HBM variable temperature bonding intelligent control method based on interface synergistic regulation

Abstract

The application belongs to the technical field of semiconductor material preparation, and is an HBM variable-temperature bonding intelligent control method based on interface energy-dislocation evolution collaborative regulation. The application avoids the technical route of simple combination of traditional heating, pressure and time, and constructs an HBM variable-temperature bonding whole-process intelligent collaborative system from the aspects of temperature field dynamic matching, interface stress global decoupling, bonding parameter online identification, defect real-time prediction, multi-physical field collaborative optimization and whole-process closed-loop self-correction. The application completely solves the technical problems of interface heat transfer mismatching, thermal stress concentration cracking, bonding parameter time-varying drift and defect difficult to inhibit in advance in the traditional HBM variable-temperature bonding, and significantly improves the bonding strength uniformity, reduces the interface cavity and crack, improves the coplanarity of the chip and the substrate, and prolongs the bonding interface life.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor materials technology, and in particular to a method for intelligent control of HBM temperature bonding based on interface synergistic regulation. Background Technology

[0002] HBM (High Bandwidth Memory), as a core storage architecture for AI computing, data centers, and high-end GPUs, relies on multi-layer vertical chip stacking and micro-bump bonding to achieve ultra-high bandwidth interconnects. Temperature-controlled bonding is the most critical process determining interface bonding strength, defect rate, and long-term reliability. In current advanced packaging mass production, traditional HBM temperature-controlled bonding processes generally suffer from three specific, non-macroscopic technical problems that cannot be solved by conventional methods: 1. Interface energy mismatch leads to poor solder wetting, high local void rate, and inconsistent bonding strength. There are significant differences in the surface energy of the passivation layer, bump metal layer, and substrate pads on the chip surface. Traditional isothermal / uniform heating cannot achieve dynamic matching of interface energy, resulting in fluctuations in solder wetting angle, local non-wetting, void accumulation, and extremely poor consistency of interface bonding strength.

[0003] 2. During temperature changes, dislocations multiply and accumulate rapidly, leading to interface embrittlement, microcracks, and early failure. The mismatch between the coefficient of thermal expansion (CTE) of the chip and the substrate generates huge cyclic stress during the temperature change stage, resulting in an exponential increase in the interface dislocation density. Dislocation entanglement and accumulation directly induce interface microcracks and brittle fracture, significantly reducing drop reliability and high-temperature storage life.

[0004] 3. The micro-area contact impedance drifts significantly with temperature. Bonding parameter mismatch leads to poor solder joints, open circuits, and electrical performance failure. The interface contact impedance drifts in real time with temperature, pressure, and oxidation level. Traditional open-loop control cannot identify and compensate for this online, resulting in unstable bonding energy input, local poor solder joints, local over-soldering, abnormal growth of intermetallic compounds (IMC), and a significantly increased risk of electrical performance failure.

[0005] The three major problems mentioned above all stem from the fact that the time-varying evolution of interface energy, dislocations, and impedance cannot be modeled and controlled. Traditional control methods that rely solely on temperature / pressure / time curves are completely inadequate and have become the core bottleneck for the mass production of high-end HBM packaging. Summary of the Invention

[0006] This invention provides an intelligent control method for HBM variable-temperature bonding based on the coordinated regulation of interface energy and dislocation evolution, comprising: S1: collecting time-series data of multiple physical quantities in the micro-region during the entire HBM variable-temperature bonding process, including interface energy, micro-region temperature, contact pressure, dislocation-related strain signals, interface contact impedance, warpage morphology, heating output, and environmental disturbance data; S2: employing an interface energy feature adaptive fusion algorithm (IEF-Fusion) to denoise, spatiotemporally align, and enhance features of multi-source weak micro-region signals, extracting interface energy, dislocation, impedance, and mechanically sensitive feature sets; S3: constructing a unified model of interface energy evolution, dislocation proliferation, and thermo-mechanical coupling based on the interface energy-dislocation coupling consistency network (IEDC-Net), realizing the interface state... S4: Dynamically optimize the heating rate, phase transition range, insulation node, and cooling slope using the dislocation suppression type variable temperature planning algorithm (DS-Temp) to achieve in-situ dislocation suppression; S5: Use the interface energy balance control algorithm (IE-Balance) to compensate the micro-region interface energy in real time, eliminating wetting unevenness and voids caused by interface energy differences; S6: Use the bonding defect self-learning suppression algorithm (Defect-Learn) to infer and predict dislocation, crack, void, and cold weld defects, achieving defect closed-loop suppression; S7: Output zone temperature, local pressure, displacement compensation, and atmosphere control collaborative commands through a multi-variable global collaborative algorithm to achieve intelligent self-learning correction throughout the entire process.

[0007] Furthermore, the modeling and solution formulas for the interface energy feature adaptive fusion algorithm (IEF-Fusion) described in S2 are as follows: in, For the first Road micro-area signal, For the interface attention function, For adaptive weights, An interface energy feature extraction operator is proposed; the interface energy and weak dislocation signal are enhanced through adaptive gradient ascent iterative solution.

[0008] Furthermore, the unified loss function of the Interface Energy-Dislocation Coupling Consistency Network (IEDC-Net) described in S3 is: in, For interface energy evolution loss, For dislocation multiplication constraint loss, For thermo-mechanical coupling loss, The loss is due to the coupling consistency of multiple physics fields; the network strictly satisfies the interface physical mechanism to achieve high-precision prediction for small samples.

[0009] Furthermore, the dynamic trajectory formula of the dislocation suppression type variable temperature programming algorithm (DS-Temp) described in S4 is as follows: in, For real-time interface temperature, The coefficient of temperature rise, For dislocation suppression weights, For pressure coupling factor, To provide an interface energy correction factor; to achieve full matching between the temperature variation path and dislocation suppression and interface energy balance.

