Quality control data analysis method and system based on dynamic distribution optimization

Abstract

The present invention discloses a quality control data analysis method and system based on dynamic distribution tuning, which relates to the field of quality control technology, and includes the following steps: obtaining wafer first data and constructing a three-dimensional feature matrix, wherein the wafer first data includes wafer surface optical image, impedance distribution data and process parameter data; extracting wavelet coefficient features based on the three-dimensional feature matrix to establish a composite distribution model, and performing dynamic baseline modeling on the normal sample feature distribution; analyzing the KL divergence of the wafer to be evaluated according to the baseline model, and adjusting the production parameters according to the KL divergence; obtaining circuit defects in the adjustment process and constructing a joint optimization function, and establishing a defect topology relationship map based on a graph neural network; constructing a closed-loop control mechanism based on the defect topology relationship map to perform feedback optimization on production. The present invention ensures the executability of optimization decisions and improves production efficiency and resource utilization through closed-loop optimization of real-time data.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and more specifically, to a quality control data analysis method and system based on dynamic distribution optimization. Background Art

[0002] In today's rapidly developing business environment, quality management has become the key to corporate competitiveness, whether in manufacturing or service industries. Data analysis, as a core tool for quality management, can help companies gain insight into product or service defects, identify and solve problems in a timely manner, optimize processes and improve efficiency, thereby improving product quality and service levels, and enhancing corporate competitiveness in the market.

[0003] Data analysis is the process of collecting, organizing, processing, analyzing and interpreting data, with the goal of discovering the value behind the data and providing a basis for decision-making. It goes beyond the simple aggregation of numbers and is a scientific method for in-depth understanding of the connotation of data. In quality management, data analysis can reveal the quality status of products or services, identify problems, and formulate improvement measures to improve the overall quality level.

[0004] At present, in the quality control scenario of the semiconductor manufacturing industry, the application of optical inspection in wafer defect detection has significant advantages, but there are also some specific problems. These problems mainly include the contradiction between detection accuracy and speed, differences in sensitivity to different types of defects, stability affected by environmental factors, and limitations in the detection of complex pattern wafers.

[0005] For example, the invention patent with announcement number: CN117709799B announces a sampling detection system and method for online quality of motor housing, which relates to the technical field of sampling detection of online quality of motor housing, including a sampling detection system, the sampling detection system includes: a detection standard module; a sampling method module; a quality detection module; a data analysis module and a result judgment module, the sampling method module adopts a flexible sampling method, which dynamically adjusts the sampling plan according to real-time production data and historical quality records. The present invention ensures that the algorithm has sufficient flexibility by adopting a flexible sampling method to adapt to changes in the production process and new quality control requirements.

[0006] For example, the precision manufacturing quality control method based on big data analysis announced by the invention patent with announcement number CN117422333A includes the following steps: S1: Collect precision manufacturing production and R&D data through the acquisition module of the acquisition device, and move it to a temporary temporary database through the source system for data integration. The optimal processing model is screened out through the screening module, and the precision manufacturing quality control management feature set is analyzed through the optimal precision manufacturing production quality control management analysis model to obtain the precision manufacturing quality control management analysis results. Based on the big data analysis of the precision manufacturing quality control management analysis results, the corresponding quality control strategies and management plans are formulated, thereby realizing the combination of big data analysis and precision manufacturing quality, improving product quality data, and using each process of the product production process to build a product quality influencing factor analysis model, and through data training, the model accurately realizes quality problems.

[0007] The above disclosed technical solutions have at least the following technical problems: the sensitivity of the existing optical detection to different types of defects on the wafer surface varies. Optical detection is usually highly sensitive to physical defects such as particle contamination and scratches, but may be less sensitive to some more subtle or specific types of defects (such as broken wires and short circuits in circuit patterns). In view of the above problems, the present invention proposes a solution. Summary of the invention

[0008] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a quality control data analysis method and system based on dynamic distribution tuning, which solves the problem of differences in sensitivity of optical inspection to wafer defects through multi-source data fusion modeling, dynamic baseline adjustment, divergence trigger parameter optimization and closed-loop control mechanism.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A quality control data analysis method based on dynamic distribution tuning includes the following steps: obtaining first wafer data and constructing a three-dimensional feature matrix, wherein the first wafer data includes an optical image of the wafer surface, impedance distribution data, and process parameter data; extracting wavelet coefficient features based on the three-dimensional feature matrix to establish a composite distribution model, and performing dynamic baseline modeling on the normal sample feature distribution; analyzing the KL divergence of the wafer to be evaluated according to the baseline model, and adjusting the production parameters according to the KL divergence; obtaining circuit defects in the adjustment process and constructing a joint optimization function, and establishing a defect topology relationship map based on a graph neural network; and constructing a closed-loop control mechanism based on the defect topology relationship map to perform feedback optimization on production.

