Wafer defect detection and regulation method and system, medium and product

Abstract

The invention discloses a wafer defect detection and regulation method and system, a medium and a product, and relates to the technical field of semiconductors. The method comprises the following steps: collecting microstructure image data and material structure spectrum data of a cutting edge of a wafer; fusing the depth geometric features of the microscopic morphology image data and the depth spectral features of the material structure spectral data to generate a composite defect feature vector; according to the composite defect feature vector, obtaining the category and the severity level of the defect, inputting the category and the severity level as input parameters into a stress simulation model based on finite element analysis to obtain a stress distribution prediction result containing a stress concentration condition; matching and outputting a regulation and control strategy identifier capable of optimizing the stress concentration condition from a preset regulation and control strategy library; and according to the regulation and control strategy identifier, a regulation and control instruction containing the specific process parameters is generated, and the regulation and control instruction is sent to downstream packaging equipment or a repair unit. By implementing the technical scheme, wafer defects can be accurately detected.

Description

Technical Field

[0001] This application relates to the semiconductor field, and more particularly to a method, system, medium, and product for wafer defect detection and control. Background Technology

[0002] With the continuous development of semiconductor technology, wafers, as the core basic material for integrated circuit manufacturing, play a crucial role in the performance and reliability of chips. Wafer manufacturing is complex, involving multiple stages, and problems in any one stage can lead to wafer defects. These defects not only affect chip yield but can also cause chip performance degradation or even failure, thereby increasing production costs and reducing production efficiency. Therefore, accurate detection and effective control of wafer defects are of great significance for improving the quality and efficiency of semiconductor manufacturing.

[0003] In existing technologies, wafer defect detection typically employs optical microscopy to directly observe the macroscopic morphology of the wafer surface to identify obvious defects. Furthermore, electron microscopy is used for more detailed observation, enabling the detection of microscopic defects. Defect control generally involves adjusting manufacturing process parameters based on experience, such as adjusting temperature and pressure, to improve wafer quality. Repair equipment is also used to locally repair detected defects to improve wafer yield.

[0004] However, existing technologies have significant drawbacks. Traditional detection methods primarily rely on analyzing single types of data, making it difficult to comprehensively and accurately reflect the characteristics and nature of wafer defects. Moreover, the control of defects lacks precise basis, often relying on experience to adjust parameters, failing to develop personalized control strategies for different types and severity of defects. This results in poor optimization of wafer stress concentration, hindering the effective improvement of wafer quality and production efficiency. Summary of the Invention

[0005] This application provides a method, system, medium, and product for wafer defect detection and control, which can accurately detect wafer defects, predict stress distribution, and provide effective control strategies to optimize stress concentration and ensure the quality of subsequent wafer packaging and use.

[0006] In a first aspect, this application provides a method for wafer defect detection and control, the method comprising: After the wafer is diced, the microscopic morphology image data and material structure spectral data of the diced edge of the wafer are acquired; The depth geometric features of the micro-morphology image data are extracted, the depth spectral features of the material structure spectral data are extracted, and the depth geometric features and the depth spectral features are fused to generate a composite defect feature vector. The composite defect feature vector is input into a preset convolutional neural network model, which outputs the defect category and severity level. The category and severity level are used as input parameters to drive a preset stress simulation model based on finite element analysis to perform simulation calculations, and the stress distribution prediction results including stress concentration are obtained. Based on the stress distribution prediction results, a control strategy identifier that can optimize the stress concentration is matched from a preset control strategy library and output. Based on the control strategy identifier, a control instruction containing specific process parameters is generated, and the control instruction is sent to the downstream packaging equipment or repair unit to drive the packaging equipment or the repair unit to perform corresponding parameter adjustments or online repair operations.

[0007] By employing the above technical solution, and integrating the depth geometric features and material spectral features of the cutting edge, the limitations of single-image detection are overcome, enabling more comprehensive and accurate identification of defect categories and assessment of their severity levels. AI-driven finite element analysis simulation models are used to simulate and predict potential stress concentration problems caused by defects, upgrading detection from post-event discovery to pre-event warning, and anticipating potential impacts on product reliability. Based on the predicted stress results, the system automatically matches and outputs optimal control strategies (such as specific process parameters) from the strategy library, and directly generates instructions to drive downstream equipment to perform adjustments or repairs, achieving a fully automated closed loop of "detection-analysis-decision-execution." This transforms traditional offline detection and adjustment, which relies on manual experience, into an online, automated, and intelligent process, significantly reducing intervention time, enabling rapid response and correction of problems, thereby improving overall production efficiency and the quality and reliability of the final product.

[0008] In some embodiments, the simulation calculation of the preset stress simulation model based on finite element analysis includes: Based on the category and severity level, the corresponding defect geometric feature parameters and material mechanical property parameters are mapped and called from the preset material parameter knowledge base. The material mechanical property parameters include elastic modulus, Poisson's ratio, yield strength and coefficient of thermal expansion. Based on the defect geometric feature parameters, a wafer geometric entity for finite element calculation is constructed, wherein the wafer geometric entity characterizes the three-dimensional morphology of the defect; The material mechanical properties and standard packaging process parameters are used as the physical input parameters of the stress simulation model, and the boundary conditions of the simulation are set. The standard packaging process parameters include the bonding temperature curve, the bonding pressure curve, and the bonding head movement trajectory. Finite element analysis is performed based on the wafer geometry, the physical input parameters, and the boundary conditions to output the quantified stress distribution at the bonding interface.

[0009] By employing the above technical solution, precise material mechanics parameters are automatically mapped and invoked based on the defect categories and levels identified by AI, and a geometric entity representing the true three-dimensional morphology of a specific defect is constructed. This makes the simulation model no longer based on idealized or general structures, but rather on high-precision modeling of the specific defect, greatly improving the realism and reliability of the simulation results. Standard packaging process parameters (such as temperature and pressure curves) are used as physical inputs and boundary conditions for the simulation, enabling the simulation to recreate the stress and heat environment of the chip during the actual packaging process. Therefore, the predicted stress distribution is no longer static, but dynamically reflects the impact of the entire packaging process on the defective wafer, accurately locating potential risk points. Abstract defect morphology is transformed into quantifiable stress distribution data (such as stress magnitude, gradient, and concentration areas). This makes assessing the harm of defects no longer rely on empirical and vague judgments, but rather on specific and precise physical quantities (such as whether the stress value exceeds the material yield strength) as the basis for decision-making, achieving scientific and accurate assessment. The output quantified stress distribution results are a direct and crucial input for the subsequent system to match the optimal control strategy from the strategy library. It ensures that the selected control strategy (such as adjusting bonding pressure or temperature) is based on a deep quantitative understanding of the physical problems caused by defects, thereby guaranteeing the accuracy and effectiveness of control and forming the cornerstone of closed-loop intelligent decision-making.

