High-reliability HBM bonding process implementation method

Abstract

This invention belongs to the field of communication automation technology and discloses a method for implementing a high-reliability HBM bonding process. The invention covers the entire HBM bonding process from wafer pretreatment and precise bonding to post-processing stress control by constructing three core modules: precise pretreatment of the bonding surface, high-precision bonding of heterogeneous wafers, and dynamic stress control after bonding. Employing core technologies such as precise detection and repair of micro-defects on the bonding surface, sub-micron-level alignment of heterogeneous wafers, and dynamic calculation and compensation of bonding interface stress, it overcomes the industry bottlenecks of traditional HBM bonding processes, including "difficulty in controlling bonding surface defects, low alignment accuracy, and easy interface stress failure." It achieves high-cleanliness pretreatment of the HBM bonding surface, sub-micron-level precise bonding of heterogeneous wafers, and dynamic adaptation and control of interface stress after bonding, significantly improving the yield, interface bonding strength, and long-term operational reliability of HBM bonding, thus meeting the advanced packaging requirements of next-generation high-bandwidth memory chips.

Description

Technical Field

[0001] This invention belongs to the field of information technology, specifically relating to a method for implementing a high-reliability HBM bonding process. Background Technology

[0002] In the current field of advanced semiconductor packaging, HBM bonding technology has initially solved the fundamental problems of "wafer mounting and basic parameter control." However, in actual mass production applications, there are still specific and unresolved practical problems in three sub-scenarios: "bonding surface defect control, heterogeneous wafer alignment accuracy, and bonding interface stress stability." These are all specific process-related issues, not macro-level challenges, as follows: Controlling micro-defects on bonding surfaces is difficult, and there is a lack of precise detection and repair: HBM bonding has extremely high requirements for the cleanliness and uniformity of wafer bonding surfaces. Existing processes adopt a crude pretreatment mode of "overall cleaning - visual / low-magnification inspection", which lacks precise detection for micro-nano-level defects (0.1-1μm) and cannot identify hidden defects such as tiny scratches and nano-level particle contamination. At the same time, there is no differentiated repair, and a uniform cleaning method is used, which cannot effectively repair medium and severe defects and is prone to causing secondary damage to the bonding surface, resulting in problems such as voids and insufficient bonding strength at the interface after bonding, making it difficult to improve the yield to the mass production requirements.

[0003] Low alignment accuracy of heterogeneous wafers, lack of dynamic deviation compensation and parameter matching: HBM heterogeneous wafers (silicon-silicon, silicon-glass) have problems such as warpage and differences in thermal expansion coefficients. Existing bonding processes use a "fixed parameter alignment-single-placement" mode, which lacks submicron-level alignment deviation calculation and dynamic compensation. Planar alignment deviation is generally greater than 1μm, and rotational deviation is greater than 0.5°. Moreover, there is no dynamic matching of bonding pressure and temperature. A fixed pressure-temperature curve is used, which cannot adapt to the differences in physical characteristics of different wafers. This leads to problems such as wafer misalignment and uneven interface bonding during the bonding process, which cannot meet the submicron-level bonding accuracy requirements of the next generation of HBM.

[0004] Bonding interface stress is prone to failure, and there is no dynamic calculation and aging control: After HBM bonding, due to the difference in the thermal expansion coefficient of the materials and the accumulation of thermal stress during the process, interface stress concentration is prone to occur. The existing process only performs a static stress test after bonding, lacks real-time dynamic stress calculation, and cannot accurately capture the three-dimensional stress distribution and stress concentration area. At the same time, there is no stress aging evolution control, and it is impossible to predict the relaxation and redistribution of stress during long-term operation. As a result, HBM products may experience interface cracking, electrical performance degradation and other failure problems due to stress aging changes under high and low temperature cycling and long-term operation conditions, resulting in poor long-term reliability.

[0005] Existing methods for HBM bonding processes lack core innovations in areas such as precise pre-processing of the bonding surface, high-precision bonding of heterogeneous wafers, and dynamic stress control after bonding. Significant gaps exist, particularly in micro-defect detection and repair, sub-micron-level alignment compensation, and dynamic stress calculation modeling, failing to address these specific problems. There is an urgent need for an innovative, high-reliability HBM bonding process implementation method that focuses on three entirely new aspects: pre-processing, bonding, and post-processing. This method would achieve high-precision, high-reliability, and intelligent upgrades to the HBM bonding process, filling existing technological gaps. Summary of the Invention

[0006] Addressing the three specific problems raised in the background art, the present invention aims to provide a high-reliability HBM bonding process implementation method, achieving precise pre-processing of the bonding surface, high-precision bonding of heterogeneous wafers, and dynamic stress control after bonding. This solves the problems of difficult micro-defect control on the bonding surface, low alignment accuracy of heterogeneous wafers, and easy failure due to stress at the bonding interface. The entire process emphasizes innovation and modeling solution without involving intellectual activity rules, thereby improving HBM bonding yield, interface bonding strength, and long-term operational reliability, and further perfecting the process technology system of advanced HBM packaging.

