Integrated packaging system and method based on glass substrate

Abstract

The invention relates to the technical field of semiconductor packaging, in particular to an integrated packaging system and method based on a glass substrate, and the method comprises the steps: collecting a time sequence process parameter set through a sensor array on a packaging production line, and generating a process stability index set through feature extraction and abnormal point screening, and inputting a pre-constructed packaging quality correlation model to obtain a multi-dimensional reliability prediction score, matching a process parameter optimization scheme, dynamically adjusting key parameters of production equipment, and starting a verification packaging process. And performing nondestructive testing on the sample to obtain an internal defect distribution diagram and interface bonding strength data, feeding the data into a packaging process tracing analysis engine in combination with a process stability index set, reversely identifying key process links and parameter deviations corresponding to defects, generating a closed-loop process correction instruction, and updating the knowledge base. According to the method, packaging process dynamic regulation and control and defect closed-loop tracing are realized, the packaging process stability and the structural reliability are improved, and the production efficiency is optimized.

Description

Technical Field

[0001] This invention relates to the field of semiconductor packaging technology, and in particular to an integrated packaging system and method based on a glass substrate. Background Technology

[0002] In existing glass substrate integrated packaging production processes, single-point sensors are often used to collect discrete process parameters. Parameter verification is performed manually or through conventional statistical methods. Packaging production is carried out according to fixed preset parameters, and packaging quality is determined by post-production inspection. Adjustments to process parameters rely on offline analysis based on human experience, and packaging quality prediction is achieved solely based on single-dimensional indicators. Defect detection depends on manual post-production verification of each process step. This traditional packaging control method lacks a systematic parameter acquisition and analysis system, process optimization execution lacks standardized matching logic, and adjustments to production parameters cannot be linked to the production process in real time.

[0003] Traditional packaging control models cannot continuously collect and process time-series parameters, nor can they transform the process operation status into a set of quantifiable process stability indicators. They lack packaging quality correlation models to perform pre-calculation of multi-dimensional reliability, cannot rely on packaging process knowledge bases to automatically match process parameter optimization schemes, and cannot dynamically adjust production equipment parameters and simultaneously start verification packaging processes.

[0004] The internal defect distribution and interface bonding strength data obtained from non-destructive testing are only used for defect judgment and cannot be linked with the set of process stability indicators for analysis. There is a lack of a packaging process traceability analysis engine to achieve reverse derivation, making it impossible to locate the key process links and parameter deviations corresponding to specific defects, and unable to generate closed-loop process correction instructions and update the packaging process knowledge base in real time. This invention aims to solve the above-mentioned technical problems in process control and defect traceability. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and to propose an integrated packaging system and method based on a glass substrate.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an integrated packaging method based on a glass substrate, comprising:

[0007] By collecting a set of time-series process parameters in the production process through a sensor array deployed on the packaging production line, feature extraction and anomaly screening are performed on the set of time-series process parameters to generate a set of process stability indicators.

[0008] The process stability index set is input into the pre-built packaging quality correlation model to calculate the multi-dimensional reliability prediction score of the packaging structure.

[0009] Based on the multi-dimensional reliability prediction score, the corresponding process parameter optimization scheme is matched in the packaging process knowledge base;

[0010] The process parameter optimization scheme is implemented to dynamically adjust the key control parameters of the production equipment and initiate a verification packaging process.

[0011] After the verification packaging process is completed, the produced sample is subjected to non-destructive testing to obtain the internal defect distribution map and interface bonding strength data of the sample.

[0012] The internal defect distribution map and interface strength data, together with the set of process stability indicators, are sent into the packaging process traceability analysis engine.

[0013] By using the packaging process traceability analysis engine to reverse-engineer, the key process steps and their parameter deviations that lead to specific defects can be identified.

[0014] Based on the identification results, targeted closed-loop process correction instructions are generated to update the corresponding entries in the packaging process knowledge base.

[0015] As a further aspect of the present invention, feature extraction and outlier screening are performed on the time-series process parameter set to generate a process stability index set, including:

[0016] The set of timing process parameters includes substrate positioning data, colloidal coating morphology data, bonding pressure waveform data, and curing temperature field data.

[0017] The set of process stability indicators includes positioning offset, glue line uniformity, pressure fluctuation coefficient, and temperature gradient standard deviation.

[0018] The substrate positioning data sequence is processed by sliding window processing. The centroid drift and standard deviation of the positioning coordinates within each window are calculated, and the positioning offset is synthesized.

[0019] Image processing is performed on the colloid coating morphology data to extract the glue line width sequence and glue line height sequence, and the ratio of the coefficient of variation of the width to the height is calculated as the glue line uniformity.

[0020] Spectral analysis is performed on the bonding pressure waveform data to extract the energy proportion outside the main frequency band, and the dispersion of the pressure peak is statistically analyzed. The pressure fluctuation coefficient is obtained by fusing the energy proportion outside the main frequency band and the dispersion of the pressure peak.

[0021] Spatial gridding analysis is performed on the curing temperature field data to calculate the difference in temperature change rate at each grid point during the curing process, and the standard deviation of the difference is taken as the standard deviation of the temperature gradient.