[0010] In some embodiments, performing finite element calculations based on the wafer geometry, the physical input parameters, and the boundary conditions, and outputting the quantized stress distribution results at the bonding interface includes: Based on the wafer geometry, the material mechanical properties of the physical input parameters, and the boundary conditions, the thermal stress field generated by the wafer structure under the bonding temperature field is calculated. The thermal stress field is used as a prestress field and superimposed with the mechanical pressure field applied by the bonding head to calculate the comprehensive stress field; Based on the comprehensive stress field, the maximum principal stress, equivalent stress, and stress concentration factor at the bonding interface are extracted. Based on the equivalent stress and defect geometric characteristics, calculate the crack propagation risk coefficient of the defect under the comprehensive stress field; The maximum principal stress, the equivalent stress, the stress concentration factor, and the crack propagation risk factor are used together as the stress distribution quantification result, and the stress concentration factor is used to characterize the stress concentration situation.

[0011] By employing the above technical solution, this method accurately simulates the real physical scenario of "thermal-mechanical" coupling during chip bonding by first calculating the thermal stress field and then superimposing it as prestress with the mechanical pressure field. This avoids the errors of considering a single load source, making the final comprehensive stress field prediction result closer to the stress state under actual process conditions. The output quantitative results (maximum principal stress, equivalent stress, and stress concentration factor) are tailored to different failure modes. The maximum principal stress is often used to assess the cracking risk of brittle materials, the equivalent stress is used to assess plastic deformation or yielding, and the stress concentration factor directly quantifies the degree to which defects amplify stress. This provides a comprehensive and accurate quantitative data foundation for subsequent reliability assessments. Going beyond static stress analysis, by calculating the crack propagation risk coefficient, this method enables the prediction of the dynamic evolution trend of defects during subsequent processes or use. This is a value-added prognostic capability that can identify defects whose current stress level is not exceeded but have a high potential propagation risk, thereby enabling early intervention. The final output is a set of highly condensed and meaningful engineering parameters, rather than a complex raw contour map. These parameters are directly compared with the material's failure threshold and process window, enabling the system to automatically, quickly, and accurately determine the severity of defects and providing a direct and reliable basis for decision-making in the next step of matching the optimal control strategy (such as adjusting temperature or pressure) from the strategy library.

[0012] In some embodiments, calculating the thermal stress field generated by the wafer structure under the bonding temperature field based on the wafer geometry, the material mechanical property parameters of the physical input parameters, and the boundary conditions specifically includes: Based on the bonding temperature curve, time-varying thermal load boundary conditions are defined on the wafer geometry. Based on the coefficient of thermal expansion in the mechanical property parameters of the material, the distribution of lattice thermal strain caused by temperature change is calculated. Based on the elastic modulus and Poisson's ratio in the material's mechanical properties, and in conjunction with the boundary conditions, the thermal stress field distribution caused by the lattice thermal strain is calculated by solving the thermoelastic constitutive equation to obtain the thermal stress field.

[0013] By employing the aforementioned technical solution, this method dynamically simulates the actual temperature change process experienced by the chip during packaging, rather than a simple static high-temperature field, by defining the time-varying thermal load based on the bonding temperature curve. This allows the calculated thermal stress to accurately reflect the real stress state caused by non-uniform heating / cooling. The lattice-scale thermal strain is calculated from the material's inherent thermal expansion coefficient and temperature change, and then the strain is converted into stress using constitutive equations based on the material's elastic modulus and Poisson's ratio. This first-principles calculation method ensures the scientific validity and high fidelity of the results, avoiding errors that may arise from empirical formulas. The calculated high-precision thermal stress field is the pre-stress field superimposed on the subsequent mechanical pressure field. Its accuracy directly determines the overall reliability of the final thermo-mechanical coupled comprehensive stress field simulation, making it a crucial and fundamental part of the entire finite element analysis process. This step specifically targets and quantifies the most common failure cause in chip packaging—thermal mismatch stress. By accurately calculating this stress field, it is possible to predict in advance the tensile stress, shear stress, etc. generated at the bonding interface due to the mismatch of thermal expansion coefficients of different materials, thereby accurately assessing the risks of warping, delamination, or cracking caused by this.

[0014] In some embodiments, the step of superimposing the thermal stress field as a prestress field with the mechanical pressure field applied by the bonding head to calculate the comprehensive stress field specifically includes: The thermal stress field is used as the initial stress state and preloaded onto the wafer geometry. Based on the bonding pressure curve and the bonding head movement trajectory, mechanical load boundary conditions are defined on the wafer geometry. Based on the material mechanical property parameters and the mechanical load boundary conditions, the stress response under the combined action of thermal prestress and mechanical load is calculated by solving the elastoplastic constitutive equation. An iterative algorithm is used to converge the stress response until a preset convergence condition is met, and the converged stress field is taken as the comprehensive stress field.

[0015] By employing the above technical solution, the thermal stress field is pre-loaded onto the model as prestress, scientifically reproducing the actual load sequence of the chip being heated first and then compressed during packaging. This ensures that the mechanical pressure is applied under a realistic initial state with thermally induced internal stress, greatly improving the fidelity of the simulation. By solving the elastoplastic constitutive equation (rather than a simple linear elastic equation), the plastic deformation stage that the material may enter under the combined action of thermal stress and mechanical pressure can be accurately simulated. This is crucial for predicting the yield, permanent deformation, and even failure of materials, and is particularly suitable for evaluating the limits and reliability of bonding processes. The use of iterative algorithms for convergent solutions is a necessary and advanced numerical processing method for addressing the geometric and material nonlinear problems arising from the superposition of thermal and mechanical stresses. This ensures that even under complex loading conditions, the final comprehensive stress field result is stable, reliable, and accurate, avoiding distortion caused by computational divergence. The effects of two mutually influencing physical fields (thermal and mechanical) are integrated into a unified comprehensive stress field. This comprehensive field is the gold standard for evaluating the performance of defects under real working conditions. It provides a unique and comprehensive data foundation for subsequent calculation of crack risk and matching control strategies, avoiding misjudgments that may be made due to evaluating different loads separately.

[0016] In some embodiments, inputting the composite defect feature vector into a preset convolutional neural network model and outputting the defect category and severity level specifically includes: The composite defect feature vector is input into the preset convolutional neural network model, which includes a shared feature encoder, a classification decoder, and a regression decoder. The shared semantic features are obtained by performing feature extraction and nonlinear transformation on the composite defect feature vector through the shared feature encoder. The shared semantic features are respectively input into the classification decoder and the regression decoder; The classification decoder outputs the category probability distribution of defects, and the category with the highest probability is determined as the final defect category; The severity level of the defect is output by the regression decoder as a continuous value. The continuous value of the severity level is normalized to obtain an intermediate value. The severity level is then determined based on the intermediate value according to a preset mapping relationship.