[0011] In a preferred embodiment, the first wafer data is obtained and a three-dimensional feature matrix is ​​constructed, specifically: an electrical signal test is performed on the wafer, an optical image of the wafer surface is obtained, and impedance distribution data and process parameter data at different positions are measured and recorded; based on timestamp and spatial position information, the collected optical image data, impedance distribution data and process parameter data are matched in time and space dimensions to establish associations between the data; the associated data are arranged and combined according to three dimensions of physical defect characteristics, electrical characteristic anomalies and process parameters, and features of the data in each dimension are extracted and quantified to construct a three-dimensional feature matrix containing multi-source information.

[0012] In a preferred embodiment, the three-dimensional feature matrix is ​​used to extract wavelet coefficient features to establish a composite distribution model, specifically: based on discrete wavelet transform, the optical image is decomposed into sub-bands of different frequencies to obtain low-frequency sub-bands and high-frequency sub-bands; in the low-frequency sub-band, the gray-level co-occurrence matrix is ​​calculated, and the energy value and contrast are calculated through the gray-level co-occurrence matrix, and the mean and variance time-domain features of the low-frequency sub-band are calculated as wavelet coefficient features; in the high-frequency sub-band, the amplitude and phase frequency-domain features are calculated as wavelet coefficient features; the extracted wavelet coefficient features are integrated with the impedance distribution data and the process parameter data to obtain the optical features and the electrical features, and the optical features are modeled using a Gaussian distribution with a time-varying mean vector, and the electrical features are modeled using a Wishart distribution; a mixed Gaussian-Wishart composite distribution model is constructed, and the feature distribution of normal samples is set to conform to the mixed Gaussian-Wishart composite distribution; the mixed Gaussian-Wishart composite distribution model parameters are calculated based on the expectation-maximization algorithm, and the mixed Gaussian-Wishart composite distribution model parameters are calculated according to the mixed Gaussian-Wishart composite distribution model parameters. The parameters of the Wishart composite distribution model are used to calculate the probability that each sample belongs to a different Gaussian component. Based on the probability, the mean vector and covariance matrix of the Gaussian distribution and the degrees of freedom and scale matrix of the Wishart distribution are updated, and the model is iterated continuously until the model converges.

[0013] In a preferred embodiment, the dynamic baseline modeling of the characteristic distribution of normal samples is specifically as follows: real-time acquisition of process parameters of the production line, the process parameters including temperature, pressure, and production speed; setting a normal fluctuation range for each process parameter, and based on the Bayesian estimation method, updating the covariance matrix according to the process parameters within the normal fluctuation range to obtain an updated mixed Gaussian-Wishart composite distribution model; and using the updated mixed Gaussian-Wishart composite distribution model as the dynamic baseline model of normal samples.

[0014] In a preferred embodiment, the KL divergence of the wafer to be evaluated is analyzed according to the baseline model, and the production parameters are adjusted according to the KL divergence, specifically: the second wavelet coefficients of several wafers to be evaluated are obtained; the second wavelet coefficients are divided into preset equal intervals, the frequency of occurrence of data in each interval is counted, and a probability distribution is constructed; the KL divergence is calculated based on the probability distribution, and the KL divergence values ​​of each dimension are weighted and summed to obtain a comprehensive KL divergence value; a KL divergence threshold is set, and compared with the comprehensive KL divergence value, and the production parameters are adjusted according to the comparison result.

[0015] In a preferred embodiment, the production parameter adjustment is specifically as follows: based on regression analysis, all factors that affect the probability distribution of wafer characteristics are obtained to form a state space; each parameter is standardized and mapped to a numerical range at the same scale; the Jacobian matrix of all parameters in the state space is calculated; the condition number of the Jacobian matrix is ​​calculated to obtain the coupling relationship between the parameters; and a safe adjustment range is set for each parameter according to the coupling relationship between the parameters, and adjustments are made.

[0016] In a preferred embodiment, the circuit defects in the adjustment process are obtained and a joint optimization function is constructed, and a defect topology relationship map is established based on a graph neural network, specifically: correlation analysis is performed on circuit defects and detection effects, correlation data of the detection effects is obtained, and a joint optimization function is constructed based on the correlation data; sample data containing physical defects and electrical defects are obtained, and the sample data is preprocessed to digitize the characteristics of the physical defects and electrical defects; each sample is used as a node of a preset basic relationship map, and the Euclidean distance between the particles and scratches of the nodes is compared with a preset first threshold to determine the node correlation; the correlation between the physical defect characteristics and the electrical defect characteristics between the nodes is calculated, and the weights of the edges in the basic relationship map are set based on the node correlation; the weights of all nodes are weighted and aggregated based on the Laplace matrix of the graph to obtain the prediction results of the relationships between all nodes, and a defect topology relationship map is drawn.

