High-precision defect detection method and system for semiconductor chip

Abstract

The invention relates to the technical field of chip optical detection, and discloses a high-precision defect detection method and system for a semiconductor chip. Initial detection data are obtained through laser scattering and photo-thermal microscopic detection, key visual features are extracted and converted into a multi-dimensional visual feature matrix, and the multi-dimensional visual feature matrix is used for detecting the defects of the semiconductor chip; according to the method, a fuzzy sensing area is positioned and potential defects are identified by using a map clustering algorithm, adaptive analysis of a supplementary visual angle is carried out and supplementary detection data is generated, and finally defect identification is carried out through the supplementary data and popularity information to obtain a chip defect result. The method effectively improves the precision and reliability of defect detection, especially has obvious advantages in recognition of complex and fuzzy regions, and solves the problem of insufficient precision of single-path optical imaging detection in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of chip optical detection, and in particular to a high-precision defect detection method and system for semiconductor chips. Background Art

[0002] In the modern chip manufacturing process, as the complexity of integrated circuits continues to increase, chip defect detection has become increasingly important. Traditional defect detection methods mainly rely on single-path optical imaging technology. Although these methods can identify some defects, for chips with small size, high density and complex structure, the detection accuracy and efficiency still have certain limitations. Summary of the Invention

[0003] The purpose of the present invention is to provide a high-precision defect detection method and system for semiconductor chips, aiming to solve the problem of insufficient accuracy of single-path optical imaging detection in the prior art.

[0004] The present invention is implemented as follows: In a first aspect, the present invention provides a high-precision defect detection method for semiconductor chips, comprising: Performing laser scattering detection and photothermal microscopy detection at an initial visual angle on the chip to be tested to obtain initial detection data of the chip to be tested; Extracting key visual features from the initial detection data, and performing feature vector conversion and data integration on the key visual features to obtain a multi-dimensional visual feature matrix; Constructing an adjacency matrix based on the multidimensional visual feature matrix, and performing cluster analysis on the adjacency matrix by using a graph clustering algorithm to locate the fuzzy perception area of ​​the chip to be tested; Combining the initial detection data with the multi-dimensional visual feature matrix to perform potential defect identification on the fuzzy perception area to obtain potential defect heat information; Performing an adaptive analysis of the potential defect heat information at a supplementary visual angle, and simultaneously constraining the adaptive analysis result based on the initial visual angle with an information gain condition to generate a supplementary visual angle; Performing laser scattering detection and photothermal microscopy detection on the chip to be tested according to the supplementary visual angle to obtain supplementary detection data of the chip to be tested; Defect identification is performed on the fuzzy perception area according to the supplementary detection data and the potential defect heat information to obtain a defect identification result of the chip to be tested.

[0005] In a second aspect, the present invention provides a high-precision defect detection system for semiconductor chips, which is used to implement the high-precision defect detection method for semiconductor chips described in any one of the first aspects, comprising: An initial detection module, used to perform laser scattering detection and photothermal microscopy detection at an initial visual angle on the chip to be tested, so as to obtain initial detection data of the chip to be tested; A feature extraction module is used to extract key visual features from the initial detection data, and perform feature vector conversion and data integration on the key visual features to obtain a multi-dimensional visual feature matrix; A cluster analysis module, configured to construct an adjacency matrix based on the multidimensional visual feature matrix, and perform cluster analysis on the adjacency matrix using a graph clustering algorithm to locate the fuzzy perception area of ​​the chip under test; A defect perception module, used for combining the initial detection data with the multi-dimensional visual feature matrix to identify potential defects in the fuzzy perception area to obtain potential defect heat information; An angle analysis module, configured to perform adaptive analysis of the potential defect heat information at a supplementary visual angle, and simultaneously constrain the adaptive analysis result based on the initial visual angle with an information gain condition to generate a supplementary visual angle; a supplementary detection module, configured to perform laser scattering detection and photothermal microscopy detection on the chip to be tested according to the supplementary visual angle to obtain supplementary detection data of the chip to be tested; A defect recognition module is used to perform defect recognition on the fuzzy perception area according to the supplementary detection data and the potential defect heat information to obtain a defect recognition result of the chip to be tested.

[0006] The present invention provides a high-precision defect detection method for semiconductor chips, which has the following beneficial effects: The present invention obtains initial detection data through laser scattering and photothermal microscopy, extracts key visual features and converts them into a multidimensional visual feature matrix, uses a graph clustering algorithm to locate fuzzy perception areas and identify potential defects, performs adaptive analysis of supplementary visual angles and generates supplementary detection data, and finally identifies defects through the supplementary data and thermal information to obtain chip defect results. This method effectively improves the accuracy and reliability of defect detection through multimodal data fusion and adaptive analysis, and has obvious advantages in the identification of complex and fuzzy areas, solving the problem of insufficient accuracy of single-path optical imaging detection in the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 This is a schematic diagram of the steps of a high-precision defect detection method for semiconductor chips provided by an embodiment of the present invention; Figure 2 It is a structural diagram of a high-precision defect detection system for semiconductor chips provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0008] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0009] The implementation of the present invention is described in detail below with reference to specific embodiments.

[0010] Reference Figure 1 、 Figure 2 As shown, a preferred embodiment of the present invention is provided.

[0011] In a first aspect, the present invention provides a high-precision defect detection method for semiconductor chips, comprising: S1: performing laser scattering detection and photothermal microscopy detection at an initial visual angle on the chip to be tested to obtain initial detection data of the chip to be tested; S2: extracting key visual features from the initial detection data, and performing feature vector conversion and data integration on the key visual features to obtain a multi-dimensional visual feature matrix; S3: constructing an adjacency matrix based on the multidimensional visual feature matrix, and performing cluster analysis on the adjacency matrix by using a graph clustering algorithm to locate the fuzzy perception area of ​​the chip to be tested; S4: combining the initial detection data with the multi-dimensional visual feature matrix to perform potential defect identification on the fuzzy perception area to obtain potential defect heat information; S5: performing adaptive analysis of the potential defect heat information at a supplementary visual angle, and performing information gain condition constraints on the adaptive analysis results according to the initial visual angle to generate a supplementary visual angle; S6: performing laser scattering detection and photothermal microscopy detection on the chip to be tested according to the supplementary visual angle to obtain supplementary detection data of the chip to be tested; S7: performing defect identification on the fuzzy sensing area according to the supplementary detection data and the potential defect heat information to obtain a defect identification result of the chip to be tested.

[0012] Specifically, in step S1 of the embodiment provided herein, an appropriate initial viewing angle is selected based on the type and size of the chip to be tested, as well as the testing requirements. This angle is typically parallel or perpendicular to the chip surface and covers the critical testing area. Selecting a representative viewing angle ensures that the test data covers a large area of ​​the chip surface, avoiding missing areas where defects may exist. This initial angle is typically the optimal testing angle verified by experimentation.

[0013] More specifically, a laser source is used to illuminate the surface of the chip to be tested, and the tiny features on its surface are detected through laser scattering. The scattered light of the laser is collected by the sensor to form detection data. Laser scattering is a commonly used method for detecting chip surface defects because it can detect changes in the surface microstructure and reflect defects such as cracks and pores. This detection method is very sensitive to changes in chip surface features and can provide high-resolution surface defect information.

[0014] More specifically, when laser irradiates the chip, the temperature of the chip surface will change due to the thermal effect. The infrared thermal imaging module is used to detect the thermal image of the chip surface and generate thermal imaging data. Photothermal microscopy technology can reveal the temperature changes on the chip surface caused by thermal effects, especially the abnormal local heat distribution caused by defects. Thermal image data helps to detect potential defect areas (such as microcracks or material defects) that are difficult to identify in conventional visual inspection.

[0015] More specifically, by combining the results of laser scattering detection and photothermal microscopy, all initial detection data are collected, including surface scattering characteristics, thermal image data and other relevant information. Combining the two detection methods can obtain more information, make up for the shortcomings of a single technology, and thus fully understand the surface condition and possible defects of the chip to be tested.

[0016] More specifically, the initial detection data is stored in a data processing system and preprocessed to remove noise, correct system errors, etc., in preparation for subsequent feature extraction and analysis. The original detection data may contain noise or errors. Preprocessing can improve the quality and accuracy of the data and ensure the effectiveness of subsequent analysis.

[0017] It is understandable that the combination of laser scattering detection and photothermal microscopy provides data in two dimensions: surface morphology and thermal diffusion. It can fully understand the state of the chip surface and greatly improve the accuracy of defect detection. Some tiny defects or thermal distribution anomalies (such as microcracks, uneven material distribution, etc.) may be difficult to detect under conventional visual inspection. Photothermal microscopy helps detect these defects through thermal imaging, which supplements the shortcomings of laser scattering detection. Both laser scattering and photothermal microscopy are non-destructive testing methods that can obtain complete detection data without destroying the chip. They are suitable for application scenarios where the integrity of the chip needs to be maintained. By combining laser scattering and photothermal microscopy detection technology, different types of defect information can be obtained, such as surface microstructure changes and thermal property changes, thereby enhancing the reliability of defect detection.

[0018] More specifically, laser scattering detection provides high-precision surface feature data and can identify tiny surface defects such as cracks and pores. Through thermal imaging technology, it can identify local temperature abnormalities on the chip surface and locate possible defects, especially in places where the thermal characteristics are different from normal areas. After combining laser scattering and photothermal microscopy detection data, the detection system can more accurately determine the type of defects, especially those subtle defects that may be missed by traditional visual inspection. The initial detection data provides the necessary input data for subsequent feature extraction, graph clustering analysis and defect identification. Through the combined analysis of multi-dimensional data, the detection results can be further optimized and the probability of false detection and missed detection can be reduced.

[0019] Specifically, in step S2 of the embodiment provided by the present invention, the initial detection data obtained from laser scattering detection and photothermal microscopy detection is preprocessed to remove noise and perform data correction. Common preprocessing methods include filtering, denoising, normalization, and standardization. The raw data may contain noise or other irrelevant information, which may affect the subsequent feature extraction and analysis process. Data preprocessing can improve the quality and accuracy of the data, ensure that the extracted features are more representative and effective, improve data quality, eliminate the interference of noise on feature extraction, and make subsequent feature extraction more accurate.

[0020] More specifically, based on the preprocessed data, key visual features are extracted from the laser scattering and photothermal microscopy data. These features may include surface texture features, thermal distribution features, shape features, edge features, texture roughness, etc. Laser scattering data features include surface details, texture changes, surface microcracks, etc. Photothermal microscopy data features include temperature distribution, hot spot areas, thermal diffusion characteristics, etc. These features can reflect tiny defects, material inhomogeneity and thermal distribution anomalies on the chip surface, and are key data for judging defect types and locating defect areas. By extracting representative visual features from multimodal detection data, the surface and thermal imaging characteristics of the chip can be fully described, which facilitates subsequent defect classification and analysis.

[0021] More specifically, the extracted key visual features are converted into vector representations in some way. Common methods include principal component analysis (PCA), linear discriminant analysis (LDA) or feature encoding (such as one-hot encoding or embedded encoding). Principal component analysis (PCA) is used to reduce feature dimensions and retain the main information of the data. Linear discriminant analysis (LDA) is used to improve the distinction between different categories. Feature encoding converts discrete features or categorical data into numerical data and high-dimensional features into low-dimensional feature vectors, which helps to reduce computational complexity while retaining the main information of the data, facilitating subsequent analysis and machine learning model processing. Feature vector conversion can reduce redundancy, improve data processing efficiency, and maintain key information of the data to avoid information loss.