[0017] By employing the aforementioned technical solution and a shared feature encoder plus dual decoder architecture, both defect classification and severity regression tasks can be simultaneously learned from a fused feature vector. This design significantly improves computational efficiency, avoids the complexity of building separate models for each task, and ensures that the feature bases upon which both tasks are based are consistent. The shared encoder performs a deep nonlinear transformation on the composite feature vector, extracting higher-level, more abstract shared semantic features. These features not only contain morphological and material information about the defect but also embody its physical essence, enabling the model to better identify and generalize to unseen and morphologically diverse defects. The classification decoder outputs a category probability distribution rather than a simple result, providing not only the most likely defect type but also information on other possible types. This output method enhances the transparency and interpretability of the model's decision-making process, facilitating engineers' assessment of the reliability of their judgments. The regression decoder directly outputs continuous values ​​of severity levels, which are then normalized and mapped to levels. This method can more precisely characterize the severity of defects, avoiding the information loss caused by simple classification. Its judgment results are data-driven, completely objective, and eliminate the subjectivity and inconsistency of human assessment.

[0018] In some embodiments, the step of extracting features and performing nonlinear transformations on the composite defect feature vector using the shared feature encoder to obtain shared semantic features specifically includes: The composite defect feature vector is processed layer by layer by a series of one-dimensional convolutional layers to extract feature representations at different levels. Based on the residual connection structure, feature representations at different levels are fused to obtain fused features; A global max pooling layer is used to reduce the dimensionality of the fused features and extract the target feature vector. The target feature vector is input into a fully connected layer for nonlinear transformation and feature compression, and the shared semantic features are output.

[0019] By employing the aforementioned technical solution, layer-by-layer processing through cascaded multi-layer one-dimensional convolutional layers enables the gradual extraction of feature representations from the original composite features, progressing from low to high levels. This deep structure significantly enhances the model's understanding and representation of complex, nonlinear defect patterns, improving feature discriminability and robustness. The use of residual connections to fuse features from different levels effectively alleviates the gradient vanishing problem in deep networks, making the model easier to train and optimize. Simultaneously, this structure combines shallow, detailed features with deep, semantic features, avoiding the loss of crucial details during deep propagation and ensuring the richness and completeness of the fused features. Dimensionality reduction using global max-pooling layers automatically focuses on and extracts the most significant and important responses in each feature channel, thereby compressing feature size and reducing computational complexity. This operation also enhances the model's tolerance to small feature shifts and changes, improving its generalization ability and anti-interference capabilities. Finally, nonlinear transformation and feature compression are performed through fully connected layers, mapping the pooled, high-dimensional features to a lower-dimensional, more abstract shared semantic feature space. This feature space has been carefully optimized to contain core information that serves both the subsequent classification and regression subtasks, providing an efficient and task-adaptive input foundation for the dual decoder.

[0020] In a second aspect, embodiments of this application provide a computer system including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the steps of the method described in any possible implementation of the first aspect.

[0021] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0022] Fourthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0023] It is understood that the computer system provided in the second aspect, the storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the method provided in this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. Simultaneously acquiring microscopic morphology images (geometric information) and material structure spectra (material information) breaks through the limitations of traditional detection methods that rely solely on image morphology. This allows the system to not only detect external defects but also to detect invisible material lattice changes or stress damage caused by cutting. Deep learning technology is used to automatically extract deep geometric and spectral features, replacing manually designed, fixed features (such as length, width, and area) in traditional image processing. This data-driven approach can capture more complex and abstract feature patterns, exhibiting stronger recognition capabilities and generalization for novel and irregular defects. 2. Combining AI recognition results with physics-based finite element analysis simulation means that the system not only points out the location of cracks, but can also simulate and predict the stress concentration at the corresponding location during the subsequent encapsulation and bonding process. This moves quality inspection from post-event detection to pre-event warning. The stress simulation model considers real process parameters (such as bonding temperature and pressure curves), making the simulation environment extremely close to the production reality. The predicted stress distribution results have extremely high reference value and can truly reflect the potential risks of defects under the final product's operating conditions. 3. Based on simulation-based stress prediction results, the system automatically matches the optimal solution from a pre-set control strategy library. This encapsulates the knowledge and best practices of domain experts, achieving automated and scientific decision-making. A complete autonomous feedback control loop is formed, from detection and analysis to the generation of control instructions containing specific process parameters, which are then directly sent to production equipment for execution (such as adjusting bonding machine parameters or triggering repair units). This achieves full-process automation and truly realizes intelligent manufacturing. Because the control strategy is matched based on the simulation results of specific defects in a particular wafer, each control is customized. This gives the production process the ability to perform personalized dynamic optimization, enabling precise compensation for minute differences in each product and maximizing yield. Attached Figure Description

[0025] Figure 1 This is a schematic flowchart of a method for wafer defect detection and control in an embodiment of this application; Figure 2 This is a schematic diagram of the process of driving a preset stress simulation model based on finite element analysis to perform simulation calculations in the embodiments of this application; Figure 3 This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application. Detailed Implementation

[0026] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0027] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0028] The following is combined Figure 1 The method of the embodiments of this application will be described below.

[0029] Figure 1 This is a flowchart illustrating a method for wafer defect detection and control in an embodiment of this application. Please refer to [link / reference]. Figure 1 A method for detecting and controlling wafer defects includes the following steps: S101. After the wafer is cut, collect the microscopic morphology image data and material structure spectral data of the cut edge of the wafer. S102. Extract the depth geometric features of the micro-morphology image data, extract the depth spectral features of the material structure spectral data, and fuse the depth geometric features and the depth spectral features to generate a composite defect feature vector. S103. Input the composite defect feature vector into a preset convolutional neural network model, output the defect category and severity level, use the category and severity level as input parameters to drive a preset stress simulation model based on finite element analysis to perform simulation calculations, obtain stress distribution prediction results including stress concentration, and based on the stress distribution prediction results, match and output the control strategy identifier that can optimize the stress concentration from the preset control strategy library. S104. Based on the control strategy identifier, generate a control instruction containing specific process parameters, and send the control instruction to the downstream packaging equipment or repair unit to drive the packaging equipment or the repair unit to perform corresponding parameter adjustment or online repair operations.

