pattern matching module, a multi-feature fusion module, and a closed-loop optimization and output module, wherein: The data acquisition and multi-source fusion module collects wafer full-layout image data and fuses process parameter data during the manufacturing process to generate a multi-source fusion data package containing image and process context information, wherein the process parameter data includes etching time and deposition thickness; The adaptive preprocessing module is connected to the data acquisition and multi-source fusion module, adaptively selects filtering parameters based on the reinforcement learning algorithm to enhance defect characteristics, and predicts potential defect-prone areas through a lightweight neural network based on historical inspection data, and outputs optimized image data; The dynamic pattern matching module is connected to the adaptive preprocessing module, dynamically generates defect pattern templates from real-time data through a clustering algorithm, and combines fast Fourier transform and generative adversarial network to perform defect matching, preferentially matching the potential defect high-incidence areas, and outputting the coordinates of suspected defects and matching confidence levels; The multi-dimensional feature fusion classification module is connected to the dynamic pattern matching module to extract the image features, physical parameter features, and process background features of suspected defects, construct a multi-dimensional feature vector, and classify the defects using a Transformer-based cross-modal learning model. At the same time, it identifies unknown defect types through unsupervised learning and outputs defect classification labels and impact assessments. The closed-loop optimization and output module is connected to the multi-dimensional feature fusion classification module, generates process adjustment suggestions based on the defect classification results through the inverse mapping algorithm, forms a closed-loop control of detection and process optimization, and outputs an interactive defect analysis report and self-optimization model parameters.

2. The full-layout defect rapid detection system based on pattern matching and classification according to claim 1 is characterized by: The data acquisition and multi-source fusion module includes a multi-modal imaging unit, which collects multi-depth image data of the entire wafer layout through optical microscopy, infrared imaging and X-ray scanning, and aligns the multi-depth image data with process parameter data through a time synchronization mechanism to generate a multi-source fusion data packet.

3. The full-layout defect rapid detection system based on pattern matching and classification according to claim 1 is characterized by: The reinforcement learning algorithm in the adaptive preprocessing module adaptively adjusts the filtering parameters based on a reward function, where the reward function is defined according to the signal-to-noise ratio and edge clarity of the defect characteristics. The lightweight neural network adopts a convolutional neural network structure and optimizes the prediction accuracy through online updates of historical detection data.

4. The full-layout defect rapid detection system based on pattern matching and classification according to claim 1 is characterized by: The clustering algorithm in the dynamic pattern matching module is the DBSCAN algorithm, which dynamically generates defect pattern templates based on the shape, size and distribution density of the defects. The generative adversarial network includes a generator and a discriminator, where the generator is used to simulate defect deformation and the discriminator is used to verify the authenticity of the matching results.

5. The full-layout defect rapid detection system based on pattern matching and classification according to claim 1 is characterized by: The dynamic pattern matching module also includes a regional focusing unit, which adjusts the priority of the matching calculation according to the potential defect high-incidence areas predicted by the adaptive preprocessing module, and accelerates the defect matching process in the high-incidence areas through GPU parallel computing.

6. The full-layout defect rapid detection system based on pattern matching and classification according to claim 1 is characterized by: The image features in the multi-dimensional feature fusion classification module include the shape, texture and grayscale distribution of the defect, the physical parameter features include the defect depth and material density, and the process background features include the defect occurrence process and equipment status. The multidimensional feature vector is input into the Transformer model after feature normalization.

7. The full-layout defect rapid detection system based on pattern matching and classification according to claim 1 is characterized by: The unsupervised learning in the multi-dimensional feature fusion classification module adopts an autoencoder structure to detect unknown defect types through reconstruction errors, and marks the detection results as abnormal data, automatically triggering the manual confirmation process.

8. The full-layout defect rapid detection system based on pattern matching and classification according to claim 1 is characterized by: The inverse mapping algorithm in the closed-loop optimization and output module is based on a Bayesian reasoning model. It deduces the cause of process deviation according to defect classification results and process background characteristics, and generates process adjustment suggestions including lithography exposure time adjustment and deposition rate optimization.

9. The full-layout defect rapid detection system based on pattern matching and classification according to claim 1 is characterized by: The closed-loop optimization and output module also includes an intelligent visualization unit, which is configured to generate a three-dimensional defect distribution heat map and a dynamic process simulation view, supporting users to adjust detection parameters and view optimization effects through an interactive interface.

10. The full-layout defect rapid detection system based on pattern matching and classification according to claim 1 is characterized by: The system also includes a self-evolving database, which is connected to the closed-loop optimization and output module and is used to store detection data, classification results and process adjustment records, and update the template generation model of the dynamic pattern matching module and the Transformer model of the multidimensional feature fusion classification module through periodic training.

Images