[0022] More specifically, the feature vectors extracted from the laser scattering data and the photothermal microscopy data are integrated to form a unified multidimensional visual feature matrix. This matrix can contain all the eigenvalues ​​corresponding to each sample (chip detection data). It is usually a two-dimensional matrix, in which each row represents a sample and each column represents a feature. The laser scattering data feature vector is combined with the photothermal microscopy data feature vector to form a complete feature representation. Integrating data from different detection methods into a unified matrix can fully utilize the advantages of multimodal data and avoid the limitations of a single data source, thereby providing more comprehensive information. Through data integration, a multidimensional feature matrix is ​​constructed that can simultaneously reflect the key information of surface and thermal imaging data, making defect identification and classification more comprehensive and accurate.

[0023] More specifically, the extracted key features can provide key information about the chip surface and thermal imaging information, helping the detection system identify potential defects. By extracting key information from the data, the chip status can be effectively described, facilitating subsequent defect identification. Feature conversion reduces data redundancy and improves computing efficiency, making feature data adaptable to machine learning model processing. Through feature vector conversion, the efficiency of data in model processing is ensured, while retaining important information, which helps to improve the performance of subsequent models.

[0024] More specifically, integrating features from different detection methods into a unified matrix can comprehensively consider surface features and thermal features, provide more comprehensive detection data, and generate a multi-dimensional feature matrix, so that the model can process rich feature data, which helps to improve the accuracy and robustness of defect detection.

[0025] It is understandable that by extracting surface features and thermal features and combining them with multimodal data, potential defects in chips can be identified more comprehensively. Even if certain defects are not obvious in a certain mode, data from another mode may reveal their abnormal characteristics. Through feature vector conversion, the redundant part in the data is reduced, making subsequent processing more efficient. The multi-dimensional visual feature matrix formed after data integration makes the feature information more systematic, facilitates the processing of machine learning algorithms, and improves overall computing efficiency.

[0026] Specifically, in step S3 of the embodiment provided by the present invention, the adjacency matrix is ​​a representation of a graph. Each element in the matrix represents the relationship between nodes (here, each detection point or sample on the chip). For each pair of samples, if their features are similar, the corresponding position in the adjacency matrix is ​​assigned a larger value, indicating their "proximity." Common calculation methods include Euclidean distance, cosine similarity, and Pearson correlation coefficient. The Euclidean distance between each pair of samples is calculated. If the distance is small, the corresponding adjacency matrix value is large, indicating that the two points are relatively similar in the feature space. Gaussian kernel functions are often used to calculate similarity. The adjacency matrix is ​​the basis of graph clustering. The adjacency matrix records the similarity information between samples and can reflect the local structure between samples. Therefore, constructing a suitable adjacency matrix is ​​the first step in cluster analysis. By defining the adjacency matrix, it is possible to capture the similarities or dissimilarities between samples and reflect the relationships between different regions on the chip surface, especially similarity in multidimensional visual feature space.

[0027] More specifically, each element (weight) in the adjacency matrix represents the similarity between samples. To enhance clustering effectiveness, these weights may need to be normalized or smoothed. Common approaches include: normalization: scaling the element values ​​in the adjacency matrix to the range [0, 1]; sparsification: removing similarities below a certain threshold and retaining only adjacencies above that threshold to reduce computational complexity. Constructing a weighted adjacency matrix can help reduce computational complexity and ensure that only nodes (detection points) with high similarity participate in clustering calculations, enhancing clustering effectiveness and computational efficiency. After normalization and sparsification, the information in the adjacency matrix will be more accurate and effective, facilitating subsequent clustering analysis.

[0028] More specifically, a Laplace matrix is ​​constructed. The Laplace matrix is ​​the basic characteristic matrix of a graph and can be calculated through the adjacency matrix and the degree matrix. The degree matrix is ​​a diagonal matrix. Each element in the matrix represents the degree of a node (that is, the number of edges connected to the node). The Laplace matrix is ​​a key input for the graph clustering algorithm. It helps describe the structure of the graph and the relationship between the nodes. Through the Laplace matrix, the local and global structural features in the data can be captured. Constructing the Laplace matrix can provide an effective graph structure representation for the graph clustering algorithm, making the subsequent clustering process more stable and effective.

[0029] More specifically, a graph clustering algorithm (such as spectral clustering) is used to analyze the Laplacian matrix. The core idea of ​​spectral clustering is to solve the eigenvalues ​​and eigenvectors of the Laplacian matrix, select the eigenvectors that can best distinguish different clusters, and use these eigenvectors to project the samples into a new space. Common steps include: calculating the eigenvalues ​​and eigenvectors of the Laplacian matrix, selecting the top k eigenvectors (usually the k with the smallest eigenvalues), forming a new feature matrix with these eigenvectors, and performing K-means clustering or other clustering algorithms on these new features. Spectral clustering can effectively process high-dimensional data. By analyzing the graph structure, it can automatically group samples according to their similarity. It is particularly suitable for processing data with complex structures. It can automatically discover potential clusters from the data and avoid the problem of artificially set cluster numbers. Through graph clustering, the areas of the chip to be tested can be accurately classified to help identify different fuzzy perception areas, especially those that are difficult to identify by traditional methods.

[0030] More specifically, based on the results of the atlas clustering algorithm, the characteristics of each cluster are analyzed, especially the clustering of fuzzy perception areas. These areas usually have similar characteristics (such as surface texture, thermal imaging characteristics, etc.) and may have defects or characteristic changes. Through the clustering results, these fuzzy perception areas can be located and further verified and analyzed. The areas obtained by clustering can help locate and analyze defective areas or special areas that may exist in the chip under test, especially fuzzy perception areas. These areas are not easily recognized under the initial visual angle detection. This difficult-to-recognize state may be due to the influence of tiny defects in the chip. Therefore, there is a risk of potential defects in the fuzzy perception area. Accurately identifying and locating the fuzzy perception area in the chip will help subsequent defect repair, quality control and performance optimization.

[0031] It can be understood that the adjacency matrix is ​​the basis of graph clustering, which reflects the similarity between chip detection data points. By calculating the similarity between samples, the adjacency matrix helps us capture the relationship between different areas in the chip during clustering. The Laplace matrix helps capture the structural information of the data in the spectral space and reveals the local and global structure between samples. Through the construction of the Laplace matrix, graph clustering can effectively cluster the data according to the structure of the graph. The graph clustering method can divide the data into different clusters through spectral features and automatically discover potential structural patterns. Through graph clustering, the fuzzy perception areas and potential defect areas in the chip can be accurately identified and divided, thereby improving the accuracy of defect location. The clustering results help locate areas with abnormalities or defects on the chip surface, especially those fuzzy areas that are difficult to identify by traditional methods. Accurately locating fuzzy perception areas will help subsequent analysis and quality control, thereby improving the reliability and performance of the chip.

[0032] Specifically, in step S4 of the embodiment provided by the present invention, the initial detection data and the multidimensional visual feature matrix are aligned and standardized. The initial detection data usually includes surface defect images, thermal imaging data, etc., which may contain noise and inconsistency. Therefore, these data need to be preprocessed, such as denoising, filtering, normalization, etc., to ensure the comparability of the data. The initial detection data may come from different sensors or different measurement methods, and the scale and characteristics of the data may be different. Standardization can eliminate the differences between different data sources and ensure the consistency of subsequent processing. The standardized data can eliminate various types of noise and errors, so that data from different sources can be effectively integrated, thereby improving the reliability of subsequent analysis.

[0033] More specifically, the initial inspection data (such as thermal imaging data and surface scan images) is fused with a multidimensional visual feature matrix. Typically, a weighted summation or other feature fusion method is used to combine each dimension of the multidimensional visual feature matrix (such as texture, brightness, color, and edges) with the initial inspection data. The fused data serves as input for further analysis. The multidimensional visual feature matrix provides multidimensional perceptual information about various areas on the chip surface, while the initial inspection data provides direct physical measurements (such as heat distribution). By fusing these two, a comprehensive understanding of the chip's state can be achieved, improving the accuracy of defect identification. The fused data provides a more comprehensive perspective. Combining visual features with actual inspection data can more accurately locate potential defect areas, especially those difficult to identify using a single data source.

[0034] More specifically, potential defect features are extracted from the fused data, and algorithms (such as edge detection, texture analysis, and thermal analysis) are used to identify potential defect areas on the chip surface. The features of these defect areas are then matched with thermal information (e.g., abnormal hot spots). Key features can be extracted using deep learning algorithms or traditional image processing techniques, such as convolutional neural networks (CNNs), SIFT, and SURF. By extracting features from the fused data, potential defects can be identified in fuzzy perception areas. These defects often appear as abnormal hot spots in thermal images or as tiny flaws in visual images. Extracting accurate features is the foundation for identifying potential defects. By extracting features and mapping them to thermal data, surface defects and potential problems can be identified, especially in fuzzy areas, reducing the possibility of missed detections and false detections.

[0035] More specifically, based on the extracted defect features, the thermal information of potential defects is calculated. The thermal information is usually determined by comparing the temperature difference between the current detection area and the normal area. Areas with large temperature abnormalities are considered to be potential defect areas, which may have problems such as material unevenness and functional failure. At this time, machine learning models (such as regression analysis, support vector machines, etc.) can be combined to analyze the relationship between defects and heat. Thermal imaging data reflects the heat distribution on the chip surface. Abnormal hot areas are usually associated with potential defects (such as current overload, short circuit or material damage, etc.). The calculation of thermal information can help accurately locate the defect area and infer its potential cause. The calculation of thermal information can intuitively identify and mark the potential defect area, especially by establishing an effective correspondence between thermal imaging images and visual images, thereby improving the accuracy of defect recognition.

[0036] More specifically, a heat threshold is set to determine whether an area has a potential defect. This threshold can be determined through experimental data or historical defect records, or it can be dynamically adjusted based on the characteristics of the data distribution. Areas with heat exceeding the threshold are considered potential defect areas. By setting the heat threshold, areas with abnormal heat can be automatically screened out, reducing manual intervention and improving the efficiency of defect detection. Through scientific threshold setting, the relationship between missed detection and false detection can be effectively balanced. By determining the heat threshold, abnormal heat areas can be accurately located, improving the accuracy and efficiency of automated detection and reducing labor costs.

[0037] More specifically, cluster analysis is performed on potential defect areas identified by heat. Graph-based clustering methods (such as spectral clustering or density clustering) are typically used to group defect areas. Each cluster represents a possible defect area. After clustering, further analysis can be performed, such as defect type classification and defect severity assessment. Clustering can help merge multiple related defect areas into a larger potential defect area, facilitating subsequent repair or quality control decisions. In addition, cluster analysis can identify the spatial distribution patterns of defects and optimize production processes. Cluster analysis can help systematically organize and evaluate potential defect areas, avoiding misjudging small, scattered defects as single defects, thereby improving the accuracy and globality of defect location.