[0030] After the wafer dicing equipment completes the dicing, the wafer is transferred to the inspection station, where a confocal microscope scans the dicing edge (scanning range: 50 μm outward from the edge) to acquire a microscopic morphology image (resolution 0.5 μm). Simultaneously, a Raman spectrometer performs a spectral scan on the dicing edge (excitation wavelength 532 nm, integration time 10 s) to analyze the thickness of the material damage layer. The microscopic morphology image data includes the three-dimensional morphology information of the dicing edge. The material structure spectral data includes the molecular vibrational spectrum of the material, used to analyze the chemical state and crystal structure changes. The thickness of the material damage layer is quantitatively analyzed through spectral characteristics (such as peak shift and peak width changes) (for example, in SiO2 layers, increased amorphization leads to significant changes in its Raman characteristic peaks). Damage depths > 5 nm are marked as abnormal; this quantitative indicator provides an objective basis for determining the severity of defects. A deep learning model is used to automatically extract deep, abstract depth geometric features from the morphology images, such as texture patterns, edge sharpness, crack orientation, and depth distribution. Similarly, deep learning models are used to extract deep spectral features from spectral data, capturing subtle spectral differences and associating them with specific material damage types (such as lattice distortion, amorphization, and contamination of chemical bonds). Deep features from different modalities (image domain and spectral domain) are concatenated or weighted and fused at the feature level to generate a unified composite defect feature vector. This fusion ensures that the final feature vector simultaneously includes both the macroscopic geometry and microscopic material state information of the defect. For example, a tiny crack (with inconspicuous geometric features) can be accurately identified as a high-risk defect if its edge material damage is severe (significant spectral features). This greatly improves the comprehensiveness and accuracy of defect characterization. A pre-trained CNN (Convolutional Neural Network) model is used to process the composite feature vector. The output includes defect category and severity level. Defect category: such as "mechanical damage," "thermal damage," "contamination." The defect cause can be accurately determined based on the defect category. Severity level: such as "minor," "moderate," "severe." The severity level is a quantified output, providing precise input parameters for subsequent simulations and decision-making. The CNN output (e.g., "moderate microcrack, damage depth 10nm") serves as the input parameter for the simulation. The FEA (Finite Element Analysis) model, based on the defect's geometric parameters (derived from morphological images) and material parameters (derived from spectral data), performs simulation calculations under realistic packaging process boundary conditions (e.g., bonding temperature, pressure curves) to obtain stress distribution prediction results. This accurately indicates the location and magnitude of stress concentration (e.g., maximum principal stress, equivalent stress) that will occur at the defect during subsequent packaging. Based on the stress concentration situation simulated by FEA, the optimal solution is matched from a pre-set control strategy library. The strategy library pre-stores the mapping relationship between various defect-stress scenarios and optimal control measures (expert knowledge base).For example: if simulation results show that the stress exceeds the threshold and there is a risk of cracking, the matching strategy is "reduce bonding temperature (reduce thermal stress) and increase bonding pressure (enhance bonding strength)"; if simulation results show that the stress does not exceed the limit but there is contamination, the matching strategy is "trigger plasma cleaning". Finally, an adjustment strategy identifier is output, which is a code or index pointing to a specific operation instruction. Based on the received adjustment strategy identifier, specific process parameter instructions can be directly generated. For example, if the strategy identifier is "Adjust_Bonding_Param", the specific instructions generated are: "Bonding temperature: 160℃; Bonding pressure: 7MPa". The instructions are sent to the packaging equipment or repair unit in real time via industrial Ethernet or high-speed bus. Packaging equipment: After receiving the instructions, it dynamically adjusts its process parameters (e.g., reducing the temperature from 180℃ to 160℃ and increasing the pressure from 5MPa to 7MPa) to compensate for the reliability risks caused by defects. Repair unit: If the instruction involves repair (e.g., plasma cleaning), the repair unit is triggered to perform the corresponding operation. After completion, the wafer enters the packaging process again.

[0031] The following is combined Figure 2 The methods of the embodiments of this application will be further explained below. Figure 2 This is a schematic diagram illustrating the process of driving a preset stress simulation model based on finite element analysis to perform simulation calculations in an embodiment of this application. Please refer to [link / reference]. Figure 2 The simulation calculations driven by the preset stress simulation model based on finite element analysis include the following steps: S201. Based on the category and the severity level, map and call the corresponding defect geometric feature parameters and material mechanical property parameters from the preset material parameter knowledge base. The material mechanical property parameters include elastic modulus, Poisson's ratio, yield strength and coefficient of thermal expansion. S202. Based on the defect geometric feature parameters, construct a wafer geometric entity for finite element calculation, wherein the wafer geometric entity characterizes the three-dimensional morphology of the defect; S203. The material mechanical property parameters and standard packaging process parameters are used as physical input parameters of the stress simulation model, and the boundary conditions of the simulation are set. The standard packaging process parameters include bonding temperature curve, bonding pressure curve and bonding head movement trajectory. S204. Perform finite element calculations based on the wafer geometry, the physical input parameters, and the boundary conditions, and output the quantified stress distribution results at the bonding interface.

