Chip advanced packaging process yield prediction method based on digital twinning

Abstract

The application discloses a chip advanced packaging process yield prediction method based on digital twinning, relates to the chip processing technical field and comprises the following steps: constructing a digital twinning body of a chip advanced packaging process; collecting multi-source heterogeneous data of a physical packaging production line in a production process in real time and synchronizing the data to the digital twinning body; driving and calibrating model parameters in the digital twinning body; performing multi-physical field coupling simulation, predicting potential defect types, positions and occurrence probabilities; generating a yield distribution map and a key influence factor analysis report; and feeding back the prediction result, yield data and analysis report to the physical production line. The application has the advantages that by constructing a digital twinning body integrating geometric, physical and process models and dynamically interacting with real-time production line data, a fundamental change from "post-event statistics" to "in-process simulation and pre-warning" is realized.

Description

Technical Field

[0001] This invention relates to the field of chip manufacturing technology, specifically to a method for predicting the yield of advanced chip packaging processes based on digital twins. Background Technology

[0002] In advanced chip packaging, such as wafer-level packaging, 2.5D / 3D packaging, and system-in-package (SiP), the process steps are complex and require extremely high precision. Traditional yield control methods mainly rely on statistical process control (SPC) based on historical production data and final offline electrical testing and destructive physical analysis. These methods are essentially "post-hoc" inspections, unable to predict and intervene in dynamic fluctuations during the production process in real time, leading to delayed problem detection, long debugging cycles, and high costs for yield improvement. While some existing physics-based finite element simulation tools can simulate the physical effects of specific process steps, the models are isolated, computationally expensive, and difficult to integrate with real-time production line data, failing to form a closed-loop prediction and optimization system covering "process parameters - multi-physics coupling effects - final defects."

[0003] With the development of Industry 4.0 technologies, digital twins, as a key enabling technology connecting the physical world and virtual space, have shown potential in the monitoring and optimization of complex systems. However, in the specific scenario of advanced chip packaging, existing technical solutions typically focus only on building digital twins of equipment status or static 3D models of products, lacking a systematic framework that can deeply integrate multi-dimensional models of "geometry-physics-process" and map dynamic production line data in real time. Specifically, how to achieve quantitative correlation prediction between microscopic defect evolution and macroscopic product yield, and how to use real-time data to drive online simulation and reverse optimization of process parameters using digital twins, remain unresolved technical bottlenecks, limiting the foresight and accuracy of yield prediction. Summary of the Invention

[0004] To address the aforementioned technical problems, this paper provides a method for predicting the yield of advanced chip packaging processes based on digital twins. This technical solution solves at least one of the technical problems mentioned in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for predicting the yield of advanced chip packaging processes based on digital twins, comprising:

[0007] A digital twin of advanced chip packaging technology is constructed, comprising a geometric model, a physical model, and a process model, which is used to perform multi-dimensional mapping of physical packaging production lines and processes in virtual space;

[0008] Real-time acquisition of multi-source heterogeneous data during the production process of the physical packaging production line, including equipment process parameters, online detection data and environmental parameters, and synchronization of the data to the digital twin;

[0009] Based on the real-time collected data, the model parameters within the digital twin are driven and calibrated to ensure that the state of the digital twin is consistent with the real-time state of the physical production line.

[0010] Using the state-consistent digital twin, multiphysics coupling simulation is performed to simulate key process steps in chip packaging and predict potential defect types, locations, and probabilities of occurrence.

[0011] Based on the simulation prediction results, calculate the overall yield of the current or next batch of products, and generate a yield distribution chart and a report on the analysis of key influencing factors.

[0012] The prediction results, yield data, and analysis reports are fed back to the physical production line to optimize process parameters, adjust control strategies, or provide early warnings of potential quality risks, forming a closed-loop control of "prediction-optimization".

[0013] Preferably, the physical model includes at least a thermodynamic model, a structural mechanics model, and an electrical model:

[0014] The thermodynamic model is used to simulate heat distribution, thermal stress, and thermal deformation during the packaging process;

[0015] The structural mechanics model is used to simulate material interface stress, package warpage, and interconnect structure reliability.

[0016] The electrical model is used to predict signal integrity, power integrity, and electromagnetic compatibility.

