Self-adaptive semiconductor packaging detection method

Abstract

The invention relates to the technical field of semiconductor packaging detection, and provides a self-adaptive semiconductor packaging detection method which comprises the following steps: synchronously acquiring packaging surface heat distribution data, three-dimensional deformation data and pin electric signal data through a multi-source sensor; performing dynamic weighted fusion on the multi-modal data based on the packaging process parameters to generate a defect feature matrix; performing dynamic threshold calibration on the defect feature matrix by adopting an adaptive threshold generation algorithm according to clustering center distribution in the historical defect pattern library; performing matching judgment on the dynamic threshold value and the real-time defect feature matrix through a multi-level logic decision model, and outputting a defect type, a position coordinate and a confidence score; and based on the defect type and the confidence score, triggering adaptive compensation control of the packaging process parameters, generating a feedback control instruction, and adjusting a packaging equipment execution mechanism in real time. The packaging quality and reliability can be improved, and the production cost and the defective rate can be reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of semiconductor packaging detection, and more particularly, to an adaptive semiconductor packaging detection method. Background Art

[0002] In semiconductor manufacturing, packaging is a crucial step in ensuring chip performance and reliability. Traditional semiconductor packaging inspection methods primarily rely on single-touch detection methods, such as optical inspection, electrical testing, or thermal imaging. While these methods can detect packaging defects to a certain extent, they have limitations. For example, optical inspection struggles to detect internal structural defects, electrical testing has limited ability to detect surface defects on the package, and thermal imaging lacks sensitivity for minute defects. Furthermore, existing inspection methods are mostly static, unable to provide real-time feedback and adjust packaging process parameters, making them difficult to adapt to the complex and changing packaging environment and high-precision requirements.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: the existing detection methods are unable to comprehensively and in real time detect multiple defect types of semiconductor packages, and are unable to adjust the packaging process parameters in real time according to the detection results, resulting in low detection efficiency and insufficient defect detection accuracy, which makes it difficult to meet the requirements of modern semiconductor packaging for high reliability and high production efficiency. Summary of the Invention

[0004] The present invention provides an adaptive semiconductor packaging detection method, comprising: S100: synchronously collecting real-time multimodal data of a semiconductor package structure through multi-source sensors, wherein the multimodal data includes package surface thermal distribution data, three-dimensional deformation data, and pin electrical signal data; S200: performing dynamic weighted fusion on the multimodal data based on packaging process parameters to generate a defect feature matrix with spatiotemporal correlation; S300: performing dynamic threshold calibration on the defect feature matrix using an adaptive threshold generation algorithm according to the cluster center distribution in the historical defect pattern library; S400: Uses a multi-level logic decision model to match dynamic thresholds with the real-time defect feature matrix and outputs defect type, location coordinates, and confidence score. S500: Based on the defect type and confidence score, trigger adaptive compensation control of packaging process parameters, generate feedback control instructions and adjust packaging equipment actuators in real time.

[0005] Furthermore, the S100 specifically includes: S101: Use a high-resolution infrared thermal imager to collect package surface temperature field data at a sampling rate of 20 frames per second to generate a thermal distribution matrix , where x, y are the coordinates of the package surface, and the temperature resolution is ≤ 0.1°C; S102: Measure the warpage and coplanarity parameters of the package structure using a laser interferometer to generate a three-dimensional deformation matrix , where z is the displacement perpendicular to the package surface, and the measurement accuracy is ≤1μm; S103: Use a multi-channel electrical probe array to synchronously detect the electrical characteristics of the package pins, including open circuit impedance , short-circuit current and leakage conductance Time series data; S104: Regarding the 、 and Perform spatiotemporal alignment, use Z-score normalization method to eliminate dimensional differences, and generate standardized multimodal datasets .

[0006] Furthermore, the dynamic weighted fusion process of S200 includes: S201: According to the thermal expansion coefficient of the packaging material , pin spacing design value , calculate the thermal-deformation coupling weight ,in The largest temperature difference in history; S202: Calculate electrical characteristic weights based on the electrical signal anomaly probability model , where k is the electrical sensitivity coefficient, is the normalized electrical signal amplitude; S203: Generate defect feature matrix F using the following formula:

[0007] in, 、 、 They are the standardized thermal, deformation and electrical signal data respectively.

[0008] Furthermore, the method for dynamically updating the electrical sensitivity coefficient k includes: When N consecutive abnormal electrical signal events are detected, press Make incremental adjustments, where is the electrical signal fluctuation variance, It is the preset adjustment step size, and N≥3.

[0009] Furthermore, the adaptive threshold generation algorithm of S300 includes: S301: Extract K cluster centers that match the current package model from the historical defect pattern library , calculate the Euclidean distance between each center and the current defect feature matrix F ; Among them, n is the feature dimension, is the value of the defect feature matrix F in the i-th feature dimension, is the value of the kth cluster center on the i-th feature dimension, is the Euclidean distance between the kth cluster center and the current defect feature matrix F; S302: Select Satisfaction The minimum distance cluster center ,in is the standard deviation of the historical cluster radius; S303: Generate dynamic threshold ,in for The mean of for The standard deviation of is the gradient impact factor, is the gradient change of the defect feature matrix in the spatial dimension.

[0010] Furthermore, the gradient change The calculation method is:

[0011] in, The feature matrix difference between adjacent detection points is calculated respectively, and the spatial step size is inversely proportional to the size of the package structure.

[0012] Furthermore, the multi-level logic decision model of S400 includes: S401: First-level judgment: If the value of at least 30% of the area in the defect feature matrix F exceeds , it is determined to be a macro defect; S402: Secondary judgment: If the macro defect is not triggered, perform wavelet transform on F to extract the high-frequency component. When , it is judged as micro crack or cavity defect; S403: Level 3 judgment: combined with electrical signal data , when satisfied When the pin is soldered, it is determined to be a pin solder joint defect; S404: Position calibration of the triggered defect type, and use of thermal-deformation cross validation method to determine the defect center coordinates .