[0032] Defect geometric characteristic parameters: Quantitative data used to define the shape and size of defects in simulations. For example, for a "microcrack," its parameters might include length, width, depth, and tip radius of curvature. Material mechanical property parameters: Inherent property parameters describing the mechanical behavior of a material under stress. Elastic modulus: The ratio of stress to strain during the elastic deformation stage of a material, characterizing the material's ability to resist elastic deformation. The higher the value, the harder the material. Poisson's ratio: The ratio of the absolute values ​​of transverse normal strain to axial normal strain when a material is under uniaxial tension or compression, characterizing the material's ability to deform laterally. Yield strength: The stress value at which a material begins to undergo significant plastic deformation. Exceeding this value, the material will undergo irreversible permanent deformation. Coefficient of thermal expansion: The relative change in size of a material per degree Celsius increase in temperature, a key parameter for calculating thermal stress. Upon receiving a specific defect description, such as {Category: Microcrack, Severity: Moderate}, the system uses (microcrack, moderate) as the joint query key to search the preset material parameter knowledge base. The knowledge base returns a pre-associated set of parameters with the key value. For example: defect geometric feature parameters: {length: 5μm, depth: 2μm, tip radius: 0.1μm}, material mechanical property parameters: {elastic modulus: 170 GPa, Poisson's ratio: 0.28, yield strength: 10 GPa, coefficient of thermal expansion: 2.6e-6 / K}. These parameters are successfully invoked, preparing for the next step of geometric modeling and physical simulation. Wafer geometry entity: A three-dimensional digital model created in computer-aided design (CAD) software or finite element analysis (FEA) software to represent the wafer and its defects. Create an idealized, defect-free 3D wafer model in the CAD software. Based on the defect geometric feature parameters, precisely create the geometric features characterizing the defects at the corresponding locations in the basic geometry (usually the cut edge) using Boolean operations (such as cutting, shelling) or feature modeling functions. For example, based on the parameters {length: 5μm, depth: 2μm}, a "V" or "U" shaped groove 5μm long and 2μm deep is created at the wafer edge to simulate microcracks. The "tip radius: 0.1μm" defines the sharpness of the groove bottom, which is crucial for stress concentration calculations. A 3D model of the wafer geometry containing accurate defect morphology is generated, which can be imported into finite element software for meshing. Standard packaging process parameters: Standard process settings used in packaging production. Bonding temperature profile: A curve describing the temperature change over time during bonding, including heating rate, peak temperature, holding time, and cooling rate. Bonding pressure profile: A curve describing the pressure change over time during bonding. Bonding head trajectory: Describes the path and velocity information of how the bonding head moves during pressure application. Boundary conditions: In finite element analysis, these are mathematical constraints used to define how the model interacts with the external environment and are prerequisites for solving differential equations. Examples include fixing certain surfaces, applying pressure, or displacement.In finite element method (FEM) software, material mechanical properties (elastic modulus, Poisson's ratio, etc.) are assigned to the material corresponding to the wafer geometry. Standard packaging process parameters are converted into software-recognizable loads. Bonding temperature profiles are applied to the model as thermal loads. Bonding pressure profiles and bonding head motion trajectories are combined to convert them into pressure or displacement loads, which are applied to the corresponding surfaces of the model (e.g., the bonding head contact surface). Constraints are imposed on the model based on the actual packaging fixtures and process conditions. For example, all degrees of freedom at the bottom of the wafer are fixed to simulate its state supported by a tray. Bonding interface: The surface on which the wafer connects to other components (e.g., the substrate, another wafer) through the bonding process. Stress distribution quantification results: Detailed data fields of the magnitude and direction of stress within the structure obtained after calculation, usually visualized as contour plots, with the numerical values ​​of specific nodes extracted. Before solving, the wafer geometry is cut into millions of tiny finite element elements (i.e., meshes). The mesh is denser in critical areas such as defects and bonding interfaces to ensure computational accuracy. Based on the input physical parameters and boundary conditions, equilibrium equations are established for each element, and this massive system of equations is solved simultaneously. This process is computationally intensive and is typically performed on high-performance computers or servers. The calculations include thermal stress analysis and mechanical stress analysis. After the calculations are completed, a stress contour map of the entire model, especially the bonding interface region, is generated. Users can obtain specific stress values ​​(e.g., MPa) at any location on the interface. The system can automatically identify and output quantitative results such as the maximum stress point and its value, and the stress concentration factor. These results directly reveal the reliability risks that defects may cause under packaging processes.

[0033] In some embodiments, performing finite element calculations based on the wafer geometry, the physical input parameters, and the boundary conditions, and outputting the quantized stress distribution results at the bonding interface includes: Based on the wafer geometry, the material mechanical properties of the physical input parameters, and the boundary conditions, the thermal stress field generated by the wafer structure under the bonding temperature field is calculated. The thermal stress field is used as a prestress field and superimposed with the mechanical pressure field applied by the bonding head to calculate the comprehensive stress field; Based on the comprehensive stress field, the maximum principal stress, equivalent stress, and stress concentration factor at the bonding interface are extracted. Based on the equivalent stress and defect geometric characteristics, calculate the crack propagation risk coefficient of the defect under the comprehensive stress field; The maximum principal stress, the equivalent stress, the stress concentration factor, and the crack propagation risk factor are used together as the stress distribution quantification result, and the stress concentration factor is used to characterize the stress concentration situation.

[0034] The finite element solver reads the bonding temperature curve, transforms it into a series of time-varying (transient analysis) or steady-state temperature values, and applies them as thermal loads to the mesh nodes of the entire wafer geometry. Based on the material's coefficient of thermal expansion, the solver calculates the thermal strain of each element due to temperature changes. Using this thermal strain as the initial strain, the solver substitutes it into the thermoelastic constitutive equation (i.e., the generalized Hooke's law), considering the elastic modulus and Poisson's ratio, to calculate the thermal stress field resulting from constrained thermal deformation. This stress field is self-equilibrium within the wafer. Multiple load cases are set up within the solver. The calculated thermal stress field is defined as the initial stress state or prestress field and locked onto each element of the model. Based on the bonding pressure curve and the bonding head's trajectory, time-varying mechanical pressure or displacement loads are defined on the contact surface between the bonding head and the wafer. The solver begins a new analysis step, applying mechanical loads on top of the existing prestress (thermal stress). Since the material may enter plasticity (considering yield strength), this is a nonlinear static analysis problem. The solver uses iterative algorithms (such as the Newton-Raphson method) to solve the system's equilibrium equations, ultimately obtaining the combined stress field under the combined action of thermal and mechanical loads. This is not a simple algebraic superposition, but rather considers the coupling results of the material's nonlinear response. Locating the bonding interface: In the post-processing software, by selecting elements or node sets, all mesh elements representing the bonding interface are precisely selected. Extracting the maximum principal stress (σmax): In the selected bonding interface region, the software is instructed to output the maximum principal stress of all nodes. This value is often used to assess the cracking risk of brittle materials (such as silicon) because cracks tend to propagate in the direction of maximum tensile stress. Calculating the equivalent stress (σvon): According to the von Mises yield criterion, the equivalent stress of each node is calculated. This value is used to assess whether the material has undergone plastic yielding or fatigue failure, and is a comprehensive indicator of the overall yield state of the material. Calculating the stress concentration factor (Kt): Kt = σmax / σnominal. Where σmax is the actual maximum stress at the defect root, and σnominal is the nominal stress under ideal conditions without defects. The coefficient is automatically calculated by comparing two simulation results (a defective model and a defect-free model) or by theoretical formulas, directly quantifying the amplification effect of the defect on stress. Defining the risk model: This is a rule-based or semi-empirical model calculation process. The system calls the corresponding risk assessment algorithm based on the defect category (e.g., microcracks). The core input parameters of this algorithm are the equivalent stress (σvon) value at the defect root and the defect's geometric characteristic parameters (e.g., crack depth a, tip curvature radius ρ). Calculating the risk coefficient (CRI): A simplified implementation is as follows: ; Here, σth is a threshold stress related to the fracture toughness of the material. A higher CRI indicates a higher risk of crack propagation under the current stress field. This coefficient provides a single, comparable risk indicator. The system packages the calculated maximum principal stress, equivalent stress, stress concentration factor, and crack propagation risk factor, along with their distribution contour maps across the entire bonding interface, into a structured data report (such as JSON or XML format), and outputs this as the stress distribution quantification result to the decision control module. This data is no longer an abstract contour map, but rather precise numerical values ​​that can be automatically interpreted and used for decision-making by the program.