[0010] Furthermore, the interface energy balance control algorithm (IE-Balance) described in S5 adopts micro-region model predictive control, using interface energy deviation, dislocation density, and contact impedance as state variables, and zone heating power and local pressure compensation as control variables, to solve for the optimal balance strategy in a rolling manner.

[0011] Furthermore, the bonding defect self-learning suppression algorithm (Defect-Learn) described in S6 uses graph attention reasoning to construct the interface energy-dislocation-temperature-defect association path, thereby realizing defect source localization and closed-loop suppression.

[0012] Furthermore, the multivariate global collaborative algorithm described in S7 is based on multi-objective consistency optimization, which unifies and coordinates the zoned thermal field, local pressure, micro-displacement, and atmospheric environment to achieve synchronized, conflict-free, and overshoot-free actions across the entire field.

[0013] Furthermore, it also includes a full-process self-learning correction mechanism, which automatically updates the model and control parameters based on real-time interface energy, dislocation status, impedance distribution, and defect indicators to improve long-term batch consistency.

[0014] Furthermore, a HBM variable-temperature bonding intelligent control system based on interface energy-dislocation evolution coordinated regulation is proposed, comprising: a micro-region multi-physical quantity sensing unit, an IEF-Fusion interface energy fusion unit, an IEDC-Net coupling modeling unit, a DS-Temp dislocation suppression temperature control unit, an IE-Balance interface energy balancing unit, a Defect-Learn defect self-learning unit, a global coordinated control unit, and a self-learning correction unit; the system is used to execute the method described in any one of claims 1-8.

[0015] Furthermore, the global collaborative control unit outputs optimal collaborative instructions for zoned heat-local force-micro displacement-atmosphere based on the interface energy state, dislocation evolution trend, coupling field distribution, and defect risk.

[0016] Beneficial effects: 1. The interface can dynamically match and balance, resulting in uniform solder wetting, significantly reduced void ratio, and greatly improved interface bonding strength.

[0017] 2. In-situ suppression of dislocation multiplication significantly reduces the risk of interface embrittlement, microcracks, and early failure.

[0018] 3. The multiphysics fields are fully coupled and consistent, and the interface evolution is predictable, controllable, and repeatable.

[0019] 4. Defects can be reasoned out and closed-loop suppressed, significantly improving bonding yield and long-term reliability.

[0020] 5. Global collaborative actions are synchronized, parameter drift is automatically compensated, and batch production consistency is extremely strong.

[0021] It features fully intelligent self-learning, eliminating the need for manual parameter tuning, and is fully compatible with high-end HBM3 / HBM4 packaging mass production. Attached Figure Description

[0022] Figure 1 Flowchart. Detailed Implementation

[0023] This embodiment addresses core issues in HBM (High Bandwidth Memory) temperature bonding processes, including uneven interface energy distribution, uncontrolled dislocation multiplication leading to numerous bonding defects, and insufficient reliability. Focusing on the core principle of coordinated regulation of interface energy and dislocation evolution, it details the specific operational procedures, logic, and efficiency-enhancing mechanisms of four key technical embodiments, comprehensively demonstrating the feasibility and superiority of this intelligent control system. The embodiments are closely interconnected and progressively build upon each other, forming a complete bonding control chain. Requiring no complex data support or formula derivation, it emphasizes practicality and technological innovation, ensuring a dual improvement in bonding quality and efficiency. Applicable to various temperature bonding production scenarios for HBM products, it possesses broad practicality and promotional value.

[0024] Example 1: Adaptive Fusion of Interface Energy Features Based on IEF-Fusion The core objective of this embodiment is to address the critical challenge of processing multi-source micro-region timing signals during HBM temperature-dependent bonding. Breaking away from the limitations of traditional signal processing methods, it utilizes the IEF-Fusion (Interface Energy Feature Fusion) algorithm to achieve efficient integration and weak feature enhancement of multi-source signals. This provides high-quality, highly reliable input signals for subsequent coupling modeling, laying the foundation for the stable operation of the entire intelligent control system. This embodiment requires no complex hardware upgrades and can be directly integrated into the signal processing module of existing bonding equipment. It is adaptable to various high-temperature and high-pressure bonding conditions, offering convenient operation and strong versatility.

[0025] 1.1 Implementation Core Ideas A smart control method for HBM variable-temperature bonding based on the coordinated regulation of interface energy and dislocation evolution includes: S1: collecting time-series data of multiple physical quantities in the micro-region during the entire HBM variable-temperature bonding process, including interface energy, micro-region temperature, contact pressure, dislocation-related strain signals, interface contact impedance, warpage morphology, heating output, and environmental disturbance data; S2: using the Interface Energy Feature Adaptive Fusion (IEF-Fusion) algorithm to denoise, align in time and space, and enhance features of multi-source weak micro-region signals, extracting interface energy, dislocation, impedance, and mechanically sensitive feature sets; S3: constructing a unified model of interface energy evolution, dislocation multiplication, and thermo-mechanical coupling based on the Interface Energy-Dislocation Coupling Consistency Network (IEDC-Net) to achieve real-time high-precision solution of interface state; S4: implementing dislocation suppression-type variable-temperature planning. The algorithm (DS-Temp) dynamically optimizes the heating rate, phase transition range, insulation nodes, and cooling slope to achieve in-situ dislocation suppression; S5: The interface energy balance control algorithm (IE-Balance) is used to compensate the micro-region interface energy in real time, eliminating wetting unevenness and voids caused by interface energy differences; S6: The bonding defect self-learning suppression algorithm (Defect-Learn) is used to infer and predict dislocation, crack, void, and cold weld defects to achieve defect closed-loop suppression; S7: The multivariate global collaborative algorithm outputs zone temperature, local pressure, displacement compensation, and atmosphere control collaborative commands to achieve intelligent self-learning correction throughout the process; The modeling and solution formula of the interface energy feature adaptive fusion algorithm (IEF-Fusion) mentioned in S2 is: where, For the first Road micro-area signal, For the interface attention function, For adaptive weights, This paper proposes an interface energy feature extraction operator. Through adaptive gradient ascent iterative solution, it enhances weak interface energy and dislocation signals. During HBM temperature bonding, multiple micro-region time-series signals need to be acquired simultaneously. These signals come from different detection modules, resulting in signal clutter, difficulty in identifying weak features, and low signal-to-noise ratios. In particular, weak features related to interface energy and dislocations are easily masked by noise under high-temperature perturbation, leading to insufficient modeling accuracy and poor control performance. This embodiment employs an adaptive fusion algorithm through a core process of signal acquisition, signal preprocessing, feature fusion, weight optimization, and feature output. It dynamically focuses on the core weak features related to interface energy and dislocations, eliminates invalid noise interference, and achieves collaborative fusion of multi-source signals, thereby improving signal representation capabilities.