[0013] Furthermore, the high frequency energy threshold The dynamic setting method is:

[0014] in, is the average high-frequency energy of historical normal samples, is the maximum high-frequency energy value of historical defect samples.

[0015] Furthermore, the adaptive compensation control of S500 includes: S501: Calculate compensation pressure for warpage defects ,in is the maximum measured deformation, The upper limit of the design allowable deformation is is the material stiffness coefficient; S502: Generate temperature gradient compensation instructions for heat accumulation defects ,in is the hotspot peak temperature, is the average temperature of the packaging area, is the heat dissipation efficiency factor; S503: For electrical defects, adjust the pressure-time curve of the bonding machine and add a delay time ,in is the nominal impedance value, is the process adjustment factor.

[0016] Furthermore, the material stiffness coefficient The method to determine is:

[0017] in, is the elastic modulus of the packaging material, t is the thickness of the packaging layer, is the bonding point contact area, Design value for pin spacing

[0018] The above-described embodiments of the present invention have at least the following beneficial effects: The adaptive semiconductor package inspection method of the present invention enables comprehensive inspection of package structures. By simultaneously collecting multimodal data such as thermal distribution, three-dimensional deformation, and electrical signals using multiple sensors, combined with dynamic weighted fusion technology, it generates a defect feature matrix with temporal and spatial correlations, thereby accurately identifying the type, location, and confidence score of package defects. This method not only detects macroscopic defects but also captures subtle defects such as microcracks or voids through techniques such as wavelet transforms, thereby improving detection sensitivity and accuracy.

[0019] Furthermore, this method can trigger adaptive compensation control of packaging process parameters in real time. Based on the detected defect type and confidence score, feedback control instructions are generated and the actuators of the packaging equipment are adjusted. This real-time feedback mechanism can effectively reduce defects caused by process deviations, optimize the packaging process, improve production efficiency and packaging quality, and reduce production costs and defective product rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which: Figure 1 A schematic flow chart of an adaptive semiconductor package detection method provided in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0022] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0023] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0024] Reference below Figure 1 , Figure 1 FIG1 is a flow chart of an adaptive semiconductor package detection method provided by an embodiment of the present invention. Figure 1 As shown, an adaptive semiconductor package detection method includes: S100: synchronously collecting real-time multimodal data of a semiconductor package structure through multi-source sensors, wherein the multimodal data includes package surface thermal distribution data, three-dimensional deformation data, and pin electrical signal data; S200: performing dynamic weighted fusion on the multimodal data based on packaging process parameters to generate a defect feature matrix with spatiotemporal correlation; S300: performing dynamic threshold calibration on the defect feature matrix using an adaptive threshold generation algorithm according to the cluster center distribution in the historical defect pattern library; S400: Uses a multi-level logic decision model to match dynamic thresholds with the real-time defect feature matrix and outputs defect type, location coordinates, and confidence score. S500: Based on the defect type and confidence score, trigger adaptive compensation control of packaging process parameters, generate feedback control instructions and adjust packaging equipment actuators in real time.

[0025] It should be noted that when using multi-source sensors to synchronously collect real-time multimodal data from semiconductor package structures, multi-source sensors refer to a combination of sensors capable of simultaneously acquiring different physical quantities, including high-resolution infrared thermal imagers, laser interferometers, and multi-channel electrical probe arrays. Multimodal data refers to a collection of data reflecting the state of the package structure from different angles, specifically including package surface thermal distribution data, three-dimensional deformation data, and pin electrical signal data. Package surface thermal distribution data refers to the distribution of the temperature field on the package surface, which can indicate whether the heat distribution within the package is uniform and whether there are localized overheating issues. Three-dimensional deformation data refers to the deformation of the package structure in three dimensions, including parameters such as warpage and coplanarity, which can determine whether the package structure has deformed due to factors such as thermal stress. Pin electrical signal data refers to the electrical characteristics of the package pins, including open-circuit impedance, short-circuit current, and leakage conductance, which are used to detect electrical defects such as cold solder joints and short circuits. The simultaneous acquisition of this data provides a comprehensive and real-time reflection of the physical state of the package structure, providing a rich information foundation for subsequent defect detection.

[0026] Specifically, a high-resolution infrared thermal imager collects temperature data on the package surface at a sampling rate of 20 frames per second, generating a thermal distribution matrix, where x and y represent the package surface coordinates, with a temperature resolution of ≤0.1°C. This means the infrared thermal imager can quickly acquire temperature information at a large number of locations on the package surface with high accuracy, accurately reflecting even subtle changes in the package surface temperature. A laser interferometer measures the warpage and coplanarity parameters of the package structure, generating a three-dimensional deformation matrix, where z represents the displacement perpendicular to the package surface, with a measurement accuracy of ≤1μm. Using the principle of laser interference, the laser interferometer can accurately measure tiny vertical displacements of the package structure, thereby determining the warpage and coplanarity of the package structure. A multi-channel electrical probe array simultaneously monitors the electrical characteristics of the package pins, collecting time-series data including open-circuit impedance, short-circuit current, and leakage conductance. This electrical characteristic data reflects the electrical state of the pins at different points in time, and analysis of this data can be used to determine whether there are any electrical faults in the pins. The collected thermal distribution matrix, three-dimensional deformation matrix, and pin electrical signal data are aligned in time and space, and the Z-score normalization method is used to eliminate dimensional differences and generate a standardized multimodal dataset. Spatiotemporal alignment refers to matching the data collected by different sensors in time and space to ensure that they correspond to the same location and time in the package structure. The Z-score normalization method is a commonly used data normalization method. By converting the data into a standard normal distribution, it eliminates differences between different data dimensions and enables comparison and fusion of different types of multimodal data on the same scale.