[0017] Preferably, the digital twin further includes a multi-scale model:

[0018] The multi-scale model includes: a microscale model for simulating the microstructure of grains, bumps, and through-silicon vias; a mesoscale model for simulating the behavior of chips, redistribution layers, and interposers; and a macroscale model for simulating the macroscopic performance of the entire package and substrate assembly.

[0019] The multi-scale models are coupled through a data interface, enabling microscopic defects to be correlated with and affect the performance prediction results of mesoscopic and macroscopic models.

[0020] Preferably, the step of using the digital twin with consistent state to perform multiphysics coupling simulation, simulating key process steps in the chip packaging process, and predicting potential defect types, locations, and probabilities of occurrence specifically includes:

[0021] Based on real-time process parameters, the corresponding process model sequence is triggered in the digital twin. The process model sequence includes a bonding model, an electroplating model, a molding compound flow model, and a photolithography model.

[0022] Run a multiphysics solver to perform coupled calculations on the multiple physical fields involved in the process, such as heat, force, electricity and current, to simulate material behavior and structural response;

[0023] Based on the results of the coupled calculation, a preset failure criterion or machine learning classifier is applied to identify abnormal regions of stress, strain, temperature or electrical parameters that exceed the threshold, and these abnormal regions are marked as potential defects.

[0024] Preferably, the multiphysics coupling simulation further includes a model adaptive learning step, which specifically includes:

[0025] The actual yield data of the final physical production line is compared with the prediction results of step S4, and the prediction error is calculated.

[0026] Based on the prediction error, the parameters or weights of the corresponding model in the digital twin are automatically corrected using the backpropagation algorithm or the Bayesian update method.

[0027] Through continuous iteration, the prediction accuracy of the digital twin is continuously improved as production data accumulates.

[0028] Preferably, the step of feeding back the prediction results, yield data, and analysis reports to the physical production line to optimize process parameters, adjust control strategies, or provide early warnings of potential quality risks, forming a closed-loop control of "prediction-optimization," specifically includes:

[0029] In the digital twin, process parameter variables to be optimized and their feasible ranges are set, with the optimization objectives being to maximize the predicted yield and / or optimize the key performance indicators.

[0030] Virtual experiments are designed to be performed in the digital space, and the digital twin is used to quickly simulate the production process and results under different parameter combinations, replacing some physical experiments.

[0031] Genetic algorithms, particle swarm optimization, or gradient descent methods are used to search for the optimal combination of process parameters that meets the optimization objective in the simulation results, and this parameter combination is then sent to the physical production line as a recommended formula.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] This invention achieves a fundamental shift in yield prediction from "post-event statistics" to "in-process simulation and pre-event early warning" by constructing a digital twin that integrates geometric, physical, and process models and dynamically interacts with real-time production line data. Its beneficial effects are specifically reflected in:

[0034] It significantly improves the real-time performance and accuracy of predictions, enabling dynamic simulation of the production line's real-time status to predict defect distribution and yield in advance, making the prediction accuracy and timeliness significantly higher than traditional statistical models.

[0035] It achieves closed-loop process optimization, which can safely and efficiently perform the "parameter adjustment-simulation verification" iteration in virtual space, quickly lock in the optimal process window, thereby shortening the R&D cycle and reducing trial and error costs;

[0036] It provides in-depth root cause analysis capabilities, intuitively revealing the mechanisms of defect formation through multiphysics coupling simulation, guiding precise process improvements rather than relying on empirical guesswork. It offers digital decision support throughout the design, production, and optimization phases of advanced chip packaging, effectively improving yield, reducing costs, and accelerating time-to-market. Attached Figure Description

[0037] Figure 1 This is a flowchart of the chip advanced packaging process yield prediction method based on digital twin proposed in this application;

[0038] Figure 2 This is an architecture diagram of the electronic devices in this solution;

[0039] Figure 3 This is a schematic diagram of the computer-readable storage medium structure in this scheme.