[0026] 1.2 Detailed Implementation Steps Step 1: Multi-source micro-area time-series signal acquisition The signal acquisition system of the bonding equipment is activated to synchronously acquire various micro-region timing signals throughout the entire HBM temperature-controlled bonding cycle, covering the four core stages of bonding: preheating, heating, holding, and cooling, ensuring the integrity and real-time performance of signal acquisition. The acquired signal types mainly include seven categories: interface energy signals, micro-region temperature signals, contact pressure signals, dislocation strain signals, interface impedance signals, warpage signals, and heating power signals. Among these, the interface energy signal primarily reflects the tightness of the bonding interface, focusing on capturing the energy state changes at the interface between the solder and the chip / substrate after melting; the dislocation strain signal is used to monitor the generation and evolution of dislocations between the chip and substrate under thermal stress; other signals are used to assist in characterizing the temperature, stress, and deformation states during the bonding process, providing a comprehensive data source for subsequent feature fusion. During acquisition, it is crucial to ensure synchronous acquisition of all types of signals to avoid feature deviations caused by signal delays, while also ensuring the continuity of signal acquisition without interruptions or missing signals.

[0027] Step 2: Signal preprocessing (spatiotemporal alignment, outlier removal, standardization) Since the acquired signals come from different detection devices, there are problems such as asynchronous acquisition time, abnormal points in the signals, and inconsistent signal dimensions. If they are directly used for feature fusion, they will seriously affect the fusion effect and subsequent modeling accuracy. Therefore, the acquired signals need to be systematically preprocessed to ensure the effectiveness and consistency of the signals.

[0028] First, spatiotemporal alignment is performed. To address the issue of inconsistent acquisition frequencies of different signals, the signal with the highest acquisition frequency is selected as the benchmark, and other signals with lower frequencies are synchronously adjusted to ensure that the acquisition times of all signals are completely consistent, eliminating the impact of time deviation. At the same time, the signals from different acquisition points in space are uniformly aligned to clarify the positional relationship of each acquisition point, providing a foundation for subsequent spatial fusion of features.

[0029] Secondly, anomaly removal is required. External electromagnetic interference and equipment operation fluctuations during the bonding process can cause abnormal changes in some signals. These anomalies can interfere with the overall characteristics of the signal. Therefore, it is necessary to identify and remove anomalies using appropriate methods, and to properly repair the missing parts after removal to ensure the continuity and stability of the signal and avoid anomalies interfering with subsequent processing.

[0030] Finally, standardization is performed. Since the various signals have different dimensions and large differences in their numerical ranges, they cannot be directly fused. Therefore, standardization is required to adjust the numerical range of all signals to the same interval, eliminate the influence of the difference in dimensions, and make the various signals fusionable, laying the foundation for subsequent weight calculation and feature fusion.

[0031] Step 3: Construct an adaptive fusion objective function for interface energy features The core of feature fusion is to construct a reasonable fusion objective function to ensure that the fused features can highlight the weakly sensitive features related to interface energy and dislocations, while also taking into account the overall signal quality. The adaptive fusion objective function for interface energy features constructed in this embodiment aims to improve the overall representational capability of the fused features, focusing on strengthening the weights of interface energy and dislocation strain signals, while also considering the contributions of other auxiliary signals to avoid a single signal dominating the fusion result.

[0032] The core of the objective function design is the introduction of a dedicated attention mechanism for interface energy. This mechanism dynamically identifies weak features related to interface energy and dislocations, and dynamically adjusts the fusion weights based on the actual characteristics of the signal, making the fusion process more targeted. Through this objective function, adaptive fusion of multi-source signals can be achieved, preserving effective information from various signals while focusing on core weak features, thus solving the problem of weak features being masked in traditional fusion methods.

[0033] Step 4: Iteratively update the weights using the adaptive gradient ascent method. To achieve the optimal solution of the fusion objective function, an adaptive gradient ascent method is used to iteratively update the fusion weights of multi-source signals. The core logic is to continuously adjust the weight allocation of various signals to optimize the representational ability of the fusion features, while automatically strengthening the weights of weakly sensitive features that are correlated with dislocations.

[0034] During the iteration process, the fusion weights are first initially allocated to ensure a uniform distribution of initial weights and avoid the impact of initial deviations on the iteration results. Then, based on the feedback from the objective function, the adjustment direction and magnitude of the weights of various signals are calculated, and the weight values ​​are dynamically updated. After each weight update, the weights need to be normalized to ensure that the sum of the weights of various signals is 1, which conforms to the fusion logic. Finally, through multiple iterations, the iteration stops and the optimal fusion weights are output when the objective function reaches a stable state.