[0022] As a further aspect of the present invention, the set of process stability indicators is input into a pre-constructed packaging quality correlation model to calculate a multi-dimensional reliability prediction score for the packaging structure, including:

[0023] The positioning offset, the uniformity of the adhesive line, the pressure fluctuation coefficient, and the standard deviation of the temperature gradient are standardized and scaled to form a feature vector;

[0024] The feature vector is input into the encapsulation quality correlation model, which is a trained deep learning network containing multiple fully connected layers.

[0025] The packaging quality correlation model outputs a multi-dimensional vector, where each dimension of the multi-dimensional vector corresponds to the mechanical strength prediction score, the airtightness prediction score, the thermal cycle life prediction score, and the electrical connection reliability prediction score, which together constitute the multi-dimensional reliability prediction score.

[0026] As a further aspect of the present invention, the training process of the packaging quality correlation model includes:

[0027] Collect a complete set of process stability indicators and their corresponding set of measured reliability data of the final product from historical packaging batches to form training sample pairs;

[0028] The process stability index set is used as input features, and the final product reliability test dataset is used as supervision labels to train the deep learning network.

[0029] The backpropagation algorithm is used to optimize the network weights and minimize the mean squared error between the multidimensional reliability prediction score of the model output and the supervision label until the model converges.

[0030] As a further aspect of the present invention, based on the multi-dimensional reliability prediction score, a corresponding process parameter optimization scheme is matched in the packaging process knowledge base, including:

[0031] Using the multi-dimensional reliability prediction score as the query condition, search for historical case records with similar prediction score patterns in the packaging process knowledge base;

[0032] Extract process parameter adjustment strategies that have successfully improved various reliability scores from the matched historical case records;

[0033] Based on the specific model and parameter range of the current production equipment, the extracted process parameter adjustment strategy is adapted and converted to generate an immediately executable process parameter optimization scheme.

[0034] As a further aspect of the present invention, the process of constructing and updating the packaging process knowledge base includes:

[0035] Record each complete packaging process, including the original set of timing process parameters, the calculated set of process stability indicators, the applied process parameter optimization schemes, and the final sample's internal defect distribution map and interface bonding strength data;

[0036] Using a set of process stability indicators and multi-dimensional reliability prediction scores as indexes, each process record is stored as a knowledge entry.

[0037] Once a new closed-loop process correction instruction is generated, the process parameter optimization scheme in the corresponding knowledge entry is corrected according to the closed-loop process instruction, and the version number is marked.

[0038] As a further aspect of the present invention, after the verification packaging process is completed, non-destructive testing is performed on the produced sample to obtain an internal defect distribution map and interface bonding strength data of the sample, including:

[0039] The sample is scanned using an X-ray tomography device to reconstruct a three-dimensional image of the sample's internal structure.

[0040] From the three-dimensional structural image, voids, cracks, delamination, and foreign object defects are identified and segmented, their location, size, and type are labeled, and the internal defect distribution map is generated;

[0041] A laser ultrasonic testing device is used to excite and receive ultrasonic signals at the sample interface. By analyzing the reflection and transmission characteristics of the ultrasonic signals, the interface bonding strength data is calculated.

[0042] As a further aspect of the present invention, the internal defect distribution map and interface bonding strength data, together with the set of process stability indicators, are sent to the packaging process traceability analysis engine, including:

[0043] The internal defect distribution map and interface bonding strength data are quantified into defect density, defect location clustering features, and average bonding strength value.

[0044] The quantified defect features are aligned with the set of process stability indicators by timestamp to construct a "process-defect" association data table;

[0045] Input the “process-defect” association data table into the packaging process traceability analysis engine.

[0046] As a further aspect of the present invention, the packaging process traceability analysis engine is used to reverse-engineer and identify the key process steps and their parameter deviations that lead to specific defects, including:

[0047] The packaging process traceability analysis engine uses a causal inference method to analyze the statistical correlation and Granger causality between each process stability index and each defect characteristic in the "process-defect" correlation data table;

[0048] Select process stability indicators that have a strong causal relationship with specific defect characteristics, and locate the process steps corresponding to the process stability indicators.

[0049] Further analysis of the original timing parameters of the aforementioned process steps in the verification process identifies the specific parameters that deviate from the standard process window and their deviation amounts.

[0050] As a further aspect of the present invention, based on the identification results, targeted closed-loop process correction instructions are generated to update the corresponding entries in the packaging process knowledge base, including:

[0051] For each identified parameter deviation, calculate the corresponding parameter compensation value or control curve correction amount based on the deviation amount and direction;

[0052] All corrections are integrated to form a structured closed-loop process correction instruction, which explicitly specifies the target equipment, target parameters, and corrected setpoints.

[0053] Based on the content of the closed-loop process correction instruction, the corresponding part of the relevant process parameter optimization scheme in the packaging process knowledge base is retrieved and updated, and the entire process data of this traceability analysis is added to the knowledge entry as a new case.