[0038] More specifically, the heat information of potential defect areas is visualized so that engineers or automated systems can intuitively observe the problem areas. Common visualization methods include heat maps, 3D images, and defect area markers. These images can be superimposed on the original chip image or thermal imaging image to help analysts quickly locate defects. Visualization allows analysts to discover potential problems more quickly and intuitively, especially in complex multi-dimensional data and fuzzy perception areas. Heat maps and other methods can more clearly mark abnormal areas. Through visualization, potential defect areas and their heat distribution can be effectively displayed, helping decision makers make accurate judgments when optimizing chip design or production, and providing a basis for further defect repair.

[0039] It is understandable that fusing data from different sources and standardizing them makes the data comparable and reduces the impact of noise on subsequent analysis. Standardized data can ensure data quality and provide stable and reliable input for subsequent defect identification. Through feature extraction and heat calculation, potential defect areas can be effectively identified, and the temperature anomaly information of these areas can be extracted. Accurate defect identification can improve the accuracy of defect detection in fuzzy perception areas, avoid missed detection and false detection, set a reasonable heat threshold, and combine clustering algorithms to locate defect areas, which helps to improve the accuracy of automated detection systems. Through heat determination and clustering analysis, defect areas can be accurately identified and located, providing a basis for subsequent repair and optimization. Through visualization technology, a clear defect area map is provided to help engineers or automated systems make decisions. Visualization helps analysts discover defects more quickly and intuitively, improving defect repair efficiency. Through these steps, potential defects in fuzzy perception areas can be efficiently identified and located, and provide an accurate basis for subsequent quality control, repair or optimization.

[0040] Specifically, in step S5 of the embodiment provided by the present invention, the effectiveness of the existing initial visual angle (such as thermal imaging, surface image, etc.) in identifying potential defects is evaluated. This is usually done by analyzing the relationship between the existing detection results and the target defect. The selection of the initial visual angle may be determined based on the characteristics of the defect (such as size, type, and position), and the shortcomings of the existing angle need to be determined. The initial visual angle is the viewing angle or sensor data selected by the system during the first detection. It may not be able to fully cover all potential defect areas, especially in complex or occluded areas. By analyzing the limitations of the existing visual angle, it can provide guidance for the subsequent selection of supplementary visual angles. Analyzing the adaptability of the initial visual angle can help locate which defects are missed or misidentified at this angle, providing data support and direction for the selection of supplementary visual angles.

[0041] More specifically, the selection criteria for supplementary visual angles typically consider the following factors: View coverage: Selecting a complementary viewing angle to the initial viewing angle to cover areas not effectively captured by the initial angle; Defect type: Determining the focus of the supplementary angle based on the characteristics of different defects (such as surface cracks, thermal damage, surface contamination, etc.); Resolution and sensitivity: Selecting a visual angle with high resolution that can capture changes in detail to improve the ability to detect small or subtle defects; The selection criteria for supplementary visual angles need to take into account multiple aspects to ensure comprehensive coverage of potential defect areas and improve detection accuracy. Selecting the right angle is crucial to improving the overall performance of the inspection system. By selecting the right supplementary visual angle, the shortcomings of the initial angle can be compensated, and the comprehensiveness and accuracy of defect recognition can be improved, especially in complex inspection scenarios.

[0042] More specifically, image processing or deep learning methods (such as convolutional neural networks) are used to analyze the adaptability of supplementary visual angles. These methods can identify defects that cannot be detected at the initial visual angle by processing image data at different angles, and determine the effectiveness of supplementary angles. The recognition effect of defects at different angles can be evaluated by analyzing the heat, shape and position of the defects. Multi-angle data analysis can effectively reveal the differences in the performance of defects at different visual angles, providing a scientific basis for the selection of supplementary visual angles. Deep learning methods can automatically learn from the data which angles are most effective for defect recognition. Through adaptive analysis methods, the contribution of each supplementary angle to defect recognition can be quantified to ensure that the most advantageous supplementary angle is selected, thereby improving the accuracy and robustness of defect recognition.

[0043] More specifically, information gain analysis is performed on the initial visual angle to evaluate the importance of feature information in the initial visual angle in defect recognition. Information gain analysis is usually based on feature selection algorithms, such as decision trees, information gain calculations, and other methods, to determine which visual features (such as heat, color, texture, etc.) contribute most to defect recognition. Information gain can help us understand the useful information contained in the initial visual angle, evaluate its contribution to defect recognition, and identify which features affect the information constraints of subsequent supplementary angles, thereby improving the efficiency of the overall system. Through information gain analysis, the feature selection and data processing processes can be optimized, thereby improving the accuracy and computational efficiency of defect recognition and avoiding redundant and invalid feature interference.

[0044] More specifically, based on the information gain result of the initial visual angle, the conditional constraint rules for the supplementary visual angle are set. Specifically, the degree of complementarity of the supplementary visual angle to the initial visual angle information can be set to constrain the difference between the information of the supplementary angle and the initial angle. For example, the validity of the new information can be ensured by constraining the data of the supplementary angle to not highly overlap with the initial angle. By setting the information gain condition constraint, it can be ensured that the information provided by the supplementary visual angle avoids redundancy while improving the detection accuracy, thereby improving the overall efficiency of the system. The information gain condition constraint helps to ensure that the supplementary angle provides valuable new information, avoids excessive redundancy and repetition of information, and improves the detection efficiency and performance of the system.

[0045] More specifically, supplementary visual angles are generated based on conditional constraint rules and adaptive analysis results. The generation of supplementary visual angles can be achieved by adjusting the camera position, adopting different types of sensors, or using different imaging technologies (such as infrared imaging, laser scanning, etc.). This process usually relies on model prediction or image synthesis technology. The supplementary visual angles need to be optimized according to the shortcomings of the initial visual angles to enhance the ability to recognize defects. By generating supplementary angles, it can be ensured that the detection system can fully capture defects from multiple perspectives. The generated supplementary visual angles can effectively make up for the shortcomings of the initial visual angles, improve the sensitivity and adaptability of the system, and enhance the comprehensiveness of defect detection.

[0046] It is understandable that analyzing the adaptability of existing visual angles and identifying the areas where defects are missed can provide a basis for the subsequent selection of supplementary angles, ensuring that the evaluation of the initial angle can provide a scientific basis for the selection and optimization of supplementary angles. By analyzing the recognition effects of defects at different angles, supplementary angles that are complementary to the initial visual angles can be selected to improve the coverage and accuracy of defect detection, optimize the selection of supplementary angles, and improve the comprehensiveness and accuracy of defect detection, especially in complex and occluded areas. Through information gain analysis and conditional constraint rules, it is ensured that the supplementary angles provide the system with effective new information rather than redundant or repeated information, generate supplementary visual angles, improve the accuracy of defect recognition, and improve the efficiency of the system. Through these steps, on the basis of existing visual angles, supplementary visual angles can be added to improve the comprehensive recognition ability of potential defects. Information gain constraints ensure that the supplementary angles effectively improve system performance, and ultimately enhance the accuracy and efficiency of defect detection.

[0047] Specifically, in step S6 of the embodiment provided by the present invention, laser scattering detection irradiates the chip surface with a laser beam and measures the intensity and angular distribution of the scattered light. When there are defects on the chip surface (such as cracks, contaminants, deformation, etc.), the pattern of the scattered light will change, and the detection system can identify defects based on these changes. Photothermal microscopy detection combines laser and infrared technology. Through a high-precision microscope system, laser is used to irradiate the chip surface to stimulate local temperature changes. Defective areas usually lead to abnormal accumulation of local heat. The thermal response is related to the location and nature of the defect and can be detected by thermal imaging or infrared imaging.

[0048] More specifically, based on the aforementioned supplementary visual angle analysis, the optimal viewing angle of laser scattering detection at the supplementary angle is determined. The supplementary visual angle is usually used to make up for the defect detection blind spot of the initial angle. Therefore, laser scattering detection needs to be performed from different angles to ensure that all potential defect areas are covered. Laser scattering detection relies on the interaction between the laser and the chip surface. Therefore, by incident laser from different visual angles, surface defects at different angles can be detected. The supplementary angle helps to cover defect areas that cannot be covered by the initial detection perspective. By performing laser scattering detection at the supplementary visual angle, defects can be identified more comprehensively, avoiding missed detections due to blind spots in the viewing angle, and improving the accuracy of defect detection.

[0049] More specifically, under the supplementary visual angle, the laser irradiation angle and the position of the detector are adjusted to obtain new laser scattering data. When collecting data, attention is paid to recording changes in the scattered light intensity, especially scattering patterns related to defects (such as abnormal changes in scattering angle and light intensity). Laser scattering detection under the supplementary visual angle can capture defect features that the initial angle fails to reveal, especially in those areas that are difficult to observe from the initial angle. Obtaining laser scattering data under the supplementary angle helps to reveal potential microscopic defects on the chip surface, such as cracks, corrosion or surface stains, and improve the comprehensiveness and accuracy of detection.

[0050] More specifically, similarly, the impact of the supplementary visual angle on photothermal microscopy is analyzed. The supplementary angle needs to select an appropriate detection angle based on the surface characteristics of the chip and possible defect areas. At this time, the focus is on the changes in thermal response under the supplementary angle, especially the temperature changes in the defect area. The imaging effect of photothermal microscopy is closely related to the angle of laser irradiation and the viewing angle of the sensor. Through the supplementary visual angle, thermal imaging data at different positions and angles can be obtained, and different defect patterns can be captured. Through photothermal microscopy at the supplementary angle, defects that cannot be detected by the initial angle can be revealed, especially tiny thermal response changes, which helps to discover hidden defect areas and improve the accuracy of chip quality assessment.

[0051] More specifically, at a supplementary visual angle, a laser is used to illuminate the chip surface and monitor its thermal response. A high-resolution infrared microscope is used to collect thermal imaging data at different angles, paying particular attention to areas with abnormal thermal response to determine whether these areas correspond to potential defects in the chip. Photothermal microscopy at a supplementary angle can obtain temperature information that is not revealed from the initial angle, which is especially important for certain defect types that are difficult to detect using laser scattering detection (such as tiny cracks or local overheating areas). Photothermal microscopy can provide more detailed thermal response data. By supplementing the visual angle, the detection of chip micro-defects can be enhanced, especially those defects on the surface that are difficult to detect using traditional methods.

[0052] More specifically, the laser scattering detection data at supplementary angles are fused with the photothermal microscopy detection data. Data fusion techniques, such as weighted averaging, neural networks, decision trees, etc., can be used to combine detection data from different sources to form a more comprehensive detection result. Laser scattering detection and photothermal microscopy detection each have their own advantages. Laser scattering is good at identifying surface deformation and tiny cracks, while photothermal microscopy detection can provide more temperature change information. Data fusion can combine the advantages of both and improve the comprehensiveness and accuracy of detection. Through data fusion, detection information from different angles can be comprehensively considered, the health status of the chip can be comprehensively evaluated, and the overall performance and reliability of the detection system can be improved.

[0053] More specifically, machine learning algorithms (such as support vector machines, random forests, convolutional neural networks, etc.) are used to analyze the fused data to automatically identify potential defects in the chip. The analysis results can help determine the type, location, and severity of the defects. Through deep learning or machine learning methods, potential defect information can be extracted from a large amount of detection data and automatically classified and identified, thereby improving detection efficiency and accuracy. Through intelligent analysis, tiny defects in the chip can be accurately identified, providing decision support for subsequent chip quality control and optimized design.