[0017] The technical effects and advantages of the quality control data analysis method and system based on dynamic distribution optimization of the present invention are as follows:

[0018] 1. The present invention realizes the deep fusion of multi-source data by comprehensively collecting high-resolution images of the wafer surface, impedance distribution data of electrical performance tests, and process parameter data, and constructing a three-dimensional feature matrix. This makes the evaluation of product quality no longer limited to a single dimension, but a comprehensive consideration from multiple angles such as physical form, electrical performance, and process.

[0019] 2. The present invention locates the process link corresponding to the root cause of the defect through the graph attention mechanism of the graph, establishes the process parameter-equipment instruction conversion matrix, and makes optimization decisions. This intelligent optimization strategy can reasonably allocate optimization resources according to the severity and impact range of the defects, and give priority to solving the problems that have the greatest impact on product quality. For process defects located at the bottom of the topological map and having a greater impact on product quality, the relevant process parameters will be adjusted first, while for some surface functional defects, they can be processed later if resources are limited. At the same time, by accurately mapping the adjustment amount derived from the topological map to the physical equipment control space, the executability of the optimization decision is ensured, blind adjustment is avoided, and production efficiency and resource utilization are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The figure is a flow chart of the quality control data analysis method based on dynamic distribution optimization of the present invention.

[0021] Figure 2 It is a schematic diagram of the structure of the quality control data analysis system based on dynamic distribution optimization of the present invention. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] Embodiment 1, Figure 1 The present invention provides a quality control data analysis method based on dynamic distribution tuning, which includes the following steps:

[0024] Acquire first wafer data and construct a three-dimensional feature matrix, wherein the first wafer data includes an optical image of a wafer surface, impedance distribution data, and process parameter data;

[0025] Based on the three-dimensional feature matrix, the wavelet coefficient features are extracted to establish a composite distribution model, and a dynamic baseline model is performed on the feature distribution of normal samples;

[0026] Analyze the KL divergence of the wafer to be evaluated based on the baseline model, and adjust the production parameters based on the KL divergence;

[0027] Obtain circuit defects during the adjustment process and construct a joint optimization function, and establish a defect topology relationship map based on the graph neural network;

[0028] Based on the defect topology relationship map, a closed-loop control mechanism is constructed to provide feedback optimization for production.

[0029] S1, obtaining first wafer data and constructing a three-dimensional feature matrix, wherein the first wafer data includes a wafer surface optical image, impedance distribution data, and process parameter data;

[0030] Optical image acquisition: Place the wafer on the workbench of the optical inspection module, adjust the focus, aperture and other parameters to ensure that the light evenly illuminates the wafer surface. Start the inspection module, scan the wafer at high resolution, and obtain surface image data. These images can clearly show the physical morphology of the wafer surface.

[0031] Electrical performance and process parameter collection: While performing optical inspection, the electrical performance test module is connected to the test point of the wafer, and by applying electrical signals, the impedance distribution data at different positions is measured and recorded. At the same time, the system automatically collects process parameter data during the production process, such as temperature, pressure, time, etc.

[0032] The obtaining of the first wafer data and constructing a three-dimensional feature matrix is ​​specifically as follows:

[0033] Acquire the optical image of the wafer surface, connect the electrical performance test module to the test point of the wafer, and measure and record the impedance distribution data and process parameter data at different positions by applying electrical signals;

[0034] Based on the timestamp and spatial position information, the collected optical image data, impedance distribution data and process parameter data are matched in time and space dimensions to establish the association between the data;

[0035] The associated data are arranged and combined according to the three dimensions of physical defect characteristics, electrical characteristic anomalies and process parameters, and the features of the data in each dimension are extracted and quantified to construct a three-dimensional feature matrix containing multi-source information.

[0036] Physical defect characteristics include scratches, cracks, particles, bubbles and stains; electrical characteristic anomalies include resistance anomalies, capacitance anomalies and voltage and current anomalies;

[0037] Image data can intuitively display the surface condition, impedance data can reflect the electrical performance, and process parameter data helps to understand the conditions of the production process. The combination of the three can more accurately analyze the relationship between wafer quality and various factors.

[0038] The benefits of three-dimensional feature matrix in data processing:

[0039] Data fusion: The 3D feature matrix can integrate multiple types of data, such as wafer surface optical images, impedance distribution data, and process parameter data, into a unified data representation. This data fusion helps eliminate information islands between data and improve the overall utilization and value of data.