[0027] Preferably, when collecting temperature field data on the package surface, a higher-resolution infrared thermal imager can be used to further improve the accuracy and spatial resolution of temperature measurement, so as to more clearly observe the temperature distribution details on the package surface. When measuring the three-dimensional deformation data of the package structure, in addition to the laser interferometer, other high-precision three-dimensional measurement equipment, such as optical coherence tomography (OCT), can also be used to provide more comprehensive deformation information. When collecting pin electrical signal data, different types of electrical probe arrays can be selected according to different package types and detection requirements to adapt to different pin layouts and electrical testing requirements. In addition, when performing spatiotemporal alignment and data standardization, advanced data processing algorithms, such as machine learning algorithms, can be combined to pre-process and extract features from the collected multimodal data to further improve the quality and availability of the data, and provide more accurate data support for subsequent defect detection and analysis.

[0028] In some embodiments, the S100 specifically includes: S101: Use a high-resolution infrared thermal imager to collect package surface temperature field data at a sampling rate of 20 frames per second to generate a thermal distribution matrix , where x, y are the coordinates of the package surface, and the temperature resolution is ≤ 0.1°C; S102: Measure the warpage and coplanarity parameters of the package structure using a laser interferometer to generate a three-dimensional deformation matrix , where z is the displacement perpendicular to the package surface, and the measurement accuracy is ≤1μm; S103: Use a multi-channel electrical probe array to synchronously detect the electrical characteristics of the package pins, including open circuit impedance , short-circuit current and leakage conductance Time series data; S104: Regarding the 、 and Perform spatiotemporal alignment, use Z-score normalization method to eliminate dimensional differences, and generate standardized multimodal datasets .

[0029] It should be noted that when using multi-source sensors to synchronously collect real-time multimodal data from semiconductor package structures, multi-source sensors refer to a combination of sensors capable of simultaneously acquiring different physical quantities, including high-resolution infrared thermal imagers, laser interferometers, and multi-channel electrical probe arrays. These sensors can reflect the state of the package structure from different angles, providing rich information for subsequent defect detection. Multimodal data encompasses package surface thermal distribution data, three-dimensional deformation data, and pin electrical signal data. The simultaneous collection of these data provides a comprehensive, real-time overview of the physical state of the package structure. Surface thermal distribution data is used to monitor internal heat uniformity and the presence of localized overheating. Three-dimensional deformation data is used to determine whether the package structure has deformed due to factors such as thermal stress. Pin electrical signal data is used to detect electrical defects such as cold solder joints and short circuits.

[0030] Specifically, a high-resolution infrared thermal imager collects temperature data on the package surface at a sampling rate of 20 frames per second, generating a thermal distribution matrix, where x and y represent the package surface coordinates, with a temperature resolution of ≤0.1°C. This means the infrared thermal imager can quickly acquire temperature information at a large number of locations on the package surface with high accuracy, accurately reflecting even subtle changes in the package surface temperature. A laser interferometer measures the warpage and coplanarity parameters of the package structure, generating a three-dimensional deformation matrix, where z represents the displacement perpendicular to the package surface, with a measurement accuracy of ≤1μm. Using the principle of laser interference, the laser interferometer can accurately measure tiny vertical displacements of the package structure, thereby determining the warpage and coplanarity of the package structure. A multi-channel electrical probe array simultaneously monitors the electrical characteristics of the package pins, collecting time-series data including open-circuit impedance, short-circuit current, and leakage conductance. This electrical characteristic data reflects the electrical state of the pins at different points in time, and analysis of this data can be used to determine whether there are any electrical faults in the pins. The collected thermal distribution matrix, three-dimensional deformation matrix, and pin electrical signal data are aligned in time and space, and the Z-score normalization method is used to eliminate dimensional differences and generate a standardized multimodal dataset. Spatiotemporal alignment refers to matching the data collected by different sensors in time and space to ensure that they correspond to the same location and time in the package structure. The Z-score normalization method is a commonly used data normalization method. By converting the data into a standard normal distribution, it eliminates differences between different data dimensions and enables comparison and fusion of different types of multimodal data on the same scale.

[0031] Preferably, when collecting temperature field data on the package surface, a higher-resolution infrared thermal imager can be used to further improve the accuracy and spatial resolution of temperature measurement, so as to more clearly observe the temperature distribution details on the package surface. When measuring the three-dimensional deformation data of the package structure, in addition to the laser interferometer, other high-precision three-dimensional measurement equipment, such as optical coherence tomography (OCT), can also be used to provide more comprehensive deformation information. When collecting pin electrical signal data, different types of electrical probe arrays can be selected according to different package types and detection requirements to adapt to different pin layouts and electrical testing requirements. In addition, when performing spatiotemporal alignment and data standardization, advanced data processing algorithms, such as machine learning algorithms, can be combined to pre-process and extract features from the collected multimodal data to further improve the quality and availability of the data, and provide more accurate data support for subsequent defect detection and analysis.

[0032] In some embodiments, the dynamic weighted fusion process of S200 includes: S201: According to the thermal expansion coefficient of the packaging material , pin spacing design value , calculate the thermal-deformation coupling weight ,in The largest temperature difference in history; S202: Calculate electrical characteristic weights based on the electrical signal anomaly probability model , where k is the electrical sensitivity coefficient, is the normalized electrical signal amplitude; S203: Generate defect feature matrix F using the following formula:

[0033] in, 、 、 These are the standardized thermal, deformation and electrical signal data respectively.

[0034] It should be noted that in the dynamic weighted fusion process, the calculation of thermal-deformation coupling weights and electrical characteristic weights are key steps, as they determine the contribution of different data to the defect feature matrix. The thermal-deformation coupling weights are calculated based on the thermal expansion coefficient of the packaging material and the design value of the pin spacing, reflecting the degree of correlation between thermal changes and deformation. The thermal expansion coefficient of the packaging material refers to the rate of change of the volume or length of the material when the temperature changes, and the design value of the pin spacing is the standard distance between the pins in the package design. The electrical characteristic weights are calculated based on the electrical signal anomaly probability model and reflect the possibility of electrical signal anomalies. By multiplying and adding these weights with the standardized data, a defect feature matrix is generated. This matrix can comprehensively reflect the thermal, deformation, and electrical characteristic states of the packaging structure, providing a basis for subsequent defect detection.