[0054] It is understandable that by performing laser scattering detection and photothermal microscopy detection based on the supplementary visual angle, the defect area on the chip surface can be fully covered and the blind area under the initial angle can be compensated. This process is supplemented and integrated through the detection data from multiple angles, which can improve the detection accuracy and reliability of chip defects. Ultimately, the combination of supplementary visual angles with laser scattering and photothermal microscopy detection makes the identification of chip defects more comprehensive and accurate, avoids missed detection and improves the accuracy of chip quality assessment.

[0055] Specifically, in step S7 of the embodiment provided by the present invention, the fuzzy perception area generally refers to an area where the detection system cannot clearly identify or it is difficult to accurately locate defects during preliminary detection. These areas may show unclear signs of defects during detection due to factors such as signal noise, lighting conditions, visual angles or resolution limitations. In order to improve the accuracy of detection, it is necessary to use supplementary detection data and thermal information to further analyze these fuzzy areas.

[0056] More specifically, supplementary inspection data usually comes from laser scattering and photothermal microscopy, which provide detailed information from all angles of the chip surface. By covering different visual angles, supplementary inspection data makes originally blurry or difficult-to-identify areas clearer. For example, laser scattering inspection can reveal tiny surface deformations and cracks, while photothermal microscopy can provide thermal response information to help identify temperature anomalies caused by defects.

[0057] More specifically, thermal information can reflect the temperature distribution in different areas of the chip surface. Defective areas usually lead to local heat accumulation, resulting in temperature fluctuations. According to the thermal images obtained by photothermal microscopy, the temperature changes in potential defective areas are more significant. This information is crucial for locating and identifying defects. Potential defects in fuzzy perception areas usually manifest as thermal anomalies different from surrounding areas.

[0058] More specifically, specific data of the fuzzy perception area is extracted from the supplementary detection data, with special attention paid to thermal information and scattering data. In laser scattering detection, changes in scattering intensity and pattern are observed. Especially when the thermal information and scattering data are consistent, these fuzzy areas may be difficult to clearly identify due to the angle or resolution of the initial detection. The detection data and thermal changes under the supplementary angle can help overcome these problems and re-examine these fuzzy areas from different angles.

[0059] More specifically, based on supplementary data, especially the thermal information obtained from photothermal microscopy, the thermal response of the fuzzy perception area is analyzed. By analyzing the thermal response pattern, such as temperature gradient, thermal wave propagation speed, temperature peak, etc., it is further confirmed whether these areas are defective areas. In the fuzzy area, the presence of defects may lead to abnormal changes in local temperature. By in-depth analysis of the thermal information, it can be clearly identified whether there is an abnormal thermal response caused by the defect.

[0060] More specifically, the data from laser scattering detection and photothermal microscopy are fused, and combined with thermal information, machine learning algorithms (such as support vector machines, convolutional neural networks, etc.) are used to identify defects in fuzzy perception areas. The features of different areas are classified and identified through the algorithm to determine whether they are potential defect areas. Data fusion can comprehensively analyze fuzzy areas from multiple dimensions (such as thermal response, scattering intensity, etc.), while machine learning algorithms can more effectively handle complex pattern recognition tasks and automatically extract valuable defect information from large amounts of data.

[0061] More specifically, after being processed by the defect recognition algorithm, the defects of the chip to be tested are located, classified and analyzed, and the defect recognition results are finally output, indicating the potential defect area of ​​the chip, the defect type (such as cracks, corrosion, contamination, etc.) and the severity of the defect. By accurately identifying defects in the fuzzy perception area, it can provide support for the subsequent quality control of the chip and ensure the reliability and accuracy of the test results.

[0062] More specifically, the identified defect areas and thermal anomaly information are visualized to generate a heat map or scatter plot, marking the location of the defect area. The thermal information can be displayed through a heat map or thermal imaging diagram. Combined with the visualization results of the scattering data, the intuitiveness of defect identification is enhanced. Graphical methods can help engineers more intuitively understand the location and characteristics of the defects, facilitating further processing and decision-making.

[0063] It is understandable that by combining supplementary detection data and thermal information to identify defects in fuzzy perception areas, the detection accuracy of chip defects can be effectively improved, especially when traditional detection methods may have blind spots or fuzzy areas. The introduction of thermal information, combined with the comprehensive analysis of laser scattering and photothermal microscopy, provides strong support for identifying tiny, deep or complex defects. Ultimately, the use of intelligent algorithms to automatically identify defects can greatly improve detection efficiency and provide accurate data support for subsequent quality control and optimization. With the assistance of thermal information, potential tiny defects can be identified to avoid missed detection. Intelligent recognition and data fusion greatly improve the processing speed and adapt to the needs of large-scale chip detection. The fuzzy perception area can make up for the limitations of traditional detection methods through comprehensive analysis of supplementary detection data and thermal information, and comprehensively improve the defect recognition capability of the chip.

[0064] The present invention provides a high-precision defect detection method for semiconductor chips, which has the following beneficial effects: The present invention obtains initial detection data through laser scattering and photothermal microscopy, extracts key visual features and converts them into a multidimensional visual feature matrix, uses a graph clustering algorithm to locate fuzzy perception areas and identify potential defects, performs adaptive analysis of supplementary visual angles and generates supplementary detection data, and finally identifies defects through the supplementary data and thermal information to obtain chip defect results. This method effectively improves the accuracy and reliability of defect detection through multimodal data fusion and adaptive analysis, and has obvious advantages in the identification of complex and fuzzy areas, solving the problem of insufficient accuracy of single-path optical imaging detection in the existing technology.

[0065] Preferably, the step of performing laser scattering detection and photothermal microscopy detection at an initial visual angle on the chip to be tested to obtain initial detection data of the chip to be tested includes: S11: placing the chip to be tested at a position to be tested, and adjusting the laser scattering excitation source and the infrared thermal imaging module to move to a specified initial visual angle; S12: analyzing the detection modes of the laser scattering excitation source and the infrared thermal imaging module respectively according to the specification design information of the chip to be tested, so as to obtain laser detection parameters and infrared detection parameters; S13: configuring parameters of the laser scattering excitation source according to the laser detection parameters to perform laser scattering detection on the chip to be tested at an initial visual angle to obtain laser scattering detection data; S14: configuring parameters of the infrared thermal imaging module according to the infrared detection parameters to perform photothermal microscopic detection of the chip to be tested at an initial visual angle to obtain photothermal microscopic detection data; S15: combining the laser scattering detection data with the photothermal microscopy detection data to obtain initial detection data.

[0066] Specifically, this step involves precisely placing the chip to be tested at a specified position on the testing equipment to ensure stability and accuracy during the testing process. The laser scattering excitation source and the infrared thermal imaging module need to be adjusted to an initial visual angle. This angle is set based on the surface characteristics of the chip, the type of defect, and the layout of the chip. Visual angle adjustment can be achieved through a precise positioning system, which usually includes a mechanical rotating platform or an electric adjustment device to adjust the laser scattering excitation source and the infrared thermal imaging module to a specified angle to ensure that the surface of the chip to be tested is fully covered, especially areas where defects may exist on the chip. The initial visual angle is usually the standard angle when designing the detection system to ensure that data from multiple different angles can be fully integrated to reduce blind spots and errors. Through precise angle settings, the laser scattering and infrared thermal imaging modules can obtain stable and consistent detection results, ensuring the comprehensiveness and accuracy of the test.

[0067] More specifically, the detection modes of the laser scattering excitation source and the infrared thermal imaging module are analyzed based on the chip's specification information (such as size, material, structure, etc.). The chip specification design information generally includes details such as circuit design, packaging structure, and material selection. The parameter configuration of the laser scattering excitation source and the infrared thermal imaging module involves selecting appropriate light wavelength, light intensity, detection angle, resolution, etc. The detection mode is optimized based on the specific characteristics of the chip. Different chips react differently to laser and thermal imaging. Therefore, the detection mode needs to be adjusted according to the chip's specification information to obtain the best detection effect. This ensures that the laser wavelength, thermal imaging resolution, and parameter settings used are most suitable for the characteristics of the chip to be tested, avoiding errors or incomplete detection due to parameter mismatch, and ensuring that the detection mode matches the chip characteristics, thereby optimizing the performance of the laser scattering and thermal imaging modules and improving the accuracy and efficiency of detection.

[0068] More specifically, the laser scattering excitation source is configured according to the designed laser detection parameters, and the laser intensity, wavelength, focusing mode and scanning mode are adjusted. At the initial visual angle, the laser scattering excitation source emits a laser beam to illuminate the surface of the chip to be tested. The laser is scattered by the surface microstructure, and the scattered light is received by the sensor and the data is recorded. Laser scattering detection technology identifies defects based on the scattering phenomenon caused by the interaction between light and the surface of the object. By adjusting the parameters of the laser source, tiny surface defects such as cracks and potholes can be accurately detected. Laser scattering detection at the initial visual angle can help determine the scattering characteristics of the chip surface, provide data support for subsequent analysis, and obtain accurate laser scattering data. By analyzing the intensity and distribution of scattered light, tiny defects or foreign matter on the chip surface can be identified to ensure the meticulousness of the detection.

[0069] More specifically, the infrared thermal imaging module is configured according to the designed infrared detection parameters, which mainly involve infrared wavelength, imaging resolution, detection accuracy, scanning mode, etc. At the initial visual angle, the module uses infrared detection to detect the heat distribution on the surface of the chip to be tested and monitor the local heat changes caused by defects. Photothermal microscopy detects defects by detecting infrared radiation generated by heat changes on the surface of the object. Defects may cause local temperature increases. Infrared thermal imaging can capture these temperature changes and display them. Configuring the parameters of the infrared thermal imaging module can help capture the thermal response caused by defects more clearly, thereby improving the accuracy of defect detection. Accurate infrared thermal imaging data helps to identify temperature fluctuations on the chip surface caused by defects, provide positioning information of defect heat anomalies, and enhance the ability to identify chip defects.

[0070] More specifically, laser scattering inspection data and photothermal microscopy inspection data are combined and processed through data fusion technology to obtain complete initial inspection data. This step may involve methods such as data standardization, denoising, and weighting. Data fusion algorithms (such as weighted averaging, decision trees, and neural networks) can be used to comprehensively consider the results of the two inspection methods to improve the accuracy of defect identification. Laser scattering and photothermal microscopy inspections provide two important information dimensions: surface defects (scattered light) and thermal anomalies (thermal response). By fusing these two data, more complete defect information can be obtained. Data fusion can improve the robustness of inspection, reduce the errors that may be caused by a single inspection method, and enhance the accuracy of the final results. By combining the data of laser scattering and photothermal microscopy inspections, more comprehensive and accurate chip defect information can be obtained, thereby providing stronger data support for subsequent defect location and classification.

[0071] More specifically, laser is used to illuminate the chip surface, and defects are detected by analyzing the intensity and pattern of scattered light. The presence of defects will change the light scattering pattern of the surface. Laser scattering detection can identify surface defects by observing this change. Infrared thermal imaging technology is used to detect temperature changes on the chip surface. Defects usually cause local temperature increases or fluctuations. Photothermal microscopy detection identifies potential defects by detecting this thermal change.

[0072] It is understandable that through the above steps, laser scattering detection and photothermal microscopy can jointly provide comprehensive data about the surface of the chip to be tested. Each step is executed to ensure the comprehensiveness, accuracy and efficiency of the data, thereby improving the accuracy of chip defect identification and providing a reliable basis for subsequent chip quality control.