[0040] Comprehensive characterization: The 3D feature matrix can comprehensively characterize the various characteristics of the wafer, including surface morphology, electrical properties, and process parameters. This comprehensive characterization helps to more accurately understand and analyze the quality status and production process of the wafer.

[0041] Feature extraction: The three-dimensional feature matrix can provide richer feature information by integrating multiple data, thereby supporting more efficient and accurate feature extraction. The feature information in the three-dimensional feature matrix can more intuitively reflect the various abnormalities and defects of the wafer, such as surface scratches, cracks, impurities, etc. This feature recognition helps to timely discover and deal with potential problems in the wafer production process, thereby improving product quality and production efficiency.

[0042] S2, extracting wavelet coefficient features based on the three-dimensional feature matrix to establish a composite distribution model, and perform dynamic baseline modeling on the feature distribution of normal samples;

[0043] The normal samples refer to wafer samples that meet production standards and quality requirements during the wafer production process and do not have any process defects or performance abnormalities.

[0044] The method of extracting wavelet coefficient features based on the three-dimensional feature matrix to establish a composite distribution model is specifically as follows:

[0045] Based on discrete wavelet transform, the optical image is decomposed into sub-bands of different frequencies to obtain low-frequency sub-bands and high-frequency sub-bands. The low- and medium-frequency sub-bands contain the structural information of the image, and the high-frequency sub-bands contain edges and textures.

[0046] In the low-frequency sub-band, the gray-level co-occurrence matrix is ​​calculated, and the energy value and contrast are calculated through the gray-level co-occurrence matrix. At the same time, the mean and variance time-domain characteristics of the low-frequency sub-band are calculated as wavelet coefficient features;

[0047] In the high frequency subband, the amplitude and phase frequency domain features are calculated as wavelet coefficient features;

[0048] The extracted wavelet coefficient features are integrated with the impedance distribution data and process parameter data to obtain the optical and electrical features. The optical features are modeled using a Gaussian distribution with a time-varying mean vector, and the electrical features are modeled using a Wishart distribution.

[0049] Construct a mixed Gaussian-Wishart compound distribution model, and set the characteristic distribution of normal samples to conform to the mixed Gaussian-Wishart compound distribution;

[0050] The parameters of the mixed Gaussian-Wishart composite distribution model are calculated based on the expectation maximization algorithm. According to the parameters of the mixed Gaussian-Wishart composite distribution model, the probability of each sample belonging to a different Gaussian component is calculated. According to the probability, the mean vector and covariance matrix of the Gaussian distribution and the degrees of freedom and scale matrix of the Wishart distribution are updated. The model is iterated continuously until convergence. The optical characteristics include:

[0051] Grayscale features: global mean, variance, and histogram statistics.

[0052] Texture features: In addition to GLCM, it also includes local binary pattern (LBP) and Gabor filter response.

[0053] Geometric features: shape descriptors (such as circularity, area), edge detection (Canny operator).

[0054] Color features: mean and variance of RGB or HSV color space.

[0055] Electrical characteristics include:

[0056] Impedance characteristics: amplitude, phase, equivalent circuit parameters (such as resistance R, inductance L, capacitance C).

[0057] Time domain characteristics: rising edge time and decay time constant of current / voltage signal.

[0058] Frequency domain characteristics: power spectral density (PSD) and resonance frequency after Fourier transform.

[0059] Statistical characteristics: mean and noise level (variance) of the electrical signal.

[0060] The Gaussian distribution modeling is specifically as follows:

[0061]

[0062] in, is the benchmark mean, To correspond to the production line process cycle, is the amplitude, is the number of frequency components, is the phase, is the time variable.

[0063] The Wishart distribution modeling is specifically as follows:

[0064]

[0065] in is the degree of freedom parameter, is the base degree of freedom parameter, is the influence coefficient of temperature deviation, which is calibrated through process DOE experiment. is the temperature deviation of the etching process,

[0066] The dynamic baseline modeling of the normal sample feature distribution is specifically as follows:

[0067] Based on sensors, real-time acquisition of process parameters of the production line, including temperature, pressure, and production speed;

[0068] Set a normal fluctuation range for each process parameter. When the collected parameters exceed this range, it is determined that the process fluctuates.

[0069] Based on the Bayesian estimation method, the covariance matrix is ​​updated according to the newly collected process parameters to obtain the updated mixed Gaussian-Wishart composite distribution model;

[0070] The updated mixed Gaussian-Wishart composite distribution model is used as the dynamic baseline model for normal samples.

