An online detection system for conductive copper paste facing AI computing power

Abstract

The present application relates to the technical field of electronic paste intelligent manufacturing and semiconductor packaging detection, and discloses a conductive copper paste online detection system oriented to AI computing power, which realizes full-process and full-dimension online collection of copper paste from wet film to solidification through a double-station multi-sensing module for wet film nondestructive testing and final inspection after sintering, breaks through the limitations of single-station detection lag and single dimension of traditional schemes, synchronously realizes defect identification, high-temperature failure risk prediction and interlayer matching evaluation, completes the management and control upgrade from post-elimination to pre-prediction, and has a cross-process closed-loop architecture with a built-in multi-field coupling nonlinear algorithm, can generate single substrate exclusive sintering process parameters, overturns traditional isolated control logic, can realize feedforward compensation of printing deviation without interrupting production, and realizes full-link data sedimentation, model iteration and product full-life-cycle quality traceability through end-side cloud cooperation.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing of electronic pastes and semiconductor packaging testing technology, and in particular to an online testing system for conductive copper paste oriented towards AI computing power. Background Technology

[0002] With the explosive development of AI large-scale models and computing power industry, AI chips are rapidly iterating towards high computing power density, high integration, and high bandwidth. HBM3 / 4 stacking and Chiplet heterogeneous packaging have become mainstream solutions. The line width / spacing of the high-density packaging substrate, which is the core interconnect carrier, has exceeded 10um / 10um. This places extreme demands on the printing accuracy, film thickness uniformity, conductivity, and internal density of conductive copper paste. Tiny defects in conductive copper paste will directly lead to abnormal interconnect impedance, signal delay, and reduced computing power in AI chips, and may even cause chip thermal failure and data center-level downtime accidents.

[0003] Currently, existing technologies for quality control in conductive copper paste production mainly fall into two categories: one is an offline sampling inspection scheme based on optical microscopes, benchtop resistance meters, and scanning electron microscopes; the other is an online surface defect screening scheme based on single-spectrum AOI visual inspection. These and existing related technologies often suffer from the following shortcomings:

[0004] 1. Existing testing solutions are mostly designed for the large linewidth conductive lines of general PCB boards, which cannot meet the micron-level high-precision testing requirements of RDL copper paste with fine lines of 8um / 8um and below for AI computing power chips. At the same time, existing solutions mostly adopt single-station final inspection or offline batch sampling inspection after sintering, which cannot cover the entire production process from the uncured state of the wet film after printing to sintering and curing. The detection of process deviations is seriously delayed, which can easily lead to the batch scrapping of high-value AI packaging substrates, resulting in economic losses of hundreds of thousands of yuan.

[0005] 2. Existing solutions can only detect geometric defects on the surface of copper paste from a single dimension. They cannot simultaneously cover the full parameter detection of the internal density and conductivity of copper paste. Furthermore, they cannot assess the compatibility of the current layer of copper paste with the upper and lower TSV interconnect structures, or predict the failure risk of copper paste circuits under long-term high-temperature operation. As a result, hidden defects such as internal voids, poor crystal boundaries, open circuits in interlayer interconnects, and electromigration failures cannot be effectively intercepted, ultimately leading to batch failure accidents such as reduced computing power and data center downtime after AI chip shipment. Summary of the Invention

[0006] The technical problem to be solved by this invention is that the existing technology has the disadvantages of insufficient scene adaptability, serious detection lag, single detection dimension, and high rate of missed detection of hidden defects. To this end, we propose an online detection system for conductive copper paste with AI computing power.

[0007] To achieve the above objectives, this application adopts the following technical solution: an online detection system for conductive copper paste for AI computing power, applied to the production control of multilayer rewiring layer RDL conductive copper paste with line width / spacing ≤8u, including a multi-sensor detection module, an AI analysis and prediction module, a process closed-loop control module, and an edge-cloud collaborative data processing module;

[0008] The multi-sensor detection module includes a non-destructive testing station for wet film after printing and a final inspection station after sintering, which are set sequentially along the production line process. They are used to acquire multi-dimensional physical parameters of copper paste in the uncured wet film state and multi-dimensional performance and structural parameters of the circuit after sintering and curing, respectively.

[0009] The AI ​​analysis and prediction module is deployed at the edge computing end to integrate all parameters obtained from the wet film non-destructive testing station and the final inspection station after sintering, and synchronously outputs the following two types of core control results, including but not limited to: first, the full-type defect identification results of the copper paste circuit after sintering; and second, the long-term operational reliability risk prediction results of the copper paste circuit under the rated working conditions of the AI ​​chip.

[0010] The process closed-loop control module adopts a cross-process collaborative closed-loop architecture and has a built-in dynamic adaptation unit for sintering process. The dynamic adaptation unit for sintering process has a built-in wet film-sintering multi-field coupling nonlinear dynamic adaptation formula. It is used to calculate and generate exclusive sintering process setting parameters for the current single substrate based on the real-time parameters of the wet film of a single substrate obtained by the wet film non-destructive testing station, combined with the failure risk prediction results and interlayer matching evaluation results output by the AI ​​analysis and prediction module. It actively compensates for the parameter deviation of the previous printing process through the subsequent sintering process.

[0011] The edge-cloud collaborative data processing module includes an edge computing layer and a cloud platform layer. The edge computing layer is used to realize real-time data processing, local inference of AI models, and real-time control of process closed loop. The cloud platform layer is used to build a digital twin of copper paste quality, accumulate full-link process and inspection data, and iteratively optimize the prediction model of the AI ​​analysis and prediction module and the process adaptation model of the process closed loop control module.

[0012] Preferably, the wet film-sintering multi-field coupled nonlinear dynamic adaptation formula is: ;

[0013] in, This is the dedicated sintering process parameter matrix for the i-th substrate. This is the sintering process reference parameter matrix. This is the process nonlinear coupling coefficient matrix. Here is the dynamic weight vector, and ⊙ is the Hadamard product operator. Let be the relative deviation vector of the wet film parameters of the i-th substrate. For reliability sensitivity coefficient, The high-temperature electromigration failure probability of the i-th substrate is output by the AI ​​analysis and prediction module. To preset the failure probability threshold, Let be the interlayer interconnection matching deviation vector of the i-th substrate. This is the vector of interlayer alignment compensation coefficients.

[0014] Preferred, proprietary sintering process parameter matrix With sintering process reference parameter matrix All are 3×1 column vectors, with vector elements representing the peak sintering temperature, sintering holding time, and sintering heating rate, respectively.

[0015] Preferably, the relative deviation vector of wet film parameters It is a 4×1 column vector, and the vector elements are, in order, the relative deviation of wet film thickness, the relative deviation of wet film linewidth, the relative deviation of copper paste viscosity, and the comprehensive deviation of production line environment. All of them are dimensionless relative deviation values, which are calculated from the real-time data collected by the wet film non-destructive testing station.

[0016] Preferably, the process nonlinear coupling coefficient matrix It is a 3×4 dimensionless matrix, and each element in the matrix is ​​calibrated by sintering orthogonal experiments and response surface methodology using AI computing power to package special conductive copper paste.

[0017] Preferred, dynamic weight vector Given a 4×1 dimensionless vector, the formula for calculating each element of the vector is: ;

[0018] in, As the benchmark weight, This is the bias sensitivity coefficient. It is the nth element of the relative deviation vector of wet film parameters.

[0019] Preferred reliability sensitivity coefficient The dimensionless value ranges from 0.2 to 1.5, determined by aging tests of copper paste circuits under rated high-temperature conditions for AI computing chips.

[0020] Preset failure probability threshold The dimensionless value ranges from 0.01 to 0.1, determined by the packaging reliability level requirements of AI computing chips.

[0021] Preferably, the interlayer interconnection matching deviation vector It is a 1×3 dimensionless vector, with elements representing the current layer copper paste pattern position deviation, sintering shrinkage deviation, and linewidth uniformity deviation, respectively.

[0022] Interlayer alignment compensation coefficient vector It is a 3×1 dimensionless vector, calibrated through alignment tests of multi-layer RDL stacked packages, used to achieve pre-optimization of interlayer interconnect matching in the sintering process.

[0023] Preferably, the full-type defect identification results of the AI ​​analysis and prediction module are generated through the built-in lightweight multimodal defect identification model;

[0024] The model adopts a hybrid architecture of YOLOv8 backbone network and Transformer context enhancement branch. The backbone network has built-in CBAM attention mechanism and multi-scale feature fusion branch.

[0025] Preferably, the cloud platform layer of the edge-cloud collaborative data processing module is used to bind the wet film inspection data, final inspection full-dimensional data, failure risk prediction results, and exclusive sintering process parameters of each substrate with the unique identification code of the substrate, and synchronize them to the enterprise manufacturing execution system (MES) to generate a full life cycle quality data package for each product, realizing forward traceability and reverse traceability.