[0073] Preferably, the steps of extracting key visual features from the initial detection data and performing feature vector conversion and data integration on the key visual features to obtain a multi-dimensional visual feature matrix include: S21: performing regional refinement processing on the laser scattering detection data in the initial detection data, so as to calculate the local curvature of the data portion corresponding to each specific position of the laser scattering detection data, and obtain a surface curvature feature value set of the chip to be tested; S22: performing a coordinated operation of a first circuit consisting of Gaussian curvature mapping and Euler characteristic extraction and a second circuit consisting of mean curvature mapping and principal curvature difference analysis on the surface curvature feature value set simultaneously to generate surface topological variable features of the chip to be tested; S23: performing fast Fourier transform on the photothermal microscopy detection data in the initial detection data to extract fundamental frequency amplitude characteristics, phase delay characteristics, and harmonic distortion rate of each specific position of the chip to be tested; S24: performing heat conduction calculation on the chip under test based on fundamental frequency amplitude characteristics, phase delay characteristics, and harmonic distortion rate at specific locations of the chip under test to obtain heat diffusion distribution information of the chip under test; S25: performing deep frequency domain slicing processing on the heat diffusion distribution information to extract microscopic thermal field distortion features; S26: Construct a chip information space according to the specification design information of the chip to be tested, and use the surface topology variable characteristics and the microscopic thermal field distortion characteristics as key visual features to map the information of corresponding positions in the chip information space, so as to vectorize the key visual features in the chip information space and obtain a multi-dimensional visual feature matrix.

[0074] Specifically, the laser scattering detection data contains the intensity information of the reflected or scattered light on the chip surface. Regional refinement processing is a local processing of this data to better analyze the surface morphology of each position. Through local curvature calculation, the geometric features of different areas of the surface (such as protrusions, depressions, edges, etc.) can be evaluated to obtain a set of characteristic values ​​of the chip surface curvature. Curvature is an important parameter for describing surface morphology and structure. Local curvature calculation can reveal changes in the microstructure of the chip surface and further provide key clues for subsequent defect detection. This refinement processing enables the detection system to more accurately capture surface changes in different areas to avoid missing any potential defects. By calculating the local curvature characteristics of the surface, more detailed and local surface feature information can be obtained, providing data support for subsequent feature extraction and analysis.

[0075] More specifically, the surface curvature feature value set is simultaneously subjected to a collaborative operation of a first circuit consisting of Gaussian curvature mapping and Euler characteristic number extraction and a second circuit consisting of mean curvature mapping and principal curvature difference analysis to generate the surface topological variable characteristics of the chip to be tested. Gaussian curvature and Euler characteristic number are indicators commonly used in geometry to describe the bending characteristics of a surface. Gaussian curvature describes the local curvature information of each point on the surface, while Euler characteristic number is a measure to describe the topological structure characteristics of the entire surface. Mean curvature and principal curvature difference are parameters to describe the overall shape characteristics of the surface, which can reflect the concavity and convexity and change trend of the surface. The collaborative operation of these curvature features can integrate multi-dimensional curvature information to obtain more accurate surface topological features. The collaborative operation of Gaussian curvature and Euler characteristic, as well as mean curvature and principal curvature difference, is to more comprehensively characterize the topological features of the chip surface. These features are crucial for judging the stability, uniformity and possible defect areas of the chip surface. Collaborative operation helps to improve the accuracy and comprehensiveness of detection through comprehensive analysis of multiple curvature features. Through the generation of surface topological variable features, the complex geometric structure of the chip surface can be accurately depicted, providing more dimensional support for defect identification and positioning.

[0076] More specifically, the photothermal microscopy data reflects the heat distribution on the chip surface. Fast Fourier transform (FFT) is a commonly used signal processing technology used to convert time domain signals into frequency domain signals. In the frequency domain, the fundamental frequency amplitude characteristics represent the main frequency components of thermal fluctuations, the phase delay characteristics represent the phase difference between the thermal response and the excitation signal, and the harmonic distortion rate reflects the presence of higher harmonics in the thermal signal. Fast Fourier transform can reveal the frequency characteristics of the thermal response on the chip surface, which is crucial for evaluating thermal distribution and judging thermal changes caused by defects. Characteristics such as fundamental frequency, phase delay and harmonic distortion can provide in-depth information on the thermal characteristics of the chip, further enhancing the sensitivity of thermal imaging detection. By extracting the frequency domain characteristics of the thermal response through fast Fourier transform, we can accurately capture tiny thermal changes on the surface and provide basic data for subsequent heat conduction solutions and defect analysis.

[0077] More specifically, fundamental frequency amplitude, phase delay, and harmonic distortion rate, as key parameters of chip thermal characteristics, can be input into the thermal conduction model, and the thermal diffusion behavior of the chip surface can be calculated using the heat conduction equation. This process requires the use of numerical calculation methods (such as the finite element method, differential method, etc.) to solve the thermal conduction characteristics of the chip surface and obtain thermal diffusion distribution information. The thermal conduction solution is based on the heat change on the chip surface to evaluate its thermal diffusion characteristics. This thermal diffusion information is of great significance for analyzing the thermal stability and potential failures of the chip under different working environments. Through the thermal conduction solution, the thermal behavior of the chip during use can be predicted, which helps to evaluate its working performance and life under different temperature conditions. By calculating the thermal diffusion information, detailed data on the thermal distribution of the chip surface can be provided, providing an effective basis for further thermal field analysis and defect identification.

[0078] More specifically, deep frequency domain slicing is a further processing of the thermal diffusion distribution information in the frequency domain, aiming to analyze the microscopic distortion characteristics of the thermal field at different frequencies. Through slicing, the detailed characteristics of the thermal field changes can be obtained, especially the tiny thermal field distortion. Frequency domain slicing can reveal tiny changes and distortions in the thermal field. These distortions are often caused by tiny defects or unevenness on the chip surface. By extracting the microscopic thermal field distortion characteristics, the sensitivity of thermal imaging analysis can be further improved, helping to identify potential defects in the chip. Deep frequency domain slicing enhances the recognition ability of microscopic thermal field distortion and provides more detailed data support for subsequent defect analysis.

[0079] More specifically, based on the chip's specification design information, a virtual "chip information space" is constructed. This space is used to accommodate various visual features of the chip. Surface topological variable features and microscopic thermal field distortion features are used as key visual features. They are mapped to specific positions in the chip information space through information mapping technology, and finally vectorized to generate a multi-dimensional visual feature matrix. By constructing the chip information space and vectorizing different visual features, complex surface features and thermal response data can be converted into machine-processable numerical forms. Vectorized expression can simplify the data processing process and facilitate subsequent analysis and defect detection. The resulting multi-dimensional visual feature matrix contains comprehensive information on the chip's surface morphology and thermal characteristics, which can effectively support subsequent defect analysis, chip quality assessment, and performance prediction.

[0080] More specifically, local curvature indicates the degree of curvature of the surface at a certain point and can be calculated using laser scattering data. Local curvature is very sensitive to small surface changes such as defects and cracks. Gaussian curvature describes the local curvature of each point on the surface, and the Euler characteristic is a measure of the overall surface topological characteristics. Both are key parameters for evaluating surface geometric properties. Fast Fourier transform (FFT) is an efficient signal processing method that converts time domain signals into frequency domain signals, thereby revealing the frequency characteristics of the signal. Heat conduction solution is to solve the process of heat propagation on the chip surface through numerical methods to evaluate the thermal characteristics of the chip, such as thermal diffusion and thermal response. Deep frequency domain slicing processing is a technology for processing heat diffusion information. By analyzing the changes in the thermal field at different frequencies, the microscopic thermal field distortion characteristics are extracted.

[0081] It's clear that through the above steps, key visual features related to chip surface morphology and thermal response can be extracted from the initial inspection data. Vectorized processing then yields a multidimensional visual feature matrix. This feature matrix not only comprehensively supports chip defect detection but also provides data for chip thermal behavior, performance evaluation, and quality control.

[0082] Preferably, the step of constructing an adjacency matrix based on the multidimensional visual feature matrix and performing cluster analysis on the adjacency matrix by using a graph clustering algorithm to locate the fuzzy perception area of ​​the chip to be tested includes: S31: performing cosine similarity calculation on the multidimensional visual feature matrix to obtain vector similarity information of the multidimensional visual feature matrix, and constructing an adjacency matrix of the multidimensional visual feature matrix based on the vector similarity information; S32: performing feature clustering analysis on the adjacency matrix using a graph clustering algorithm, and performing vector clustering processing on the multidimensional visual feature matrix according to the feature clustering result, so as to divide the multidimensional visual feature matrix into a plurality of clusters; S33: Calculating the cluster density of each cluster, and performing cluster density difference amplitude analysis on adjacent clusters to obtain a cluster density characteristic value and an adjacent difference characteristic value of each cluster; S34: simulating the overall clustering environment based on the cluster density characteristic value of each cluster in the multidimensional visual feature matrix, and assigning a numerical value of information clarity based on the cluster density characteristic value to each cluster according to the simulation result, so as to generate information clarity of each cluster; S35: Adaptively adjusting the information clarity of each cluster based on the adjacent difference characteristic value of each cluster, and judging the adjusted information clarity according to a preset threshold, so as to locate the corresponding fuzzy perception area on the chip to be tested according to the cluster that does not meet the preset threshold.

[0083] Specifically, a cosine similarity calculation is performed on the multidimensional visual feature matrix to obtain vector similarity information of the multidimensional visual feature matrix, and an adjacency matrix of the multidimensional visual feature matrix is ​​constructed based on the vector similarity information. The cosine similarity between each vector in the multidimensional visual feature matrix is ​​calculated. Cosine similarity is an indicator that measures the directional similarity of two vectors. The value range is -1 to 1. The larger the value, the more similar the two vectors are. Then, an adjacency matrix is ​​constructed based on the calculated similarity information. Each element in the adjacency matrix represents the similarity between two vectors. Each row and column of the matrix corresponds to a visual feature vector. Cosine similarity can effectively measure the similarity between feature vectors and further determine the feature similarity of different regions. In this way, the similarity relationship between different regions on the chip surface can be captured. The adjacency matrix can intuitively represent the relationship between features and is the basis of graph clustering analysis. Through cosine similarity calculation and adjacency matrix construction, the information in the multidimensional visual feature matrix can be converted into a graph form, which facilitates subsequent clustering analysis. The adjacency matrix provides the necessary data structure for the graph clustering algorithm.

[0084] More specifically, a graph clustering algorithm (such as a spectral clustering algorithm) is used to perform cluster analysis on the multidimensional visual feature matrix based on the similarity information of the adjacency matrix. The graph clustering algorithm finds the natural clustering structure in the data by performing eigenvalue decomposition on the adjacency matrix, clusters the feature vectors into several clusters, and then divides the multidimensional visual feature matrix. The graph clustering algorithm can effectively identify the group structure in the data and cluster similar areas together based on feature similarity, which helps to accurately distinguish the feature differences between different areas on the chip surface. Feature clustering can simplify the subsequent analysis process and improve the detection accuracy of fuzzy areas. Through graph clustering, different areas of the chip can be divided into different clusters according to their visual features, so that each cluster represents an area with similar features, which provides a basis for further analysis and positioning of fuzzy areas.