[0071] The covariance matrix is ​​updated according to the newly collected process parameters, specifically:

[0072]

[0073] in, is the covariance matrix at time t, is the base covariance matrix, is the Hadamard product, and are the coefficients corresponding to temperature and pressure, and are the deviations of temperature and pressure from the reference values, is an exponential function.

[0074] Monitor the changes in model parameters (such as the mean vector and covariance matrix of the Gaussian distribution, and the degrees of freedom and scale matrix of the Wishart distribution) in each iteration. Calculate the difference between the parameters in two consecutive iterations (you can use metrics such as Euclidean distance). When these differences are all less than a specific threshold, the model is considered to have converged. For example, for the mean vector, if the mean vector is between the nth iteration and the n+1th iteration, and other parameters also meet similar conditions, the model is considered to have converged. This indicates that the parameters are no longer changing significantly and the model is becoming stable.

[0075] The mixed Gaussian-Wishart composite distribution can better describe the distribution of complex data, where the Gaussian distribution is used to characterize the central tendency and dispersion of the data, and the Wishart distribution is used to describe the covariance structure of the data.

[0076] Decompose the image: Taking the three-layer decomposition as an example, the first decomposition divides the image into a low-frequency subband LL1 and three high-frequency subbands LH1, HL1, and HH1. The low-frequency subband LL1 retains the main contour and low-frequency information of the image, and the high-frequency subbands LH1, HL1, and HH1 contain high-frequency details in the horizontal, vertical, and diagonal directions, respectively. Then, LL1 is decomposed for the second time to obtain LL2, LH2, HL2, HH2, and so on.

[0077] The mean represents the average brightness of the image, and the variance reflects the discreteness of the brightness.

[0078] S3, analyzing the KL divergence of the wafer to be evaluated according to the baseline model, and adjusting the production parameters according to the KL divergence;

[0079] The KL divergence of the wafer to be evaluated is analyzed according to the baseline model, and the production parameters are adjusted according to the KL divergence, specifically:

[0080] Obtaining second wavelet coefficients of a plurality of wafers to be evaluated;

[0081] Divide the second wavelet coefficients into preset equal intervals, count the frequency of data in each interval, and construct a probability distribution;

[0082] Calculate the KL divergence based on the probability distribution, and perform weighted summation on the KL divergence values ​​of each dimension to obtain the comprehensive KL divergence value;

[0083] Set the KL divergence threshold and compare it with the comprehensive KL divergence value, and adjust the production parameters based on the comparison results.

[0084] The parameter adjustment process is specifically as follows:

[0085] Based on regression analysis, all factors that affect the probability distribution of wafer features are obtained to form parameter dimensions of the state space; the factors include exposure time, light intensity, lens focal length, frequency and amplitude of the test signal, motor speed, valve opening, temperature, humidity, and air pressure;

[0086] Each parameter is standardized and mapped to a numerical range on the same scale;

[0087] Compute the Jacobian matrix of all parameters in the state space;

[0088] Calculate the condition number of the Jacobian matrix and obtain the coupling relationship between the parameters;

[0089] A safe adjustment range is set for each parameter according to the coupling relationship between the parameters, and adjustments are made.

[0090] During the parameter adjustment process, the parameter value is monitored in real time to ensure that it is always within the safe adjustment range. Once the parameter value is found to be close to or beyond the safety boundary, the strategy is adjusted immediately to prevent parameter oscillation and ensure the stable operation of the system.

[0091] The Jacobian matrix describes the first-order partial derivative of a vector function and can reflect the impact of a change in one parameter on other parameters.

[0092] The condition number of the Jacobian matrix is ​​calculated to obtain the coupling relationship between the parameters, which is specifically:

[0093] Specifically, the larger the condition number, the closer the matrix is ​​to being singular, that is, there may be a strong linear correlation between the parameters and the coupling relationship is complex.

[0094] The KL divergence is specifically calculated as follows:

[0095]

[0096] The comprehensive KL divergence value is specifically calculated as follows:

[0097]

[0098] in, is the KL divergence of the probability distributions X and Y, is the number of elements in the sample space, is the probability of event i in the probability distribution X, is the probability of event i in the probability distribution Y, is the weight of the j-th dimension, is the KL divergence value of the jth dimension, is the comprehensive KL divergence value, and n is the number of feature dimensions.

[0099] Exposure time adjustment: If the overall optical image of the current batch is dark or bright, and is significantly different from the baseline model, the exposure time can be adjusted. If the image is dark, it means underexposure, increase the exposure time by 10% - 20%; if the image is bright, it means overexposure, reduce the exposure time by 10% - 20%. After adjustment, re-acquire the image and compare the difference between the new image and the baseline model. If it still does not meet the requirements, the exposure time can be fine-tuned again, with each adjustment range of 5%.