[0054] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0055] By leveraging sensor arrays deployed on the packaging production line, a set of time-series process parameters is collected. Feature extraction and anomaly screening are performed on these parameters to generate a quantified set of process stability indicators. This set of indicators is then input into a pre-built packaging quality correlation model, which calculates a multi-dimensional reliability prediction score for the packaging structure. Based on this score, corresponding process parameter optimization schemes are matched against a packaging process knowledge base. The execution of these schemes dynamically adjusts key control parameters of the production equipment, simultaneously initiating the verification packaging process. The collection of time-series parameters aligns with the continuous production process, feature extraction and anomaly screening refine parameter processing dimensions, the process stability indicator set visually presents the process operation status, and the multi-dimensional reliability prediction score covers different performance dimensions of the packaging structure. The matching of process parameter optimization schemes relies on the knowledge base for standardized integration, and the dynamic adjustment of production equipment parameters and the verification packaging process are seamlessly connected.

[0056] Non-destructive testing is performed on samples produced from the verification packaging process to obtain internal defect distribution maps and interface bonding strength data. These data, along with a set of process stability indicators, are simultaneously fed into the packaging process traceability analysis engine. Through reverse deduction analysis, key process steps that cause specific defects are identified, and the parameter deviations of the corresponding process steps are clarified. Based on the identification results, targeted closed-loop process correction instructions are generated. These instructions directly update the corresponding entries in the packaging process knowledge base. Non-destructive testing data and process parameters are fused, and reverse deduction directly points to the process source of the defect. The identification of parameter deviations and defect types are precisely correlated. The generation of closed-loop process correction instructions and the updating of the knowledge base form a closed-loop control, and the process correction actions directly affect the parameter system of packaging production. Attached Figure Description

[0057] Figure 1 This is a flowchart of an integrated packaging method based on a glass substrate according to the present invention;

[0058] Figure 2 A flowchart for calculating multi-dimensional reliability prediction scores;

[0059] Figure 3 Flowchart for matching process parameter optimization schemes;

[0060] Figure 4 This is a graph showing the relationship between interfacial bonding strength and ultrasonic reflection coefficient.

[0061] Figure 5 The effect diagram of the execution of the closed-loop process correction instruction. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0063] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0064] See Figure 1A sensor array deployed on the packaging production line collects a set of time-series process parameters during the production process. Feature extraction and anomaly screening are performed on these parameters to generate a set of process stability indicators. This set of indicators is input into a pre-built packaging quality correlation model to calculate a multi-dimensional reliability prediction score for the packaging structure. Based on this score, a corresponding process parameter optimization scheme is matched in the packaging process knowledge base. The optimization scheme is executed, dynamically adjusting key control parameters of the production equipment and initiating a verification packaging process. After the verification packaging process is completed, non-destructive testing is performed on the produced samples to obtain internal defect distribution maps and interface bonding strength data. These maps, along with the process stability indicator set, are fed into a packaging process traceability analysis engine. The engine is used to reverse-engineer and identify key process steps and parameter deviations leading to specific defects. Based on the identification results, targeted closed-loop process correction instructions are generated, updating the corresponding entries in the packaging process knowledge base.

[0065] In one embodiment of the present invention, on a glass substrate mounting production line, a sensor array collects a set of time-series process parameters in real time. These parameters include substrate positioning data, adhesive coating morphology data, bonding pressure waveform data, and curing temperature field data. The process stability index set includes positioning offset, adhesive line uniformity, pressure fluctuation coefficient, and temperature gradient standard deviation. A sliding window processing method is applied to the substrate positioning data sequence, calculating the centroid drift and standard deviation of the positioning coordinates within each window to synthesize the positioning offset. In a specific implementation, the substrate positioning data sequence comes from an encoder and a vision positioning system. Each window contains coordinate data from 50 consecutive sampling points. The centroid drift is the average magnitude of the offset vectors of all coordinate points within the window relative to the theoretical position, and the standard deviation is the dispersion of these offset vector magnitudes. Data comparison shows that during normal production, the average centroid drift of the substrate positioning data sequence is less than 0.1 mm and the standard deviation is less than 0.05 mm. During abnormal production, the average centroid drift of the substrate positioning data sequence may rise to 0.5 mm and the standard deviation may exceed 0.2 mm.

[0066] In practice, image processing is performed on the colloid coating morphology data to extract the adhesive line width and height sequences. The ratio of the coefficient of variation of the width to the height is calculated as the adhesive line uniformity. The colloid coating morphology data is obtained from a high-speed line scan camera, acquiring a cross-sectional image every 0.1 mm along the adhesive line path. From each image, the adhesive line width and height values ​​are extracted to form the width and height sequences. The adhesive line uniformity is calculated using the following formula:

[0067]

[0068] in: Indicates the uniformity of the adhesive lines. The coefficient of variation represents the line width sequence. This represents the coefficient of variation for the glue line height sequence, defined as the ratio of the sequence's standard deviation to its mean. In some embodiments, the coefficient of variation for the glue line width sequence... The coefficient of variation for the glue line width sequence is calculated as the standard deviation of the glue line width sequence divided by the mean of the glue line width sequence. The calculation is the standard deviation of the adhesive line height sequence divided by the mean of the adhesive line height sequence. In data comparison, the uniformity of the adhesive lines during uniform coating is considered. Approximately 1, uniformity of adhesive lines when unevenly coated. Deviation 1.