[0026] The technical effects and advantages of this invention are as follows:

[0027] In this invention, a multi-sensor detection module is formed by the wet film non-destructive testing station and the final inspection station after sintering, which are set up along the production line process. This enables full-process, multi-dimensional online data acquisition of conductive copper paste from the uncured wet film after printing to the sintering and curing process. This breaks through the industry limitations of traditional solutions, which can only achieve single-station surface defect detection after sintering, and have a single detection dimension and serious lag. Relying on the AI ​​analysis and prediction module deployed at the edge computing end, it realizes the synchronous output of all types of defects in copper paste circuits, long-term high-temperature operation failure risk prediction, and interlayer interconnect matching evaluation. This completes the upgrade of the control paradigm from post-defect removal to pre-risk prediction. It adopts a cross-process collaborative closed-loop architecture with a built-in wet film-sintering multi-field coupling nonlinear dynamic adaptation formula. Taking the real-time parameters of the wet film, the failure risk prediction results, and the interlayer matching evaluation results as inputs, the core algorithm directly generates the sintering process parameters specific to each substrate. This subverts the traditional solution where printing problems can only be adjusted by adjusting the printing process and sintering problems can only be adjusted by adjusting the sintering process. The isolated control logic, without interrupting the continuous production cycle, achieves proactive feedforward compensation for front-end printing parameter deviations through the back-end sintering process. Simultaneously, relying on the edge-cloud collaborative data processing module, it realizes end-to-end process data accumulation, continuous iterative optimization of AI models and algorithm coefficients, and full lifecycle quality traceability. Ultimately, in high-density packaging scenarios for AI computing chips with linewidths of 8um / 8um and below, it achieves a significant reduction in the missed detection rate of tiny copper paste defects, a significant improvement in overall product yield and batch consistency, and an exponential decrease in failure rate during long-term high-temperature operation. It also significantly improves the overall OEE of the production line, perfectly adapting to the high-reliability, high-precision, and high-density packaging mass production requirements of core AI computing scenarios such as HBM high-bandwidth storage, automotive-grade AI chips, and high-computing-power GPU Chiplet heterogeneous integration. This fundamentally solves the core pain points of traditional copper paste control solutions in advanced packaging scenarios: insufficient adaptability, inadequate control precision, and unreliable reliability. Attached Figure Description

[0028] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts:

[0029] Figure 1 This is a schematic diagram of the process of an online detection system for conductive copper paste for AI computing power according to the present invention;

[0030] Figure 2 This is a control logic diagram of an online detection system for conductive copper paste oriented towards AI computing power according to the present invention;

[0031] Figure 3 This is an AI model inference flowchart for an online detection system for conductive copper paste with AI computing power according to the present invention. Detailed Implementation

[0032] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0033] Example 1: Refer to Figure 1-3 As shown, this invention provides a technical solution: an online detection system for conductive copper paste oriented towards AI computing power, comprising a multi-sensor detection module, an AI analysis and prediction module, a process closed-loop control module, and an edge-cloud collaborative data processing module, characterized in that:

[0034] The multi-sensor detection module includes a non-destructive testing station for wet film after printing and a final inspection station after sintering, which are set sequentially along the production line process. They are used to acquire multi-dimensional physical parameters of copper paste in the uncured wet film state and multi-dimensional performance and structural parameters of the circuit after sintering and curing, respectively.

[0035] The AI ​​analysis and prediction module, deployed at the edge computing end, is used to integrate all parameters obtained from the wet film non-destructive testing station and the final inspection station after sintering, and synchronously output the following two types of core control results, including but not limited to: first, the full-type defect identification results of the copper paste circuit after sintering; second, the long-term operational reliability risk prediction results of the copper paste circuit under the rated working conditions of the AI ​​chip.

[0036] The process closed-loop control module adopts a cross-process collaborative closed-loop architecture and incorporates a dynamic adaptation unit for the sintering process. This dynamic adaptation unit incorporates a wet film-sintering multi-field coupling nonlinear dynamic adaptation formula, which is as follows:

[0037] It is used to calculate and generate exclusive sintering process setting parameters adapted to the current single substrate based on the real-time parameters of the wet film obtained by the wet film non-destructive testing station, combined with the failure risk prediction results and interlayer matching evaluation results output by the AI ​​analysis and prediction module, and actively compensate for the parameter deviation of the previous printing process through the subsequent sintering process.

[0038] The edge-cloud collaborative data processing module includes an edge computing layer and a cloud platform layer. The edge computing layer is used to realize real-time data processing, local inference of AI models, and real-time control of process closed loop. The cloud platform layer is used to build a digital twin of copper paste quality, accumulate full-link process and inspection data, and iteratively optimize the prediction model of the AI ​​analysis and prediction module and the process adaptation model of the process closed loop control module.

[0039] This embodiment applies to an 8-layer stacked wafer-level packaging production line for HBM4 high-bandwidth memory chips. This production line primarily serves the high-bandwidth memory packaging of GPU chips used for AI large-scale model training, and is a core production link in AI computing infrastructure. The HBM4 memory chip, as the core storage carrier of the AI ​​computing system, adopts an 8-layer DRAM bare die vertical stacking architecture. Interlayer interconnection is achieved through TSVs and multiple redistribution layers (RDLs). The design linewidth / spacing of its core interconnect lines reaches 8µm / 8µm, the number of RDL stacking layers is ≥4, and the TSV interconnect density exceeds 1×10⁻⁶. 6 The requirement of one piece per cm² places extremely high demands on the production quality of conductive copper paste.

[0040] The on-site implementation and production line deployment of this embodiment strictly adhere to the cleanliness, anti-static, and high stability requirements of semiconductor packaging production lines. The system's deployment and mass production operation are completed in five core stages, as detailed below:

[0041] Phase 1: The core objective of this phase is to complete the physical deployment of the system hardware on the production line, achieving seamless integration with existing production equipment without altering the original production process and cycle time. Specific implementation steps are as follows:

[0042] Along the production line, a wet film non-destructive testing station is deployed at the discharge end of the fully automatic screen printing machine, and a final inspection station is deployed at the discharge end of the vacuum hot air circulating sintering furnace. Both stations use independent marble vibration damping bases and are rigidly connected to the production line conveyor belt to ensure that the vibration amplitude during the testing process is ≤1um, meeting the high-precision testing requirements of 8um linewidth. The production line distance between the two stations is matched with the length of the sintering furnace, ensuring that the system has completed the calculation and distribution of sintering process parameters before the substrate enters the sintering furnace after the wet film testing is completed, achieving dynamic process adaptation without delay.

[0043] Within the wet film non-destructive testing station, a multispectral AOI imaging unit, a line laser triangulation thickness gauge array unit, a resonant online viscosity detection unit, and an environmental sensing unit are installed sequentially. The multispectral AOI imaging unit employs a dual-camera symmetrical layout, covering the entire 12-inch wafer substrate for inspection without blind spots. The line laser triangulation thickness gauge array unit uses a lateral full-width scanning layout to acquire film thickness distribution data across the entire substrate area without sampling interruptions. The resonant online viscosity detection unit is directly installed in the screen printing machine's feed line, enabling real-time online detection of the copper paste viscosity without the need for offline sampling. The environmental sensing unit is deployed in the four corners of the testing station to achieve uniform acquisition of temperature, humidity and cleanliness throughout the station. In the final inspection station after sintering, the second multispectral AOI imaging unit, the array-type micro-pitch four-probe resistance detection unit, the high-frequency ultrasonic micro-detection unit and the second environmental sensing unit are installed in sequence. The array-type micro-pitch four-probe detection unit adopts a gantry-type motion architecture, which can achieve accurate detection of any point on the substrate with a positioning accuracy of ≤2um. The high-frequency ultrasonic micro-detection unit adopts a water immersion focusing probe to achieve non-destructive testing of the internal structure without damaging the copper paste circuit.

[0044] The trigger control terminals of all sensing units are connected to the motion encoders of the production line servo motors at the hardware level to build a synchronous trigger architecture for the entire system. The position pulse signals of the encoders are simultaneously connected to the controllers of all sensing units to achieve precise synchronization of all detection actions with the movement of the production line substrate. The time synchronization accuracy is ≤1ms, ensuring that the data collected by different sensing units correspond to the same spatial position of the substrate, laying the foundation for subsequent multimodal data fusion.

[0045] An industrial-grade edge computing server cluster is deployed in the local data center of the production line, including two real-time inference servers and one PLC control server. All servers use industrial-grade redundant power supplies to meet the 24-hour uninterrupted operation requirements of the production line. The edge computing servers communicate with the sensor units of the two inspection stations, the printing press, and the control system of the sintering furnace through gigabit industrial Ethernet. Data transmission adopts the real-time industrial Ethernet protocol, with an end-to-end transmission latency of ≤10ms, ensuring the real-time issuance of process control commands.

[0046] The edge computing terminal is connected to the enterprise's private cloud platform via a 5G industrial private network to build a data transmission channel for end-edge-cloud collaboration. Data transmission adopts AES-256 encryption to ensure the security of production data. The cloud platform deploys a copper paste quality digital twin system to realize the accumulation, analysis and model iteration optimization of data from the entire production line. At the same time, it connects with the enterprise's MES, ERP and PLM systems to achieve full-process integration of production data.