[0100] Light intensity adjustment: If the image contrast is significantly different from the baseline model, it may be a light intensity problem. When the image contrast is low and the picture is blurry, increase the light intensity by 20% - 30%; if the contrast is high and the image details are severely lost, reduce the light intensity by 20% - 30%. After adjustment, detect the image features again and further optimize the light intensity based on the results.

[0101] Lens focus adjustment: If the image appears blurry or distorted, it may be that the lens focus is inaccurate. Fine-tune the lens focus by 0.5-1 mm at a time, and compare the clarity and feature integrity of the image before and after the adjustment until the image meets the baseline model standard.

[0102] Test signal frequency adjustment: When the impedance distribution data is abnormal and the analysis shows that the test signal frequency is inappropriate, adjust the frequency according to the electrical performance test theory and historical experience. For example, if the current frequency is 1kHz, try to adjust the frequency to 1.5kHz or 0.8kHz, retest and calculate the KL divergence. If the requirement is still not met, continue to adjust the frequency within a certain range, with each adjustment amplitude of 0.2kHz.

[0103] Adjustment of the test signal amplitude: If the test results fluctuate greatly and deviate significantly from the baseline model, adjust the test signal amplitude. If the amplitude is too high and causes data saturation, reduce the amplitude by 10% - 20%; if the amplitude is too low and the signal is not obvious, increase the amplitude by 10% - 20%. After adjustment, test again, compare the results with the baseline model, and continue to fine-tune the amplitude if necessary.

[0104] Furthermore, based on historical data and production experience, a KL divergence threshold is pre-set. Specifically, after statistical analysis of the KL divergence values ​​of a large number of normal batches and a small number of abnormal batches, it is found that the KL divergence values ​​of normal batches are basically below 0.2, while the KL divergence values ​​of abnormal batches are generally higher than 0.3, so the threshold is set to 0.25. This threshold is not fixed and needs to be regularly re-evaluated and adjusted according to the actual situation in the production process, such as equipment updates, process improvements, etc.

[0105] S4, obtains circuit defects during the adjustment process and constructs a joint optimization function, and establishes a defect topology relationship map based on a graph neural network;

[0106] The circuit defects include disconnection and short circuit;

[0107] The circuit defects in the adjustment process are obtained and a joint optimization function is constructed, and a defect topology relationship map is established based on a graph neural network, specifically:

[0108] Conduct correlation analysis on circuit defects and detection effects, obtain correlation data of detection effects and build a joint optimization function based on the correlation data;

[0109] Acquire sample data containing physical defects and electrical defects, and pre-process the sample data to digitize the features of the physical defects and electrical defects;

[0110] Each sample is used as a node of a preset basic relationship map, and the node correlation is determined by comparing the Euclidean distance between particles and scratches between nodes with a preset first threshold;

[0111] Calculate the correlation between physical defect characteristics and electrical defect characteristics between nodes, and set the weight of the edge in the basic relationship graph based on the node correlation;

[0112] Based on the Laplace matrix of the graph, the weights of all nodes are weighted and aggregated to obtain the prediction results of the relationships between all nodes, and a defect topology relationship map is drawn.

[0113] The joint optimization function is specifically:

[0114]

[0115] in, and are the weight coefficients of disconnection and short circuit, is the detection error of the broken wire, is the short circuit detection error, is the convolution kernel size, is the incident angle of light, is the polarization parameter, which controls the polarization state of the light reflected from the surface.

[0116] Furthermore, if two nodes are related in position and have high feature similarity, the edge weight is assigned to 0.8-1.0; if they are only related in position or only have similar features, the weight is assigned to 0.5-0.8; if the relationship between the two is weak, the weight is assigned to 0.1-0.5. For nodes with no obvious relationship, the edge is not connected.

[0117] S5, based on the defect topology relationship map, build a closed-loop control mechanism to provide feedback optimization for production.

[0118] Based on the defect topology relationship map, a closed-loop control mechanism is constructed to perform feedback optimization on production, specifically:

[0119] Based on the defect topology relationship map, the intrinsic relationship between different defects and process parameters and equipment status is analyzed to obtain the key influencing factors and propagation paths of each defect;

[0120] Generate a parameter adjustment priority list based on the hierarchical position of the defect type in the topology map;

[0121] Use the graph attention mechanism of the graph to locate the process link corresponding to the root cause of the defect;

[0122] Map the adjustment amount derived from the topological map to the physical equipment control space and establish the process parameter-equipment instruction conversion matrix;

[0123] Make optimization decisions based on the process parameter-equipment instruction conversion matrix and feed back the optimization decisions to the production line in real time;

[0124] After each process adjustment, a new KL divergence value is calculated. If the divergence value exceeds the threshold, the process parameters will be continuously optimized until the divergence value is reduced to a preset range.