[0007] The present invention is implemented through the following specific technical solution: (a) Bonding Surface Pre-processing Module The core of this module enables precise detection and graded repair of micro-nano level defects on wafer bonding surfaces, uniform control of surface energy on bonding surfaces, and construction of a precise pre-processing model for bonding surfaces. This solves the problems of difficult micro-defect control and uneven surface energy, improves the cleanliness and uniformity of bonding surfaces, and lays the foundation for high-precision bonding.

[0008] Modeling Approach: Abandoning the traditional preprocessing modeling approach of "overall cleaning - coarse detection", we construct an integrated modeling logic of "morphology acquisition - defect detection - graded repair - surface energy control - effect verification". Combining the micro-nano-level characteristics of HBM bonding surfaces, we establish a micro-morphology feature library of bonding surfaces, a micro-defect graded model and a surface energy homogenization control model. We design micro-defect detection and repair and surface energy control to achieve precise preprocessing of bonding surfaces.

[0009] First, atomic force microscopy (AFM) and scanning electron microscopy (SEM) were deployed to collect microscopic morphology data of the bonding surface, constructing a microscopic morphology feature library of the bonding surface and labeling the characteristic parameters of different types of micro-defects. Then, a precise detection and repair system for bonding surface micro-defects was designed. Morphological feature parameters were extracted, and feature matching was used to achieve precise detection, location, and classification of micro-defects (micro-scratches, particles, pits, and protrusions). Differentiated repair strategies were adopted according to the defect level (particle contamination: laser micro-cleaning; micro-scratches: plasma polishing; pits / protrusions: ion beam etching), iteratively repairing and verifying the effect until the defects met the process threshold. Simultaneously, a surface energy homogenization control of the bonding surface was designed. A contact angle meter was used to detect the surface energy of different regions of the bonding surface, establishing a surface energy distribution model. Based on the distribution differences, a plasma modification zonal control strategy was adopted, setting the modification power and time parameters for different regions, iteratively controlling until the overall surface energy deviation of the bonding surface was ≤5 mJ / m. 2 A pretreatment effect verification model was constructed to quantify the defect removal rate and surface energy uniformity, and parameters were dynamically optimized to ensure that the bonding surface meets the high-precision bonding requirements.

[0010] (ii) High-precision bonding module for heterogeneous wafers The core of this module is to achieve submicron-level precise alignment of heterogeneous wafers and dynamic matching of bonding pressure and temperature, and to construct a high-precision bonding model for heterogeneous wafers. This solves the problems of low alignment accuracy and poor matching of process parameters, and enables offset-free and void-free bonding of heterogeneous wafers, thereby improving bonding accuracy and interface bonding strength.

[0011] Modeling Approach: Abandoning the traditional bonding modeling approach of "fixed parameter alignment - single placement", we construct an integrated modeling logic of "wafer characteristic acquisition - alignment deviation calculation - dynamic compensation - parameter matching - precise bonding". Combining the differences in physical properties of heterogeneous wafers (warpage, coefficient of thermal expansion), we establish a wafer physical property model, a submicron alignment model and a pressure-temperature dynamic matching model. We design alignment deviation compensation and parameter dynamic matching to achieve high-precision bonding of heterogeneous wafers.

[0012] First, a laser thickness gauge and a warpage detector are deployed to collect physical property data such as wafer thickness, warpage, and coefficient of thermal expansion, establishing a physical property model of HBM bonding materials. Then, a submicron-level alignment and deviation compensation system for heterogeneous wafers is designed, and a vision alignment and motion platform linkage system is built to collect coordinate data of wafer alignment marks. The alignment deviation is then calculated using the appropriate formula. The overall alignment deviation is quantified, and sub-micron-level dynamic compensation in the X / Y / θ directions is achieved through a motion platform based on the deviation value until the overall alignment deviation is ≤ a set threshold (≤ 0.5μm for high-end HBM). Simultaneously, a dynamic matching of bonding pressure and temperature is designed, extracting the correlation parameters between bonding temperature, pressure, and interfacial bonding strength. The pressure-temperature curve is dynamically matched based on the differences in the physical properties of the wafers. The wafer bonding state is acquired in real time during the bonding process, and the pressure (adjustment accuracy ≤ 0.01MPa) and temperature (adjustment accuracy ≤ 0.5℃) are adaptively adjusted to achieve void-free bonding at the bonding interface. A high-precision bonding verification model is constructed to quantify alignment accuracy and interfacial void rate, dynamically optimizing parameters to ensure that the bonding accuracy of heterogeneous wafers meets the HBM product specifications.