[0085] More specifically, the density of each cluster is calculated, that is, the concentration of feature points within each cluster. Cluster density is typically determined by the number of points within the cluster and the spatial distribution of the cluster. The cluster density differences between different clusters are calculated. By analyzing the amplitude of the cluster density differences, the density characteristic values ​​and their differences between the different clusters are obtained. Cluster density is an indicator of the concentration of regional features. By calculating cluster density, regions with different feature densities on the chip surface can be identified. Analyzing cluster density differences helps determine whether the "boundaries" between clusters are clear, thereby discovering possible fuzzy areas on the surface. The calculation of cluster density and cluster density differences can provide a specific basis for subsequent fuzzy area positioning. High-density areas may correspond to clear areas, while low-density areas may correspond to fuzzy areas.

[0086] More specifically, based on the cluster density eigenvalues ​​of each cluster, the distribution of the entire cluster environment is simulated, the information clarity of each cluster in the overall environment is evaluated, and information clarity is assigned to each cluster based on the cluster density eigenvalues. Clusters with higher density may represent clear areas, while clusters with lower density may represent fuzzy areas. The density eigenvalues ​​of clusters reflect the information concentration of the cluster. By assigning information clarity to clusters, the visual information quality of each area can be intuitively evaluated. This step helps to provide a quantitative basis for fuzzy area positioning, so that clusters with lower clarity can be accurately marked, and the clarity score of each cluster is obtained, which can provide a basis for subsequent fuzzy area identification. A high-definition area means that the information quality of the area is higher, while a low-definition area may contain fuzzy or unclear features.

[0087] More specifically, based on the adjacent difference eigenvalues ​​of the clusters, adaptive adjustment of information clarity is performed. This means that if the density difference between a cluster and its adjacent clusters is large, the information clarity of the cluster will be adjusted accordingly. A preset threshold is set to judge the information clarity. Clusters below the threshold are considered to be fuzzy areas, and then the fuzzy perception areas are located. In actual chip detection, the definition of fuzzy areas is often closely related to the contrast differences of the surrounding areas. The adaptive adjustment of adjacent differences can make the determination of fuzzy areas more sensitive and accurate. The preset threshold helps to standardize the judgment of fuzzy areas in actual applications and ensure the reliability of detection. By adaptively adjusting the information clarity and combining it with threshold judgment, the fuzzy areas on the chip can be accurately identified, which helps to locate potential defects or problem areas in quality control and provide an efficient defect detection tool.

[0088] It can be understood that this technical solution accurately divides the visual features of the chip surface into multiple clusters through technical means such as cosine similarity, graph clustering, and cluster density analysis, and successfully locates the fuzzy perception area through analysis of information clarity and proximity differences. Through the effective combination of these steps, efficient detection of chip surface defects and precise positioning of fuzzy areas can be achieved, thereby providing strong support for chip quality control.

[0089] Preferably, the step of combining the initial detection data with the multi-dimensional visual feature matrix to identify potential defects in the fuzzy perception area to obtain potential defect heat information includes: S41: extracting corresponding information from the initial detection data and the multidimensional visual feature matrix according to the fuzzy perception area, so as to extract first identification information and second identification information corresponding to the fuzzy perception area from the initial detection data and the multidimensional visual feature matrix; S42: performing probability analysis of each defect type on the first identification information according to a pre-built chip defect manifestation analysis model to obtain a probability value of each defect type corresponding to the fuzzy perception area; S43: Digitally reproducing the corresponding defect types in the fuzzy perception areas according to the probability values ​​of the defect types in the fuzzy perception areas, so as to construct a digital model of defect representation; S44: extracting features from the digital model of defect manifestation of each defect type, and performing feature matching processing on the second identification information based on the extracted features to generate verification parameters for each defect type; S45: Combining the probability value of each defect type corresponding to the fuzzy perception area with the verification parameter, the potential defect heat information is obtained.

[0090] Specifically, based on the fuzzy perception area, information corresponding to the area is extracted from the initial detection data and the multidimensional visual feature matrix respectively. The initial detection data may contain some sensor data or other device measurement data, while the multidimensional visual feature matrix contains visual features in the image or video data. The extracted first identification information and second identification information correspond to certain key features of the fuzzy perception area respectively, which are used for subsequent analysis and processing. The fuzzy perception area represents certain unclear areas on the chip surface, which may have potential defects. Therefore, by extracting information related to the area from different data sources, the characteristics of the area can be analyzed more comprehensively. The first identification information and the second identification information can provide different types of data support for subsequent defect identification and positioning. By extracting information related to the fuzzy perception area, the necessary data basis can be provided for subsequent defect identification. The extracted information can help effectively locate the potential defect area and further analyze the defect type.

[0091] More specifically, pre-built defect manifestation analysis models are used. These models exhibit different characteristics and manifestations according to different types of defects (such as cracks, scratches, missing materials, etc.). According to the first identification information (features extracted from the initial detection data and the multidimensional visual feature matrix), a probability analysis of each defect type is performed. The model will calculate the probability value for each defect type based on the relationship between the first identification information and the defect type. Various defect types usually exhibit specific visual characteristics on the chip surface. By establishing a defect manifestation analysis model, the possible defect types and their probabilities in the area can be predicted based on the characteristics of the fuzzy perception area. The probability analysis can provide a basis for subsequent decision-making, ensure that the identified defect types and probabilities are more accurate, and obtain the probability values ​​of each defect type corresponding to the fuzzy perception area. These probability values ​​provide a quantitative basis for determining the defect type, which facilitates subsequent digital reproduction and feature matching.

[0092] More specifically, the fuzzy perception area is digitally reproduced according to the probability values ​​of each defect type obtained in step S42, which means that the expression of the area under different defect types is simulated through the probability value. Specifically, digital reproduction is to express the potential defects of the fuzzy perception area as a digital model through mathematical modeling or simulation of each defect type, thereby forming a digital model of defect expression. Digital reproduction can make the potential defect expression of the fuzzy perception area clearer. After conversion into a digital model, subsequent feature extraction and matching can be easier. Constructing a digital model of defect expression helps to fully understand the visual expression of defects and can provide a basis for further verification parameter generation. Through digital reproduction, the defect expression of the fuzzy perception area is digitized and standardized, making subsequent analysis and verification more efficient and accurate.

[0093] More specifically, by performing feature extraction on the digital models of each defect type, the key features of the defect manifestation, such as shape, size, texture, etc., are extracted, and the extracted features are feature matched with the second identification information to determine whether the defect manifestation matches the features of the actual detection area. Feature matching can be performed through various algorithms (such as similarity matching, machine learning, etc.). Feature extraction helps to extract the key information that best represents the defect from complex digital models. These features can help us compare with the actually detected features to confirm the defect type. Generating verification parameters through feature matching can enhance the judgment and verification of the defect type and improve the accuracy of recognition. The generated verification parameters can be used to confirm the specific type and characteristics of the defect, providing more accurate information for subsequent defect detection.

[0094] More specifically, the probability value of each defect type obtained in step S42 is combined with the verification parameter obtained in step S44, and the heat information of the potential defect is calculated based on the relationship between the two. The heat information is a comprehensive evaluation based on the probability value and the verification parameter, reflecting the possibility and severity of defects in the fuzzy perception area. Combining the probability value and the verification parameter can provide more comprehensive defect heat information. Taking into account the probability of occurrence of the defect type and the accuracy of the verification, the defect risk of the area can be effectively judged. The heat information can provide a basis for subsequent decision-making (such as defect repair, quality control, etc.), help optimize the production process and improve detection efficiency. By combining the probability and verification parameters, the potential defect heat information of the fuzzy perception area is obtained. This information can provide accurate guidance for chip quality inspection and defect analysis, and help improve the efficiency of defect location and repair.

[0095] As you can understand, this technical solution uses a multi-step process, starting with information extraction from the fuzzy sensing area, proceeding through defect type probability analysis, digital reproduction, feature extraction and matching, and ultimately generating potential defect heat information by combining probability values ​​and verification parameters. This method can efficiently identify and locate potential defects on the chip surface, providing a reliable basis for defect management and quality control. This process can significantly improve the accuracy and efficiency of defect identification and optimize defect monitoring and repair strategies during the production process.

[0096] Preferably, the step of performing adaptive analysis of the potential defect heat information at a supplementary visual angle and constraining the adaptive analysis result with an information gain condition according to the initial visual angle to generate a supplementary visual angle includes: S51: Acquire detection function information and visual angle range of a laser scattering excitation source for performing laser scattering detection and an infrared thermal imaging module for performing photothermal microscopy detection; S52: configuring a plurality of candidate visual angles based on the visual angle range, and analyzing optical detection characteristics caused by viewing angle changes for each of the candidate visual angles according to the detection function information, so as to generate optical detection characteristics for each of the candidate visual angles; S53: performing a potential defect detection effect analysis on the potential defect heat information according to the optical detection characteristics of each of the candidate visual angles to obtain an adaptability parameter of each of the candidate visual angles; S54: performing a coherence analysis on the detection effect of each candidate visual angle relative to the initial visual angle to generate an information gain weight of each candidate visual angle as a supplementary visual angle; S55: Sort the candidate visual angles according to the adaptability parameter, and analyze the feasibility of the visual angles in accordance with the sorting order and the corresponding information gain weights, so as to generate supplementary visual angles.

[0097] Specifically, information from two primary inspection modules—the laser scattering excitation source and the infrared thermal imaging module—is acquired and analyzed. These modules, used for laser scattering inspection and photothermal microscopy inspection, respectively, offer different physical principles and inspection capabilities. It is also necessary to determine the visual angle ranges of these modules, i.e., the angular ranges over which they can effectively inspect. Laser scattering inspection and photothermal microscopy are technical means for addressing different defect characteristics, such as surface irregularities and temperature variations. They have different detection capabilities and applicable angular ranges. Clarifying these inspection functions and visual angle ranges provides the necessary technical parameters for subsequent visual angle configuration and inspection characteristic analysis, and contributes to a comprehensive understanding of the capabilities and limitations of the inspection modules.

[0098] More specifically, based on the previously acquired visual angle range, multiple candidate visual angles are configured. These candidate visual angles are angles at which defects may be further detected. The optical detection characteristics of each candidate visual angle are analyzed, and the impact of different angles on laser scattering detection and infrared thermal imaging modules is analyzed. Different viewing angles may have different effects on detection capabilities, so it is necessary to evaluate the optical characteristics of each visual angle. Different detection angles will affect the optical characteristics of the detection, such as reflection, scattering, absorption and other phenomena. Therefore, configuring candidate visual angles and analyzing their optical characteristics can provide an effective angle selection basis for subsequent defect detection. Optical detection characteristics directly affect the sensitivity and accuracy of defect recognition. Accurate characteristic analysis can help understand the impact of different visual angles on defect detection results. The optical detection characteristics of the candidate visual angles are obtained, providing data support for subsequent adaptive parameter calculations and helping to optimize angle selection.

[0099] More specifically, based on the optical detection characteristics of each candidate visual angle, the detection effect of the potential defect heat information is analyzed. Specifically, different visual angles will have different detection effects on heat information, so it is necessary to analyze the detection effect of potential defects at each angle. The result of the analysis is to generate adaptability parameters for each candidate visual angle. These parameters reflect the detection adaptability of each visual angle under specific heat information. Each candidate visual angle may respond differently to the defect heat information, especially under different materials and defect types. The detection effect may vary greatly. The adaptability parameters can quantify the degree of adaptability of each visual angle to the defect heat information, and provide a quantitative basis for subsequent visual angle screening. The adaptability parameters of each candidate visual angle can quantify the adaptability of each angle under potential defect heat information, and help screen out the most suitable visual angle.