[0125] The hierarchical position is that the bottom layer process defect is greater than the middle layer structure defect and the surface layer function defect;

[0126] Combined with online test data (such as voltage, current, frequency, etc.) and image data, the covariance matrix and baseline in the model are continuously optimized to adapt to fluctuations in the production process and ensure that the process parameters always operate in the best state. If the defect rate is reduced but still higher than the standard value after adjusting the process parameters, the data is analyzed again and other related parameters are adjusted in a coordinated manner, thereby starting a new round of optimization cycle to achieve continuous improvement of the production process and continuous improvement of quality.

[0127] Embodiment 2, a quality control data analysis system based on dynamic distribution optimization, is characterized by comprising the following modules:

[0128] Data acquisition module: used to obtain first wafer data and construct a three-dimensional feature matrix, wherein the first wafer data includes an optical image of the wafer surface, impedance distribution data, and process parameter data;

[0129] Wavelet feature extraction and composite distribution modeling module: used to extract wavelet coefficient features based on the three-dimensional feature matrix to establish a composite distribution model and perform dynamic baseline modeling on the feature distribution of normal samples;

[0130] Parameter adjustment trigger module: used to analyze the KL divergence of the wafer to be evaluated according to the baseline model and adjust the production parameters according to the KL divergence;

[0131] Circuit defect topology relationship map establishment module: used to obtain circuit defects during the adjustment process and build a joint optimization function, and establish a defect topology relationship map based on the graph neural network;

[0132] Closed-loop control and production feedback optimization module: used to build a closed-loop control mechanism based on the defect topology relationship map and perform feedback optimization on production.

[0133] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0134] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0135] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0136] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0137] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0138] Finally: 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, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A quality control data analysis method based on dynamic distribution optimization, characterized in that: The following steps are involved: Acquire first wafer data and construct a three-dimensional feature matrix, wherein the first wafer data includes an optical image of a wafer surface, impedance distribution data, and process parameter data; Based on the three-dimensional feature matrix, the wavelet coefficient features are extracted to establish a composite distribution model, and a dynamic baseline model is performed on the feature distribution of normal samples; The establishment of the composite distribution model is specifically as follows: Decomposing the optical image into low-frequency and high-frequency sub-bands based on discrete wavelet transform; In the low-frequency sub-band, the gray-level co-occurrence matrix is ​​calculated, and the energy value and contrast are calculated through the gray-level co-occurrence matrix. At the same time, the mean and variance time-domain characteristics of the low-frequency sub-band are calculated as wavelet coefficient features; In the high frequency subband, the amplitude and phase frequency domain features are calculated as wavelet coefficient features; The wavelet coefficient characteristics are integrated with the impedance distribution data and the process parameter data to obtain the optical and electrical characteristics. The optical characteristics are modeled using the Gaussian distribution with a time-varying mean vector, and the electrical characteristics are modeled using the Wishart distribution. A mixed Gaussian-Wishart composite distribution model is constructed based on the Gaussian distribution model and the Wishart distribution model, and the characteristic distribution of normal samples is set to conform to the mixed Gaussian-Wishart composite distribution; Calculate the parameters of the mixed Gaussian-Wishart composite distribution model based on the expectation maximization algorithm, calculate the probability of each sample belonging to a different Gaussian component according to the model parameters, update the mean vector and covariance matrix of the Gaussian distribution and the degrees of freedom and scale matrix of the Wishart distribution according to the probability, and iterate continuously until the model converges; Analyze the KL divergence of the wafer to be evaluated based on the baseline model, and adjust the production parameters based on the KL divergence; Obtain circuit defects during the adjustment process and construct a joint optimization function, and establish a defect topology relationship map based on the graph neural network; Based on the defect topology relationship map, a closed-loop control mechanism is constructed to provide feedback optimization for production.

2. The quality control data analysis method based on dynamic distribution tuning according to claim 1, characterized in that: The obtaining of the first wafer data and constructing a three-dimensional feature matrix is ​​specifically as follows: Conduct electrical signal tests on the wafer, obtain optical images of the wafer surface, measure and record impedance distribution data and process parameter data at different locations; Based on the timestamp and spatial position information, the collected optical image data, impedance distribution data and process parameter data are matched in time and space dimensions to establish the association between the data; The associated data are arranged and combined according to the three dimensions of physical defect characteristics, electrical characteristic anomalies and process parameters, and the features of the data in each dimension are extracted and quantified to construct a three-dimensional feature matrix containing multi-source information.