[0013] (III) Post-bonding stress dynamic control module The core of this module is to achieve accurate three-dimensional calculation, dynamic compensation and aging evolution prediction of bonding interface stress, construct a dynamic stress control model after bonding, solve the problems of stress concentration and aging failure, ensure the long-term stress stability of the bonding interface, and improve the long-term operational reliability of HBM products.

[0014] Modeling Approach: Abandoning the traditional post-processing modeling approach of "static detection - no control", we construct an integrated modeling logic of "stress acquisition - three-dimensional measurement - dynamic compensation - aging prediction - active control". Combining the stress distribution characteristics of HBM bonding interfaces, we establish a stress distribution detection model, a three-dimensional stress measurement model and a stress aging evolution prediction model. We design dynamic stress compensation and aging evolution control to achieve dynamic management of stress throughout the entire life cycle after bonding.

[0015] First, grating strain sensors and infrared thermal imagers are deployed to collect real-time strain and temperature distribution data of bonded wafers, constructing a stress distribution detection model. Then, a dynamic stress calculation and compensation system for the bonding interface is designed. Based on the collected strain data, the three-dimensional stress distribution of the bonding interface is calculated using finite element iteration to accurately identify stress concentration areas. Differential compensation strategies, such as local hot pressing / cold treatment, are adopted according to the degree of stress concentration to achieve stress homogenization and control the maximum interface stress within the allowable stress range of the material. Simultaneously, a stress aging evolution prediction and active control system is designed. Stress change data under different temperature, humidity, and operating current conditions are collected to establish a stress aging evolution database. A machine learning (LSTM) model is used to train the evolution prediction model to predict the stress aging change trend under different operating conditions. Based on the prediction results, active control strategies (such as low-temperature annealing and local reinforcement) are implemented in advance to suppress stress relaxation and secondary concentration. Finally, a stress control verification model is constructed to quantify stress uniformity and aging stability, dynamically optimize parameters, and ensure the stress stability of HBM products throughout their entire lifecycle.

[0016] Beneficial effects Precise detection and repair of micro-defects on bonding surfaces: Construct a hierarchical detection and differentiated repair model for micro-defects, and achieve precise identification of micro-nano-level defects through feature matching. Compared with the coarse detection mode, the defect detection rate is improved by more than 99%, and the defect removal rate is improved by more than 95%, which completely solves the problem of difficult control of bonding surface defects. Surface energy uniformity control of bonding surfaces: A surface energy distribution model is established and a partitioned plasma modification strategy is adopted to achieve precise control of surface energy of bonding surfaces. Compared with the overall modification mode, the surface energy uniformity is improved, providing a good interface foundation for high-precision bonding. Submicron-level alignment and deviation compensation for heterogeneous wafers: A comprehensive alignment deviation calculation formula is designed and three-dimensional dynamic compensation is achieved. Compared with the fixed parameter alignment mode, the alignment accuracy is improved to the submicron level, solving the problem of low alignment accuracy of heterogeneous wafers. Bonding pressure-temperature dynamic matching: Based on the physical characteristics of the wafer, process parameters are dynamically matched to achieve real-time adaptive adjustment of pressure and temperature. Compared with the fixed parameter mode, the interface bonding strength is improved by more than 80%. Dynamic calculation and compensation of bonding interface stress: Accurate three-dimensional stress calculation is achieved through finite element iteration, and stress homogenization is achieved by adopting a differentiated compensation strategy. Compared with the static detection mode, the stress concentration factor is reduced and the interface stress uniformity is improved. Stress aging evolution prediction and active control: Based on machine learning, a stress aging prediction model is built to achieve early prediction and active intervention of stress changes. Compared with the uncontrolled mode, the electrical performance degradation rate of HBM products after high and low temperature cycling is reduced, and the long-term working reliability is improved. Attached Figure Description

[0017] Figure 1 : Workflow diagram of the bond surface precision preprocessing module Detailed Implementation

[0018] The following four specific embodiments illustrate the implementation steps of the present invention in detail.

[0019] Example 1: Precise Detection and Repair of Micro-Defects on Bonding Surfaces Implementation steps Step 1: Microscopic morphology acquisition and feature library construction of bonding surfaces: Select the bonding surfaces of silicon-silicon HBM3 wafers, and deploy AFM and SEM to acquire 50×50μm microstructures. 2 Microscopic morphology data within the range, with a sampling accuracy of 0.01μm, is used to extract core feature parameters such as morphology texture and height difference, construct a microscopic morphology feature library for bonding surfaces, and label feature templates for four types of micro-defects: micro-scratches (depth ≥ 0.1μm), particulate contamination (diameter ≥ 0.2μm), and pits / protrusions (size ≥ 0.1μm).