[0035] Specifically, the calculation of the thermal-deformation coupling weight takes into account the package material's thermal expansion coefficient, the historical maximum temperature difference, and the designed pin pitch. A larger thermal expansion coefficient indicates more significant material deformation during temperature fluctuations; the historical maximum temperature difference reflects the range of temperature fluctuations; larger temperature differences indicate more significant thermal impacts; and smaller designed pin pitches indicate greater inter-pin influence. Therefore, the thermal-deformation coupling weight integrates these factors to reflect the extent of thermal influence on deformation. The electrical characteristic weight is calculated based on an electrical signal anomaly probability model, taking into account the electrical sensitivity coefficient and normalized electrical signal amplitude. The electrical sensitivity coefficient is an adjustment parameter used to adjust the weight based on the degree of electrical signal anomaly; the normalized electrical signal amplitude reflects the degree of anomaly. A larger electrical signal amplitude results in a larger electrical characteristic weight, indicating a higher probability of anomaly. By multiplying and summing these weights with the standardized thermal, deformation, and electrical signal data, a defect signature matrix is generated. This matrix integrates the effects of thermal, deformation, and electrical characteristics, providing a more comprehensive representation of the defect characteristics of the package structure.

[0036] Preferably, when calculating the thermal-deformation coupling weight, it can be adjusted according to the specific characteristics of the packaging material. For example, for materials with a higher thermal expansion coefficient, the calculated value of the weight can be appropriately increased to more accurately reflect the impact of thermal changes on deformation. When calculating the electrical characteristic weight, it can be optimized according to the package type and common defect type. For example, if a certain package type is more prone to short-circuit failures, the weight corresponding to the short-circuit current can be appropriately increased to improve the detection sensitivity of short-circuit defects. In addition, the electrical sensitivity coefficient can be dynamically adjusted according to the fluctuation of the electrical signal during the actual detection process. For example, when multiple consecutive electrical signal abnormality events are detected, the electrical sensitivity coefficient can be dynamically updated according to certain rules, thereby improving the response capability to electrical signal abnormalities. This dynamic adjustment mechanism can better adapt to complex and changeable packaging environments and improve the accuracy and reliability of defect detection.

[0037] In some embodiments, the method for dynamically updating the electrical sensitivity coefficient k includes: When N consecutive abnormal electrical signal events are detected, press Make incremental adjustments, where is the electrical signal fluctuation variance, It is the preset adjustment step size, and N≥3.

[0038] It should be noted that in the dynamic update method for the electrical sensitivity coefficient, the electrical sensitivity coefficient is a key parameter used to adjust the impact of electrical signal anomalies on defect detection. When multiple consecutive electrical signal anomalies are detected, it indicates that the current electrical sensitivity coefficient is insufficient to reflect the electrical signal anomaly, and therefore needs to be dynamically adjusted. The electrical signal fluctuation variance is a statistic that measures the degree of change in the electrical signal, reflecting the fluctuation of the electrical signal over a period of time. The preset adjustment step size is a pre-set parameter used to control the amplitude of each adjustment of the electrical sensitivity coefficient. This dynamic update mechanism ensures that the electrical sensitivity coefficient can be adjusted promptly based on actual detection conditions, thereby improving the accuracy and adaptability of defect detection.

[0039] Specifically, the dynamic update method of the electrical sensitivity coefficient takes into account the number of consecutive occurrences of electrical signal abnormal events, the electrical signal fluctuation variance, and the preset adjustment step size. When N consecutive electrical signal abnormal events are detected, it means that the current electrical sensitivity coefficient needs to be increased to better reflect the abnormality of the electrical signal. The electrical signal fluctuation variance is used to measure the degree of fluctuation of the electrical signal. The larger the fluctuation variance, the higher the degree of abnormality of the electrical signal, so the electrical sensitivity coefficient needs to be increased accordingly. The preset adjustment step size determines the amplitude of each adjustment and can be set according to actual detection requirements. For example, if electrical signal abnormal events occur frequently, the preset adjustment step size can be appropriately increased to speed up the adjustment of the electrical sensitivity coefficient. Through this dynamic update method, the electrical sensitivity coefficient can be adjusted in a timely manner according to the actual detection situation, thereby improving the accuracy and adaptability of defect detection.

[0040] Preferably, the dynamic update method of the electrical sensitivity coefficient can be further refined. For example, a threshold can be set, and the electrical sensitivity coefficient will only be adjusted when the number of consecutive occurrences of electrical signal abnormal events exceeds the threshold, so as to avoid frequent adjustments of the electrical sensitivity coefficient due to accidental electrical signal abnormal events. In addition, different adjustment strategies for the electrical sensitivity coefficient can be implemented according to the type and severity of the electrical signal abnormal events. For example, for severe short-circuit faults, the adjustment step size can be appropriately increased to quickly improve the electrical sensitivity coefficient; while for minor leakage faults, the adjustment step size can be appropriately reduced to avoid over-adjustment. In addition, a machine learning algorithm can be introduced to automatically adjust the update strategy of the electrical sensitivity coefficient by learning and analyzing historical data, thereby further improving the accuracy and adaptability of defect detection.

[0041] In some embodiments, the adaptive threshold generation algorithm of S300 includes: S301: Extract K cluster centers that match the current package model from the historical defect pattern library , calculate the Euclidean distance between each center and the current defect feature matrix F ; Among them, n is the feature dimension, is the value of the defect feature matrix F in the i-th feature dimension, is the value of the kth cluster center on the i-th feature dimension, is the Euclidean distance between the kth cluster center and the current defect feature matrix F; S302: Select Satisfaction The minimum distance cluster center ,in is the standard deviation of the historical cluster radius; S303: Generate dynamic threshold ,in for The mean of for The standard deviation of is the gradient impact factor, is the gradient change of the defect feature matrix in the spatial dimension.