[0035] In some embodiments, calculating the thermal stress field generated by the wafer structure under the bonding temperature field based on the wafer geometry, the material mechanical property parameters of the physical input parameters, and the boundary conditions specifically includes: Based on the bonding temperature curve, time-varying thermal load boundary conditions are defined on the wafer geometry. Based on the coefficient of thermal expansion in the mechanical property parameters of the material, the distribution of lattice thermal strain caused by temperature change is calculated. Based on the elastic modulus and Poisson's ratio in the material's mechanical properties, and in conjunction with the boundary conditions, the thermal stress field distribution caused by the lattice thermal strain is calculated by solving the thermoelastic constitutive equation to obtain the thermal stress field.

[0036] The system reads the bonding temperature profile, which describes the complete temperature change over time experienced by the wafer during the packaging process. For example, starting at room temperature (25°C), the temperature is increased to 180°C at a rate of 10°C / min, held for 300 seconds, and then allowed to cool naturally. In the finite element method (FEM) software, this temperature-time profile is set as an analysis load step. This means the simulation does not calculate a single static temperature field, but rather simulates a transient thermal process. This temperature profile is applied to the entire model or a specific surface of the wafer geometry. The software assigns the corresponding temperature value as a thermal load to each node of the model at each time point. This step transforms the actual manufacturing process parameters into thermal input conditions that the FEM software can understand for calculation. It ensures that the simulated thermal environment is consistent with the real manufacturing process. The coefficient of thermal expansion is retrieved from the material parameters. This coefficient is an inherent property of the material, representing the relative change in material length per unit temperature change (unit: 1 / °C or 1 / K). For each element in the model, the finite element solver calculates its free thermal strain based on the temperature change at that element node (ΔT = current temperature - initial reference temperature) and the material's coefficient of thermal expansion (α). The formula is: ε_thermal = α * ΔT, where ε_thermal is the thermal strain, a dimensionless quantity. After this calculation is completed across the entire model, a lattice thermal strain distribution field is obtained. This field describes how much deformation each element would undergo if it could freely expand or contract. This step calculates the amount of free expansion / contraction of the material without any constraints. In reality, wafers in packaging devices are constrained by fixtures, adjacent materials, etc., and cannot freely deform according to the thermal strain calculated in the previous step. This deformation inconsistency leads to the generation of internal stress. Boundary conditions in the model are used to define these constraints (e.g., fixing certain surfaces). The solver calls the material's elastic modulus (E) and Poisson's ratio (ν). These two parameters define the stress-strain relationship of the material within its elastic range. The solver uses thermal strain (ε_thermal) as the initial strain and substitutes it into the constitutive equations (i.e., the generalized Hooke's law) describing the behavior of linear elastic materials for calculation. The simplified physical idea is: total strain = mechanical strain + thermal strain. Due to boundary constraints, the total strain is limited; therefore, to compensate for the deformation caused by thermal strain, corresponding mechanical strain must be generated internally, thus producing stress. The formula for calculating stress (taking one-dimensional as an example) is: σ = E * (ε_total - ε_thermal), where ε_total represents the total strain. In the three-dimensional case, the calculation is more complex and requires consideration of the Poisson effect. By solving the equilibrium equations of the entire model, the software calculates the stress generated in each element due to the constraint of thermal deformation, and finally outputs the thermal stress field distribution of the entire wafer. This is usually displayed in the form of a contour plot, and stress values ​​at any point can be extracted.

[0037] In some embodiments, the step of superimposing the thermal stress field as a prestress field with the mechanical pressure field applied by the bonding head to calculate the comprehensive stress field specifically includes: The thermal stress field is used as the initial stress state and preloaded onto the wafer geometry. Based on the bonding pressure curve and the bonding head movement trajectory, mechanical load boundary conditions are defined on the wafer geometry. Based on the material mechanical property parameters and the mechanical load boundary conditions, the stress response under the combined action of thermal prestress and mechanical load is calculated by solving the elastoplastic constitutive equation. An iterative algorithm is used to converge the stress response until a preset convergence condition is met, and the converged stress field is taken as the comprehensive stress field.

[0038] The entire data of the equilibrium thermal stress field calculated in the previous analysis step (thermal stress analysis) (i.e., the stress value at each integration point) is imported as the initial conditions into the current new analysis step. This means that a stable, self-balancing internal stress field already exists inside the wafer model before the mechanical load is applied. This internal stress field is generated by the previous heating process and constitutes the current pre-stress state of the wafer. This step accurately restores the actual physical sequence: during packaging, the wafer always undergoes heating first, and then receives mechanical pressure from the bonding head in a state where internal thermal stress (pre-stress) has been generated. This is not a simple algebraic addition of stresses, but rather establishes the initial reference state for the mechanical load. The system reads the bonding pressure curve (describing how the pressure changes over time) and the bonding head motion trajectory (describing the displacement and angle of the bonding head), defines the contact relationship between the bonding head and the wafer surface (e.g., surface-to-surface contact) based on the motion trajectory, and specifies its motion mode. The pressure curve is converted into a pressure load and applied to the reference point or contact surface of the bonding head. Depending on the actual fixture conditions, necessary constraints (such as fixing constraints) are applied to the bottom of the wafer or other parts to prevent rigid body displacement. This step accurately maps the second physical field—the mechanical load—into the finite element model, preparing for the simulation of key mechanical actions in the bonding process. An elastoplastic constitutive model considering the material's yield strength is used, defining the elastic behavior of the material before reaching its yield strength and its plastic behavior after yielding (requiring input parameters such as the hardening curve). The solver begins a new analysis step, progressively applying mechanical loads based on the existing thermal prestress. For each load increment, the solver needs to: determine whether the material at each point in the model is in an elastic or plastic state; and calculate the new stress-strain response caused by the mechanical load increment according to the elastoplastic constitutive equations; this response is a result coupled with the initial thermal stress, rather than being calculated independently. This is the most critical step. It acknowledges the fact that in high-stress defect regions, the material may have already undergone or is close to plastic yielding. Using an elastoplastic model allows for a more realistic prediction of material behavior (such as permanent deformation), resulting in accurate and reliable stress results. This is the core difference between high-end simulation and basic simulation. Due to material and contact nonlinearities, the system cannot be solved directly. A solver (such as using the Newton-Raphson iteration method) performs multiple iterations within a single load increment step. Each iteration calculates an unbalanced force and adjusts the displacement field to attempt to bring the system back to equilibrium. After each iteration, the residuals (unbalanced force), displacement corrections, etc., are checked to ensure they are within a preset convergence tolerance. If convergence is achieved, the solution is accepted, and the process proceeds to the next load increment step. If convergence fails, the process continues with another iteration or the load increment step is reduced and retried until convergence is achieved. Once all mechanical load increment steps have been successfully applied and solved, the resulting stress field within the entire model is the composite stress field.This field is the final result of the complex, nonlinear coupling of thermal and mechanical stresses, taking into account the plasticity stage of the material. It most realistically reflects the actual stress state of the wafer during the bonding process and serves as the standard basis for reliability assessment and decision-making.