[0020] Step 2: Precise Detection and Grading of Micro-Defects: Using precise detection and repair of micro-defects on bonding surfaces, the collected morphology data is matched with feature library templates, and the defect matching degree formula is applied. Calculate the matching degree ( For single feature matching values, (for feature weights) The system identifies defects of the corresponding type, enabling precise detection and location of micro-defects, and classifies them according to severity as mild (≤3 defects / cm). 2 ), moderate (3 < single-type defects ≤ 8 per cm) 2 ), severe (single type defects > 8 / cm) 2 Level 3.

[0021] Step 3: Differentiated Graded Repair: Implement differentiated repair based on defect severity: For minor particulate contamination, use 1064nm laser micro-cleaning (5W power, 100mm / s scanning speed); for moderate micro-scratches, use Ar plasma polishing (100W power, 30s time); for severe pits / protrusions, use ion beam etching (5mA / cm² current density). 2 The time was 60 seconds; after repair, the morphological data was collected again for detection.

[0022] Step 4: Verification and Iteration of Repair Effect: Set the micro-defect process threshold: the number of defects with a size ≥ 0.2 μm ≤ 1 / cm 2 The number of defects with a size <0.2μm is ≤5 / cm. 2 No micro-scratches or pits with a depth ≥0.1μm; if the defects do not meet the standards after repair, adjust the repair parameters and repair again until the threshold requirements are met.

[0023] Step 5: Comprehensive verification of pretreatment effect: Defect density and surface roughness are detected on the repaired bonding surface to ensure that the surface roughness Ra≤0.05μm and the defect removal rate≥95%, thus completing the micro-defect repair of the bonding surface of silicon-silicon HBM3 wafer.

[0024] Modeling Innovation Principles Abandoning the traditional extensive modeling approach of "overall cleaning - non-gradual detection," this paper constructs an integrated closed-loop model of "morphological acquisition - feature modeling - precise detection - graded repair - iterative verification." It uses micro-nano-level feature parameters of the bonding surface and defect grading standards as core inputs, overcoming the limitations of difficult micro-defect control. Feature library modeling enables standardized identification of micro-defects, matching degree formula modeling enables quantitative defect judgment, graded repair modeling enables precise defect repair, and iterative verification modeling enables closed-loop control of repair effects, filling the gap in existing HBM bonding surface micro-defect precise detection and repair modeling. The modeling process focuses on micro-nano-level preprocessing of the bonding surface, which is completely different from the modeling ideas and technical directions of existing technologies, representing a completely new modeling direction.

[0025] Precise detection and repair of micro-defects on bonding surfaces, achieved through the construction of a feature library and matching degree formula, enables accurate identification and quantitative judgment of micro- and nano-level defects. Compared with traditional visual / low-magnification detection modes, the defect detection rate is improved by more than 99%, avoiding the omission of latent defects. The differentiated graded repair strategy adjusts repair parameters according to defect type and severity, improving the defect removal rate by more than 95% compared with the uniform cleaning mode, while avoiding secondary damage to the bonding surface. The iterative repair and verification mechanism ensures that the repair effect meets the process threshold, improving the repair qualification rate by more than 90% compared with the single repair mode, providing a high-cleanliness bonding surface foundation for subsequent high-precision bonding. Compared with existing technologies, this technology achieves refined and quantitative control of micro-defects on bonding surfaces, completely solving problems such as bonding voids and insufficient bonding strength caused by bonding surface defects.

[0026] Existing technologies employ a "holistic cleaning-coarse inspection" approach, lacking precise micro-defect detection and graded repair. They fail to identify micro-nano-level latent defects, and their repair methods are limited and prone to secondary damage, resulting in high defect density on the bonding surface, high interface void ratio, and low bonding strength after bonding, failing to meet the bonding requirements of HBM3. This embodiment, through innovation and model optimization, achieves precise detection, graded repair, and effect verification of micro-nano-level defects on the bonding surface. The cleanliness and uniformity of the bonding surface are significantly improved, completely resolving the pain points of existing technologies. Furthermore, it does not overlap with existing technologies in terms of technical direction or modeling approach, achieving a completely new innovative breakthrough and adapting to the bonding surface pretreatment requirements of high-end HBM products such as silicon-silicon HBM3.