[0040] The numbers on the map are:

[0041] 500 - Electronic device; 501 - Bus; 502 - CPU; 503 - ROM; 504 - RAM; 505 - Communication port; 506 - Input / output component; 507 - Hard disk; 508 - User interface; 600 - Computer-readable storage medium. Detailed Implementation

[0042] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0043] Reference Figure 1 As shown, a method for predicting the yield of advanced chip packaging processes based on digital twins includes:

[0044] A digital twin of advanced chip packaging processes is constructed, comprising a geometric model, a physical model, and a process model. This digital twin is used to map the physical packaging production line and processes in a multi-dimensional virtual space. The geometric model accurately represents the spatial shape and dimensions of all structures, including the chip, substrate, bumps, and redistribution layers. The physical model (such as thermal, mechanical, and electrical models) describes material properties and multi-physics behavior. The process model defines the flow logic of specific steps such as bonding, electroplating, and molding. The integration of these three models forms a virtual replica that fully and multi-dimensionally reflects the real-world process, laying the model foundation for subsequent simulation and prediction.

[0045] Real-time acquisition of multi-source heterogeneous data from the physical packaging production line during the production process, including equipment process parameters, online monitoring data, and environmental parameters, is synchronized to the digital twin. Equipment process parameters (such as temperature, pressure, and alignment accuracy) provide process inputs, online monitoring data (such as optical inspection (AOI) and online metering) reflect intermediate results, and environmental parameters (such as temperature, humidity, and cleanliness) constitute manufacturing boundary conditions. Synchronizing this real-time, multi-source data to the digital twin ensures that the virtual world can dynamically perceive the real state of the physical world, which is a prerequisite for achieving real-time simulation and accurate prediction.

[0046] Based on real-time acquired data, the model parameters within the digital twin are driven and calibrated to ensure that the state of the digital twin remains consistent with the real-time state of the physical production line. By using real-time data as input to drive the model's operation (e.g., using the actual heating temperature as the boundary condition of the thermal model) and utilizing partial data (such as detection results) to back-calibrate the model's initial parameters or correct model errors, the simulation state of the digital twin can continuously track and approximate the actual operating state of the physical production line. This dynamic calibration mechanism overcomes the biases caused by model idealization and is a core element in achieving high-precision prediction.

[0047] By utilizing a consistent digital twin, multiphysics coupled simulations can be performed to model key process steps in chip packaging and predict potential defect types, locations, and probabilities. Within a calibrated, consistent digital twin, coupled calculations involving heat and stress, fluid and curing can be performed to proactively simulate physical processes such as warpage due to thermal expansion mismatch, porosity caused by unbalanced molding compound flow, and interconnect failure due to electromigration. By analyzing stress concentration areas, temperature anomalies, or regions exceeding electrical performance limits in the simulation results, potential defect patterns and spatial distributions can be identified, and their probability of occurrence can be extrapolated based on historical data or models.

[0048] Based on simulation prediction results, the overall yield of the current or next batch of products is calculated, and a yield distribution map and a key influencing factor analysis report are generated. The overall yield is calculated by integrating the predicted probabilities of various defects. The yield distribution map (e.g., based on the chip's position on the wafer) can visually display weak areas. The key influencing factor analysis uses methods such as sensitivity analysis to quantify the impact of different process parameters (such as bonding pressure and curing temperature profile) on the final defects and yield, thereby identifying the key parameters that need to be prioritized for control and providing direction for precise optimization.

[0049] The predicted results, yield data, and analysis reports are fed back to the physical production line to optimize process parameters, adjust control strategies, or provide early warnings of potential quality risks, forming a closed-loop "prediction-optimization" control system. This completes the loop from virtual prediction to physical intervention. Feedback can manifest in several ways: providing early warnings of potential batch quality risks before problems occur; directly setting up equipment based on optimized parameter combinations (such as better temperature profiles) provided by virtual simulations during process development or adjustment; or dynamically adjusting the focus of Statistical Process Control (SPC) based on influencing factor analysis reports. This transforms production control from reactive to preventative, and from experience-driven to data- and model-driven.