[0042] It should be noted that the adaptive threshold generation algorithm dynamically calibrates the threshold of the defect feature matrix based on the distribution of cluster centers in the historical defect pattern library. The historical defect pattern library is a database that stores previously detected defect feature data. By performing cluster analysis on this data, cluster centers for different defect types can be determined. The distribution of cluster centers reflects the location and range of different defect types in the feature space. The adaptive threshold generation algorithm uses the Euclidean distance between these cluster centers and the current defect feature matrix to select the closest cluster center and generate a dynamic threshold based on this cluster center. This method dynamically adjusts the threshold based on historical data, thereby improving the accuracy and adaptability of defect detection.

[0043] Specifically, the adaptive threshold generation algorithm first extracts K cluster centers that match the current package model from the historical defect pattern library. These cluster centers are obtained by clustering analysis of historical defect data, and each cluster center represents a specific defect type. Then, the Euclidean distance between each cluster center and the current defect feature matrix is calculated. The smaller the Euclidean distance, the more similar the current defect feature is to the defect type represented by the cluster center. The minimum distance cluster center that meets certain conditions is selected. For example, when the Euclidean distance is less than 3 times the standard deviation of the historical cluster radius, it is considered that the current defect feature is close enough to the defect type represented by the cluster center. Finally, a dynamic threshold is generated based on the mean, standard deviation and gradient influence factor of the cluster center. The gradient influence factor is an adjustment parameter used to adjust the threshold according to the gradient change of the defect feature matrix in the spatial dimension, so as to better adapt to the defect detection needs of different packaging structures.

[0044] Preferably, the adaptive threshold generation algorithm can be further refined. For example, when selecting cluster centers, more statistical indicators can be introduced to evaluate the similarity between the cluster centers and the current defect feature matrix, rather than just the Euclidean distance. Furthermore, the gradient influence factor can be dynamically adjusted based on the complexity of the package structure. For complex package structures, the value of the gradient influence factor can be appropriately increased to improve the detection sensitivity of minor defects. Furthermore, machine learning algorithms can be introduced to automatically adjust the parameters of the threshold generation algorithm by learning and analyzing historical defect data, thereby further improving the accuracy and adaptability of defect detection. For example, algorithms such as support vector machines (SVMs) or neural networks can be used to optimize the selection of cluster centers and the generation of dynamic thresholds.

[0045] In some embodiments, the gradient variation The calculation method is:

[0046] in, The feature matrix difference between adjacent detection points is calculated respectively, and the spatial step size is inversely proportional to the size of the package structure.

[0047] It should be noted that the calculation method for gradient variation is based on the spatial variation of the defect signature matrix. The gradient variation reflects the rate of change of the defect signature in space and is an important parameter for assessing defect distribution and severity. By calculating the gradient variation, the location and extent of the defect can be more accurately determined, thereby improving defect detection accuracy. The spatial step size refers to the spatial interval used in gradient calculation. It is inversely proportional to the package size, meaning that for smaller packages, the spatial step size should be smaller to ensure calculation accuracy.

[0048] Specifically, the calculation of the gradient change involves calculating the difference of the defect feature matrix in three spatial directions. First, the gradient is approximately calculated by the difference of the feature matrices of adjacent detection points. For example, in the x-direction, the difference of the feature values of two adjacent points is calculated and divided by the spatial step to obtain the gradient in the x-direction. Similarly, the gradients in the y-direction and z-direction are calculated. Then, the gradient values in these three directions are squared and squared to obtain the magnitude of the gradient change. This method can comprehensively consider the changes in the defect characteristics in the three directions, thereby more comprehensively reflecting the spatial distribution characteristics of the defects. The choice of spatial step should be determined according to the specific size of the packaging structure. For example, for a smaller packaging structure, the spatial step can be set to 0.1mm, while for a larger packaging structure, the spatial step can be appropriately increased to 1mm.

[0049] Preferably, the calculation method of the gradient change can be further refined. For example, when calculating the gradient, a more advanced numerical differentiation method, such as the central difference method, can be used to improve the accuracy of the gradient calculation. The central difference method approximates the gradient by calculating the average difference between two adjacent points, which can more accurately reflect the changes in the defect characteristics. In addition, the concept of adaptive spatial step size can be introduced to dynamically adjust the spatial step size according to the local changes in the defect feature matrix. For example, in areas where the defect characteristics change greatly, the spatial step size can be reduced to improve the calculation accuracy; in areas where the defect characteristics change less, the spatial step size can be appropriately increased to improve the calculation efficiency. In addition, image processing technology, such as edge detection algorithm, can be combined to pre-process the defect feature matrix to enhance the calculation effect of the gradient change.

[0050] In some embodiments, the multi-level logic decision model of S400 includes: S401: First-level judgment: If the value of at least 30% of the area in the defect feature matrix F exceeds , it is determined to be a macro defect; S402: Secondary judgment: If the macro defect is not triggered, perform wavelet transform on F to extract the high-frequency component. When , it is judged as micro crack or cavity defect; S403: Level 3 judgment: combined with electrical signal data , when satisfied When the pin is soldered, it is determined to be a pin solder joint defect; S404: Position calibration of the triggered defect type, and use of thermal-deformation cross validation method to determine the defect center coordinates .

[0051] It should be noted that the multi-level logic decision model is a hierarchical judgment method used to match dynamic thresholds with real-time defect feature matrices to determine the type of defect. This model gradually narrows the range of defect types through step-by-step judgment, thereby improving the accuracy and reliability of the judgment. The first-level judgment is mainly based on the overall characteristics of the defect feature matrix to determine whether there are macro defects; the second-level judgment extracts high-frequency components through wavelet transform to detect micro defects such as microcracks or voids; the third-level judgment combines electrical signal data to further confirm the specific type of electrical defect. Finally, the center coordinates of the defect are determined through the thermal-deformation cross-validation method, providing accurate defect location information for subsequent process adjustments.