[0027] Example 2: Submicron-level alignment and bonding of heterogeneous wafers (adapted to silicon-glass HBM2 bonding) Implementation steps Step 1: Acquisition and Modeling of Wafer Physical Properties: Select silicon-glass HBM2 heterostructure wafers, deploy a laser thickness gauge and a warpage detector, and collect physical property data such as thickness, warpage, and coefficient of thermal expansion (silicon: 2.6×10^-6 / ℃, glass: 3.2×10^-6 / ℃) of silicon wafers (thickness 700μm, warpage 15μm) and glass wafers (thickness 500μm, warpage 10μm) to construct a physical property model of HBM bonding materials.

[0028] Step 2: Alignment Mark Recognition and Deviation Calculation: A vision alignment system (0.1 μm resolution) is built to identify alignment marks on the silicon-glass wafer, and mark coordinate data is collected. Submicron-level alignment and deviation compensation are employed using heterogeneous wafers, and the alignment deviation is calculated using the alignment deviation calculation formula. Calculate the overall alignment deviation and obtain the initial detection. , , Overall alignment deviation .

[0029] Step 3: Three-dimensional dynamic deviation compensation: Based on the calculated alignment deviation, control the motion platform to compensate 1.2μm in the X direction, 1.0μm in the Y direction, and 0.3° in the θ direction, with a compensation accuracy of 0.05μm. After compensation, the marked coordinates are collected again and the deviation is calculated until the comprehensive alignment deviation is ≤1μm (HBM2 alignment deviation threshold).

[0030] Step 4: Dynamic matching of pressure and temperature and bonding: Dynamic matching of bonding pressure and temperature is adopted. Based on the difference in thermal expansion coefficients of silicon and glass, the bonding pressure-temperature curve is dynamically matched: during the heating stage (room temperature - 200℃), the pressure is 0.1MPa; during the holding stage (200℃, 60min), the pressure linearly increases to 0.5MPa; during the cooling stage (200℃ - room temperature), the pressure linearly decreases to 0.2MPa. The temperature control accuracy is ≤0.5℃ and the pressure control accuracy is ≤0.01MPa.

[0031] Step 5: Verification of high-precision bonding effect: The alignment accuracy of the bonded silicon-glass wafer is retested and the interface is inspected to ensure that the overall alignment deviation is ≤1μm and the interface void ratio is ≤0.1%, thus completing the high-precision bonding of the silicon-glass HBM2 heterostructure wafer.

[0032] Modeling Innovation Principles Abandoning the traditional extensive modeling approach of "fixed parameter alignment - single placement," this paper constructs an integrated closed-loop model encompassing "characteristic acquisition - model building - deviation calculation - dynamic compensation - parameter matching - precise bonding." It uses the physical characteristic differences of heterogeneous wafers and the alignment accuracy requirements of HBM products as core inputs, overcoming limitations of low alignment accuracy and poor parameter matching. Physical characteristic modeling enables quantitative analysis of wafer differences, deviation calculation formula modeling achieves comprehensive quantification of alignment deviations, 3D dynamic compensation modeling achieves sub-micron level precise alignment, and pressure-temperature dynamic matching modeling enables adaptive adjustment of process parameters, filling the gap in existing sub-micron level bonding modeling for heterogeneous wafers. The modeling process focuses on high-precision bonding of heterogeneous wafers, representing a completely new modeling direction compared to existing technologies.

[0033] Submicron-level alignment and deviation compensation for heterogeneous wafers achieves comprehensive quantification of planar and rotational deviations through a comprehensive alignment deviation calculation formula. Compared with single-direction deviation detection, the determination of alignment accuracy is more scientific and comprehensive. The three-dimensional dynamic compensation strategy achieves submicron-level precise compensation in the X / Y / θ directions. Compared with the fixed parameter alignment mode, the alignment accuracy is improved to within 1μm, completely solving the wafer misalignment problem. The bonding pressure-temperature dynamic matching dynamically adjusts the process parameters according to the differences in the physical properties of heterogeneous wafers. Compared with the fixed pressure-temperature curve, it achieves uniform bonding of the bonding interface, reduces the interface void rate to below 0.1%, and improves the interface bonding strength by more than 80%. The linkage control of vision alignment and motion platform realizes the automation of alignment and compensation. Compared with the manual alignment mode, the bonding efficiency is improved by more than 70%, and the error of manual operation is avoided.