[0069] In practice, spectral analysis is performed on the bonding pressure waveform data to extract the energy proportion outside the main frequency band, and the dispersion of the pressure peak is statistically analyzed. The pressure fluctuation coefficient is obtained by fusing the energy proportion outside the main frequency band and the dispersion of the pressure peak. The bonding pressure waveform data comes from a dynamic pressure sensor with a sampling frequency of 1. The pressure curves of the entire bonding process were recorded, and the spectrum analysis was performed using Fast Fourier Transform, with the dominant frequency band defined as 0.5. up to 5 The range, specifically the energy proportion outside the main frequency band, is the ratio of the sum of the energy of all frequency components outside the main frequency band to the total energy in the frequency domain. The dispersion of pressure peaks is characterized by calculating the standard deviation of all pressure peaks. During the fusion process, the energy proportion outside the main frequency band and the standard deviation of pressure peaks are linearly combined using preset weights to form the pressure fluctuation coefficient. Optionally, the weights are set based on process experience. In data comparison, the energy proportion outside the main frequency band in the stable bonding process is less than 0.1, and the standard deviation of pressure peaks is less than 5. The energy outside the dominant frequency band of the unstable bonding process may account for more than 0.3%, and the standard deviation of the pressure peak may be greater than 20. .

[0070] In practice, the curing temperature field data is spatially gridded for analysis. The difference in temperature change rate at each grid point during the curing process is calculated, and its standard deviation is taken as the temperature gradient standard deviation. The curing temperature field data is obtained from an infrared thermal imager, with a resolution of 10... The temperature distribution image on the substrate surface is acquired by frequency acquisition and the image is divided into... The grid is defined, with each grid point representing the average temperature of a region. The rate of temperature change is the first derivative of the temperature at each grid point with time. The standard deviation of the temperature gradient is calculated using the difference method; it is the standard deviation of the rate of temperature change across all grid points during the curing process. In some embodiments, the curing process is divided into heating, holding, and cooling stages. The rate of temperature change is positive during the heating stage and negative during the cooling stage. The standard deviation of the temperature gradient comprehensively reflects the temperature uniformity of the entire curing process. Data comparison shows that the standard deviation of the temperature gradient is less than 2 when heating is uniform. When heating is uneven, the standard deviation of the temperature gradient may exceed 5. .

[0071] See Figure 2 In one embodiment of the present invention, after generating the set of process stability indicators, the positioning offset, glue line uniformity, pressure fluctuation coefficient, and temperature gradient standard deviation are standardized and scaled to form a feature vector. The standardization and scaling adopts the min-max normalization method, mapping each process stability indicator to the [0,1] interval. The original value range of the positioning offset is 0 to 1 mm, the original value range of the glue line uniformity is 0.5 to 2.0, the original value range of the pressure fluctuation coefficient is 0 to 1, and the original value range of the temperature gradient standard deviation is 0 to 10°C / s. Data comparison shows that when the positioning offset increases from 0.05 mm to 0.8 mm, its normalized feature value increases from 0.05 to 0.8. In some embodiments, the standardization and scaling can also adopt the Z-score method to process the data into a distribution with a mean of 0 and a standard deviation of 1.

[0072] In practice, the standardized and scaled feature vector is input into the encapsulation quality correlation model (EMC). The EMC is a trained deep learning network containing multiple fully connected layers. Its network structure includes an input layer, three hidden layers, and an output layer. The input layer has four neurons, the same dimension as the feature vector. The three hidden layers have 128, 64, and 32 neurons respectively, and the output layer has four neurons. The EMC outputs a multi-dimensional vector. Each dimension of this vector corresponds to a mechanical strength prediction score, an airtightness prediction score, a thermal cycle life prediction score, and an electrical connection reliability prediction score, collectively forming a multi-dimensional reliability prediction score. Each prediction score is a value between 0 and 1, with scores closer to 1 indicating higher reliability. In the data comparison, when a set of feature vectors representing process fluctuations is input, such as a high normalized value for positioning offset and a low normalized value for glue line uniformity, the mechanical strength prediction score output by the packaging quality correlation model may be 0.3 and the airtightness prediction score may be 0.5. However, when a set of feature vectors representing process stability is input, the mechanical strength prediction score may be 0.9 and the airtightness prediction score may be 0.95.