[0047] Phase Two: The core objective of this phase is to complete the accuracy calibration of all hardware units and the initialization of system parameters to ensure the accuracy of the test data and the stability of system operation. The specific implementation steps are as follows:

[0048] A high-precision checkerboard calibration plate of 200mm×200mm was used to calibrate the spatial coordinates of all imaging units and detection units at the two detection stations, and a unified substrate pixel coordinate system was established to ensure that the spatial positioning error of all sensing units is ≤3um. For the detection requirement of 8um linewidth, distortion correction was performed on the AOI imaging unit to ensure that the imaging distortion rate is ≤0.01% across the entire area, thus avoiding linewidth detection errors caused by imaging distortion.

[0049] A high-precision oscilloscope is used to synchronously calibrate the trigger signals of all sensing units, and the trigger delay parameters of each unit are adjusted to ensure that the trigger time deviation of all sensing units is ≤1ms. Dynamic verification is performed on a substrate running continuously on the production line to ensure that the data collected by different sensing units are completely aligned in the time dimension without any misalignment.

[0050] For each sensing unit, corresponding standard samples were used for accuracy calibration and traceability. Specifically, the line laser triangulation thickness measurement array unit was calibrated using standard thickness sheets to ensure film thickness measurement accuracy of ±0.1µm; the array-type micro-pitch four-probe unit was calibrated using standard sheet metal; and the high-frequency ultrasonic microscopic inspection unit was calibrated using standard artificial defect samples to ensure that the detection resolution of internal void defects was ≤30µm. All calibration processes were recorded in complete form, which meets the quality management requirements of semiconductor packaging production lines.

[0051] Based on the material properties and process requirements of the HBM4 packaging copper paste, the system's process baseline parameters are initialized, including the sintering process baseline parameter matrix, wet film parameter target values, failure probability threshold, and adaptation baseline element weights. The sintering process baseline parameters are set in conjunction with the recommended process and historical mass production data provided by the copper paste manufacturer. The wet film parameter target values ​​are fully matched with the HBM4 packaging design drawings, and the failure probability threshold is set according to the reliability level requirements of the HBM4 chip, ensuring that the system initialization parameters are fully adapted to the production needs of this scenario.

[0052] Phase Three: The core objective of this phase is to complete the transfer learning warm-up of the AI ​​model and the construction of cross-domain adaptation benchmarks, specifically tailored to the characteristics of the HBM4 packaging scenario. This ensures that the system's AI analysis and prediction capabilities fully match the needs of this scenario. The specific implementation steps are as follows:

[0053] We collected historical mass production data, including copper paste inspection data, process data, and reliability aging data for HBM4 packaging, to construct a scenario-specific dataset. The dataset contains over 200,000 defect image samples of 8µm linewidth copper paste, over 100,000 sets of data corresponding to wet film and sintering processes, and over 5,000 sets of reliability data from high-temperature aging tests. It covers all typical defect types, process deviation scenarios, and failure modes in HBM4 packaging, providing a data foundation for model transfer learning.

[0054] Based on the lightweight multimodal defect recognition model, multi-sensor feature-high temperature impedance drift mapping model, and interlayer alignment matching model of this invention, transfer learning pre-training is performed using a scenario-based dataset, the backbone network weights of the model are frozen, the head network and fully connected layers of the model are optimized for the specific features of the HBM4 scenario, and the feature extraction weights and loss function of the model are adjusted, so that the model can quickly adapt to the detection and prediction needs of the HBM4 scenario and solve the problem of insufficient model generalization ability in small sample scenarios.

[0055] The system collects architectural characteristics, copper paste material characteristics, cross-layer interconnect structural characteristics, and inherent properties of copper paste lines from different memory nodes on the HBM4 packaging production line. Through a transfer learning preheating module, it performs correlation mapping to establish the corresponding correlation information between memory architecture characteristics, copper paste material characteristics, and inherent properties of the lines. It extracts key characteristics and attributes that affect the quality and migration adaptability of HBM4 copper paste cross-layer interconnects. The system then uses transfer learning preheating analysis to determine the compatibility correlation strength of each element, selects core elements for the adaptation benchmark, and supplements auxiliary elements such as network transmission characteristics and memory node load characteristics. Finally, it forms a copper paste migration adaptation benchmark framework adapted to the HBM4 8-layer stacking scenario, providing a benchmark basis for subsequent cross-process adaptation.

[0056] The preheated model was validated using a test dataset to ensure that the accuracy of identifying 8µm linewidth micro-defects, predicting high-temperature failure risks, and assessing interlayer matching all met mass production requirements. For deviations discovered during the validation process, the model parameters and feature extraction logic were further optimized to ensure stable operation of the model in HBM4 mass production scenarios.

[0057] Phase Four: The core objective of this phase is to complete the functional integration and testing of all modules in the system, verify the feasibility and stability of the cross-process closed-loop control logic, and ensure that the linkage between the system and production line equipment fully meets mass production requirements. The specific implementation steps are as follows:

[0058] The multi-sensor detection module, AI analysis and prediction module, process closed-loop control module, and edge-cloud collaborative data processing module were individually functionally verified to ensure that all functions of each module could be implemented normally and that the output results were accurate and stable. For the multi-sensor detection module, the integrity of full-area detection, the synchronization of data acquisition, and the accuracy of detection data were verified. For the AI ​​analysis and prediction module, the real-time performance of model inference and the accuracy of output results were verified. For the process closed-loop control module, the accuracy of process parameter calculation, the real-time performance of command issuance, and the linkage with the sintering furnace control system were verified. For the edge-cloud collaborative module, the stability of data transmission, the data analysis capabilities of the cloud platform, and the feasibility of model iteration were verified.

[0059] After completing the single-module verification, the full-process linkage verification between modules is carried out. The verification is conducted to ensure the smoothness of the entire process link from the collection and transmission of data by the multi-sensor detection module to the inference of the AI ​​analysis and prediction module, to the execution of the process closed-loop control module and the accumulation of data by the edge-cloud collaboration module. This ensures that the data interaction between the modules is without delay, loss, or error, and that the collaborative operation logic between the modules fully meets the design requirements.

[0060] The key focus of this invention is to verify the core cross-process feedforward closed-loop control logic. By manually setting different degrees of wet film parameter deviation, the system is verified to accurately calculate the corresponding sintering process adjustment parameters from the wet film detection data, and to effectively compensate for the wet film deviation after being sent to the sintering furnace, ensuring that the quality of the sintered copper paste meets the design requirements. Through verification in multiple sets of different deviation scenarios, the parameters and weights of the core algorithm are optimized to ensure that the closed-loop control logic can achieve effective compensation under different deviation scenarios without overcompensation or undercompensation issues.

[0061] After completing the logic verification in the laboratory environment, a small-batch trial run was conducted on the production line. 100 HBM4 packaging substrates were put into full-process production to verify the stability, detection accuracy and closed-loop control effect of the system in the actual operating environment of the production line. The yield data and quality data of the trial run batch and the historical mass production batch were compared to verify the quality improvement effect brought by the system. Optimization and adjustment were made for the details that occurred during the trial run to ensure that the system is fully adapted to the mass production operation requirements of the production line.

[0062] Phase Five: The core objective of this phase is to complete the full production line integration of the system, achieving stable 24 / 7 uninterrupted operation, while establishing a continuous iterative optimization mechanism. The specific implementation steps are as follows:

[0063] The system is fully integrated into the mass production process of the production line, replacing the original manual sampling and offline testing solutions. All substrates on the production line undergo full-process testing and closed-loop control by this system, achieving full inspection and control at the mass production level. System operation training is provided to production line operators, and standardized operating procedures and anomaly handling specifications are established to ensure that production line personnel can use the system proficiently.

[0064] The system's operational status is monitored 24 / 7, including the operating status of hardware devices, the stability of model inference, the accuracy of process control, and the integrity of data transmission. An anomaly warning mechanism is set up so that an alarm is triggered immediately in the event of an anomaly, and maintenance personnel are notified to handle it in a timely manner, ensuring the uninterrupted and stable operation of the system in the mass production environment.

[0065] Through the edge-cloud collaborative platform, we continuously accumulate full-volume testing data, process data, quality data, and reliability data from the production line. We regularly iterate and optimize the coefficients of the AI ​​model and core algorithms to continuously improve the system's testing accuracy, prediction precision, and process compensation effect. At the same time, based on the iterative upgrades of the HBM4 packaging process, we synchronously adjust the system's parameters and adaptation logic to ensure that the system always adapts to the latest production requirements.