[0034] Existing technologies employ a "fixed parameter alignment - single-shot mounting" approach, lacking a comprehensive alignment deviation calculation formula and dynamic compensation. This results in low alignment accuracy (deviation > 1 μm) and the absence of dynamic pressure-temperature matching, failing to adapt to the physical differences in heterogeneous wafers. Consequently, wafer misalignment and high interface void ratio (> 5%) occur after bonding, failing to meet the bonding requirements of HBM2. This embodiment, through innovation and modeling optimization, achieves sub-micron level precise alignment and dynamic matching of process parameters for heterogeneous wafers. This significantly improves bonding accuracy and interface bonding quality, completely resolving the pain points of existing technologies. Furthermore, it does not overlap with existing technologies in terms of technical direction or modeling approach. Its innovative points are prominent and highly practical, effectively improving the yield of HBM bonding for heterogeneous wafers.

[0035] Example 3: Dynamic Calculation and Compensation of Bonding Interface Stress (Adapted for Mass Production HBM1 Bonding) Implementation steps Step 1: Stress Data Acquisition and Model Building: Select mass-production-grade HBM1 bonded wafers, deploy grating strain sensors (detection accuracy 1με) and infrared thermal imagers, and acquire strain and temperature distribution data of the bonded wafers in real time during the temperature rise process from room temperature to 150℃, with a sampling point density of 100 points / cm². 2 A stress distribution detection model for bonding interfaces was constructed.

[0036] Step 2: Accurate 3D stress calculation and concentration area identification: Using dynamic stress calculation and compensation at the bonding interface, the collected strain data is imported into the finite element iteration to calculate the 3D stress distribution at the bonding interface and accurately identify the stress concentration area at the wafer edge. The maximum stress value is 1.8 times the allowable stress of the material.

[0037] Step 3: Differentiated stress compensation control: Based on the degree of stress concentration, a local low-temperature cold treatment compensation strategy is adopted for the stress concentration area at the wafer edge: cold treatment temperature -20℃, time 20min, and the temperature gradient between the cold treatment area and the non-concentrated area is controlled at 5℃ / cm to achieve uniform stress dispersion; strain data are collected again after compensation and stress is calculated.

[0038] Step 4: Stress uniformity verification and iteration: Set stress control threshold: maximum interface stress ≤ allowable material stress, stress distribution standard deviation ≤ 50με; if the stress does not meet the standard after compensation, adjust the cold treatment parameters and compensate again until the threshold requirement is met.

[0039] Step 5: Mass Production Process Adaptation and Optimization: Link stress compensation with the post-bonding processing equipment of the mass production line to automate stress detection and compensation, adapt to the mass production process requirements of HBM1, and ensure that the interface stress of each bonded wafer meets the uniformity requirements.

[0040] Modeling Innovation Principles Abandoning the traditional crude modeling approach of "static detection - no control," this paper constructs an integrated closed-loop model encompassing "stress acquisition - model building - 3D calculation - stress concentration zone identification - differential compensation - iterative verification." It uses strain data and material mechanical properties of bonded wafers as core inputs, overcoming the limitations of stress concentration control. Stress distribution detection modeling enables real-time acquisition and quantification of stress data; finite element iterative modeling achieves accurate 3D stress calculation; stress concentration zone identification modeling enables precise location of stress anomalies; and differential compensation modeling achieves uniform stress control, filling the gap in existing HBM bonding interface stress dynamic calculation and compensation modeling. The modeling process focuses on the dynamic control of post-bonding stress, representing a completely new modeling direction, distinct from existing modeling approaches and technical directions.

[0041] The dynamic calculation and compensation of bonding interface stress achieves real-time, high-precision acquisition of stress data through grating strain sensors, which can accurately capture the dynamic change process of stress compared with traditional static stress detection. Finite element iteration realizes three-dimensional stress calculation of the bonding interface. Compared with two-dimensional stress detection, the stress distribution identification is more comprehensive, and the positioning accuracy of stress concentration areas is improved by more than 90%. Differentiated local compensation strategy implements precise control for stress concentration areas. Compared with the overall stress control mode, the stress compensation efficiency is improved by more than 80%, and secondary stress disturbance in non-concentration areas is avoided. Iterative compensation and verification mechanism ensures that stress uniformity meets the process threshold. Compared with the non-iterative mode, the stress control qualification rate is improved by more than 95%. It will be linked with mass production equipment to realize automated control. Compared with manual control, the ability to adapt to mass production is greatly improved, and the error of manual operation is avoided.

[0042] Existing technologies only perform a single static stress test after bonding, lacking real-time dynamic stress calculation and differentiated compensation. This fails to accurately capture three-dimensional stress distribution and stress concentration areas, and also hinders effective stress control, resulting in severe stress concentration in the bonded wafers. Consequently, interface cracking easily occurs during subsequent packaging and testing, making it difficult to achieve yields suitable for mass production. This embodiment, through innovation and model optimization, achieves real-time calculation, accurate identification, and dynamic compensation of bonding interface stress, significantly improving interface stress uniformity. It completely solves the pain points of existing technologies and has no overlap with their technical direction or modeling approach. It effectively adapts to the mass production process requirements of HBM1, improving mass production yield.