[0073] In practical implementation, the training process of the packaging quality correlation model includes collecting a complete set of process stability indicators and their corresponding final product reliability test data sets from historical packaging batches to form training sample pairs. The number of historical packaging batches is 10,000, with each batch corresponding to one set of process stability indicators and one set of final product reliability test data sets. The final product reliability test data set includes measured values ​​of mechanical strength, airtightness, thermal cycle life, and electrical connection reliability. These measured values ​​are obtained through destructive physics analysis, helium mass spectrometry leak detection, temperature cycling tests, and on-resistance tests, and are also normalized to the [0,1] interval as supervision labels. In practical implementation, the set of process stability indicators is used as input features, and the final product reliability test data set is used as supervision labels to train the deep learning network. A stochastic gradient descent optimizer is used during training, with a learning rate set to 0.001. The backpropagation algorithm is used to optimize the network weights, minimizing the mean square error between the multi-dimensional reliability prediction score output by the model and the supervision labels until the model converges. The loss function is... Defined as:

[0074]

[0075] in: This represents the total number of training samples. Indicates the first The first sample Supervision label values ​​in each dimension, The packaging quality correlation model represents the first... The first sample The model outputs predicted scores across multiple dimensions. It's understood that the convergence criterion is that the loss function value on the training set no longer significantly decreases over 10 consecutive training epochs. Optionally, dropout can be used during training to prevent overfitting. In some embodiments, training samples are divided into training, validation, and test sets in a 7:2:1 ratio. The trained package quality correlation model is used to calculate multi-dimensional reliability prediction scores for package structures online. Optionally, the package quality correlation model is periodically retrained using new production and test data for updates.

[0076] See Figure 3 In one embodiment of the present invention, after calculating the multi-dimensional reliability prediction score of the packaging structure, the multi-dimensional reliability prediction score is used as a query condition to search for historical case records with similar prediction score patterns in the packaging process knowledge base. The multi-dimensional reliability prediction score is a four-dimensional vector. ,in Represents the mechanical strength prediction score. Represents the airtightness prediction score. Represents the thermal cycling lifetime prediction score. This represents the predicted score for electrical connection reliability. Finding similar predicted score patterns involves calculating the Euclidean distance between the query vector and the predicted score vectors corresponding to each historical case record in the knowledge base. To achieve this, the distance calculation formula is:

[0077]

[0078] in: The Euclidean distance represents the predicted score vector of a historical case record in the knowledge base. The smaller the value, the higher the similarity. Data comparison shows that when the current predicted score vector is [0.3, 0.9, 0.8, 0.7], it may match a historical predicted score vector in the knowledge base with a value of [0.35, 0.88, 0.82, 0.72]. Their Euclidean distance is... The similarity score was 0.07, and therefore it was judged as similar.

[0079] In practice, process parameter adjustment strategies that have successfully improved various reliability scores are extracted from matched historical case records. Each historical case record contains complete packaging process data, and the process parameter adjustment strategies are stored in structured data format. Specifically, they record a set of parameter adjustment instructions for equipment such as substrate positioning platforms, dispensing valves, bonding heads, and curing ovens. Examples include adjusting the servo gain parameters of the substrate positioning platform, modifying the opening time and pressure curve of the dispensing valve, optimizing the pressing speed and holding time of the bonding head, or adjusting the set temperature and heating rate of each temperature zone of the curing oven. Based on the specific model and parameter range of the current production equipment, the extracted process parameter adjustment strategies are adapted to generate an immediately executable process parameter optimization scheme. Adaptation involves scaling the absolute parameter values ​​in the historical strategies according to the technical specifications of the current equipment, or mapping instructions that depend on specific hardware interfaces to instruction codes that the current equipment control system can recognize. In some embodiments, the current maximum operating pressure of the dispensing valve is 0.7 MPa, while the recommended pressure value in the historical strategy is 0.8 MPa. The adaptation conversion will scale the recommended pressure value to 0.7 MPa and adjust the dispensing duration accordingly to ensure consistent total dispensing volume.

[0080] In practical implementation, the construction and updating process of the packaging process knowledge base includes recording each complete packaging process, including the original set of timing process parameters, the calculated set of process stability indicators, the applied process parameter optimization schemes, and the final sample's internal defect distribution map and interface bonding strength data. Each process record is assigned a unique identifier and associated with a timestamp. Using the set of process stability indicators and multi-dimensional reliability prediction scores as indexes, each process record is stored as a knowledge entry. The index is built by concatenating the four values ​​of the process stability indicator set and the four values ​​of the multi-dimensional reliability prediction score into an eight-dimensional feature vector, and the database's indexing function is used to accelerate similarity retrieval based on this vector. When a new closed-loop process correction instruction is generated, the process parameter optimization scheme part in the corresponding knowledge entry is corrected according to the closed-loop process correction instruction, and a version number is marked. For example, if the initial version number of a knowledge entry is V1.0, after updating according to the closed-loop process correction instruction, its process parameter optimization scheme part is replaced by the new scheme, and the version number is updated to V1.1. The initial process parameter optimization scheme is still retained as a historical version for traceability.

[0081] In one embodiment of the present invention, after the verification packaging process is completed, the produced sample is subjected to non-destructive testing to obtain the internal defect distribution map and interface bonding strength data of the sample. An X-ray computed tomography (CT) scanner is used to scan the sample and reconstruct a three-dimensional structural image of the sample's interior. The voltage of the CT scanner is set to 160 kV, the current to 150 μA, and the scanning resolution to 5 μm per pixel. By rotating the sample and acquiring two-dimensional projection images from multiple angles, a filtered back-projection algorithm is used to reconstruct the complete three-dimensional volume data of the sample's interior. Cavities, cracks, delamination, and foreign matter defects are identified and segmented from the three-dimensional structural image, and their location, size, and type are labeled to generate an internal defect distribution map. The identification and segmentation process is based on threshold segmentation and region growing algorithms. A grayscale threshold range is set to distinguish between materials and defects. For each connected region, its volume, equivalent diameter, centroid coordinates, and defect type are calculated. The internal defect distribution map is stored in a three-dimensional coordinate mapping file format, which includes the type code, location coordinates (X, Y, Z), and equivalent diameter (D) of each defect. Referring to Table 1, a partial internal defect statistics of an example sample are shown.