[0050] In some preferred embodiments, the construction and implementation of the physical model specifically includes: First, establishing a parameterized thermodynamic finite element model. This model is based on the physical properties of the packaging materials (such as chips, substrates, molding compounds, and under-bump metal layers), such as specific heat capacity, thermal conductivity, and coefficient of thermal expansion, as well as the process temperature curves (such as reflow soldering curves and molding curing temperatures) obtained in real time from the physical production line. By solving the coupled equations of heat conduction and thermal stress, the three-dimensional transient temperature field distribution of the package throughout the process, the resulting thermal stress field, and the final thermally induced deformation are calculated. Subsequently, the output of this thermodynamic model (especially temperature distribution and thermal deformation) will serve as key boundary conditions to drive a high-fidelity structural mechanics model. This model further incorporates the constitutive relations of materials (such as elastoplasticity and creep characteristics), interface properties (such as adhesion energy and friction coefficient), and complex geometries (such as through-silicon vias and redistribution layers). Through nonlinear static or transient dynamic analysis, it accurately simulates package warpage, interface delamination risk, stress concentration and fatigue life of solder joints or bumps caused by thermal mismatch between materials, as well as the mechanical reliability of micro-interconnect structures. Finally, based on the final geometric deformation (warping, displacement) and stress state calculated by the aforementioned model, a three-dimensional full-wave electromagnetic simulation or circuit parameter extraction electrical model is constructed. The influence of physical deformation on interconnect width, spacing, and dielectric layer thickness, as well as the modulation effect of stress on electrical parameters such as carrier mobility and dielectric constant, are parametrically mapped to the mesh and properties of the electrical model. This allows for the prediction of signal integrity (such as insertion loss, return loss, and crosstalk), impedance and noise of the power distribution network, and overall electromagnetic compatibility characteristics of the packaged interconnect system under real physical conditions. This achieves closed-loop simulation and prediction from process conditions to multi-physics effects, and ultimately to electrical performance and potential failures.

[0051] In some preferred embodiments, the digital twin also includes a multi-scale model:

[0052] The multi-scale models include: microscale models for simulating the microstructure of grains, bumps, and through-silicon vias; mesoscale models for simulating the behavior of chips, redistribution layers, and interposers; and macroscale models for simulating the macroscopic performance of the entire package and substrate assembly.

[0053] Multiscale models are coupled through data interfaces, enabling microscopic defects to correlate with and affect mesoscopic and macroscopic performance prediction results.

[0054] Specifically, the construction and coupling steps of the multi-scale model include: First, at the microscale, molecular dynamics simulations or high-resolution finite element methods are used to establish refined models of the internal structure of grains, the microstructure of bumps (such as copper pillar bumps and micro solder balls), the sidewall morphology of through-silicon vias, and the intermetallic compounds at the interface. The intrinsic properties of the material at the nanometer / micrometer scale are input, and key parameters of its micromechanical and failure behaviors, such as local stress concentration, interface crack initiation, and electromigration void formation, are simulated and output. Second, at the mesoscale, an equivalent model representing a single chip, redistribution layer network, or silicon interposer is constructed. This model does not completely replicate all microscopic details, but uses homogenization methods or surrogate models to take the equivalent material properties calculated by the microscale model (such as the effective elastic modulus after considering microscopic defects, anisotropic thermal conductivity, and equivalent circuit parameters) and key boundary conditions (such as the interface strength transferred from the microscale model) as input, thereby efficiently simulating chip-level thermo-mechanical coupling deformation, current density distribution and temperature rise of interconnects, and the mechanical response of redistribution layers under load. Finally, at the macroscopic scale, a finite element model or simplified lumped parameter model is established, encompassing the entire package, substrate, heat sink, and even the system motherboard. This model uses the chip's equivalent properties, heat source distribution, and initial warpage conditions calculated from the mesoscopic model as input to perform macroscopic simulations of thermal distribution, structural warpage, modal vibration, and system-level electrical performance at the full package level. Key state variables and equivalent parameters are automatically transferred between models at different scales via pre-defined data interfaces. For example, the probability of silicon via interface defects predicted at the microscopic scale is transformed into the statistical distribution increment of interconnect resistance in the mesoscopic model, leading to the prediction of voltage drop degradation in macroscopic power integrity simulation. This achieves quantitative correlation and traceability from microscopic defects to system-level performance degradation.

[0055] In some preferred embodiments, using a state-consistent digital twin, multiphysics coupling simulation is performed to simulate key process steps in the chip packaging process and predict potential defect types, locations, and probabilities of occurrence. Specifically, this includes:

[0056] Based on real-time process parameters, the corresponding process model sequence is triggered in the digital twin. The process model sequence includes bonding model, electroplating model, molding compound flow model and photolithography model.