[0035] Through this iterative process, the optimal weights will automatically tilt towards interface energy and dislocation strain signals, significantly increasing the proportion of these two types of signals in the fused features. This achieves automatic enhancement of weakly sensitive features related to interface energy and dislocations, effectively solving the problem of weak features being difficult to identify in traditional signal processing.

[0036] Step 5: Output high signal-to-noise ratio fused features and input them into the IEDC-Net coupled model. The optimal fusion weights obtained through iteration are applied to the preprocessed multi-source signal to calculate a high signal-to-noise ratio fusion feature sequence for the entire cycle. The fusion features can accurately capture the subtle changes in interface energy and dislocation evolution, while eliminating invalid noise interference, thus significantly improving signal quality.

[0037] After the fused features are output, they are input to the subsequent IEDC-Net interface-dislocation coupling modeling unit via a dedicated data interface, serving as the core input data for modeling and providing reliable support for subsequent coupled modeling and state solving. This ensures the real-time output of the fused features and seamless integration with the subsequent modeling process, guaranteeing the continuous operation of the entire intelligent control system.

[0038] 1.3 Efficiency Enhancement Principle Traditional signal processing methods have significant limitations. They cannot effectively identify weakly sensitive features related to interface energy and dislocations. Especially in the high-temperature disturbance environment of HBM temperature bonding, the signal-to-noise ratio is extremely low, and weak features are masked by a large amount of noise, resulting in insufficient accuracy of subsequent modeling and failing to provide a reliable basis for control strategies.

[0039] The IEF-Fusion algorithm in this embodiment achieves intelligent fusion of multi-source signals by combining a dedicated attention mechanism for interfaces with adaptive weight iteration. The attention mechanism can dynamically focus on weak features related to interface integration and dislocation evolution, automatically identifying and amplifying the signal contribution of these weak features; the adaptive weight iteration can dynamically adjust the weight allocation of various signals according to the actual characteristics of the signals, further highlighting core weak features and suppressing interference from invalid noise.

[0040] This algorithm significantly enhances the signal representation capability, enabling the fused features to accurately reflect the true state of interface energy and dislocation evolution. It effectively addresses the pain points of traditional signal processing, providing reliable input support for subsequent high-precision coupling modeling and laying the foundation for precise control of the entire intelligent control system. Furthermore, the algorithm requires no complex hardware support and can be directly integrated into existing equipment, reducing application costs and enhancing the practicality and promotional value of the technology.

[0041] Example 2: Interface Energy-Dislocation Coupling Consistency Modeling Based on IEDC-Net This embodiment aims to address the core issues of traditional modeling methods, such as the inability to simultaneously characterize interface energy and dislocation evolution, and the resulting state inconsistencies caused by independent calculations of multiple fields. By constructing IEDC-Net (Interface Energy-Dislocation Coupling Consistency Network), it achieves unified modeling and synchronous solution of multiple field states, including interface energy, dislocations, thermal, and mechanical fields. This provides consistent and reliable prediction data for subsequent dislocation suppression, interface equilibrium, and defect suppression, improving the modeling accuracy and control reliability of the entire control system. This embodiment features a simple modeling logic, requiring no complex numerical calculations, emphasizing the consistency of multi-field coupling, and is adaptable to the modeling needs of various HBM temperature-dependent bonding systems.

[0042] 2.1 Implementation Core Ideas The HBM variable-temperature bonding process is a complex multi-physics coupling process. There are close interactions between interface energy evolution, dislocation multiplication, heat conduction, and force balance. Traditional modeling methods often adopt a multi-field independent modeling approach, calculating each physical field separately. This leads to contradictions in the state results of different physical fields, which cannot accurately reflect the true evolution state of the bonding interface, and thus affect the formulation and execution of subsequent control strategies.

[0043] The core idea of ​​this embodiment is to construct a unified interface energy-dislocation coupling consistency network, which incorporates the physical constraints of interface energy evolution, dislocation propagation, heat conduction, and force balance into the same network framework. Through unified loss function constraints, the synchronous solution of multiple field states is achieved, ensuring that the state results of interface energy, dislocation density, temperature field, and stress field are completely self-consistent, eliminating the state contradictions of traditional modeling methods, and providing a unified and reliable interface evolution basis for subsequent real-time control.

[0044] 2.2 Detailed Implementation Steps Step 1: Establish the core equations for various physical processes Based on the physical process of HBM temperature-dependent bonding, we established equations for interface energy evolution, dislocation multiplication kinetics, heat conduction, and force balance, clarifying the core laws of various physical processes and providing a foundation for subsequent coupled modeling.

[0045] The interface energy evolution equation mainly describes the dynamic change of interface energy during the bonding process. It combines factors such as bonding temperature, contact pressure, and interface state to reflect the evolution trend of interface energy with the bonding process. The dislocation multiplication kinetic equation describes the generation, multiplication, and evolution of dislocations. It combines factors such as thermal stress and interface energy changes to capture the dynamic changes in dislocation density. The heat conduction equation describes the heat transfer and distribution during the bonding process, reflecting the dynamic changes in temperature in different regions. The force balance equation describes the force balance state during the bonding process. It combines factors such as contact pressure, thermal stress, and deformation to ensure the stability of the force state.

[0046] The establishment of various equations does not require complex numerical derivation. The focus is on highlighting the core laws of the physical process, ensuring that the equations can accurately reflect the evolution characteristics of various physical quantities, and providing reliable theoretical support for the subsequent construction of coupled networks.

[0047] Step 2: Construct an interface-dislocation coupling consistency network and a unified loss function Based on the core equations of various physical processes, the IEDC-Net interface energy-dislocation coupling consistency network is constructed. This network integrates four physical processes—interface energy evolution, dislocation multiplication, heat conduction, and force equilibrium—into a single network framework, achieving unified modeling of multi-field coupling. The core design idea of ​​the network is to break through the limitations of traditional independent multi-field modeling and achieve collaborative solving of various physical quantities through optimized network structure design.