[0066] The multi-sensor detection module is the data input source for the entire system and the foundation for achieving full-process quality control. In this embodiment, the module is divided into two core units along the production line process: a wet film non-destructive testing station and a post-sintering final inspection station. They operate synchronously according to the production line's production rhythm. The specific operation process is as follows:

[0067] After the screen printing machine completes the copper paste printing on the substrate, the substrate automatically enters the wet film non-destructive testing station via a conveyor belt. The production line motion encoder provides real-time feedback on the substrate's position. When the substrate reaches the testing position, all sensing units are simultaneously triggered to start the testing. First, the multispectral AOI imaging unit performs a full-area scan of the substrate, acquiring the graphic geometric parameters of the wet film copper paste lines, including key information such as line width, line spacing, pattern position, pinholes, burrs, short circuits, and open circuits. Simultaneously, multispectral imaging distinguishes the copper paste wet film from the substrate substrate, avoiding testing errors caused by substrate reflection. Simultaneously, the linear laser triangulation thickness array unit performs a transverse full-area scan of the substrate. The system performs full-area scanning to collect thickness distribution data of the wet film, generating a wet film thickness heat map to identify defects such as uneven film thickness and insufficient paste. The resonant viscosity online detection unit collects the viscosity data of the current printing batch of copper paste in real time and transmits it to the system synchronously. The environmental sensing unit collects the temperature, humidity, and cleanliness data in the detection station in real time as the basis for environmental deviation correction. After preliminary preprocessing and normalization, all collected wet film data is tagged with a unique substrate identification code and transmitted in real time to the AI ​​analysis and prediction module at the edge computing end via industrial Ethernet. At the same time, it is also transmitted to the process closed-loop control module to provide input for subsequent process dynamic adaptation.

[0068] After the substrate is sintered and cured in the sintering furnace, it is automatically conveyed to the final inspection station. Hardware-triggered detection is also achieved through a production line encoder. First, the second multispectral AOI imaging unit performs full-area imaging of the sintered copper paste circuitry, collecting surface morphology defect data, comparing the changes in the pattern before and after sintering, and calculating the circuit sintering shrinkage rate. Simultaneously, the array-type micro-pitch four-probe resistance detection unit performs precise sheet resistance measurement of the copper paste circuitry on the substrate according to preset detection points, collecting room-temperature conductivity data, and evaluating the sintering density of the copper paste. In addition to conductivity uniformity, the high-frequency ultrasonic microscopic inspection unit scans key interconnect areas and analyzes the internal density of the copper paste through high-frequency ultrasonic echo signals to identify hidden defects such as internal voids, delamination, and poor sintering boundaries. The second environmental sensing unit collects environmental data from the sintering station for process parameter correction. All sintered inspection data are also tagged with a unique substrate identification code and transmitted to the AI ​​analysis and prediction module at the edge computing end for defect identification, reliability prediction, and interlayer matching evaluation, while also serving as feedback for iterative optimization of process parameters.

[0069] All data collected from the two workstations will undergo unified spatiotemporal alignment preprocessing at the edge computing end. Based on the timestamp of the production line encoder, all sensor data of the same substrate will be aligned to the same time dimension to ensure the temporal consistency of the data. Through the spatial coordinate calibration matrix, all sensor data will be mapped to a unified substrate pixel coordinate system to ensure that the data of different sensor units are completely corresponding in the spatial dimension, providing a high-quality data source for subsequent multimodal feature fusion.

[0070] The AI ​​analysis and prediction module is the core of the entire system's decision-making. Deployed in a real-time inference server at the edge computing end, it completes multi-modal feature fusion, defect identification, long-term failure risk prediction, and inter-layer matching assessment based on the full amount of data collected by the multi-sensor detection module, providing a decision-making basis for process closed-loop control. The specific operation flow of this module in this embodiment is as follows:

[0071] The multimodal feature fusion submodule receives full data from the wet film and final inspection stations. First, it performs min-max normalization on the sensor data of different dimensions to eliminate the impact of dimensional differences on feature fusion. Then, it uses an attention mechanism weighted fusion method to perform weighted fusion on the geometric features, thickness features, viscosity features, environmental features, as well as the morphological features, conductivity features, and internal density features after sintering. For the HBM48 layer stacking scenario, it focuses on strengthening the weights of interlayer alignment-related features, sintering shrinkage-related features, and reliability-related features. Finally, it generates a 128-dimensional copper paste full life cycle feature matrix, providing a unified feature input for the subsequent three submodules.

[0072] The lightweight multimodal defect recognition submodule uses multispectral AOI images as the core input and integrates film thickness, resistance, and ultrasonic testing data for defect-assisted judgment. Through a lightweight multimodal defect recognition model optimized by HBM4 scene transfer learning, it achieves accurate recognition of all types of defects in copper paste circuits. The model first preprocesses the input image, extracts multi-scale features through the backbone network, enhances the feature response of small defects by combining the CBAM attention mechanism, and captures the global context information of the circuit through the Transformer context enhancement branch. Finally, it outputs the type, location, size, and severity level of the defect, and classifies the substrate for defect level, providing a basis for subsequent substrate current distribution. For 8μm linewidth micro-defects in the HBM4 scene, the model has specifically optimized the feature receptive field and small target detection branch to ensure accurate recognition of defects such as micro-pinholes and linewidth deviations, without missed or false detections.

[0073] The high-temperature failure risk prediction submodule is the core design of this invention for meeting the long-term reliability requirements of HBM4. Taking the copper paste full life cycle feature matrix as input, it outputs the long-term operational failure risk prediction results of the copper paste circuit under the rated high temperature conditions of the AI ​​chip through a multi-sensor feature-high-temperature impedance drift mapping model optimized by transfer learning. The model first extracts core features from the feature matrix, such as film thickness uniformity, internal density, room temperature conductivity, linewidth uniformity, and sintering process parameters. Through a fully connected neural network, it performs nonlinear mapping and outputs the impedance drift rate and electromigration failure probability of the copper paste circuit after 1000 hours of aging under 125℃ high-temperature conditions. At the same time, it divides the failure probability into three risk levels: high, medium, and low. For the high reliability requirements of the HBM4 chip, the model is specifically optimized for the recognition accuracy of low failure probability scenarios to ensure accurate identification of potential early failure risks, providing a basis for subsequent process optimization and reliability management.

[0074] The interlayer matching evaluation submodule is a dedicated design adapted to the HBM 48-layer stack architecture. Its core is an interlayer alignment matching model optimized for specific scenarios. The model takes the current layer's copper paste pattern position deviation, measured sintering shrinkage rate, and linewidth uniformity as inputs. Combined with the TSV interconnect structure design coordinates and aperture parameters of the upper and lower RDLs, it calculates the interconnect alignment deviation between the current layer and the upper and lower layers. It predicts the risk of open circuits and impedance mismatch in interlayer interconnects after sintering. For the cumulative deviation problem of 8-layer stacking, the model also combines the detection data of the previous few copper paste layers to calculate the cumulative interlayer deviation and predict the alignment risk after multi-layer stacking in advance. When the alignment deviation exceeds the preset threshold, a defect diversion warning is immediately triggered. At the same time, the printing pattern pre-compensation parameters are output to the printer for subsequent substrate printing position and pattern size compensation, thus avoiding interlayer interconnect failure problems after multi-layer stacking from the root.

[0075] The process closed-loop control module is the core of the entire system. Deployed in the PLC control server at the edge computing end, it adopts a cross-process collaborative closed-loop architecture. The core is the dynamic adaptation unit for the sintering process, and it also includes a wet film feedforward pre-control unit and a printing parameter iterative optimization unit. The specific operation flow of this module in this embodiment is as follows:

[0076] The wet film feedforward control unit performs pre-control on copper paste incoming material and environmental fluctuations. It receives copper paste viscosity detection data and production line environmental data in real time. Based on the adaptation benchmark constructed by transfer learning, it pre-adjusts process parameters such as printing speed, squeegee gap, and printing pressure of the printer. For example, when the copper paste viscosity is detected to be higher than the target value, it automatically reduces the printing speed and increases the squeegee gap to compensate for the thinner film caused by the increased viscosity. When the ambient humidity is detected to be lower, it automatically adjusts the solvent evaporation compensation parameters of the printer to avoid crack defects caused by the wet film drying too quickly. It reduces parameter deviations in the printing process from the source and improves the stability of wet film quality.

[0077] The dynamic adaptation unit for the sintering process is the core of this invention's cross-process feedforward control and a key innovation distinguishing it from traditional solutions. Its core is a multi-field coupled nonlinear dynamic adaptation formula for wet film-sintering. This unit receives the wet film parameters of a single substrate from the wet film non-destructive testing station in real time. Combining the failure risk prediction results and interlayer matching evaluation results from the AI ​​analysis and prediction module, it calculates and generates the substrate's specific sintering process parameters using the core formula. These parameters include three core parameters: peak sintering temperature, holding time, and heating rate. These parameters are then sent to the sintering furnace's control system in real time. Based on the received parameters, the sintering furnace dynamically adjusts the sintering process for the substrate. The sintering process curve is adjusted to compensate for parameter deviations in the preceding printing process. For example, when the wet film thickness is too thick, the peak temperature is appropriately increased and the holding time is extended to ensure that the copper paste in the thick film area is completely sintered and dense. When the wet film linewidth is too small, the heating rate is appropriately reduced and the holding time is shortened to reduce the linewidth shrinkage caused by sintering shrinkage and ensure that the linewidth after sintering meets the design requirements. The entire calculation and parameter distribution process is completed before the substrate enters the sintering furnace. The parameter adjustment response time is ≤2s, and the continuous production cycle of the production line is not interrupted at all. This achieves true personalized process adaptation and cross-process feedforward compensation for a single substrate.