[0057] Run the multiphysics solver to perform coupled calculations on multiple physical fields such as heat, force, electricity and current involved in the process, and simulate material behavior and structural response;

[0058] Based on the results of the coupled calculations, a preset failure criterion or machine learning classifier is applied to identify abnormal regions of stress, strain, temperature, or electrical parameters that exceed the threshold, and these abnormal regions are marked as potential defects.

[0059] Specifically, the methods for performing multiphysics coupling simulation using digital twins include: First, the system automatically calls and sequentially executes the corresponding process model sequence in the digital twin based on the real-time process formula synchronized from the production line. For example, the bonding model simulates the bump eutectic or hot-press bonding process based on the actual temperature-pressure-time curve, including the growth of intermetallic compounds and interface reactions; the electroplating model simulates the electrodeposition morphology and uniformity of copper pillars or redistribution layers based on current density, additive concentration, and convection conditions; the molding model simulates the non-Newtonian fluid filling process of molding compound, leading edge fusion line, and curing shrinkage based on transmission pressure, mold temperature, and material rheological properties; and the photolithography model simulates the chemical amplification, pattern formation, and dimensional changes of photoresist based on exposure dose, focal length, and post-baking conditions. Secondly, an integrated multiphysics solver is driven to perform sequential or strongly coupled numerical solutions for the aforementioned processes. For example, in the molding process, computational fluid dynamics (simulating material flow) and heat transfer (simulating curing exothermics) are coupled; in the reflow soldering process, heat transfer (temperature field), thermal stress (warpage), and diffusion dynamics (solder interface reaction) are coupled, thereby accurately simulating the behavioral evolution of materials and their final structural response under complex physical fields. Finally, based on the full-field data (such as stress, strain, temperature gradient, and current density distribution) generated by the coupled calculations, the system applies preset physical failure criteria (such as von Mises stress exceeding the material yield strength, interface peeling energy exceeding the adhesion strength, and local temperature exceeding the material degradation threshold) or uses a machine learning classifier (such as neural networks or support vector machines) trained on historical defect data to scan and evaluate the simulation area. It automatically identifies abnormal areas that exceed safety thresholds and highlights these areas in the 3D model as potential defects such as cracks, voids, delamination, or electromigration. The probability of occurrence can be estimated based on the degree to which these defects exceed the threshold, combined with a statistical model, ultimately generating a visualized defect prediction map.

[0060] In some preferred embodiments, the multiphysics coupling simulation further includes a model adaptive learning step, which specifically includes:

[0061] The actual yield data of the final physical production line is compared with the prediction results of step S4, and the prediction error is calculated.

[0062] Based on the prediction error, the parameters or weights of the corresponding model in the digital twin are automatically corrected using the backpropagation algorithm or the Bayesian update method.

[0063] Through continuous iteration, the prediction accuracy of digital twins can be continuously improved as production data accumulates.

[0064] Specifically, the adaptive learning steps of the model include: after each production batch is completed, the system automatically collects the actual defect distribution and yield data obtained from the final electrical performance test, automatic optical inspection, and slicing analysis of that batch. This data is then precisely spatially aligned and quantitatively compared with the predicted defect map and predicted yield generated by the digital twin during the synchronous simulation of that batch's production process. This allows for the calculation of prediction errors (such as false negative rate and false positive rate) for different defect types and the prediction deviation of the overall yield. Subsequently, based on this error data, the system uses optimization algorithms to automatically calibrate the underlying model parameters of the digital twin: for example, for a defect classifier built from a neural network, the backpropagation algorithm is used to adjust the weights and biases of network nodes in reverse according to the error; for models based on physical formulas but containing uncertain parameters (such as coefficients of the material constitutive equation), a Bayesian update method is used, taking actual observation data as new evidence to iteratively update the prior probability distribution of the model parameters, thereby obtaining a more accurate posterior estimate. This closed-loop learning process of "production-simulation-comparison-correction" will continue to cycle as production data accumulates, enabling the multiphysics simulation model, defect prediction model, and even process model in the digital twin to gradually approach and adapt to the dynamic characteristics of the real production line, ultimately achieving autonomous and continuous improvement in prediction accuracy.