[0052] Specifically, in the first-level judgment, if at least 30% of the region values in the defect feature matrix exceed the dynamic threshold, the defect is identified as a macro defect. A region here refers to a continuous portion of the defect feature matrix with a certain spatial range. The dynamic threshold is derived from an adaptive threshold generation algorithm and serves as the critical value for distinguishing between normal and defective states. In the second-level judgment, the defect feature matrix is subjected to a wavelet transform to extract the high-frequency component. Wavelet transform is a method that can analyze signals simultaneously in time and frequency. High-frequency components reflect rapidly changing portions of the signal and are typically associated with microscopic defects such as microcracks or voids. When the high-frequency energy exceeds a set high-frequency energy threshold, the defect is identified as a microcrack or void defect. In the third-level judgment, the defect is identified as a pin solder joint defect based on the electrical signal data. For example, if the open-circuit impedance in the electrical signal data exceeds a certain threshold and the high-frequency energy is below another threshold, the defect is identified as a pin solder joint defect. Finally, the defect center coordinates are determined using a thermal-deformation cross-validation method. This method leverages the correlation between thermal distribution and deformation data and calculates the product of the two to determine the defect center location.

[0053] Preferably, each step of the multi-level logical decision model can be further refined. For example, in the first-level judgment, more characteristic parameters can be introduced to improve the judgment accuracy of macro defects. In addition to the proportion of regional values exceeding the dynamic threshold, statistical parameters such as the overall mean and standard deviation of the defect feature matrix can also be considered. In the second-level judgment, the basis function selection and decomposition layer number of the wavelet transform can be optimized according to the specific defect characteristics. For example, for certain types of microcracks, specific wavelet basis functions can be selected to improve the extraction effect of high-frequency components. In the third-level judgment, more electrical signal features, such as the rate of change of leakage conductance, can be combined to more accurately judge the pin solder joint defect. In addition, in the thermal-deformation cross-validation method, a weighting factor can be introduced to weight the contribution of the thermal distribution and deformation data according to the reliability of the two, so as to more accurately determine the defect center coordinates.

[0054] In some embodiments, the high frequency energy threshold The dynamic setting method is:

[0055] in, is the average high-frequency energy of historical normal samples, is the maximum high-frequency energy value of historical defect samples.

[0056] It should be noted that the dynamic setting method for the high-frequency energy threshold is determined based on the high-frequency energy statistical characteristics of historical normal samples and historical defective samples. This method can dynamically adjust the threshold based on actual detection data, thereby improving the accuracy and adaptability of defect detection. Historical normal samples refer to sample data collected under normal packaging processes, and their high-frequency energy reflects the signal characteristics under normal conditions. Historical defective samples refer to sample data with known defects, and their high-frequency energy reflects the signal characteristics under defective conditions. By combining the high-frequency energy characteristics of these two types of samples, a dynamic high-frequency energy threshold can be generated to distinguish between normal and defective conditions.

[0057] Specifically, the dynamic setting method for the high-frequency energy threshold takes into account the historical average high-frequency energy value of normal samples and the historical maximum high-frequency energy value of defective samples. The historical average high-frequency energy value of normal samples reflects the typical level of high-frequency energy under normal conditions, while the historical maximum high-frequency energy value of defective samples represents the upper limit of high-frequency energy under defective conditions. By taking a weighted average of these two values, a dynamic high-frequency energy threshold can be obtained. For example, the average high-frequency energy value of normal samples can be multiplied by 0.7, the maximum high-frequency energy value of defective samples can be multiplied by 0.3, and then these two weighted values can be added together to obtain the final high-frequency energy threshold. This weighting method can be adjusted according to actual detection needs to better balance the ability to distinguish between normal and defective conditions.

[0058] Preferably, the method for dynamically setting the high-frequency energy threshold can be further refined. For example, the high-frequency energy statistical characteristics of historical normal samples and historical defective samples can be calculated separately according to different packaging types and process conditions, thereby generating more targeted dynamic thresholds. In addition, machine learning algorithms can be introduced to automatically adjust the weighting coefficients by learning and analyzing historical data to further improve the accuracy and adaptability of threshold setting. For example, a neural network algorithm can be used to train historical data, and the weighting coefficients can be dynamically adjusted based on the training results to better adapt to the defect detection needs under different packaging structures and process conditions.

[0059] In some embodiments, the adaptive compensation control in S500 includes: S501: Calculate compensation pressure for warpage defects ,in is the maximum measured deformation, The upper limit of the design allowable deformation is is the material stiffness coefficient; S502: Generate temperature gradient compensation instructions for heat accumulation defects ,in is the hotspot peak temperature, is the average temperature of the packaging area, is the heat dissipation efficiency factor; S503: For electrical defects, adjust the pressure-time curve of the bonding machine and add a delay time ,in is the nominal impedance value, is the process adjustment factor.

[0060] It should be noted that adaptive compensation control is a method that adjusts packaging process parameters in real time based on defect detection results. The goal is to optimize the packaging process and reduce defects through a feedback mechanism. Adaptive compensation control includes targeted compensation measures for warpage, thermal stress, and electrical defects. Warpage refers to bending and deformation of the package structure caused by factors such as thermal stress; thermal stress refers to localized overheating caused by uneven heat distribution within the package; and electrical defects refer to problems such as poor electrical connections on the pins. By calculating compensation pressure, providing temperature gradient compensation instructions, and adjusting the bonder's pressure-time curve, these defects can be effectively corrected to improve packaging quality.