[0078] The printing parameter iterative optimization unit performs closed-loop optimization for systematic deviations in the printing process. It receives quality data from the final inspection station after sintering and analysis results from the AI ​​analysis and prediction module in real time. Through a PID iterative algorithm, it reversely corrects the process parameters of the printing press. For example, when a systematic deviation of excessive wet film linewidth occurs on multiple consecutive substrates, the squeegee pressure and screen parameters of the printing press are automatically adjusted to correct the linewidth deviation. When interlayer alignment deviation occurs on multiple consecutive substrates after sintering, the pattern alignment parameters of the printing press are automatically adjusted for pre-compensation. Through continuous iterative optimization by this unit, the systematic deviations in the printing process are continuously reduced, the batch consistency of wet film quality is improved, and the process compensation pressure on individual substrates is reduced.

[0079] The edge-cloud collaborative data processing module is the core of the entire system's iterative optimization. It is divided into an edge computing layer and a cloud platform layer, achieving decoupling between real-time control and global optimization, while ensuring the system's continuous iteration capability. In this embodiment, the specific operation flow of this module is as follows:

[0080] The edge computing layer is deployed locally on the production line and is responsible for computing and control tasks with high real-time requirements, including real-time data preprocessing, local AI model inference, real-time closed-loop process control, and local data caching. All detection, inference, and control tasks that require real-time response are completed at the edge layer without relying on the cloud network, ensuring that the system can still operate normally when the production line is offline and guaranteeing the continuity of production. At the same time, the edge layer cleans, desensitizes, and compresses all the collected data before uploading it to the cloud platform according to a preset cycle, providing a data foundation for global analysis and model iteration in the cloud.

[0081] The cloud platform layer is deployed on the enterprise's private cloud, and its core responsibilities include non-real-time global optimization and iteration tasks, such as building a digital twin of copper paste quality, accumulating end-to-end data, iteratively optimizing models, managing multi-production line collaboration, and tracing data throughout the entire lifecycle. The cloud platform receives all data uploaded from the edge layer, builds a digital twin of HBM4 packaged copper paste quality, restores the entire production process of the production line, and enables backtracking and root cause analysis of quality issues. Based on the accumulated full data, the coefficients of the AI ​​model and core algorithm are iteratively optimized regularly, and the optimized model weights and algorithm coefficients are distributed to the edge layer to achieve continuous system evolution. At the same time, the cloud platform supports parallel access from multiple production lines, enabling collaborative management and experience sharing of production line data across different production bases. In addition, the cloud platform connects to the enterprise's MES system, binding the inspection data, process parameters, and failure risk prediction results of each substrate with the substrate's unique identification code, achieving full lifecycle quality data traceability from raw materials to shipment, fully complying with the quality management standards of the semiconductor packaging industry.

[0082] In this embodiment, considering the characteristics of an 8-layer stacked HBM4 memory chip, the core wet film-sintering multi-field coupling nonlinear dynamic adaptation formula of this invention has been specifically extended. Based on the original formula, an inter-layer stacking cumulative deviation correction factor and a multi-layer sintering shrinkage collaborative correction term have been added, solving the interconnect failure problem caused by inter-layer cumulative deviation in the 8-layer stacked architecture. The extended complete formula is as follows: ;

[0083] in, This is the dedicated sintering process parameter matrix for the i-th substrate. This is the sintering process reference parameter matrix. This is the process nonlinear coupling coefficient matrix. Here is the dynamic weight vector, and ⊙ is the Hadamard product operator. Let be the relative deviation vector of the wet film parameters of the i-th substrate. For reliability sensitivity coefficient, The high-temperature electromigration failure probability of the i-th substrate is output by the AI ​​analysis and prediction module. To preset the failure probability threshold, Let be the interlayer interconnection matching deviation vector of the i-th substrate. This is the vector of interlayer alignment compensation coefficients.

[0084] Compared to the original formula, this embodiment extends and adds a cumulative deviation correction matrix for interlayer stacking. This matrix is ​​a 3×1 column vector, consistent with the dimension of the sintering process parameter adjustment, and its specific form is as follows: ;

[0085] The physical meaning and calculation formulas for each element are as follows:

[0086] Peak temperature cumulative deviation correction coefficient, used to correct the influence of cumulative deviation of multilayer stacking on the sintering peak temperature, the calculation formula is: ;

[0087] In the formula, This represents the cumulative alignment deviation of the preceding layers already completed on the current substrate. The sensitivity coefficient of cumulative deviation to peak temperature is calibrated through HBM4 multilayer stacking sintering test. The dimensionless value ranges from 0.1 to 0.8. The core function of this coefficient is to appropriately adjust the sintering peak temperature when the cumulative alignment deviation of the previous layer is too large, control the sintering shrinkage rate of the current layer, offset the cumulative deviation of the previous layer, and avoid the deviation from exceeding the design threshold after multilayer stacking.

[0088] : Cumulative Deviation Correction Factor for Insulation Time, used to correct the impact of cumulative deviation in multi-layer stacking on insulation time. The calculation formula is: ;

[0089] In the formula, This represents the cumulative sintering shrinkage deviation of the preceding layers on the current substrate. The coefficient is a sensitivity factor of cumulative shrinkage rate to holding time. It is calibrated through experiments and has a dimensionless value range of 0.2-1.0. The core function of this coefficient is to precisely control the sintering shrinkage rate of the current layer by adjusting the holding time, so as to achieve coordinated control of the shrinkage rate of multi-layer stacking and ensure the precise alignment of the interlayer interconnection structure.

[0090] : The cumulative deviation correction factor for heating rate is used to correct the effect of cumulative deviation in multilayer stacking on the heating rate. The calculation formula is as follows: ;

[0091] In the formula, The coefficient is the sensitivity coefficient of cumulative deviation to heating rate. It is calibrated by experiment and has a dimensionless value range of 0.05-0.5. The core function of this coefficient is to control the grain growth process of copper paste sintering by adjusting the heating rate, optimize the dimensional stability of copper paste circuits, and further reduce the cumulative deviation of multilayer stacking.

[0092] Meanwhile, this embodiment addresses the interlayer interconnection matching deviation vector in the formula. The design has been extended. In addition to the original graphic position deviation, sintering shrinkage deviation, and linewidth uniformity deviation, two new elements have been added: cumulative alignment deviation of the preceding layer and TSV aperture deviation. The extended vector is a 1×5 dimensionless vector, and the corresponding interlayer alignment compensation coefficient vector has been simultaneously optimized to a 5×1 column vector. This further improves the accuracy of interlayer matching control and fully adapts to the requirements of HBM48-layer stacking scenarios.

[0093] In this embodiment, when the system is running in mass production on the HBM4 packaging production line, it executes the following complete steps to achieve full-process detection, prediction, control, and optimization of a single substrate. The specific steps are as follows:

[0094] The 12-inch wafer substrate is fed into the screen printing machine through the production line feeder. The system's wet film feedforward pre-control unit collects the viscosity data of the copper paste and the production line environment data in real time. Based on the adaptation benchmark constructed by transfer learning, the printing process parameters of the printing machine are pre-adjusted to complete the pre-control before printing and reduce printing deviations from the source.

[0095] The screen printing machine completes the screen printing of copper paste on the RDL layer of the substrate according to the pre-adjusted process parameters, and transfers the designed circuit pattern onto the substrate. After printing, the substrate is automatically sent to the wet film non-destructive testing station by conveyor belt.

[0096] After the substrate enters the wet film non-destructive testing station, the production line encoder synchronously triggers all sensing units to complete the online detection of all parameters of the substrate wet film, including graphic geometric parameters, wet film thickness distribution, copper paste viscosity, environmental data, etc. After preprocessing and spatiotemporal alignment, all detection data are transmitted to the edge computing end in real time and synchronously distributed to the AI ​​analysis and prediction module and the process closed-loop control module.

[0097] The sintering process dynamic adaptation unit of the process closed-loop control module receives wet film detection data and calculates and generates sintering process parameters specific to the substrate, including peak temperature, holding time, and heating rate, by using the extended wet film-sintering multi-field coupling nonlinear dynamic adaptation formula and combining it with the cumulative deviation data of the preceding layers. These parameters are then sent to the control system of the vacuum hot air sintering furnace in real time. The sintering furnace completes the pre-adjustment of the process parameters and waits for the substrate to enter.

[0098] After the substrate exits the wet film inspection station, it enters the sintering furnace. The sintering furnace completes the sintering and curing of copper paste according to the process curve generated specifically for the substrate. Through dynamic adjustment of the sintering process, the wet film deviation of the previous printing is compensated, and the sintering shrinkage rate is controlled to optimize the interlayer alignment and matching.

[0099] After sintering, the substrate enters the final inspection station after sintering. The production line encoder synchronously triggers all sensing units to complete the full-dimensional inspection of the copper paste after sintering, including surface morphology, conductivity, internal density, environmental data, etc. After preprocessing, all inspection data is transmitted to the AI ​​analysis and prediction module at the edge computing end.