[0065] In some preferred embodiments, the prediction results, yield data, and analysis reports are fed back to the physical production line to optimize process parameters, adjust control strategies, or provide early warnings of potential quality risks, forming a closed-loop control of "prediction-optimization." Specifically, this includes:

[0066] In the digital twin, the process parameter variables to be optimized and their feasible ranges are set, and the optimization objectives are to maximize the predicted yield and / or optimize the key performance indicators.

[0067] Virtual experiments are designed and executed in the digital space, and digital twins are used to quickly simulate production processes and results under different parameter combinations, replacing some physical experiments.

[0068] Genetic algorithms, particle swarm optimization, or gradient descent methods are used to search for the optimal combination of process parameters that meets the optimization objective in the simulation results, and this parameter combination is then sent to the physical production line as a recommended formula.

[0069] Specifically, the methods for forming a "prediction-optimization" closed-loop control include: First, process engineers or the system set the key process parameter variables to be optimized (such as peak temperature and holding time for reflow soldering, injection pressure and mold temperature for molding) and their physically feasible upper and lower limits in the optimization module interface of the digital twin. Maximizing the predicted yield is explicitly set as the primary optimization objective, while optional specific performance indicators (such as maximum warpage less than a certain threshold or minimizing critical interconnect resistance) can be used as constraints or one of multiple objectives. Next, the optimization engine automatically executes efficient virtual experimental design in the digital space. For example, it uses the Latin hypercube sampling method to generate a series of representative parameter combination sample points in a given multidimensional parameter space. Then, it drives the calibrated digital twin to quickly perform a full-process simulation from process to final performance and defect prediction for each parameter combination sample. This obtains massive amounts of "process parameter-prediction result" mapping data within hours or minutes, greatly replacing time-consuming and costly physical experiments. Finally, based on the dataset generated by the virtual simulation, the optimization engine employs intelligent optimization algorithms (such as genetic algorithms and particle swarm optimization) for global optimization. These algorithms iteratively generate, evaluate, and evolve parameter combinations in the parameter space, ultimately searching for the optimal combination of process parameters that maximizes the predicted yield and satisfies all constraints. The system automatically encapsulates this optimal parameter combination into a standard process recipe file, which is then distributed to the corresponding equipment controller on the physical production line via the manufacturing execution system, directly driving the next batch of production. This completes a fully automated closed-loop control process from virtual space prediction and optimization to physical space execution.

[0070] Furthermore, the method according to the embodiments of this application can also be achieved by means of... Figure 2 The architecture of the electronic device shown is used to implement this. For example... Figure 2 As shown, the electronic device 500 may include a bus 501, one or more CPUs 502, ROM 503, RAM 504, a communication port 505 connected to a network, input / output components 506, a hard disk 507, etc. The storage device in the electronic device 500, such as ROM 503 or hard disk 507, may store a chip advanced packaging process yield prediction method based on digital twins provided in this application. The electronic device 500 may also include a user interface 508. Of course, Figure 2 The architecture shown is merely exemplary and can be omitted as needed when implementing different devices. Figure 2 One or more components in the illustrated electronic device.

[0071] Figure 3 This is a schematic diagram of a computer-readable storage medium structure provided in one embodiment of this application. Figure 3The diagram illustrates a computer-readable storage medium 600 according to one embodiment of this application. The computer-readable storage medium 600 stores computer-readable instructions. When executed by a processor, the computer-readable instructions can perform a digital twin-based advanced chip packaging process yield prediction method according to an embodiment of this application, as described with reference to the above figures. The computer-readable storage medium 600 includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0072] In summary, the advantages of this invention are: by constructing a digital twin that integrates geometric, physical, and process models and dynamically interacting with real-time data from the production line, a fundamental shift in yield prediction is achieved from "post-event statistics" to "in-process simulation and pre-event early warning".