[0043] Example 4: Stress aging evolution prediction and active control (adapted to automotive-grade HBM4 bonding) Implementation steps Step 1: Stress aging evolution data acquisition: Select automotive-grade HBM4 bonded wafers, simulate automotive-grade operating conditions (temperature -40℃~125℃ cycle, humidity 85%, operating current 1A), and collect stress change data within 1000 hours, including stress relaxation rate, stress redistribution characteristics, etc., to construct a stress aging evolution database.

[0044] Step 2: Machine learning prediction model training: Using stress aging evolution prediction and active control, with operating parameters (temperature, humidity, current) as input and stress change trend as output, the stress aging evolution prediction model is trained using LSTM machine learning, with the training set accounting for 80% and the validation set accounting for 20%, and the model prediction accuracy ≥90%.

[0045] Step 3: Time-bound trend prediction and proactive control strategy formulation: The automotive-grade full life cycle operating parameters (10 years, temperature -40℃~125℃ cycle, humidity 85%) are input into the trained prediction model to predict that HBM4 bonded wafers will have a stress relaxation rate ≥20% and secondary stress concentration at the edges in the 5th year; Based on the prediction results, a proactive control strategy is formulated: add a low-temperature annealing treatment (100℃, 120min) during the packaging process, and use a nano-coating to locally reinforce the wafer edges.

[0046] Step 4: Implementation and Effect Verification of Active Control: The established active control strategy is implemented on the HBM4 bonded wafer. The automotive-grade full life cycle conditions are simulated again, stress change data is collected, and the control effect is verified to ensure that the stress relaxation rate is ≤10% and there is no secondary stress concentration throughout the entire life cycle.

[0047] Step 5: Automotive-grade reliability verification: The regulated HBM4 product is subjected to automotive-grade reliability testing (1000 cycles of high and low temperatures and 1000 hours of damp heat aging). After the test, the electrical performance and interface bonding status of the product are checked to ensure that the electrical performance degradation rate is ≤5% and that there is no cracking or delamination at the interface.

[0048] Modeling Innovation Principles Abandoning the traditional extensive modeling approach of "no prediction and no active control," this paper constructs an integrated closed-loop model encompassing "data acquisition, database construction, model training, trend prediction, strategy formulation, active control, and effect verification." It uses automotive-grade operating parameters and stress aging variation characteristics as core inputs, overcoming the limitations of stress aging failure. The aging evolution database modeling achieves standardized storage of stress variation data; LSTM machine learning modeling enables accurate prediction of stress aging trends; active control strategy modeling allows for early intervention based on prediction results; and automotive-grade verification modeling achieves full lifecycle verification of the control effect, filling the gap in existing HBM bond stress aging evolution prediction and control modeling. The modeling process focuses on the long-term stress stability of automotive-grade HBM, representing a completely different modeling approach and technical direction from existing technologies, thus paving the way for a new modeling direction.

[0049] Stress aging evolution prediction and proactive regulation: By constructing a large-scale stress aging evolution database, sufficient training data is provided for machine learning models. Compared with training on small sample data, the model prediction accuracy is improved to over 90%, enabling accurate prediction of stress change trends throughout the entire life cycle. LSTM machine learning can capture the nonlinear characteristics of stress aging changes, improving prediction accuracy by over 80% compared to traditional linear prediction models and avoiding trend misjudgment. Based on the prediction results, proactive regulation strategies achieve "early intervention" in stress aging problems. Compared to the traditional "post-event remediation" mode, it can suppress stress relaxation and secondary concentration at the root, improving the long-term operational reliability of HBM products by over 90%. Dedicated regulation strategies are formulated for the harsh operating conditions of automotive-grade products. Compared to the general regulation mode, the automotive-grade reliability adaptability of the products is improved by over 85%, meeting the full life cycle operational requirements of automotive-grade products.

[0050] Existing technologies lack stress aging evolution prediction and proactive control, making it impossible to predict stress change trends in HBM products during long-term operation or to implement early intervention. This leads to failures in automotive-grade HBM products under harsh conditions such as high and low temperature cycling and long-term operation due to stress aging changes, resulting in interface cracking, electrical performance degradation, and other issues that fail to meet automotive-grade reliability requirements. This embodiment, through innovation and model optimization, achieves accurate prediction and proactive control of stress aging evolution, significantly improving the long-term stress stability and automotive-grade reliability of HBM4 products. It completely solves the pain points of existing technologies and has no overlap with existing technologies in terms of technical direction and modeling ideas. The innovation is clear and highly practical, effectively supporting the application of HBM products in the automotive electronics field.