[0100] The AI ​​analysis and prediction module integrates full data from wet film and final inspection to generate a copper paste full life cycle feature matrix. It simultaneously outputs the defect identification results, long-term operation failure risk prediction results, and interlayer matching evaluation results of the substrate through a defect identification model, a high-temperature failure risk prediction model, and an interlayer matching evaluation model. Based on the output results, the substrate is graded and sorted according to quality. Substrates with defects exceeding the standard enter the rework or scrap channel, while qualified substrates enter the next process.

[0101] The printing parameter iterative optimization unit of the process closed-loop control module receives final inspection data and AI analysis results, analyzes the systematic deviations in the printing process, and corrects the printing machine's process parameters in reverse through iterative algorithms to optimize the subsequent substrate printing process and continuously improve batch consistency.

[0102] The full-volume testing data, process data, AI analysis results, and quality grading results of the substrate are uploaded to the cloud platform through the edge layer and stored in conjunction with the substrate's unique identification code. The cloud platform constructs a digital twin of the copper paste quality, accumulates full-chain data, and regularly iterates and optimizes the AI ​​model and core algorithm coefficients. The optimized parameters are then sent to the edge layer to achieve continuous system evolution. At the same time, all data is synchronized to the enterprise's MES system to achieve full lifecycle quality traceability for the product.

[0103] Qualified substrates that have completed the entire process of testing and control are unloaded by a material unloading machine and enter the next packaging process; substrates with defects or risks exceeding the standard are sorted according to preset rules to ensure that only products that meet the quality requirements enter the subsequent stages.

[0104] In terms of quality control effectiveness, the system achieves full-process, multi-dimensional online inspection of copper paste from wet film to sintering and curing, completely changing the limitations of traditional solutions that can only detect visible surface defects. It not only achieves 100% inspection of surface defects such as short circuits and open circuits, but also enables online evaluation of implicit indicators such as the internal density, interlayer compatibility, and long-term reliability of the copper paste. The failure rate of minute defects with an 8µm linewidth is reduced to near zero, and the interception rate of implicit defects reaches over 99.9%. Through cross-process feedforward compensation, real-time dynamic compensation for wet film deviations is achieved, and the linewidth deviation of the copper paste lines after sintering is controlled within ±1µm. Within this range, film thickness uniformity has been improved by over 80%, and batch consistency of copper paste sintering density has been significantly improved. This fundamentally solves the problem of poor quality after sintering caused by wet film deviation. Targeting the core pain point of HBM4 8-layer stacking, the system achieves accurate prediction and pre-compensation of interlayer interconnect alignment deviation through pre-assessment of interlayer matching and correction of cumulative deviation. The open circuit defect rate of interlayer TSV interconnects has been reduced by over 95%, completely solving the yield bottleneck caused by cumulative deviation of multi-layer stacking. The overall product yield has increased from 88% in the traditional solution to over 99.5%, achieving a qualitative leap in the mass production yield of HBM4 packaging.

[0105] In terms of reliability improvement, the system's pioneering online prediction of long-term high-temperature operation failure risks enables full reliability inspection of each substrate. This completely changes the lagging mode of traditional solutions that can only verify reliability through offline sampling aging tests. It can accurately identify potential early failure risks such as electromigration and impedance drift during the production process, intercept high-risk products in advance, and prevent failed products from flowing into downstream stages. Through the linkage optimization of sintering process and reliability prediction, the system can dynamically adjust the sintering process for high-risk substrates, optimize the microstructure of copper paste, and improve its anti-electromigration ability and high-temperature stability. The final HBM4 chip shipped has a long-term operating failure rate of less than 0.05% under high temperature and high load conditions at 125℃, which is lower than 0.8% of the traditional solution. This fully meets the extreme requirements of AI large model training servers for long-term stable operation of memory chips, and significantly reduces the downtime risk and maintenance costs of downstream application scenarios.

[0106] In terms of production efficiency and operational effectiveness, the system's cross-process feedforward closed-loop control completely breaks the traditional mindset of adjusting printing to resolve printing problems. Deviations in the printed wet film do not require machine downtime to adjust printing machine parameters; effective compensation can be achieved through dynamic adaptation of the subsequent sintering process. Downtime caused by process adjustments has been reduced by more than 90%, and the overall OEE of the production line has increased from 75% in the traditional solution to 93%, significantly improving capacity utilization. At the same time, the system achieves full-process automation of detection, analysis, and control, replacing the traditional manual offline sampling and manual process adjustment mode. The frequency of manual intervention on the production line has been reduced from more than 15 times a day to less than 2 times a day, significantly reducing labor costs and quality risks caused by human error. In addition, the system's edge-cloud collaborative architecture enables traceability of quality data throughout the product lifecycle, fully complying with the quality management standards of the semiconductor industry. It provides complete data support for product quality certification, customer complaint handling, and process iteration, significantly improving the quality management level of the production line.

[0107] In terms of scenario adaptation and flexible production, the system achieves rapid scenario adaptation through transfer learning. It can flexibly adapt to the production needs of HBM4 packages with different layers, linewidths, and materials. When the production line switches product models, the system can quickly optimize the model and adaptation benchmark through transfer learning, without the need for long-term production line debugging. The product changeover time is shortened by more than 80%, perfectly adapting to the rapid iteration pace of the AI ​​computing chip industry. At the same time, the system's personalized process adaptation capability for a single substrate can perform exclusive process optimization for the actual state of each substrate, breaking the limitation of the traditional solution of using a uniform process curve for the whole batch. Even in the face of uncertainties such as fluctuations in incoming materials and environmental fluctuations, it can ensure the stability of product quality and greatly improve the anti-interference capability and flexible production level of the production line.

[0108] Example 2: The application scenario of this example is a 5µm linewidth heterogeneous integrated packaging production line for automotive-grade AI chips. The automotive-grade AI chips produced by this production line are mainly used in L2-L4 level autonomous driving domain controllers, which are the core computing units of intelligent connected vehicles and are directly related to the driving safety of vehicles. Therefore, the reliability, stability and consistency of the chips are required to meet extremely high standards that far exceed those of consumer-grade chips.

[0109] The on-site implementation and production line deployment of this embodiment strictly adhere to the IATF 16949 quality management system requirements of the automotive electronics industry, as well as the high reliability, high stability, and traceability requirements of semiconductor packaging production lines. The system deployment and automotive-grade mass production certification are completed in six core stages. The specific implementation steps are as follows:

[0110] Phase 1: Deconstructing Automotive-Grade Scenario Requirements and Customizing System Solutions

[0111] The core objective of this phase is to complete the customized design of the system solution to meet the specific requirements of automotive-grade AI chips, ensuring that all functions and performance indicators of the system fully meet the requirements for automotive-grade certification and mass production. The specific implementation steps are as follows:

[0112] By collaborating with the quality, technology, and certification teams of automakers, chip design companies, and packaging plants, we comprehensively analyzed the full-dimensional requirements for copper paste quality control in automotive-grade AI chips. These requirements included functional safety requirements, ASIL-B / D level safety objectives, AEC-Q100 Grade 1 testing standards, IATF 16949 quality management requirements, PPM-level defect control objectives, and full lifecycle data traceability requirements. This resulted in a detailed requirements specification document, which serves as the basis for system design.

[0113] Based on the deconstructed automotive-grade requirements, this system solution is customized and optimized. On the hardware side, all hardware devices are industrial-grade products that meet automotive-grade industrial environment requirements, possessing wide-temperature range operation capabilities, vibration resistance, and interference resistance. A redundant backup architecture is also designed to avoid production line downtime due to single-point failures. On the software side, the software architecture is designed according to functional safety standards, adding anomaly handling, fault self-checking, and data backup mechanisms to ensure system stability and reliability. In terms of algorithms, the AI ​​model and core algorithms are optimized to meet the automotive-grade wide-temperature range and high reliability requirements, adding dedicated functions such as wide-temperature range reliability prediction, thermal shock performance evaluation, and zero-defect control. Regarding data management, a full lifecycle data traceability system compliant with automotive-grade requirements is designed to ensure that all data is tamper-proof, traceable, and auditable, meeting certification requirements.

[0114] We organized industry experts and customer teams to conduct multiple rounds of reviews of customized solutions to ensure that the solutions fully cover all automotive-grade requirements. At the same time, we developed a detailed verification plan, including hardware reliability verification, software functional safety verification, algorithm accuracy verification, mass production stability verification, and certification compliance verification, and clearly defined the test standards, test methods and acceptance indicators for each verification.