[0073] 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 predicting the yield of advanced chip packaging processes based on digital twins, characterized in that, include: A digital twin of advanced chip packaging technology is constructed, comprising a geometric model, a physical model, and a process model, which is used to perform multi-dimensional mapping of physical packaging production lines and processes in virtual space; Real-time acquisition of multi-source heterogeneous data during the production process of the physical packaging production line, including equipment process parameters, online detection data and environmental parameters, and synchronization of the data to the digital twin; Based on the real-time collected data, the model parameters within the digital twin are driven and calibrated to ensure that the state of the digital twin is consistent with the real-time state of the physical production line. Using the state-consistent digital twin, multiphysics coupling simulation is performed to simulate key process steps in chip packaging and predict potential defect types, locations, and probabilities of occurrence. Based on the simulation prediction results, calculate the overall yield of the current or next batch of products, and generate a yield distribution chart and a report on the analysis of key influencing factors. The prediction results, yield data, and analysis reports are fed back to the physical production line to optimize process parameters, adjust control strategies, or provide early warnings of potential quality risks, forming a closed-loop control of "prediction-optimization".

2. The method for predicting the yield of advanced chip packaging processes based on digital twins according to claim 1, characterized in that, The physical model includes at least a thermodynamic model, a structural mechanics model, and an electrical model: The thermodynamic model is used to simulate heat distribution, thermal stress, and thermal deformation during the packaging process; The structural mechanics model is used to simulate material interface stress, package warpage, and interconnect structure reliability. The electrical model is used to predict signal integrity, power integrity, and electromagnetic compatibility.

3. The method for predicting the yield of advanced chip packaging processes based on digital twins according to claim 2, characterized in that, The digital twin also includes a multi-scale model: The multi-scale model includes: a microscale model for simulating the microstructure of grains, bumps, and through-silicon vias; a mesoscale model for simulating the behavior of chips, redistribution layers, and interposers; and a macroscale model for simulating the macroscopic performance of the entire package and substrate assembly. The multi-scale models are coupled through a data interface, enabling microscopic defects to be correlated with and affect the performance prediction results of mesoscopic and macroscopic models.

4. The method for predicting the yield of advanced chip packaging processes based on digital twins according to claim 3, characterized in that, The process of using the digital twin with consistent state to perform multiphysics coupling simulation, simulating key process steps in chip packaging, and predicting potential defect types, locations, and probabilities of occurrence specifically includes: Based on real-time process parameters, a corresponding process model sequence is triggered in the digital twin. The process model sequence includes a bonding model, an electroplating model, a molding compound flow model, and a photolithography model. Run a multiphysics solver to perform coupled calculations on the multiple physical fields involved in the process, such as heat, force, electricity and current, to simulate material behavior and structural response; Based on the results of the coupled calculation, a preset failure criterion or machine learning classifier is applied to identify abnormal regions of stress, strain, temperature or electrical parameters that exceed the threshold, and these abnormal regions are marked as potential defects.

5. The method for predicting the yield of advanced chip packaging processes based on digital twins according to claim 4, characterized in that, The multiphysics coupling simulation further includes a model adaptive learning step, which specifically involves: The actual yield data of the final physical production line is compared with the prediction results of step S4, and the prediction error is calculated. Based on the prediction error, the parameters or weights of the corresponding model in the digital twin are automatically corrected using the backpropagation algorithm or the Bayesian update method. Through continuous iteration, the prediction accuracy of the digital twin is continuously improved as production data accumulates.

6. The method for predicting the yield of advanced chip packaging processes based on digital twins according to claim 5, characterized in that, The step of feeding back the prediction results, yield data, and analysis reports to the physical production line to optimize process parameters, adjust control strategies, or provide early warnings of potential quality risks, forming a closed-loop control of "prediction-optimization," specifically includes: In the digital twin, process parameter variables to be optimized and their feasible ranges are set, with the optimization objectives being to maximize the predicted yield and / or optimize the key performance indicators. Virtual experiments are designed to be performed in the digital space, and the digital twin is used to quickly simulate the production process and results under different parameter combinations, replacing some physical experiments. Genetic algorithms, particle swarm optimization, or gradient descent methods are used to search for the optimal combination of process parameters that meets the optimization objective in the simulation results, and this parameter combination is then sent to the physical production line as a recommended formula.

7. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a digital twin-based advanced chip packaging process yield prediction method as described in any one of claims 1-6.

8. A computer-readable storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by the processor, they implement the chip advanced packaging process yield prediction method based on digital twins as described in any one of claims 1-6.

Images