[0039] In some embodiments, inputting the composite defect feature vector into a preset convolutional neural network model and outputting the defect category and severity level specifically includes: The composite defect feature vector is input into the preset convolutional neural network model, which includes a shared feature encoder, a classification decoder, and a regression decoder. The shared semantic features are obtained by performing feature extraction and nonlinear transformation on the composite defect feature vector through the shared feature encoder. The shared semantic features are respectively input into the classification decoder and the regression decoder; The classification decoder outputs the category probability distribution of defects, and the category with the highest probability is determined as the final defect category; The severity level of the defect is output by the regression decoder as a continuous value. The continuous value of the severity level is normalized to obtain an intermediate value. The severity level is then determined based on the intermediate value according to a preset mapping relationship.

[0040] The composite defect feature vector is first fed into a shared feature encoder. The shared feature encoder typically consists of multiple fully connected layers and activation functions (such as ReLU) stacked together, performing nonlinear transformations and dimensionality reduction on these features. Its purpose is to eliminate redundant information and extract highly abstract shared semantic features that best represent the essence of the defect. These features simultaneously contain information related to both "what kind of defect it is" and "how severe it is." This avoids building two separate models for classification and regression tasks, significantly reducing computational resource consumption and training time. Ensuring that classification and regression tasks are based on the same set of high-value features avoids decision discrepancies caused by different feature sources, improving the model's generalization ability and robustness. The shared semantic features output from the shared encoder are copied twice and fed to the classification decoder and regression decoder respectively. Each decoder consists of several fully connected layers, with its network structure and parameters independently optimized according to its respective task objective. The classification decoder computes through the network and ultimately outputs a class probability distribution. For example, for a three-class classification problem (mechanical damage / thermal damage / contamination), the output might be a vector of the form [0.02, 0.17, 0.81]. The system applies the argmax function to find the class with the largest probability value in the probability distribution. In the example above, the maximum probability is 0.81, corresponding to the index of class 2 (contamination), so the final defect category of the sample is determined to be "contamination". The output probability distribution provides the confidence level of the model's judgment, not just the result. This increases the transparency and interpretability of the decision, facilitating manual review or risk assessment in subsequent processes. The regression decoder calculates through the network and outputs a continuous value. This value is a raw, unfiltered severity estimate (e.g., an output value of 8.65). Since the range of this raw value is uncertain, it needs to be normalized to map it to a fixed interval (e.g., [0, 1]). For example, the sigmoid function converts 8.65 to 0.92. This 0.92 is the intermediate value. The system converts the normalized intermediate value into the final severity level according to a preset mapping relationship. Example mapping: [0, 0.3) corresponds to Level 1: Slight; [0.3, 0.7) corresponds to Level 2: Moderate; [0.7, 1] corresponds to Level 3: Severe. Based on this mapping, 0.92 is determined to be "Level 3: Severe".

[0041] In some embodiments, the step of extracting features and performing nonlinear transformations on the composite defect feature vector using the shared feature encoder to obtain shared semantic features specifically includes: The composite defect feature vector is processed layer by layer by a series of one-dimensional convolutional layers to extract feature representations at different levels. Based on the residual connection structure, feature representations at different levels are fused to obtain fused features; A global max pooling layer is used to reduce the dimensionality of the fused features and extract the target feature vector. The target feature vector is input into a fully connected layer for nonlinear transformation and feature compression, and the shared semantic features are output.

[0042] The encoder consists of multiple (cascaded) one-dimensional convolutional layers. Each layer convolves the input and then introduces nonlinearity through an activation function (such as ReLU). The lower convolutional layers, closer to the input, have smaller kernels and narrower receptive fields, primarily capturing local, fine-grained feature patterns (e.g., a specific spectral peak or image edge fragment). As data progresses, deeper convolutional layers (through stacking) have larger receptive fields, enabling the combination of lower-level features into more global, abstract feature patterns (e.g., combining edges into crack shapes, or correlating multiple spectral peaks to determine material phase transitions). Through this hierarchical structure, the network can automatically learn feature representations from data at different levels, from simple to complex and from local to global, thus achieving a deep understanding of the nature of defects. Residual connections are a core structure of modern deep networks like ResNet. Instead of simply passing data layer by layer, it skips one or more layers of input (or output of a previous layer) and directly adds it element-wise with the output of the next layer. Feature fusion: Residual connections are used to fuse low-level features (from shallower layers, rich in detail but low in abstraction) and high-level features (from deeper layers, semantically abstract but potentially losing detail). For example, the output of layer 3 is added to the output of layer 5. This makes the network easier to train, allowing for the construction of deeper encoders without gradient problems. It ensures that important details captured in the lower layers (such as subtle geometric features of defects) are not lost when extracting high-level semantic features. This fusion guarantees the richness and comprehensiveness of the final features. After convolution and feature fusion, the resulting feature map may still have high dimensionality. Global max pooling is a simple operation: for each feature channel of the feature map, it takes the maximum value among all values ​​in that channel as the representative of that channel. Assuming the fused feature has 512 channels, after global max pooling, each channel retains only one maximum value, resulting in a target feature vector of length 512. Compressing the feature map into a compact vector significantly reduces the number of parameters required for subsequent computations, retaining only the most active and significant responses in each feature channel. This makes the feature vector insensitive to changes in feature position, enhancing the model's robustness. The resulting target feature vector is then input into one or more fully connected layers. Each neuron in a fully connected layer is connected to all outputs of the previous layer, undergoing complex nonlinear combinations and transformations through weight matrices and activation functions. This process can be viewed as a further refinement of the features, removing redundancy and retaining the core information most useful for subsequent classification and regression tasks. After processing by the fully connected layers, the final output is a feature vector with lower dimensionality and higher information density—this is the shared semantic feature.

[0043] Through comparative testing of 1000 232-layer 3D NAND wafers, the packaging yield of this invention increased from 92% to 97%, rework costs decreased from $100 per wafer to $60, and the production cycle was shortened from 24 hours to 18 hours, verifying the effectiveness of the technology. Compared with the prior art, the beneficial effects of this invention are as follows: 1) Significantly improved packaging yield: The packaging failure rate caused by dicing defects decreased from the traditional 30% to below 8%, and the overall SSD chip yield increased from 92% to 97%; 2) Significantly reduced costs: Rework caused by dicing defects (such as re-dicing or scrapping) was reduced, saving approximately $40 in rework costs per wafer (based on a rework cost of $100 per wafer); 3) Shortened production cycle: Dicing and packaging processes are carried out in tandem, eliminating the need for intermediate storage and long-distance transportation, shortening the production cycle from the traditional 24 hours to 18 hours; 4) Enhanced process reliability: By dynamically adjusting packaging parameters, the impact of dicing defects on bonding strength is compensated, and the solder ball failure rate is reduced by 15%-20%.