[0115] Phase Two: The core objective of this phase is to complete the physical deployment of the system hardware on the automotive-grade packaging production line. The hardware selection and installation fully comply with the requirements of the automotive electronics production environment. The specific implementation steps are as follows:

[0116] The automotive-grade packaging production line has a cleanliness level of Class 100 and extremely high requirements for anti-static properties. Therefore, the layout of the two testing stations strictly follows the cleanliness and anti-static specifications of the production line. The station chambers are made of all-stainless steel anti-static material and equipped with FFU high-efficiency filter units to ensure that the cleanliness inside the station is consistent with the production line. All equipment is reliably grounded and equipped with static elimination devices to prevent static electricity from damaging the copper paste circuit and testing equipment. The station layout follows the production process of the production line. The wet film non-destructive testing station is deployed at the output end of the high-precision screen printing machine, and the final inspection station after sintering is deployed at the output end of the vacuum sintering furnace. Both stations are seamlessly connected to the production line conveyor belt without changing the original production process.

[0117] To meet the testing requirements of 5µm linewidth, the wet film non-destructive testing station is equipped with an ultra-high resolution multispectral AOI imaging unit with a resolution of 0.5µm / pixel, ensuring accurate testing of 5µm linewidth; the measurement accuracy of the line laser triangulation thickness array unit is improved to ±0.05µm, achieving nanometer-level film thickness testing; at the same time, a copper paste thermal expansion coefficient testing unit is added to provide data support for wide temperature range reliability assessment.

[0118] The final inspection station after sintering is equipped with a 30µm pitch four-probe resistance detection unit, which is suitable for sheet resistance detection of single lines with a line width of 5µm; it is also equipped with a 20MHz high-frequency ultrasonic microscopic detection unit, which improves the lateral resolution to 20µm, enabling the detection of even smaller internal defects; at the same time, a wide-temperature-range impedance testing unit is added to realize online detection of the impedance characteristics of copper paste lines at different temperatures, providing data support for wide-temperature-range reliability prediction. All sensing units adopt an industrial-grade wide-temperature-range design, which can operate stably for a long time in the production line environment, and are equipped with redundant backup. The key sensing units adopt a one-for-one backup architecture to avoid production line downtime caused by equipment failure.

[0119] PLC control servers and edge computing servers that meet functional safety requirements are deployed locally on the production line. All controllers have passed SIL2 functional safety certification and have functions such as fault self-diagnosis, redundancy switching, and emergency shutdown. The control system is connected to the production line's safety interlock system. When the system malfunctions, it can immediately trigger a production line safety warning to prevent unqualified products from flowing into the next process. All control commands adopt a dual verification mechanism to ensure the accuracy of process parameters and avoid product scrapping due to command errors.

[0120] Deploy a local data storage system that meets automotive-grade requirements, employing RAID5 redundant array storage to ensure secure storage of production data. The data storage period is no less than 15 years, meeting the traceability requirements for the entire lifecycle of automotive electronic products. Build an industrial-grade redundant ring network to achieve communication connections between all devices. Data transmission adopts the TSN time-sensitive network protocol to ensure the real-time and deterministic nature of data transmission. Data transmission between the edge, device, and cloud is encrypted using national cryptographic algorithms, and a data anti-tampering mechanism is set up. Once production data is written, it cannot be modified, ensuring the authenticity and traceability of the data.

[0121] Phase Three: The core objective of this phase is to complete the high-precision calibration and automotive-grade parameter initialization of all hardware units, ensuring the accuracy, repeatability, and traceability of the test data. All calibration processes will generate complete and auditable records, meeting the requirements of the IATF 16949 system. The specific implementation steps are as follows:

[0122] A 12-inch high-precision calibration board was used to perform nanometer-level spatial coordinate calibration on all imaging and detection units, establishing a unified substrate coordinate system to ensure spatial positioning error ≤1µm. For the detection requirement of 5µm linewidth, sub-pixel-level distortion correction was performed on the AOI imaging unit to ensure full-frame imaging distortion rate ≤0.005%. A high-precision time synchronizer was used to calibrate the triggering system of all sensing units to ensure time synchronization accuracy ≤0.5ms. Detailed calibration data, calibration personnel, and calibration time were recorded for all calibration processes, forming a traceable calibration report.

[0123] For each sensing unit, metrological calibration is performed using standard samples calibrated by the National Institute of Metrology to ensure the traceability of the test data. For example, the four-probe unit uses a standard sheet resistance sheet certified by the National Institute of Metrology for multi-point calibration to ensure measurement accuracy across the entire range, while the laser thickness measurement unit uses a standard thickness sheet for full-range calibration to ensure nanometer-level measurement accuracy. All calibration processes are documented with complete calibration records, including calibration equipment, standard samples, calibration data, and environmental conditions. At the same time, a regular calibration plan is developed to ensure the long-term stability of equipment accuracy and meet the consistency requirements of automotive-grade mass production.

[0124] Combining the design requirements of automotive-grade AI chips, the automotive-grade characteristics of copper paste materials, and the testing standards of AEC-Q100 certification, the initialization system's process baseline parameters, defect judgment thresholds, failure probability thresholds, and wide-temperature-range reliability assessment standards were established. Among these, the failure probability threshold was set to 0.01 according to the automotive-grade zero-defect requirement, which is far lower than the requirements for consumer-grade chips. The upper and lower limits of the process parameters were strictly set according to automotive-grade process specifications to ensure that all process fluctuations were within a controllable range. At the same time, all initialization parameters underwent joint review by the design, process, and quality teams, resulting in traceable parameter approval records.

[0125] Phase Four: The core objective of this phase is to complete the transfer learning adaptation of the AI ​​model and the construction of a wide-temperature-range adaptation benchmark to meet the high reliability requirements of automotive-grade wide-temperature-range scenarios. This will ensure that the model can accurately predict the wide-temperature-range reliability of copper paste circuits and meet the zero-defect control requirements for automotive-grade applications. The specific implementation steps are as follows:

[0126] We collected historical mass-produced automotive-grade AI chip copper paste inspection data, process data, temperature cycling test data, high-temperature aging test data, and failure analysis data from the production line to construct an automotive-grade exclusive dataset. The dataset contains over 300,000 defect image samples of 5µm linewidth copper paste, over 150,000 sets of wet film and sintering process data, over 8,000 sets of wide-temperature-range temperature cycling test data, and over 6,000 sets of long-term high-temperature aging test data. It covers all defect types, process deviation scenarios, and failure modes in automotive-grade scenarios, especially early failure modes in wide-temperature-range environments, providing a comprehensive data foundation for model transfer learning.

[0127] Based on the AI ​​model architecture of this invention, transfer learning and full optimization are performed using automotive-grade dedicated datasets. For the detection requirements of 5µm linewidth micro-defects, the backbone network of the lightweight multimodal defect recognition model is optimized, adding a small target detection branch to enhance the feature extraction capability of micro-defects and reduce the false negative rate. For the wide-temperature-range reliability prediction requirements, the multi-sensor feature-high-temperature impedance drift mapping model is optimized, adding feature dimensions such as wide-temperature-range impedance characteristics, thermal expansion coefficient, and thermal shock resistance, optimizing the model's nonlinear mapping capability, and improving the prediction accuracy of wide-temperature-range failure risk. For the automotive-grade zero-defect requirement, the model's loss function is optimized, reducing the weight penalty for false negatives to ensure the core goal of zero false negatives. After model training, extensive scenario-based verification is conducted to ensure stable operation under various extreme conditions and meet automotive-grade functional safety requirements.

[0128] Information on the wide-temperature-range characteristics of automotive-grade copper paste materials, the thermal shock resistance of the packaging architecture, the operating conditions of the automotive environment, and the inherent properties of the copper paste circuitry are collected. Through a transfer learning preheating processing module, correlation mapping is performed to establish the corresponding correlation information between material characteristics, architecture characteristics, operating conditions, and circuit properties. Key characteristic items and key attribute items affecting the wide-temperature-range reliability of copper paste are extracted. The compatibility correlation strength of each element is determined through transfer learning preheating analysis, and an automotive-grade wide-temperature-range adaptation benchmark framework is constructed to provide a benchmark for subsequent process adaptation and optimization, ensuring that the copper paste circuitry has stable performance in a wide temperature range of -40℃ to 155℃.

[0129] In accordance with the AEC-Q100 testing standards, the model's accuracy, stability, and robustness were comprehensively verified and tested, including extreme condition testing, boundary condition testing, and anti-interference testing. The verification results show that the model's accuracy in identifying 5µm micro-defects is ≥99.9%, the false negative rate is 0, and the accuracy in predicting failure risks over a wide temperature range is ≥98%, fully meeting the automotive-grade zero-defect control requirements. All verification data were compiled into a complete verification report, which serves as supporting material for product certification.

[0130] Phase 5: The core objective of this phase is to complete the functional integration and testing of all modules of the system, verify the stability and reliability of the closed-loop control logic, and simultaneously complete the integration with the production line safety interlock system to ensure that the system fully meets automotive-grade mass production requirements. The specific implementation steps are as follows:

[0131] The four core modules were individually functionally verified. This verification not only validated the functionality but also focused on the modules' fault handling capabilities, anomaly response capabilities, and redundancy switching capabilities. For example, for the multi-sensor detection module, abnormal scenarios such as sensor failure and data loss were simulated to verify the module's anomaly warning and fault switching capabilities. For the AI ​​analysis and prediction module, abnormal scenarios such as abnormal input data and model inference failure were simulated to verify the module's anomaly handling and degraded operation capabilities. For the process closed-loop control module, scenarios such as incorrect command issuance and equipment communication interruption were simulated to verify the module's safety interlocking and emergency shutdown capabilities. This ensured that no unqualified products would be released under any abnormal scenario, thus meeting functional safety requirements.