[0082] Table 1: Statistical Table of Internal Defects in Samples

[0083]

[0084] A laser ultrasonic testing device is used to excite and receive ultrasonic signals at the sample interface. By analyzing the reflection and transmission characteristics of the ultrasonic signals, the interface bonding strength data is calculated. The laser ultrasonic testing device uses a 10 nanosecond Nd:YAG laser to generate ultrasonic waves, and an interferometer is used to receive the reflected signals. The interface bonding strength data is obtained by measuring the ultrasonic reflection coefficient R at the interface and applying empirical formulas.

[0085]

[0086] in: This represents the interfacial bonding strength value calculated through inversion. and These are fitting parameters related to the acoustic impedance of the material. This is the measured ultrasonic reflection coefficient. In some embodiments, at a well-bonded interface between the glass and the silicon chip, the ultrasonic reflection coefficient... Typically below 0.1, the calculated interfacial bond strength value The ultrasonic reflection coefficient is relatively high at weakly bonded interfaces. The calculated interfacial bond strength value may be greater than 0.3. Lower.

[0087] In practical implementation, the internal defect distribution map and interface bonding strength data, along with a set of process stability indicators, are fed into the packaging process traceability analysis engine. First, the internal defect distribution map and interface bonding strength data are quantified into defect density, defect location clustering characteristics, and average bonding strength value. The formula is calculated by counting the total number of defects per unit volume. ,in It is the total number of defects. The effective detection volume of the sample is used. The defect location clustering features are obtained by calculating the local spatial density of the defect centroid coordinates. The average bonding strength value is the arithmetic mean of the interface bonding strength values ​​of multiple measurement points. In practice, the quantified defect features are aligned with the process stability index set by timestamp to construct a "process-defect" association data table. The process stability index set includes positioning offset, glue line uniformity, pressure fluctuation coefficient, and temperature gradient standard deviation. Each index is associated with the time period of sample production, and the quantified defect features are associated with the sample as a whole. The alignment operation is based on the sample production batch number and the time logic of the process flow.

[0088] See Figure 4This is a graph showing the relationship between interfacial bonding strength and ultrasonic reflection coefficient. As the ultrasonic reflection coefficient R increases from 0.05 to 0.40, the interfacial bonding strength rapidly decreases from approximately 70 MPa to near 0 MPa, exhibiting a clear negative correlation. The fluctuations in the curve represent noise in the simulated measured data, reflecting uncertainties in actual testing. When the reflection coefficient R < 0.1, the interfacial bonding strength is generally higher than 40 MPa, belonging to a high-quality bonding region; when the reflection coefficient R > 0.3, the interfacial bonding strength is generally lower than 10 MPa, belonging to a weak bonding or failure region; the intermediate region is a transition zone where the bonding strength continuously decreases with increasing reflection coefficient. On the production line, the R value can be quickly obtained through laser ultrasonic testing, allowing for rapid determination of interfacial bonding strength by referring to this graph, eliminating the need for destructive testing and significantly improving testing efficiency.

[0089] In one embodiment of the present invention, after the "process-defect" association data table is input into the packaging process traceability analysis engine, the engine is used to reverse-engineer and identify the key process steps and their parameter deviations that lead to specific defects. The packaging process traceability analysis engine employs a causal inference method to analyze the statistical correlation and Granger causality between each process stability index and each defect characteristic in the "process-defect" association data table. The statistical correlation is calculated using the Pearson correlation coefficient. The evaluation is conducted using the following formula:

[0090]

[0091] in: and These represent a certain process stability index value and a certain defect characteristic value from the same batch of samples, respectively. and It is their mean. This refers to the sample size. The Granger causality test, by establishing a vector autoregressive model, examines whether one time series variable helps predict another. Data comparison shows that the Pearson correlation coefficient between the pressure fluctuation coefficient and the void / defect density... When the absolute value is greater than 0.8 and the Granger test p-value is less than 0.05, the packaging process traceability analysis engine determines that the pressure fluctuation coefficient is the Granger factor of the void defect density.

[0092] In practical implementation, the packaging process traceability analysis engine filters out process stability indicators that have a strong causal relationship with specific defect characteristics, locates the process steps corresponding to the process stability indicators, and defines a strong causal relationship as simultaneously satisfying the absolute value of statistical correlation. Furthermore, the Granger criterion p-value is <0.05. For example, the packaging process traceability analysis engine identified a strong causal relationship between void defect density and pressure fluctuation coefficient. The pressure fluctuation coefficient originates from the analysis of bonding pressure waveform data, thus pinpointing the bonding process stage. Further analysis of the original timing process parameters in the verification flow of the bonding process stage identifies specific parameters deviating from the standard process window and their deviations. The original timing process parameters are the raw sensor data sequences without feature extraction. For the bonding process stage, the original timing process parameters include the pressure-time curve of the bonding head, and the standard process window specifies that the bonding pressure should be within the target value. Within ±5%, the pressure fluctuation during the holding stage should be less than 2%. By comparing the pressure-time curves recorded in the verification process, the packaging process traceability analysis engine identified that the actual pressure fluctuation during the holding stage reached 8%, with a deviation of +6% relative to the standard process window.