[0061] Specifically, for warpage defects, the compensation pressure is calculated based on the measured maximum deformation, the design allowable upper limit of deformation, and the material stiffness coefficient. The measured maximum deformation refers to the maximum bending of the package structure determined through testing; the design allowable upper limit of deformation refers to the maximum bending limit allowed in the package design; and the material stiffness coefficient reflects the material's ability to resist deformation. The product of these three parameters yields the compensation pressure used to correct warpage. For heat accumulation defects, the generation of temperature gradient compensation instructions takes into account the hotspot peak temperature, the average package area temperature, and the heat dissipation efficiency factor. The hotspot peak temperature refers to the temperature at the highest point within the package; the average package area temperature is the average temperature across the entire package area; and the heat dissipation efficiency factor reflects the heat dissipation capacity of the package material. The difference and product of these parameters are calculated to generate the temperature gradient compensation instructions used to adjust heat dissipation. For electrical defects, when adjusting the bonder's pressure-time curve, the added delay time is calculated based on the nominal impedance value and the process adjustment factor. The nominal impedance value is the impedance value of the pin in the normal connection state; the process adjustment factor is an adjustment parameter set based on actual process conditions. By calculating the difference and product of these two parameters, the adjustment amount for the bonder's pressure-time curve can be determined.

[0062] Preferably, the various steps of the adaptive compensation control can be further refined. For example, when calculating the compensation pressure for warpage defects, more material characteristic parameters, such as elastic modulus and Poisson's ratio, can be introduced to more accurately reflect the mechanical behavior of the material. For thermal aggregation defects, the temperature gradient compensation instructions can be more accurately calculated by combining heat flow analysis and simulating the heat conduction process inside the package. For electrical defects, a real-time monitoring mechanism can be introduced to dynamically adjust the pressure-time curve during the bonding process to ensure the reliability of the pin connection. In addition, an intelligent control system, such as a fuzzy logic controller or a PID controller, can be introduced to automatically adjust the compensation parameters according to real-time detection data, thereby further improving the automation and intelligence level of the packaging process.

[0063] In some embodiments, the material stiffness coefficient The method to determine is:

[0064] in, is the elastic modulus of the packaging material, t is the thickness of the packaging layer, is the bonding point contact area, Design value for pin spacing.

[0065] It should be noted that the material stiffness coefficient is determined based on the physical properties of the packaging material. It is a parameter calculated to characterize the material's ability to resist deformation. This coefficient is used in adaptive compensation control to calculate the compensating pressure for warpage defects, ensuring that the package structure maintains a good shape and stability during processing. The elastic modulus of the packaging material is an indicator of the material's stiffness during the elastic deformation stage, reflecting the material's ability to resist deformation under stress. The package layer thickness is the thickness of the material layer within the package structure and affects the overall stiffness of the material. The bond contact area refers to the contact area between the pin and the packaging material, which affects the force transfer efficiency. The pin pitch design value is the standard distance between pins in the package design, which affects the overall layout and stability of the package structure. Through the comprehensive calculation of these parameters, a reasonable material stiffness coefficient can be obtained to guide the optimization of the packaging process.

[0066] Specifically, the calculation of the material stiffness coefficient involves the elastic modulus of the packaging material, the thickness of the packaging layer, the bond point contact area, and the designed pin pitch. The elastic modulus of the packaging material is an inherent material property and can usually be obtained from material manuals or experimental testing. The thickness of the packaging layer can be determined based on the actual packaging design and is generally in the micron to millimeter range. The bond point contact area depends on the size and shape of the pins and can be calculated through actual measurements or design parameters. The designed pin pitch is a key parameter in packaging design and is usually determined based on the chip's pin layout and packaging process requirements. By substituting these parameters into the calculation formula, the material stiffness coefficient can be obtained. For example, for a common semiconductor packaging material with an elastic modulus of 100 GPa, a packaging layer thickness of 0.5 mm, a bond point contact area of 0.1 mm², and a designed pin pitch of 1 mm, the corresponding material stiffness coefficient can be obtained through calculation.

[0067] Preferably, the calculation method for the material stiffness coefficient can be further refined. For example, the temperature dependence of the material can be taken into account, as the elastic modulus of the packaging material changes with temperature. In this case, the elastic modulus can be measured at different temperatures and weighted averaged according to the actual operating temperature range to obtain a more accurate material stiffness coefficient. In addition, for complex packaging structures, finite element analysis (FEA) can be used to simulate the mechanical behavior of the material, thereby more accurately determining the material stiffness coefficient. FEA can take into account factors such as the material's nonlinear properties, geometry, and boundary conditions, providing a more comprehensive assessment of mechanical properties. Through these methods, the accuracy and reliability of the material stiffness coefficient can be further improved, thereby optimizing the packaging process and improving packaging quality.

[0068] The above-mentioned embodiments of the present invention have the following beneficial effects: The adaptive semiconductor packaging detection method of the present invention can achieve comprehensive detection of the packaging structure. By synchronously collecting multimodal data such as thermal distribution, three-dimensional deformation, and electrical signals through multi-source sensors, combined with dynamic weighted fusion technology, it can generate a defect feature matrix with spatiotemporal correlation, thereby accurately identifying the type, location, and confidence score of the packaging defect. This method can not only detect macro defects, but also capture micro defects such as microcracks or voids through technologies such as wavelet transform, thereby improving the sensitivity and accuracy of detection. At the same time, the dynamic weighted fusion process can calculate the thermal-deformation coupling weight based on the thermal expansion coefficient of the packaging material and the pin spacing design value, and calculate the electrical characteristic weight based on the electrical signal abnormality probability model, further optimizing the generation of the defect feature matrix and improving the reliability of detection.

[0069] Furthermore, this method can trigger adaptive compensation control of packaging process parameters in real time, generating feedback control instructions and adjusting the actuators of the packaging equipment based on the detected defect type and confidence score. This real-time feedback mechanism can effectively reduce defects caused by process deviations, optimize the packaging process, improve production efficiency and packaging quality, and reduce production costs and defective product rates. The adaptive threshold generation algorithm can dynamically calibrate the threshold based on the distribution of cluster centers in the historical defect pattern library. The multi-level logic decision model can match and judge the dynamic threshold with the real-time defect feature matrix, further improving the accuracy and adaptability of defect detection. At the same time, the dynamic setting method of the high-frequency energy threshold and the targeted compensation control for warpage, thermal accumulation, and electrical defects can effectively address different types of packaging defects and improve overall packaging quality.