[0048] Simultaneously, all physical constraints are incorporated into a unified loss function. Through the constraint effect of the loss function, it is ensured that the interface energy, dislocation density, temperature field, stress field, and other state results output by the network satisfy the constraints of various physical equations, achieving consistency across multiple fields. The design of the loss function emphasizes the self-consistency of multiple fields, avoiding contradictions between different physical field states and ensuring the accuracy and reliability of network modeling.

[0049] Step 3: Input real-time micro-area data to train the network, and output a unified state. The high signal-to-noise ratio fusion features output from Example 1 are used as input data for the IEDC-Net network. Simultaneously, real-time micro-region state data from the bonding process is incorporated to train the network. During training, network parameters are continuously adjusted to ensure that the network's output state matches the actual bonding state, thereby guaranteeing the network's modeling accuracy.

[0050] After network training, it can receive input micro-region data in real time and synchronously output unified state results for interface energy, dislocation density, temperature field, and stress field, achieving millisecond-level synchronous solution of multiple field states. The output state results can accurately reflect the true evolution state of the bonding interface, and the various physical quantities are mutually coordinated and completely self-consistent, avoiding the problem of contradictory multiple field states in traditional modeling methods.

[0051] Step 4: Provide consistent prediction basis for subsequent control modules The unified state results output by the network are synchronously transmitted to the subsequent dislocation suppression and temperature control unit, interface energy equalization unit, and defect suppression unit via the data interface, serving as the core basis for these modules to formulate control strategies. Because the output state results are consistent and reliable, it ensures that the strategy formulation of each subsequent control module is based on a unified interface evolution state, avoiding control chaos caused by inconsistent state bases.

[0052] Meanwhile, the network can update the status results in real time. As the bonding process progresses, it dynamically outputs the latest interface evolution status, ensuring that subsequent control strategies can adapt to changes in the interface status in a timely manner, thereby improving the real-time performance and accuracy of control.

[0053] 2.3 Efficiency Enhancement Principle The core drawback of traditional modeling methods is that they cannot simultaneously characterize the synergistic relationship between interface energy and dislocation evolution. The use of independent calculations for multiple fields leads to contradictions in the state results of different physical fields. For example, the calculated results of the temperature field do not match the calculated results of the stress field, and the evolution of dislocation density is not synchronized with the changes in interface energy. These contradictions can cause deviations in the formulation of subsequent control strategies and affect the bonding quality.

[0054] The IEDC-Net coupling consistency model in this embodiment integrates multiple physical processes such as interface energy, dislocations, heat, and force through a unified network framework and loss function constraints, achieving simultaneous solution of multiple field states. The network can automatically coordinate the mutual influence between various physical quantities, ensuring that the output interface energy, dislocation density, temperature field, and stress field states are completely self-consistent, eliminating state contradictions in traditional modeling methods.

[0055] Meanwhile, the model achieves millisecond-level synchronous solving, enabling real-time output of the interface evolution state and providing timely and reliable data for subsequent real-time control. Compared to traditional modeling methods, this model significantly improves modeling accuracy and response speed, accurately capturing subtle changes in interface evolution. It provides a unified predictive basis for subsequent dislocation suppression, interface equalization, and defect suppression, ensuring the control accuracy and reliability of the entire control system and effectively reducing bonding defects caused by modeling deviations.

[0056] Example 3: Dislocation Suppression-Based Variable Temperature Programming The core objective of this embodiment is to address the problems of increased dislocation multiplication and interface embrittlement caused by traditional fixed heating curves. By employing the DS-Temp (Dislocation Suppression-Type Variable Temperature Planning) algorithm, with dislocation suppression as the core objective, it dynamically plans the temperature path of the bonding process, blocking dislocation multiplication channels at the source of the thermal history. Simultaneously, it maintains interfacial energy equilibrium and stress smoothing, improving the toughness and fatigue resistance of the bonding interface and ensuring the stability of bonding quality. This embodiment is simple to operate and can be directly adapted to the temperature control module of existing bonding equipment without complex hardware modifications.

[0057] 3.1 Implementation Core Ideas Dislocation multiplication is one of the core factors affecting the quality of HBM temperature-dependent bonding. Traditional bonding processes use fixed heating curves and cannot dynamically adjust the temperature strategy according to the dislocation evolution state. At high temperatures, dislocations are prone to multiply in large quantities, which can lead to defects such as interface embrittlement and cracks, thus reducing bonding reliability.

[0058] The core idea of ​​this embodiment is to dynamically calculate the optimal temperature trajectory for dislocation suppression based on the real-time dislocation density, interface energy, stress, and temperature distribution output by IEDC-Net, using the DS-Temp algorithm. The temperature strategy during the heating, holding, and cooling stages is adjusted in real time according to the dislocation multiplication trend. While suppressing dislocation initiation and multiplication, it ensures balanced interface energy and smooth stress, avoiding interface damage caused by abrupt stress changes, and achieving a dual improvement in dislocation suppression and bonding quality.

[0059] 3.2 Detailed Implementation Steps Step 1: Obtain real-time interface and multi-field status data The IEDC-Net coupling consistency model acquires real-time multi-field state data, including dislocation density, interface energy, stress, and temperature distribution, during the bonding process, ensuring data real-time performance and accuracy. This data forms the core basis for dynamically planning temperature-varying trajectories, accurately reflecting the current dislocation evolution state, interface energy distribution, and stress and temperature distribution at the bonding interface, providing reliable support for adjusting temperature strategies.