[0100] More specifically, the detection effect of each candidate visual angle is compared with the initial visual angle, and the coherence between them is analyzed. Specifically, the coherence of detection effect refers to whether the detection results of different visual angles are complementary or overlapping. An information gain weight is generated to reflect the information gain that the candidate visual angle can provide relative to the initial visual angle. Angles with high coherence will have greater information gain weights. Different visual angles may bring different information gains. Especially when the detection effect of the initial visual angle is limited, the supplementary visual angle can improve the accuracy of defect identification. By analyzing the coherence between angles, it is possible to avoid selecting redundant angles, optimize detection efficiency and effect, and generate information gain weights, which can better measure the contribution of the candidate visual angles to the supplementary information. Through this analysis, the most effective visual angle can be screened out for subsequent applications.

[0101] More specifically, the candidate visual angles are sorted according to their adaptability parameters. Visual angles with higher adaptability parameters are ranked first, indicating that they perform better under specific heat information. Combined with the information gain weights of the sorted visual angles, a feasibility analysis of the visual angles is performed, ultimately generating a list of supplementary visual angles to supplement and optimize the defect detection performance of the initial visual angles. By sorting the candidate visual angles, those with good adaptability and high information gain can be prioritized, further improving the accuracy of defect recognition. Combined with the analysis of adaptability parameters and information gain weights, the actual application effects of the supplementary visual angles can be better evaluated, the final detection strategy can be optimized, and a set of supplementary visual angles is generated that can effectively supplement the deficiencies of the initial visual angles, thereby improving the comprehensiveness and accuracy of defect recognition.

[0102] It can be understood that this technical solution, through a series of analysis and optimization steps, combines the detection functions and visual angles of laser scattering and infrared thermal imaging modules to conduct an adaptive analysis of the supplementary visual angle for potential defect heat information. Through the execution of each step, from the analysis of the optical characteristics of the selected angle to the calculation of the information gain weight, the optimal supplementary visual angle is finally obtained. This process optimizes the defect detection effect, improves the system's adaptability and accuracy under different conditions, can more effectively cope with the diversity of chip defects, and improve the efficiency and accuracy of defect detection.

[0103] Preferably, the step of performing laser scattering detection and photothermal microscopy detection on the chip to be tested according to the supplementary visual angle to obtain supplementary detection data of the chip to be tested includes: S61: Adjust the laser scattering excitation source and the infrared thermal imaging module to move to a specified supplementary visual angle; S62: driving the laser scattering excitation source and the infrared imaging module to perform laser scattering detection and photothermal microscopy detection on the chip to be tested, respectively, to obtain supplementary detection data of the chip to be tested.

[0104] Specifically, based on the analysis results of the previous step, a set of supplementary visual angles are selected. These visual angles have been evaluated as being able to effectively supplement the deficiencies of the initial visual angles and have high detection value. The relative positions of the laser scattering excitation source and the infrared thermal imaging module are adjusted so that they can be positioned according to the set supplementary visual angles, ensuring that the laser and infrared thermal imaging equipment can perform detection at appropriate angles. By precisely adjusting the equipment position, it is ensured that the detection module can perform detection from multiple supplementary visual angles, providing more comprehensive data. In this way, the coverage capability of the detection system at different visual angles can be enhanced, and the recognition rate of defects can be improved.

[0105] More specifically, the laser scattering excitation source and the infrared thermal imaging module respectively inspect the chip to be tested. The laser scattering excitation source will use laser to scan the chip surface to detect tiny defects that may exist on the surface, while the infrared thermal imaging module will perform photothermal microscopy to capture the temperature distribution on the chip surface, thereby detecting potential internal defects or temperature anomalies. Laser scattering and photothermal microscopy inspections are performed at supplementary visual angles to obtain supplementary inspection data at these angles. By inspecting from multiple supplementary visual angles, the supplementary inspection data can provide more detailed and comprehensive defect information, especially through the combination of laser scattering and infrared thermal imaging modules, which can make up for the defect types or details that may not be captured by the initial visual angle. These supplementary inspection data can effectively improve the recognition rate and accuracy of chip defects, especially in the detection of complex or tiny defects, and can provide richer information support.

[0106] It can be understood that this technical solution precisely adjusts the detection angle of the laser scattering excitation source and the infrared thermal imaging module, and combines these two technical means to conduct comprehensive inspection of the chip to be tested, and obtains more accurate inspection data from a supplementary visual perspective. In this way, it can make up for the shortcomings of the initial visual angle, improve the comprehensiveness and accuracy of chip defect identification, and ultimately optimize the chip inspection process and improve inspection quality and efficiency.

[0107] Preferably, the step of performing defect identification on the fuzzy sensing area according to the supplementary detection data and the potential defect heat information to obtain a defect identification result of the chip to be tested includes: S71: extracting key detection information from the supplementary detection data, and determining the clarity of the key detection information to obtain the clarity of the supplementary detection data; S72: When the information clarity of the supplementary inspection data meets the preset standard, the supplementary inspection data is processed for each defect type to obtain a defect identification feature, and the defect identification feature is verified in combination with the potential defect heat information to obtain a defect identification result; S73: When the information clarity of the supplementary detection data does not meet the preset standard, performing wavelet transform processing on the supplementary detection data to extract texture features; S74: performing cluster analysis on the texture features, and limiting the defect expression data of the supplementary detection data according to the cluster analysis result, so as to determine the defect expression data in the supplementary detection data; S75: Separating the defect expression data from the remaining data, extracting edge features from the defect expression data using an edge detection algorithm, and performing feature matching processing on the edge features for each defect type based on the potential defect heat information to obtain a feature matching degree for each defect type; S76: Combining the probability value of each defect type in the potential defect heat information with the verification parameter, a comprehensive weighted analysis is performed on the feature matching degree of each defect type to obtain a defect recognition result.

[0108] Specifically, key visual features such as edges, textures, and morphologies are extracted from the supplementary inspection data. These features can help identify and distinguish different types of defects, and the clarity of the extracted visual features is judged. This can be accomplished by comparing indicators such as feature contrast and edge clarity to ensure the recognizability of the features in the image. Extracting key visual features and judging their clarity is the first step in the defect recognition process. Only clear and easy-to-distinguish features can provide accurate support for subsequent defect recognition. Information clarity judgment can help filter out data of poor quality, avoid erroneous defect recognition results, and ensure that only clear supplementary inspection data is used for subsequent processing, thereby improving the accuracy and efficiency of subsequent defect recognition. If the clarity is insufficient, subsequent processing steps (such as wavelet transform) will be automatically performed to improve data quality.

[0109] More specifically, when the clarity of the supplementary inspection data meets the preset standards, the data is started to be identified for various types of defects. This can be accomplished through machine learning algorithms, deep learning models, or specific image processing techniques (such as morphological analysis). The identified defect features are verified in combination with the potential defect heat information to ensure that the defect features are consistent with the heat information and enhance the accuracy of recognition. Identifying different types of defects is a key link, and incorporating heat information can improve the accuracy of defect recognition, because certain types of defects may be more obvious under specific temperature distributions. The verification process ensures that the identified defects are not only visually apparent, but also can be verified for authenticity through heat data, thereby avoiding misidentification and providing accurate defect recognition results. At the same time, the combination of heat information improves the accuracy of recognition, which can reduce misidentification, especially in areas where defects are not obvious.

[0110] More specifically, if the clarity of the supplementary inspection data does not meet the standards, wavelet transform is used for data processing. Wavelet transform is a technology that can capture data features at multiple scales and can extract the texture features of the image. This process helps to extract detailed information from fuzzy or low-quality data, especially texture and slight changes. It is usually used to enhance image features. For data with low information clarity, direct defect identification may not obtain reliable results. Wavelet transform can enhance data quality through multi-resolution analysis, especially enhance the clarity of texture features. Extracting texture features helps to identify subtle defects, especially surface defects and material inhomogeneities. Processing low-definition data through wavelet transform improves the texture feature expression of the image, which helps to accurately identify subsequent defects, especially when the data is unclear.

[0111] More specifically, cluster analysis is performed on the extracted texture features, using clustering algorithms (such as K-means clustering and DBSCAN) to group similar features, forming clusters of defect data with similar properties. Based on the results of the cluster analysis, specific defect type data is identified from the supplementary inspection data to clarify which data represents actual defects. Cluster analysis can help distinguish different types of defects, especially those manifested as texture anomalies. Through cluster analysis, potential defect areas can be effectively extracted from large amounts of data, and defect expression data can be determined. This helps to filter out practical defect information from the inspection data, and can extract more specific and representative defect expression data from the data, providing a clear target for subsequent defect matching and feature recognition.

[0112] More specifically, the identified defect expression data is separated from other data, and its edge features are extracted using an edge detection algorithm (such as Canny edge detection). Based on the potential defect heat information, feature matching is performed on the extracted edge features to find edge features that match each defect type, and the feature matching degree is calculated. Edge detection helps to clarify the boundaries of the defect, especially when the shape and edges of the defect are not clear, it can help identify the specific defect edges. The combination of feature matching and heat information can improve the accuracy of defect recognition and avoid misjudgment. Through the combination of edge feature extraction and heat information, the edge features of the defect are accurately extracted, and matching information is provided for subsequent defect type recognition, thereby enhancing the accuracy of defect recognition.

[0113] More specifically, the probability value of each defect type in the potential defect heat information is combined with the feature matching degree, and a comprehensive evaluation of each defect type is performed through weighted analysis. The verification parameters are used to adjust the results of the weighted analysis to ensure that the final recognition result is more accurate. The comprehensive weighted analysis combines the heat information and the feature matching degree, and can give a final defect recognition result based on the probability of defect occurrence and its matching degree with the image features. The verification parameters further optimize the recognition result to ensure that the final identified defect type and location are more accurate. Through comprehensive analysis and weighted processing, the final accurate defect recognition result is obtained, which improves the overall accuracy and reliability of defect detection, especially under the combined effect of different defect types and heat information.

[0114] It can be understood that through a series of image processing and data analysis steps, this technical solution can extract key features from supplementary detection data, optimize data quality, perform accurate defect identification, and improve identification accuracy by combining potential defect heat information. Through this process, various defects in the chip to be tested can be effectively identified, and a more comprehensive and accurate defect analysis can be made based on feature matching and heat information.

[0115] Reference Figure 2 As shown, in a second aspect, the present invention provides a high-precision defect detection system for semiconductor chips, which is used to implement the high-precision defect detection method for semiconductor chips described in any one of the first aspects, comprising: An initial detection module, used to perform laser scattering detection and photothermal microscopy detection at an initial visual angle on the chip to be tested, so as to obtain initial detection data of the chip to be tested; A feature extraction module is used to extract key visual features from the initial detection data, and perform feature vector conversion and data integration on the key visual features to obtain a multi-dimensional visual feature matrix; A cluster analysis module, configured to construct an adjacency matrix based on the multidimensional visual feature matrix, and perform cluster analysis on the adjacency matrix using a graph clustering algorithm to locate the fuzzy perception area of ​​the chip under test; A defect perception module, used for combining the initial detection data with the multi-dimensional visual feature matrix to identify potential defects in the fuzzy perception area to obtain potential defect heat information; An angle analysis module, configured to perform adaptive analysis of the potential defect heat information at a supplementary visual angle, and simultaneously constrain the adaptive analysis result based on the initial visual angle with an information gain condition to generate a supplementary visual angle; a supplementary detection module, configured to perform laser scattering detection and photothermal microscopy detection on the chip to be tested according to the supplementary visual angle to obtain supplementary detection data of the chip to be tested; A defect recognition module is used to perform defect recognition on the fuzzy perception area according to the supplementary detection data and the potential defect heat information to obtain a defect recognition result of the chip to be tested.