[0093] In practical implementation, based on the identification results, targeted closed-loop process correction instructions are generated, and the corresponding entries in the packaging process knowledge base are updated. For each identified parameter deviation, the corresponding parameter compensation value or control curve correction amount is calculated based on its deviation magnitude and direction. For example, if the identified pressure fluctuation deviation during the holding pressure stage is +6%, a proportional gain coefficient for the pressure control loop is generated based on the control model. The parameter compensation value is reduced by 15%, or a new, filtered, and smoother pressure setpoint curve is generated as the control curve correction amount. All corrections are integrated into a structured closed-loop process correction instruction, encapsulated in JSON or XML format, explicitly specifying the target equipment, target parameters, and corrected setpoint values. Optionally, the closed-loop process correction instruction may also include the effective batch range or conditions. In some embodiments, the correction amount may involve multiple parameters, such as simultaneously adjusting pressure control parameters and bond head descent speed; these corrections are integrated into a single instruction. Based on the content of the closed-loop process correction instruction, the corresponding parts of the relevant process parameter optimization schemes in the packaging process knowledge base are retrieved and updated. The retrieval is achieved by matching the process steps and parameter names that cause the defect. It is understood that the update operation ensures the continuous evolution of the knowledge base. All data from this traceability analysis, including the original "process-defect" association data table, causal analysis results, identified parameter deviations, and the generated closed-loop process correction instruction, are appended as new cases to the updated knowledge entries. The new cases are linked as sub-documents and include timestamps and version identifiers.

[0094] See Figure 5This is a graph showing the execution effect of a closed-loop process correction command. It illustrates the trends of various process stability indicators over time after the correction. The pressure fluctuation coefficient rapidly decreased from 8.0 before correction, dropping to 1.8 after 24 hours, representing the largest decrease and indicating a significant improvement in bonding pressure control. The temperature gradient standard deviation gradually decreased from 5.0 to 2.5, indicating a continuous improvement in temperature field uniformity during the curing stage. The positioning offset slowly decreased from 2.5 to 1.4, indicating a steady improvement in substrate positioning accuracy. The adhesive line uniformity slowly increased from 0.3 to 0.6, representing a slight improvement in the uniformity of adhesive coating. All process stability indicators show a continuous optimization trend, proving the effectiveness of the closed-loop process correction command. The first 4 hours represent the most significant improvement in indicators, after which the improvement level off, indicating that the process system gradually reaches a new stable state after correction.

[0095] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An integrated packaging method based on a glass substrate, characterized in that, The method includes: By collecting a set of time-series process parameters in the production process through a sensor array deployed on the packaging production line, feature extraction and anomaly screening are performed on the set of time-series process parameters to generate a set of process stability indicators. The process stability index set is input into the pre-built packaging quality correlation model to calculate the multi-dimensional reliability prediction score of the packaging structure. Based on the multi-dimensional reliability prediction score, the corresponding process parameter optimization scheme is matched in the packaging process knowledge base; The process parameter optimization scheme is implemented to dynamically adjust the key control parameters of the production equipment and initiate a verification packaging process. After the verification packaging process is completed, the produced sample is subjected to non-destructive testing to obtain the internal defect distribution map and interface bonding strength data of the sample. The internal defect distribution map and interface strength data, together with the set of process stability indicators, are sent into the packaging process traceability analysis engine. By using the packaging process traceability analysis engine to reverse-engineer, the key process steps and their parameter deviations that lead to specific defects can be identified. Based on the identification results, targeted closed-loop process correction instructions are generated to update the corresponding entries in the packaging process knowledge base.

2. The integrated packaging method based on a glass substrate as described in claim 1, characterized in that, Feature extraction and outlier screening are performed on the time-series process parameter set to generate a process stability index set, including: The set of timing process parameters includes substrate positioning data, colloidal coating morphology data, bonding pressure waveform data, and curing temperature field data. The set of process stability indicators includes positioning offset, glue line uniformity, pressure fluctuation coefficient, and temperature gradient standard deviation. The substrate positioning data sequence is processed by sliding window processing. The centroid drift and standard deviation of the positioning coordinates within each window are calculated, and the positioning offset is synthesized. Image processing is performed on the colloid coating morphology data to extract the glue line width sequence and glue line height sequence, and the ratio of the coefficient of variation of the width to the height is calculated as the glue line uniformity. Spectral analysis is performed on the bonding pressure waveform data to extract the energy proportion outside the main frequency band, and the dispersion of the pressure peak is statistically analyzed. The pressure fluctuation coefficient is obtained by fusing the energy proportion outside the main frequency band and the dispersion of the pressure peak. Spatial gridding analysis is performed on the curing temperature field data to calculate the difference in temperature change rate at each grid point during the curing process, and the standard deviation of the difference is taken as the standard deviation of the temperature gradient.