[0060] During the data acquisition process, it is necessary to ensure the smooth transmission of data to avoid untimely temperature strategy adjustments due to data delays. At the same time, a simple validity assessment should be performed on the acquired data to remove invalid data and ensure the accuracy of temperature planning.

[0061] Step 2: Dynamically calculate the optimal temperature-dependent trajectory for dislocation suppression Based on the acquired real-time state data, the optimal temperature-varying trajectory for dislocation suppression is dynamically calculated using the DS-Temp algorithm. The calculation of the temperature-varying trajectory takes dislocation suppression as the core objective, and combines the constraints of interface energy equilibrium and stress smoothing. It comprehensively considers key parameters such as heating rate, holding temperature, holding time, and cooling slope to formulate the optimal temperature path that adapts to the current dislocation evolution state.

[0062] During the calculation, the focus is on analyzing the evolution trend of dislocation density. If an upward trend in dislocation density is found, the temperature parameters are adjusted in a timely manner. By optimizing the heating rate and adjusting the holding temperature, the initiation and proliferation of dislocations are suppressed. At the same time, the balanced distribution of interface energy and the gradual change of stress are taken into account to avoid uneven distribution of interface energy or abrupt changes in stress due to temperature adjustment, thus ensuring the stability of the bonding interface.

[0063] Step 3: Adjust the temperature strategy based on the dislocation multiplication trend The calculated optimal temperature trajectory is transformed into a specific temperature control strategy, adjusting the temperature parameters in real time for the three stages of heating, holding, and cooling based on the dislocation multiplication trend. During the heating stage, the heating rate is adjusted according to the dislocation multiplication; if multiplication is rapid, the heating rate is appropriately reduced to prevent a rapid temperature increase that could lead to a large number of dislocations. During the holding stage, the holding temperature and holding time are adjusted based on the interface energy and dislocation density state to suppress dislocation multiplication while promoting a balanced distribution of interface energy. During the cooling stage, the cooling slope is optimized to avoid excessively rapid cooling that could cause abrupt stress changes, thereby reducing dislocation generation and the appearance of interface cracks.

[0064] The temperature strategy needs to be adjusted in real time to respond to changes in the dislocation evolution state, ensuring that the adjustment of each temperature parameter can specifically suppress dislocation multiplication, while ensuring the stability of interface energy and stress.

[0065] Step 4: Maintain interfacial energy balance and stress smoothness During temperature strategy adjustments, maintaining interfacial energy equilibrium and stress smoothness are always considered crucial constraints to prevent uneven interfacial energy distribution or abrupt stress changes caused by temperature adjustments. By monitoring interfacial energy and stress states in real time, if significant deviations in interfacial energy or abrupt stress changes are detected, temperature parameters are adjusted promptly. This is achieved by optimizing holding time and adjusting the cooling slope to compensate for interfacial energy differences, alleviate stress concentration, ensure that interfacial energy remains in equilibrium, and maintain smooth stress changes to prevent interfacial damage.

[0066] Step 5: Output dislocation suppression optimal temperature command The adjusted temperature control strategy is translated into specific temperature commands, which are then output to the temperature control module of the bonding equipment via a data interface to control the operation of the heating module, achieving precise temperature control during the bonding process. The output of temperature commands must be real-time and accurate, synchronized with the progress of the bonding process, ensuring that temperature control can adapt promptly to changes in dislocation evolution states and achieve the goal of dislocation suppression.

[0067] Meanwhile, the temperature control effect is fed back in real time, and the temperature command is further fine-tuned based on the feedback results to ensure that the dislocation suppression effect reaches the optimal level, while ensuring the stability of bonding quality.

[0068] 3.3 Efficiency Enhancement Principle Traditional HBM variable temperature bonding uses a fixed heating curve, which cannot dynamically adjust the temperature strategy according to the dislocation evolution state. In high-temperature environments, the fixed heating rate and holding temperature can easily lead to a large number of dislocations multiplying. The accumulation of dislocations will exacerbate interface embrittlement, reduce the toughness and fatigue resistance of the interface, and thus cause bonding defects such as cracks and voids, affecting the bonding reliability.

[0069] The DS-Temp algorithm in this embodiment, with dislocation suppression as its core objective, breaks through the limitations of traditional fixed temperature curves. By acquiring dislocation evolution state data in real time, it dynamically plans the optimal temperature-changing trajectory, blocking dislocation multiplication channels from the source of the thermal process. The algorithm can adjust heating, holding, and cooling strategies in real time according to the dislocation multiplication trend, specifically suppressing the initiation and multiplication of dislocations, and significantly reducing the interface dislocation density.

[0070] Meanwhile, during temperature adjustment, the algorithm balances interfacial energy equilibrium and stress smoothing, avoiding uneven interfacial energy distribution and abrupt stress changes caused by temperature adjustment. This effectively reduces interfacial damage and improves interfacial toughness and fatigue resistance. Compared to traditional fixed temperature profiles, this algorithm significantly reduces bonding defect rates, improves the stability and reliability of bonding quality, and extends the lifespan of HBM products, all without requiring additional hardware costs, making it highly practical and economical.

[0071] Example 4: Interface Balancing and Defect Suppression Based on IE-Balance and Defect-Learn This embodiment aims to address the problems of uneven interfacial energy distribution and difficulty in predicting and suppressing defects during HBM temperature bonding. Through the synergistic effect of the IE-Balance (interfacial energy balancing) algorithm and the Defect-Learn (defect self-learning) algorithm, it achieves precise control of interfacial energy balance and pre-prediction and in-situ suppression of bonding defects, further improving bonding quality, reducing defect generation, and ensuring the stability and reliability of the bonding process. This embodiment combines interfacial balancing and defect suppression to form a dual guarantee, adapting to various complex bonding conditions.