[0116] In this embodiment, for the specific implementation of each module in the above system embodiment, please refer to the above method embodiment, which will not be repeated here.

[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A high-precision defect detection method for semiconductor chips, characterized in that: include: Performing laser scattering detection and photothermal microscopy detection at an initial visual angle on the chip to be tested to obtain initial detection data of the chip to be tested; Extracting key visual features from the initial detection data, and performing feature vector conversion and data integration on the key visual features to obtain a multi-dimensional visual feature matrix; Constructing an adjacency matrix based on the multidimensional visual feature matrix, and performing cluster analysis on the adjacency matrix by using a graph clustering algorithm to locate the fuzzy perception area of ​​the chip to be tested; Combining the initial detection data with the multi-dimensional visual feature matrix to perform potential defect identification on the fuzzy perception area to obtain potential defect heat information; Performing an adaptive analysis of the potential defect heat information at a supplementary visual angle, and simultaneously constraining the adaptive analysis result based on the initial visual angle with an information gain condition to generate a supplementary visual angle; Performing laser scattering detection and photothermal microscopy detection on the chip to be tested according to the supplementary visual angle to obtain supplementary detection data of the chip to be tested; Defect identification is performed on the fuzzy perception area according to the supplementary detection data and the potential defect heat information to obtain a defect identification result of the chip to be tested.

2. The high-precision defect detection method for semiconductor chips according to claim 1, wherein: The steps of performing laser scattering detection and photothermal microscopy detection at an initial visual angle on the chip to be tested to obtain initial detection data of the chip to be tested include: The chip to be tested is placed at a position to be tested, and the laser scattering excitation source and the infrared thermal imaging module are adjusted to move to a specified initial visual angle; Analyzing the detection modes of the laser scattering excitation source and the infrared thermal imaging module respectively according to the specification design information of the chip to be tested to obtain laser detection parameters and infrared detection parameters; Configuring parameters of the laser scattering excitation source according to the laser detection parameters to perform laser scattering detection on the chip to be tested at an initial visual angle to obtain laser scattering detection data; Configuring parameters of the infrared thermal imaging module according to the infrared detection parameters to perform photothermal microscopic detection of the chip to be tested at an initial visual angle to obtain photothermal microscopic detection data; The laser scattering detection data is combined with the photothermal microscopy detection data to obtain initial detection data.

3. The high-precision defect detection method for semiconductor chips according to claim 1, wherein: The steps of extracting key visual features from the initial detection data and performing feature vector conversion and data integration on the key visual features to obtain a multi-dimensional visual feature matrix include: performing regional refinement processing on the laser scattering detection data in the initial detection data to calculate the local curvature of the data portion corresponding to each specific position of the laser scattering detection data, thereby obtaining a surface curvature feature value set of the chip to be tested; Simultaneously, performing a collaborative operation of a first circuit consisting of Gaussian curvature mapping and Euler characteristic extraction and a second circuit consisting of mean curvature mapping and principal curvature difference analysis on the surface curvature feature value set to generate surface topological variable features of the chip to be tested; Performing a fast Fourier transform on the photothermal microscopy detection data in the initial detection data to extract the fundamental frequency amplitude characteristics, phase delay characteristics, and harmonic distortion rate of each specific position of the chip to be tested; Solving the heat conduction of the chip under test based on the fundamental frequency amplitude characteristics, phase delay characteristics, and harmonic distortion rate of each specific position of the chip under test to obtain heat diffusion distribution information of the chip under test; Performing deep frequency domain slicing processing on the heat diffusion distribution information to extract microscopic thermal field distortion features; A chip information space is constructed according to the specification design information of the chip to be tested, and the surface topology variable characteristics and the microscopic thermal field distortion characteristics are used as key visual features to map the information of the corresponding positions in the chip information space, so as to vectorize the key visual features in the chip information space and obtain a multi-dimensional visual feature matrix.

4. The high-precision defect detection method for semiconductor chips according to claim 1, wherein: The steps of constructing an adjacency matrix based on the multidimensional visual feature matrix and performing cluster analysis on the adjacency matrix by using a graph clustering algorithm to locate the fuzzy perception area of ​​the chip to be tested include: performing cosine similarity calculation on the multidimensional visual feature matrix to obtain vector similarity information of the multidimensional visual feature matrix, and constructing an adjacency matrix of the multidimensional visual feature matrix based on the vector similarity information; Performing feature clustering analysis on the adjacency matrix using a graph clustering algorithm, and performing vector clustering processing on the multidimensional visual feature matrix based on the feature clustering result, so as to divide the multidimensional visual feature matrix into a plurality of clusters; Calculating the cluster density of each cluster respectively, and performing cluster density difference amplitude analysis on each adjacent cluster to obtain a cluster density characteristic value and a neighboring difference characteristic value of each cluster; Simulating the overall clustering environment in combination with the cluster density eigenvalues ​​of each cluster in the multidimensional visual feature matrix, and assigning numerical values ​​of information clarity based on the cluster density eigenvalues ​​to each cluster according to the simulation results, so as to generate information clarity of each cluster; The information clarity of each cluster is adaptively adjusted based on the adjacent difference characteristic value of each cluster, and the adjusted information clarity is judged according to a preset threshold, so as to locate the corresponding fuzzy perception area on the chip to be tested according to the clusters that do not meet the preset threshold.

5. The high-precision defect detection method for semiconductor chips according to claim 1, wherein: The step of combining the initial detection data with the multi-dimensional visual feature matrix to identify potential defects in the fuzzy perception area to obtain potential defect heat information includes: extracting corresponding information from the initial detection data and the multidimensional visual feature matrix according to the fuzzy perception area, so as to extract first identification information and second identification information corresponding to the fuzzy perception area from the initial detection data and the multidimensional visual feature matrix; Performing a probability analysis of each defect type on the first identification information according to a pre-built chip defect manifestation analysis model to obtain a probability value of each defect type corresponding to the fuzzy perception area; According to the probability value of each defect type corresponding to the fuzzy perception area, the corresponding defect type is digitally reproduced in the fuzzy perception area to construct a digital model of defect representation; Extracting features from the digital model of defect manifestations of each defect type, and performing feature matching processing on the second identification information based on the extracted features to generate verification parameters for each defect type; The potential defect heat information is obtained by combining the probability value of each defect type corresponding to the fuzzy perception area and the verification parameter.

6. The high-precision defect detection method for semiconductor chips according to claim 1, wherein: The steps of performing adaptive analysis of the potential defect heat information at a supplementary visual angle and constraining the adaptive analysis result based on the initial visual angle to generate a supplementary visual angle include: Acquiring detection function information and visual angle range of a laser scattering excitation source for performing laser scattering detection and an infrared thermal imaging module for performing photothermal microscopy detection; Configuring a plurality of candidate visual angles based on the visual angle range, and analyzing optical detection characteristics caused by viewing angle changes for each of the candidate visual angles according to the detection function information, so as to generate optical detection characteristics for each of the candidate visual angles; performing a potential defect detection effect analysis on the potential defect heat information according to the optical detection characteristics of each of the candidate visual angles to obtain an adaptability parameter for each of the candidate visual angles; Performing a coherence analysis on the detection effect of each of the candidate visual angles relative to the initial visual angle to generate an information gain weight brought by each of the candidate visual angles as a supplementary visual angle; The candidate visual angles are sorted according to the adaptability parameter, and the feasibility of the visual angles is analyzed in combination with the corresponding information gain weights according to the sorting order to generate supplementary visual angles.

7. The high-precision defect detection method for semiconductor chips according to claim 1, wherein: The step of performing laser scattering detection and photothermal microscopy detection on the chip to be tested according to the supplementary visual angle to obtain supplementary detection data of the chip to be tested includes: Adjust the laser scattering excitation source and the infrared thermal imaging module to move to a specified supplementary visual angle; The laser scattering excitation source and the infrared imaging module are driven to perform laser scattering detection and photothermal microscopy detection on the chip to be tested respectively, so as to obtain supplementary detection data of the chip to be tested.

8. The high-precision defect detection method for semiconductor chips according to claim 5, wherein: The step of performing defect identification on the fuzzy sensing area according to the supplementary detection data and the potential defect heat information to obtain a defect identification result of the chip to be tested includes: Extracting key detection information from the supplementary detection data, and determining the clarity of the key detection information to obtain the clarity of the supplementary detection data; When the information clarity of the supplementary inspection data meets the preset standard, the supplementary inspection data is processed for each defect type identification to obtain a defect identification feature, and the defect identification feature is verified in combination with the potential defect heat information to obtain a defect identification result; When the information clarity of the supplementary detection data does not meet the preset standard, performing wavelet transform processing on the supplementary detection data to extract texture features; performing a cluster analysis on the texture features, and limiting the defect expression data of the supplementary detection data according to the cluster analysis result, so as to determine the defect expression data in the supplementary detection data; Separating the defect expression data from the remaining data, extracting edge features from the defect expression data using an edge detection algorithm, and performing feature matching processing on the edge features for each defect type based on the potential defect heat information to obtain a feature matching degree for each defect type; Combined with the probability value of each defect type in the potential defect heat information and the verification parameter, a comprehensive weighted analysis is performed on the feature matching degree of each defect type to obtain a defect recognition result.

9. A high-precision defect detection system for semiconductor chips, characterized in that: A method for high-precision defect detection of a semiconductor chip according to any one of claims 1 to 8, comprising: An initial detection module, used to perform laser scattering detection and photothermal microscopy detection at an initial visual angle on the chip to be tested, so as to obtain initial detection data of the chip to be tested; A feature extraction module is used to extract key visual features from the initial detection data, and perform feature vector conversion and data integration on the key visual features to obtain a multi-dimensional visual feature matrix; A cluster analysis module, configured to construct an adjacency matrix based on the multidimensional visual feature matrix, and perform cluster analysis on the adjacency matrix using a graph clustering algorithm to locate the fuzzy perception area of ​​the chip under test; A defect perception module, used for combining the initial detection data with the multi-dimensional visual feature matrix to identify potential defects in the fuzzy perception area to obtain potential defect heat information; An angle analysis module, configured to perform adaptive analysis of the potential defect heat information at a supplementary visual angle, and simultaneously constrain the adaptive analysis result based on the initial visual angle with an information gain condition to generate a supplementary visual angle; a supplementary detection module, configured to perform laser scattering detection and photothermal microscopy detection on the chip to be tested according to the supplementary visual angle to obtain supplementary detection data of the chip to be tested; A defect recognition module is used to perform defect recognition on the fuzzy perception area according to the supplementary detection data and the potential defect heat information to obtain a defect recognition result of the chip to be tested.

Images