[0051] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for implementing a high-reliability HBM bonding process, characterized in that, Includes the following steps: S1: Precise preprocessing of bonding surfaces: Collect micro-morphology data of wafer bonding surfaces to build a feature library. Through precise detection and repair of bonding surface micro-defects and uniform control of bonding surface surface energy, the precise identification and graded repair of micro-defects and the uniform optimization of bonding surface surface energy are achieved, resulting in a high-cleanliness uniform bonding surface. S2: High-precision bonding of heterogeneous wafers. A wafer alignment and bonding linkage system is built. Through submicron-level alignment and deviation compensation of heterogeneous wafers and dynamic matching of bonding pressure and temperature, submicron-level precise alignment of wafers and dynamic adaptation of bonding process parameters are achieved, thus completing high-precision bonding of heterogeneous wafers. S3: Dynamic stress control after bonding. Real-time acquisition of stress distribution data of bonded wafers. Through dynamic calculation and compensation of bonding interface stress, prediction and active control of stress aging evolution, accurate calculation and dynamic compensation of interface stress, early prediction and active intervention of stress aging changes are achieved, ensuring long-term stress stability of bonding interface. In step S2, the submicron-level alignment and deviation compensation of the heterogeneous wafer includes an alignment deviation calculation formula, which is as follows: The constraints are , To account for alignment deviations, For planar coordinate alignment deviation, To compensate for rotational alignment deviation, The alignment deviation threshold is set according to the HBM product specifications; high-end HBM products... Standard HBM products .

2. The method according to claim 1, characterized in that, The precise detection and repair of micro-defects on the bonding surface in step S1 includes the following sub-steps: extracting the core feature parameters of the micro-morphology of the bonding surface, establishing a micro-defect grading standard, achieving precise detection and location of micro-defects through feature matching, adopting a differentiated repair strategy according to the defect level, and iteratively verifying the repair effect until the defect meets the process threshold.

3. The method according to claim 1, characterized in that, The surface energy homogenization control of the bonding surface in step S1 includes the following sub-steps: detecting the surface energy values ​​of different regions of the bonding surface, establishing a surface energy distribution model, adopting a plasma modification zonal control strategy, setting modification parameters for different regions, and iteratively controlling until the overall surface energy deviation of the bonding surface is ≤5mJ / m. 2 .

4. The method according to claim 1, characterized in that, The bonding pressure-temperature dynamic matching in step S2 includes the following sub-steps: establishing a physical property model of the HBM bonding material, extracting the correlation parameters between bonding temperature, pressure and interfacial bonding strength, and dynamically matching the pressure-temperature curve according to the wafer thickness and warpage to achieve real-time adaptive adjustment of parameters during the bonding process.

5. The method according to claim 1, characterized in that, The dynamic calculation and compensation of bonding interface stress in step S3 includes the following sub-steps: constructing a stress distribution model based on grating strain sensing data, calculating the three-dimensional stress distribution of the bonding interface through finite element iteration, and adopting a differentiated compensation strategy of local hot pressing / cold treatment according to the stress concentration area to achieve uniform control of stress.

6. The method according to claim 1, characterized in that, The stress aging evolution prediction and active control in step S3 includes the following sub-steps: establishing a stress aging evolution database, using machine learning to train an evolution prediction model, predicting the stress aging change trend under different working conditions, and implementing active control strategies in advance based on the prediction results to suppress stress relaxation and concentration.

7. The method according to any one of claims 1-6, characterized in that, The process threshold for micro-defects on the bonding surface is: the number of defects with a size ≥ 0.2 μm ≤ 1 / cm. 2 The number of defects with a size <0.2μm is ≤5 / cm. 2 No micro-scratches or pits with a depth of ≥0.1μm.

8. The method according to any one of claims 1-6, characterized in that, In the dynamic matching of bonding pressure and temperature, the pressure control accuracy is ≤0.01MPa and the temperature control accuracy is ≤0.5℃, achieving a void-free bonding interface.

9. The method according to any one of claims 1-6, characterized in that, The method can be applied to the bonding process of 2D / 3D stacked HBM1-HBM4 and subsequent upgraded products, and is suitable for heterogeneous wafer bonding scenarios such as silicon-silicon and silicon-glass.

10. A high-reliability HBM bonding process implementation system, characterized in that, It includes a bonding surface precision preprocessing module, a heterogeneous wafer high-precision bonding module, a post-bonding stress dynamic control module, and a central control unit. The central control unit is communicatively connected to the three functional modules and executes the method described in any one of claims 1-9 to achieve intelligent and efficient control of the entire HBM bonding process.

Images