3. The quality control data analysis method based on dynamic distribution tuning according to claim 2 is characterized in that: The dynamic baseline modeling of the normal sample feature distribution is specifically as follows: Real-time acquisition of process parameters of the production line, including temperature, pressure, and production speed; A normal fluctuation range is set for each process parameter, and based on a Bayesian estimation method, a covariance matrix is ​​updated according to the process parameters within the normal fluctuation range to obtain an updated mixed Gauss-Wishart composite distribution model; The updated mixed Gauss-Wishart composite distribution model is used as the dynamic baseline model for normal samples.

4. The quality control data analysis method based on dynamic distribution tuning according to claim 3 is characterized in that: The KL divergence of the wafer to be evaluated is analyzed according to the baseline model, and the production parameters are adjusted according to the KL divergence, specifically: Obtaining second wavelet coefficients of a plurality of wafers to be evaluated; Divide the second wavelet coefficients into preset equal intervals, count the frequency of data in each interval, and construct a probability distribution; Calculate the KL divergence based on the probability distribution, and perform weighted summation on the KL divergence values ​​of each dimension to obtain the comprehensive KL divergence value; Set the KL divergence threshold and compare it with the comprehensive KL divergence value, and adjust the production parameters based on the comparison results.

5. The quality control data analysis method based on dynamic distribution tuning according to claim 4 is characterized in that: The production parameter adjustment is specifically as follows: Based on regression analysis, all factors that affect the probability distribution of wafer features are obtained to form a state space; Each parameter is standardized and mapped to a numerical range on the same scale; Compute the Jacobian matrix of all parameters in the state space; Calculate the condition number of the Jacobian matrix and obtain the coupling relationship between the parameters; A safe adjustment range is set for each parameter according to the coupling relationship between the parameters, and adjustments are made.

6. The quality control data analysis method based on dynamic distribution tuning according to claim 5, characterized in that: The circuit defects in the adjustment process are obtained and a joint optimization function is constructed, and a defect topology relationship map is established based on a graph neural network, specifically: Conduct correlation analysis on circuit defects and detection effects, obtain correlation data of detection effects and build a joint optimization function based on the correlation data; Acquire sample data containing physical defects and electrical defects, and pre-process the sample data to digitize the features of the physical defects and electrical defects; Each sample is used as a node of a preset basic relationship map, and the node correlation is determined by comparing the Euclidean distance between particles and scratches between nodes with a preset first threshold; Calculate the correlation between physical defect characteristics and electrical defect characteristics between nodes, and set the weight of the edge in the basic relationship graph based on the node correlation; Based on the Laplace matrix of the graph, the weights of all nodes are weighted and aggregated to obtain the prediction results of the relationships between all nodes, and a defect topology relationship map is drawn.

7. The quality control data analysis method based on dynamic distribution tuning according to claim 6, characterized in that: The joint optimization function is specifically: in, and are the weight coefficients of disconnection and short circuit, is the detection error of the broken wire, is the short circuit detection error, is the convolution kernel size, is the incident angle of light, is the polarization parameter, which controls the polarization state of the light reflected from the surface.

8. The quality control data analysis method based on dynamic distribution optimization according to claim 7, characterized in that: The KL divergence is specifically calculated as follows: The comprehensive KL divergence value is specifically calculated as follows: in, is the KL divergence of the probability distributions X and Y, is the number of elements in the sample space, is the probability of event i in the probability distribution X, is the probability of event i in the probability distribution Y, is the weight of the j-th dimension, is the KL divergence value of the jth dimension, is the comprehensive KL divergence value, and n is the number of feature dimensions.

9. A system using the quality control data analysis method based on dynamic distribution optimization according to any one of claims 1 to 8, characterized in that: Includes the following modules: Data acquisition module: used to obtain first wafer data and construct a three-dimensional feature matrix, wherein the first wafer data includes an optical image of the wafer surface, impedance distribution data, and process parameter data; Wavelet feature extraction and composite distribution modeling module: used to extract wavelet coefficient features based on the three-dimensional feature matrix to establish a composite distribution model and perform dynamic baseline modeling on the feature distribution of normal samples; Parameter adjustment trigger module: used to analyze the KL divergence of the wafer to be evaluated according to the baseline model and adjust the production parameters according to the KL divergence; Circuit defect topology relationship map establishment module: used to obtain circuit defects during the adjustment process and build a joint optimization function, and establish a defect topology relationship map based on the graph neural network; Closed-loop control and production feedback optimization module: used to build a closed-loop control mechanism based on the defect topology relationship map and perform feedback optimization on production.

Images