[0070] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, server, mobile phone, or tablet.

[0071] The above descriptions merely illustrate some preferred embodiments of the present invention and the underlying technical principles. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present invention is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. An adaptive semiconductor package detection method, characterized in that: The following steps are involved: S100: synchronously collecting real-time multimodal data of a semiconductor package structure through multi-source sensors, wherein the multimodal data includes package surface thermal distribution data, three-dimensional deformation data, and pin electrical signal data; S200: performing dynamic weighted fusion on the multimodal data based on packaging process parameters to generate a defect feature matrix with spatiotemporal correlation; S300: performing dynamic threshold calibration on the defect feature matrix using an adaptive threshold generation algorithm according to the cluster center distribution in the historical defect pattern library; S400: Uses a multi-level logic decision model to match dynamic thresholds with the real-time defect feature matrix and outputs defect type, location coordinates, and confidence score. S500: Based on the defect type and confidence score, trigger adaptive compensation control of packaging process parameters, generate feedback control instructions and adjust the packaging equipment actuator in real time.

2. The adaptive semiconductor package detection method according to claim 1, wherein: The S100 specifically includes: S101: Use a high-resolution infrared thermal imager to collect package surface temperature field data at a sampling rate of 20 frames per second to generate a thermal distribution matrix; S102: Measure the warpage and coplanarity parameters of the package structure using a laser interferometer to generate a three-dimensional deformation matrix; S103: Use a multi-channel electrical probe array to synchronously detect the electrical characteristics of the package pins, including open circuit impedance , short-circuit current and leakage conductance Time series data; S104: performing spatiotemporal alignment on the heat distribution matrix, the three-dimensional deformation matrix, and the time series data, using a Z-score normalization method to eliminate dimensional differences, and generating a standardized multimodal data set.

3. The adaptive semiconductor package detection method according to claim 1, wherein: The dynamic weighted fusion process of S200 includes: S201: According to the thermal expansion coefficient of the packaging material , pin spacing design value , calculate the thermal-deformation coupling weight as shown in the following formula; ,in The largest temperature difference in history. is the thermal-deformation coupling weight; S202: Calculate the electrical characteristic weight based on the electrical signal abnormality probability model, as shown in the following formula; , where k is the electrical sensitivity coefficient, is the normalized electrical signal amplitude, is the electrical characteristic weight; S203: Generate defect feature matrix F using the following formula: ； in, 、 、 These are the standardized thermal, deformation and electrical signal data respectively.

4. The adaptive semiconductor package detection method according to claim 3, wherein: The method for dynamically updating the electrical sensitivity coefficient k includes: When N consecutive abnormal electrical signal events are detected, press Make incremental adjustments, where is the electrical signal fluctuation variance, It is the preset adjustment step size, and N≥3.

5. The adaptive semiconductor package detection method according to claim 1, wherein: The adaptive threshold generation algorithm of S300 includes: S301: Extract K cluster centers that match the current package model from the historical defect pattern library , calculate the Euclidean distance between each center and the current defect feature matrix F ; Among them, n is the feature dimension, is the value of the defect feature matrix F in the i-th feature dimension, is the value of the kth cluster center on the i-th feature dimension, is the Euclidean distance between the kth cluster center and the current defect feature matrix F; S302: Select Satisfaction The minimum distance cluster center ,in, is the standard deviation of the historical cluster radius; S303: Generate dynamic threshold ,in for The mean of for The standard deviation of is the gradient impact factor, is the gradient change of the defect feature matrix in the spatial dimension.

6. The adaptive semiconductor package detection method according to claim 5, wherein: The gradient change The calculation method is: ； in, The feature matrix difference between adjacent detection points is calculated respectively, and the spatial step size is inversely proportional to the size of the package structure.

7. The adaptive semiconductor package detection method according to claim 1, wherein: The multi-level logic decision model of S400 includes: S401: First-level judgment: If the value of at least 30% of the area in the defect feature matrix F exceeds , it is determined to be a macro defect; S402: Secondary judgment: If the macro defect is not triggered, perform wavelet transform on F to extract the high-frequency component. When , it is judged as micro crack or cavity defect; S403: Level 3 judgment: combined with electrical signal data , when satisfied When the pin is soldered, it is determined to be a pin solder joint defect; S404: Position calibration is performed on the triggered defect type, and the defect center coordinates are determined using a thermal-deformation cross-validation method.

8. The adaptive semiconductor package detection method according to claim 7, wherein: The high-frequency energy threshold The dynamic setting method is: ； in, is the average high-frequency energy of historical normal samples, is the maximum high-frequency energy value of historical defect samples, is the high frequency energy threshold.

9. The adaptive semiconductor package detection method according to claim 1, wherein: The adaptive compensation control of S500 includes: S501: Calculate compensation pressure for warpage defects ,in, is the maximum measured deformation, The upper limit of the design allowable deformation is is the material stiffness coefficient, Compensation pressure for correcting warpage defects; S502: Generate temperature gradient compensation instructions for heat accumulation defects ,in is the hotspot peak temperature, is the average temperature of the packaging area, is the heat dissipation efficiency factor, It is the temperature gradient compensation instruction; S503: For electrical defects, adjust the pressure-time curve of the bonding machine and add a delay time ,in is the nominal impedance value, is the process adjustment coefficient, A new delay time is added when adjusting the pressure-time curve of the bonding machine. is the open circuit impedance.

10. The adaptive semiconductor package detection method according to claim 9, wherein: The material stiffness coefficient The method for determining is shown in the following formula: ； in, is the elastic modulus of the packaging material, t is the thickness of the packaging layer, is the bonding point contact area, Design value for pin spacing.

Images