3. The integrated packaging method based on a glass substrate as described in claim 2, characterized in that, The set of process stability indicators is input into a pre-built packaging quality correlation model to calculate a multi-dimensional reliability prediction score for the packaging structure, including: The positioning offset, the uniformity of the adhesive line, the pressure fluctuation coefficient, and the standard deviation of the temperature gradient are standardized and scaled to form a feature vector; The feature vector is input into the encapsulation quality correlation model, which is a trained deep learning network containing multiple fully connected layers. The packaging quality correlation model outputs a multi-dimensional vector, where each dimension of the multi-dimensional vector corresponds to the mechanical strength prediction score, the airtightness prediction score, the thermal cycle life prediction score, and the electrical connection reliability prediction score, which together constitute the multi-dimensional reliability prediction score.

4. The integrated packaging method based on a glass substrate as described in claim 3, characterized in that, The training process of the packaging quality correlation model includes: Collect a complete set of process stability indicators and their corresponding set of measured reliability data of the final product from historical packaging batches to form training sample pairs; The process stability index set is used as input features, and the final product reliability test dataset is used as supervision labels to train the deep learning network. The backpropagation algorithm is used to optimize the network weights and minimize the mean squared error between the multidimensional reliability prediction score of the model output and the supervision label until the model converges.

5. The integrated packaging method based on a glass substrate as described in claim 4, characterized in that, Based on the multi-dimensional reliability prediction score, corresponding process parameter optimization schemes are matched in the packaging process knowledge base, including: Using the multi-dimensional reliability prediction score as the query condition, search for historical case records with similar prediction score patterns in the packaging process knowledge base; Extract process parameter adjustment strategies that have successfully improved various reliability scores from the matched historical case records; Based on the specific model and parameter range of the current production equipment, the extracted process parameter adjustment strategy is adapted and converted to generate an immediately executable process parameter optimization scheme.

6. The integrated packaging method based on a glass substrate as described in claim 5, characterized in that, The process of building and updating the packaging process knowledge base includes: Record each complete packaging process, including the original set of timing process parameters, the calculated set of process stability indicators, the applied process parameter optimization schemes, and the final sample's internal defect distribution map and interface bonding strength data; Using a set of process stability indicators and multi-dimensional reliability prediction scores as indexes, each process record is stored as a knowledge entry. Once a new closed-loop process correction instruction is generated, the process parameter optimization scheme in the corresponding knowledge entry is corrected according to the closed-loop process instruction, and the version number is marked.

7. The integrated packaging method based on a glass substrate as described in claim 6, characterized in that, After the verification packaging process is completed, the produced sample is subjected to non-destructive testing to obtain the internal defect distribution map and interface bonding strength data of the sample, including: The sample is scanned using an X-ray tomography device to reconstruct a three-dimensional image of the sample's internal structure. From the three-dimensional structural image, voids, cracks, delamination, and foreign object defects are identified and segmented, their location, size, and type are labeled, and the internal defect distribution map is generated; A laser ultrasonic testing device is used to excite and receive ultrasonic signals at the sample interface. By analyzing the reflection and transmission characteristics of the ultrasonic signals, the interface bonding strength data is calculated.

8. The integrated packaging method based on a glass substrate as described in claim 7, characterized in that, The internal defect distribution map and interface strength data, along with the set of process stability indicators, are fed into the packaging process traceability analysis engine, including: The internal defect distribution map and interface bonding strength data are quantified into defect density, defect location clustering features, and average bonding strength value. The quantified defect features are aligned with the set of process stability indicators by timestamp to construct a "process-defect" association data table; Input the "process-defect" association data table into the packaging process traceability analysis engine.

9. The integrated packaging method based on a glass substrate as described in claim 8, characterized in that, By using the packaging process traceability analysis engine to reverse engineer the process, the key process steps and their parameter deviations that lead to specific defects are identified, including: The packaging process traceability analysis engine uses a causal inference method to analyze the statistical correlation and Granger causality between each process stability index and each defect characteristic in the "process-defect" correlation data table; Select process stability indicators that have a strong causal relationship with specific defect characteristics, and locate the process steps corresponding to the process stability indicators. Further analysis of the original timing parameters of the aforementioned process steps in the verification process identifies the specific parameters that deviate from the standard process window and their deviation amounts.

10. The integrated packaging method based on a glass substrate as described in claim 9, characterized in that, Based on the identification results, targeted closed-loop process correction instructions are generated to update the corresponding entries in the packaging process knowledge base, including: For each identified parameter deviation, calculate the corresponding parameter compensation value or control curve correction amount based on the deviation amount and direction; All corrections are integrated to form a structured closed-loop process correction instruction, which explicitly specifies the target equipment, target parameters, and corrected setpoints. Based on the content of the closed-loop process correction instruction, the corresponding part of the relevant process parameter optimization scheme in the packaging process knowledge base is retrieved and updated, and the entire process data of this traceability analysis is added to the knowledge entry as a new case.

Images