[0072] 4.1 Implementation Core Ideas During HBM temperature bonding, uneven distribution of interfacial energy can lead to uneven solder wetting, which in turn can cause defects such as voids and cold solder joints. At the same time, defects such as dislocations, cracks, and voids are hidden and sudden, and traditional methods are difficult to predict in advance. They can only be dealt with after they occur, which cannot effectively suppress defects.

[0073] The core idea of ​​this embodiment is to achieve real-time equilibrium control of interface energy through the IE-Balance algorithm, improve solder wetting uniformity, and reduce the causes of defects from the source. At the same time, the Defect-Learn algorithm is used to construct a defect inference path, realize the pre-judgment of various bonding defects, output in-situ suppression strategies in advance, and achieve active suppression of defects. This forms a dual guarantee of interface equilibrium control and pre-defect suppression, and comprehensively improves bonding quality.

[0074] 4.2 Detailed Implementation Steps Step 1: Real-time acquisition of interface and defect-related status data Real-time acquisition of state data such as interface energy deviation, dislocation density, impedance drift, and stress concentration during the bonding process. These data come from the IEDC-Net coupling consistency model and real-time detection module, which can accurately reflect the distribution of interface energy, dislocation evolution, interface contact state, and stress concentration, providing a reliable basis for interface energy balance control and defect prediction.

[0075] During data acquisition, ensure the real-time nature and integrity of the data, focusing on capturing changes in key parameters related to defects, such as interface energy deviation and impedance drift, to promptly detect anomalies in the interface state and provide timely support for subsequent control and prediction.

[0076] Step 2: Implement UI performance balance control using the IE-Balance algorithm. The IE-Balance interface energy balancing algorithm is adopted. Based on the real-time acquired interface energy deviation data, the micro-area model predictive control method is used to adjust the heat power and local pressure in the bonding process in zones, so as to achieve precise balance control of interface energy.

[0077] In practice, based on the distribution of interfacial energy, the bonding interface is divided into multiple micro-regions. For each micro-region with interfacial energy deviation, the corresponding heat power and local pressure are adjusted: for micro-regions with low interfacial energy, the heat power is appropriately increased and the local pressure is adjusted to promote solder wetting and improve interfacial energy; for micro-regions with high interfacial energy, the heat power is appropriately decreased and the local pressure is adjusted to avoid interface damage caused by excessively high interfacial energy. Through zonal adjustment, a balanced energy distribution is achieved across the entire bonding interface, improving solder wetting uniformity and reducing defects such as voids and incomplete soldering from the source.

[0078] Step 3: Real-time compensation for interface energy differences to improve wetting uniformity During the interface energy equalization control process, the interface energy changes in each micro-region are monitored in real time. Based on the dynamic changes in interface energy deviation, the adjustment parameters of heat power and local pressure are adjusted in a timely manner to achieve real-time compensation for interface energy differences. Through continuous compensation and adjustment, the interface energy is ensured to remain in a balanced state, avoiding interface energy deviations caused by external disturbances, equipment fluctuations, and other factors. This further improves the wetting uniformity of the solder, ensures a tight bond at the bonding interface, and reduces the possibility of defects.

[0079] Step 4: Defect-Learn algorithm constructs defect inference path The Defect-Learn self-learning algorithm, based on graph attention networks, is employed to construct a complete inference path from interface energy to dislocations, temperature, and defects. This inference path can establish a correlation between state parameters such as interface energy deviation, dislocation density, temperature distribution, and stress concentration and various bonding defects (dislocations, cracks, voids, and incomplete welds). Through self-learning capabilities, it continuously optimizes the inference logic and improves the accuracy of defect prediction.

[0080] The algorithm learns from a large amount of bonding process data to understand the intrinsic relationship between different state parameters and defect generation. Based on real-time collected state data, it can quickly infer the type, location and probability of defects that may occur in the current bonding process, thus achieving advance prediction of defects.

[0081] Step 5: Predict defect risks and output in-situ suppression strategies Based on the inference results of the Defect-Learn algorithm, the system predicts the risk of defects such as dislocations, cracks, voids, and cold solder joints in the bonding process in real time. According to the type, location, and probability of the defects, targeted in-situ suppression strategies are formulated. For example, if a void risk is predicted in a certain micro-region, the IE-Balance algorithm adjusts the thermal power and local pressure of that micro-region to improve solder wetting and suppress void formation. If a rapid dislocation multiplication is predicted, the system feeds back to the DS-Temp algorithm to adjust the temperature strategy, suppressing dislocation multiplication and preventing crack formation.

[0082] The in-situ suppression strategy is output to the corresponding control module via a data interface to actively suppress defects, prevent their further development, and ensure bonding quality. Simultaneously, the algorithm records relevant data on defect prediction and suppression, continuously optimizing prediction accuracy and suppression strategies through self-learning to improve the effectiveness of defect suppression.

[0083] 4.3 Efficiency Enhancement Principle In traditional bonding processes, uneven distribution of interfacial energy and difficulty in predicting defects are the main reasons for the high bonding defect rate. Traditional methods can only remedy defects after they occur, failing to prevent defects from occurring in the first place. At the same time, the lack of targeted control over interfacial energy makes it difficult to achieve precise equilibrium, resulting in uneven solder wetting and further aggravating defect formation.

[0084] This embodiment achieves a dual improvement in interface balance and defect suppression through the synergistic effect of the IE-Balance and Defect-Learn algorithms. The IE-Balance algorithm uses a partitioned adjustment method to achieve precise balance control of interface energy, compensate for interface energy differences in real time, improve solder wetting uniformity, reduce the causes of defects such as voids and cold solder joints from the source, and improve the quality of interface bonding.

[0085] The Defect-Learn algorithm constructs a complete defect reasoning path, enabling proactive defect prediction. This allows for the early detection of defect risks and the prevention of defect generation and development. Simultaneously, through its self-learning capability, it continuously optimizes prediction accuracy and suppression strategies, ensuring the targeted and effective nature of defect suppression.