[0044] The methods for wafer defect detection and control in the embodiments of this application have been described above. The computer system in the embodiments of this application will be described in detail below in conjunction with the above methods for wafer defect detection and control.

[0045] Please see Figure 3 This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application.

[0046] In some embodiments, the computer system 300 includes a computer device, which may be a terminal device. The computer device includes a processor 301, a memory 302, a sensor module 303, a communication module 304, an input device 305, and an output device 306 connected via a system bus. The processor 301 of the computer device provides computing and control capabilities. The memory 302 of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database is used to store data.

[0047] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0048] In some embodiments of this application, a computer-readable storage medium is provided, including instructions that, when executed on the computer system 300, cause the computer system 300 to perform a wafer defect detection and control method according to an embodiment of this application.

[0049] In some embodiments of this application, a computer program product is also provided, which, when run on a computer system 300, causes the computer system 300 to execute a wafer defect detection and control method according to an embodiment of this application.

[0050] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0051] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0052] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for detecting and controlling wafer defects, characterized in that, The method comprises the following steps: After wafer cutting is completed, microscopic morphology image data and material structure spectrum data of a cutting edge of the wafer are collected; Deep geometric features of the microscopic morphology image data are extracted, deep spectrum features of the material structure spectrum data are extracted, the deep geometric features are fused with the deep spectrum features, and a composite defect feature vector is generated; The composite defect feature vector is input into a preset convolutional neural network model, and a category and a severity level of a defect are output; the category and the severity level are taken as input parameters to drive a preset stress simulation model based on finite element analysis to perform simulation calculation, so that a stress distribution prediction result containing a stress concentration situation is obtained; based on the stress distribution prediction result, a control strategy identifier capable of optimizing the stress concentration situation is matched and output from a preset control strategy library; According to the control strategy identifier, a control instruction containing specific process parameters is generated, and the control instruction is sent to a downstream packaging device or a repair unit to drive the packaging device or the repair unit to perform corresponding parameter adjustment or online repair operation.

2. The method of claim 1, wherein, The driving of the preset stress simulation model based on finite element analysis to perform simulation calculation comprises: According to the category and the severity level, corresponding defect geometric feature parameters and material mechanical property parameters are mapped and called from a preset material parameter knowledge base, the material mechanical property parameters including elastic modulus, Poisson's ratio, yield strength and thermal expansion coefficient; According to the defect geometric feature parameters, a wafer geometric entity for finite element calculation is constructed, the wafer geometric entity representing a three-dimensional morphology of the defect; The material mechanical property parameters and standard packaging process parameters are taken as physical input parameters of the stress simulation model, and a simulated boundary condition is set, the standard packaging process parameters including a bonding temperature curve, a bonding pressure curve and a bonding head motion trajectory; Based on the wafer geometric entity, the physical input parameters and the boundary condition, finite element calculation is performed, and a stress distribution quantization result at a bonding interface is output.

3. The method of claim 2, wherein, The finite element calculation based on the wafer geometric entity, the physical input parameters and the boundary condition to output the stress distribution quantization result at the bonding interface comprises: Based on the wafer geometric entity, the material mechanical property parameters of the physical input parameters and the boundary condition, a thermal stress field generated by a wafer structure under a bonding temperature field is calculated; The thermal stress field is taken as a pre-stress field, and a mechanical pressure field applied by a bonding head is superimposed to calculate a comprehensive stress field; Based on the comprehensive stress field, a maximum principal stress, an equivalent stress and a stress concentration coefficient at the bonding interface are extracted; According to the equivalent stress and the defect geometric feature, a crack propagation risk coefficient of the defect under the comprehensive stress field is calculated; The maximum principal stress, the equivalent stress, the stress concentration coefficient and the crack propagation risk coefficient are taken as the stress distribution quantization result together, and the stress concentration coefficient is used to represent the stress concentration situation.

4. The method of claim 3, wherein, calculating a thermal stress field generated by the wafer structure under a bonding temperature field based on the wafer geometric entity, the material mechanics performance parameters of the physical input parameters, and the boundary conditions, specifically comprising: defining a thermal load boundary condition varying with time on the wafer geometric entity according to the bonding temperature curve; calculating a thermal lattice strain distribution caused by temperature variation based on the thermal expansion coefficient in the material mechanics performance parameters; calculating a thermal stress field distribution caused by the thermal lattice strain by solving a thermoelastic constitutive equation based on the elastic modulus and Poisson's ratio in the material mechanics performance parameters and the boundary conditions to obtain the thermal stress field.

5. The method of claim 4, wherein, superimposing the thermal stress field as a prestress field on a mechanical pressure field applied by a bonding head to calculate a comprehensive stress field, specifically comprising: preloading the thermal stress field as an initial stress state on the wafer geometric entity; defining a mechanical load boundary condition on the wafer geometric entity according to the bonding pressure curve and the bonding head motion trajectory; calculating a stress response under the combined action of thermal prestress and mechanical load by solving an elastoplastic constitutive equation based on the material mechanics performance parameters and the mechanical load boundary condition; convergently solving the stress response by using an iterative algorithm until a preset convergence condition is met, and taking the stress field obtained after convergence as the comprehensive stress field.

6. The method of claim 1, wherein, inputting the composite defect feature vector into a preset convolutional neural network model to output the category and severity level of the defect, specifically comprising: inputting the composite defect feature vector into the preset convolutional neural network model, the preset convolutional neural network model comprising a shared feature encoder, a classification decoder, and a regression decoder; obtaining shared semantic features by feature extraction and nonlinear transformation of the composite defect feature vector through the shared feature encoder; inputting the shared semantic features into the classification decoder and the regression decoder, respectively; outputting a category probability distribution of the defect through the classification decoder, and determining the category with the maximum probability as the final defect category; outputting a severity level continuous value of the defect through the regression decoder, normalizing the severity level continuous value to obtain an intermediate value, and determining the severity level based on the intermediate value according to a preset mapping relationship.

7. The method of claim 6, wherein, obtaining shared semantic features by feature extraction and nonlinear transformation of the composite defect feature vector through the shared feature encoder, specifically comprising: extracting feature representations at different levels by layer-by-layer processing of the composite defect feature vector through a series of multi-layer one-dimensional convolutional layers; fusing the feature representations at different levels to obtain fused features according to a residual connection structure; extracting a target feature vector by dimension reduction of the fused features through a global maximum pooling layer; inputting the target feature vector into a fully connected layer for nonlinear transformation and feature compression to output the shared semantic features.

8. A computer system comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1 to 7. The processor executes the computer program to implement the steps of the method of any one of claims 1-7.

9. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the method of any one of claims 1-7.

10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions, when executed by the processor, implement the steps of the method of any one of claims 1-7.

Images