[0086] The two algorithms work together to form a complete defect control chain of prevention-prediction-suppression, which not only solves the problem of uneven interface energy distribution but also achieves active defect suppression, significantly reducing the bonding defect rate and improving the stability and reliability of bonding quality. Compared with traditional methods, this embodiment has a more significant defect suppression effect and better interface bonding quality, which can effectively improve the performance and service life of HBM products while reducing production costs and rework rates.

[0087] The four implementation examples are closely linked and work synergistically to form a complete intelligent control process for HBM variable-temperature bonding based on the coordinated regulation of interface energy and dislocation evolution. Example 1 achieves the fusion of multi-source signals and weak feature enhancement, providing reliable input for modeling; Example 2 constructs a multi-field coupling consistency model, outputting a unified interface evolution state; Example 3 dynamically plans the variable-temperature trajectory based on state data to suppress dislocation proliferation; Example 4 achieves interface energy balancing and defect pre-suppression, ensuring bonding quality. The entire process requires no complex formula derivation or data support, is highly practical and widely applicable, effectively solves the core pain points of traditional HBM variable-temperature bonding, improves bonding quality and efficiency, and has significant engineering application value.

[0088] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for intelligent control of HBM variable-temperature bonding based on interface collaborative regulation, characterized in that, include: S1: Collect time-series data of multiple physical quantities in the micro-region during the entire HBM variable-temperature bonding process, including interface energy, micro-region temperature, contact pressure, dislocation-related strain signals, interface contact impedance, warpage morphology, heating output, and environmental disturbance data; S2: Employ the Interface Energy Feature Adaptive Fusion Algorithm (IEF-Fusion) to denoise, spatiotemporally align, and enhance features of multi-source weak micro-region signals, extracting interface energy, dislocation, impedance, and mechanically sensitive feature sets; S3: Construct a unified model of interface energy evolution, dislocation multiplication, and thermo-mechanical coupling based on the Interface Energy-Dislocation Coupling Consistency Network (IEDC-Net) to achieve real-time high-precision solution of interface states; S4: Utilize the Dislocation Suppression-Type Variable-Temp Programming Algorithm (DS-Temp) to control the heating rate. The following steps are implemented: S5: Dynamic optimization of dislocation rate, phase transition range, insulation node, and cooling slope to achieve in-situ dislocation suppression; S6: Real-time compensation of micro-region interface energy using the interface energy balance control algorithm (IE-Balance) to eliminate wetting unevenness and voids caused by interface energy differences; S7: Inference and prediction of dislocation, crack, void, and cold weld defects using the bonding defect self-learning suppression algorithm (Defect-Learn) to achieve defect closed-loop suppression; S8: Outputting zone temperature, local pressure, displacement compensation, and atmosphere control collaborative commands through a multi-variable global collaborative algorithm to achieve intelligent self-learning correction throughout the entire process; The modeling and solution formulas for the interface energy feature adaptive fusion algorithm (IEF-Fusion) mentioned in S2 are as follows: in, For the first Road micro-area signal, For the interface attention function, For adaptive weights, An interface energy feature extraction operator is proposed; the interface energy and weak dislocation signal are enhanced through adaptive gradient ascent iterative solution.

2. The method according to claim 1, characterized in that, The unified loss function of the interface energy-dislocation coupling consistency network described in S3 is: in, For interface energy evolution loss, For dislocation multiplication constraint loss, For thermo-mechanical coupling loss, The loss is due to the coupling consistency of multiple physics fields; the network strictly satisfies the interface physical mechanism to achieve high-precision prediction for small samples.

3. The method according to claim 1, characterized in that, The dynamic trajectory formula of the dislocation suppression variable-temperature programming algorithm described in S4 is: in, For real-time interface temperature, The coefficient of temperature rise, For dislocation suppression weights, For pressure coupling factor, To provide an interface energy correction factor; to achieve full matching between the temperature variation path and dislocation suppression and interface energy balance.

4. The method according to claim 1, characterized in that, The interface energy balance control algorithm described in S5 adopts micro-region model predictive control, with interface energy deviation, dislocation density, and contact impedance as state variables, and zone heating power and local pressure compensation as control variables, and solves the optimal balance strategy in a rolling manner.

5. The method according to claim 1, characterized in that, The bonding defect self-learning suppression algorithm described in S6 uses graph attention reasoning to construct the interface energy-dislocation-temperature-defect association path, thereby achieving defect source localization and closed-loop suppression.

6. The method according to claim 1, characterized in that, The multivariate global collaborative algorithm described in S7 is based on multi-objective consistency optimization, which coordinates the zoned thermal field, local pressure, micro-displacement, and atmospheric environment to achieve synchronized, conflict-free, and overshoot-free action across the entire field.

7. The method according to claim 1, characterized in that, It also includes a full-process self-learning correction mechanism that automatically updates the model and control parameters based on real-time interface energy, dislocation status, impedance distribution, and defect indicators to improve long-term batch consistency.

8. A variable-temperature bonding intelligent control system for HBM based on the coordinated regulation of interface energy and dislocation evolution, characterized in that, include: The system comprises a micro-area multi-physical quantity sensing unit, an IEF-Fusion interface energy fusion unit, an IEDC-Net coupling modeling unit, a DS-Temp dislocation suppression and temperature control unit, an IE-Balance interface energy balancing unit, a Defect-Learn defect self-learning unit, a global collaborative control unit, and a self-learning correction unit; the system is used to execute the method described in any one of claims 1-8; the global collaborative control unit outputs optimal collaborative instructions for zoned heat-local force-micro-displacement-atmosphere based on the interface energy state, dislocation evolution trend, coupling field distribution, and defect risk.

Images