[0132] After completing the single-module verification, the entire system's interconnected process is verified. This verifies the smoothness of the entire data flow from collection, analysis, control to traceability. Simultaneously, the system's integration with the production line safety interlock system is verified. Scenarios of excessive defects and reliability risks are simulated to verify the system's ability to accurately identify and immediately trigger production line diversion, intercepting non-conforming products and ensuring they do not flow into the next process. Severe system failure scenarios are also simulated to verify the system's ability to immediately trigger production line safety warnings, notify operators, and maintain the production line's safety status. All verifications generate complete test records, meeting functional safety certification requirements.

[0133] The key focus is on validating the system's cross-process closed-loop control logic. Addressing the precise control requirements for sintering shrinkage in ultra-fine 5µm lines, multiple wet film deviation scenarios were used to verify whether the system can dynamically adjust the sintering process to precisely control the sintering shrinkage rate, ensuring that the linewidth deviation after sintering is controlled within ±0.5µm. For wide-temperature-range reliability optimization, the system's ability to dynamically adjust the sintering process based on failure risk prediction results was verified, optimizing the microstructure of the copper paste and improving its wide-temperature-range stability and thermal shock resistance. Continuous verification on thousands of substrates ensured the stability, consistency, and repeatability of the closed-loop control logic, meeting the requirements for automotive-grade mass production.

[0134] After completing laboratory validation, small-batch trial production was conducted in accordance with the automotive industry's PPAP requirements. 500 automotive-grade AI chip substrates were used for full-process production to comprehensively verify the system's stability, testing accuracy, closed-loop control effect, and data traceability capabilities in a mass production environment. The trial-produced products underwent full AEC-Q100 Grade 1 testing, including reliability tests such as temperature cycling, high-temperature storage, high-temperature and high-humidity, and electromigration. All test results met the certification requirements. At the same time, all data from the trial production was traceable throughout the entire lifecycle and fully complied with the IATF16949 system requirements. The PPAP approval was successfully obtained, and the system is ready for mass production.

[0135] Phase Six: The core objective of this phase is to complete the full production line integration of the system, establish a continuous iteration and change management mechanism that meets automotive-grade requirements, ensure the long-term stable operation of the system, and meet the mass production control requirements of automotive-grade products. The specific implementation steps are as follows:

[0136] The system is fully integrated into the mass production process of the production line, replacing the original testing and control solutions, realizing full-process control at the mass production level, and releasing standardized SOPs, including system operation specifications, equipment maintenance specifications, anomaly handling specifications, change management specifications, etc. Comprehensive training and assessment are conducted for production line operators, quality personnel, and maintenance personnel to ensure that all personnel can master the relevant specifications.

[0137] Establish a 24 / 7 system operation monitoring mechanism to monitor the operating status of hardware equipment, the stability of model inference, the accuracy of process control, and the integrity of data storage in real time. Set up a multi-level early warning mechanism to trigger an alarm immediately in case of an anomaly. Maintenance personnel should handle the situation according to the SOP specifications, formulate regular equipment maintenance and calibration plans, and complete the daily maintenance and periodic calibration of the equipment according to the plan to ensure the long-term stability of equipment accuracy. All maintenance and calibration operations are fully recorded, traceable, and auditable.

[0138] A strict change management mechanism has been established. Any changes to system parameters, models, or algorithms must be applied for, reviewed, verified, and approved in accordance with the change management process before they can be implemented on the production line. This ensures that all changes are under control and do not affect the stability of product quality. At the same time, through the edge-cloud collaborative platform, all data in the mass production process is continuously accumulated, and models and algorithms are regularly optimized and iterated to continuously improve the detection accuracy and prediction accuracy of the system. All iteration processes generate complete verification reports and change records to meet automotive-grade change control requirements.

[0139] All production data, testing data, process data, and quality data of the system are bound to the unique identification code of the substrate and stored encrypted on both local and cloud platforms for a period of no less than 15 years, covering the entire life cycle of automotive products. The data is tamper-proof, traceable, and auditable, and the entire production process data of any product can be retrieved at any time, meeting the traceability requirements of automakers and regulatory authorities, while providing complete data support for the root cause analysis of quality problems.

[0140] In this embodiment, the system has been specifically optimized for the scenario requirements of automotive-grade AI chips. It is divided into four core modules that run sequentially. Each module has added automotive-grade-specific functions and operating logic. The specific operating flow is as follows:

[0141] The multi-sensor detection module has undergone dedicated hardware and process optimization to meet the detection requirements of automotive-grade 5µm linewidth and wide temperature range reliability requirements. It is divided into a wet film non-destructive testing station and a post-sintering final inspection station. The specific operation process is as follows:

[0142] After screen printing, the substrate enters the wet film non-destructive testing station. The production line encoder enables hardware-level synchronous triggering, and the ultra-high resolution multispectral AOI imaging unit performs a full-area scan of the substrate, acquiring the graphic geometric parameters of the 5µm linewidth copper paste lines, including linewidth, line spacing, pattern position, micro-pinholes, burrs, etc., with an imaging resolution of 0.5µm / pixel, ensuring accurate identification of minute deviations in the 5µm linewidth. Simultaneously, the nanoscale line laser triangulation thickness measurement array unit acquires the full-area film thickness distribution data of the wet film, with a measurement accuracy of ±0.05µm, accurately identifying minute non-uniformities in film thickness. The resonant viscosity online detection unit acquires the viscosity data of the incoming copper paste in real time, the thermal expansion coefficient detection unit acquires the thermal expansion characteristics data of the copper paste, and the environmental sensing unit acquires the temperature, humidity, and cleanliness data of the production line. After all wet film inspection data is preprocessed and spatiotemporally aligned, a unique DPM code tag is affixed to the substrate, and it is transmitted to the edge computing end in real time, providing input for subsequent process adaptation and reliability prediction.

[0143] After sintering and solidification, the substrate enters the final inspection station after sintering, simultaneously triggering all sensing units to start detection. The ultra-high resolution second multispectral AOI imaging unit collects surface morphology data of the copper paste lines after sintering, compares the graphic changes before and after sintering, and accurately calculates the sintering shrinkage rate. The 30µm pitch ultra-micro pitch four-probe resistance detection unit performs precise sheet resistance measurement on a single 5µm linewidth copper paste line, collects room temperature conductivity data, and evaluates the sintering density and conductivity uniformity. The 20MHz high-frequency ultrasonic microscopy detection unit scans the copper paste lines to identify hidden defects such as internal voids and delamination larger than 20µm. The wide temperature range impedance testing unit performs impedance characteristic tests on key lines at different temperatures, collects impedance change data in the range of -40℃ to 155℃, and evaluates its wide temperature range impedance stability. All final inspection data are also bound to the substrate DPM code and transmitted to the AI ​​analysis and prediction module at the edge computing end.

[0144] All collected data undergoes automotive-grade preprocessing and integrity verification to ensure its authenticity, accuracy, and completeness. For missing or abnormal data, a data anomaly warning is immediately triggered, and a backup sensing unit is activated to re-collect the data, ensuring that misjudgments are not caused by data anomalies. All data is accompanied by metadata such as timestamps, device numbers, and operator numbers to ensure full traceability and meet automotive-grade data management requirements.

[0145] The AI ​​analysis and prediction module has been specifically optimized to meet the core requirements of automotive-grade wide-temperature-range high reliability and zero defects. Its core components include a multi-modal feature fusion submodule, a zero-defect-oriented defect identification submodule, a wide-temperature-range failure risk prediction submodule, and an automotive-grade inter-layer matching assessment submodule. The specific operation flow is as follows:

[0146] The multimodal feature fusion submodule receives full data from wet film and final inspection, performs normalization processing, and then uses an attention mechanism for weighted fusion. For automotive-grade scenarios, it focuses on strengthening the weights of wide-temperature-range impedance characteristics, thermal expansion characteristics, and thermal shock resistance-related features to generate a 156-dimensional copper paste full lifecycle feature matrix, providing feature input for subsequent submodules. At the same time, it performs outlier verification on the feature data to ensure the accuracy of the feature data and avoid prediction errors caused by feature anomalies.

[0147] The zero-defect-oriented defect identification submodule is based on a lightweight multimodal defect identification model optimized for automotive-grade scenarios. It has been specifically optimized for tiny defects with a linewidth of 5µm. The model uses multi-scale feature fusion and small target enhancement branches, combined with multi-sensor data for auxiliary judgment, to achieve accurate identification of all types of defects. In order to meet the automotive-grade zero-defect requirements, the model has optimized the decision threshold to reduce the false detection rate to 0. At the same time, through a dual-model cross-inference mechanism, the false detection rate is further reduced. The model outputs the type, location, size, and severity level of the defect. In addition, according to the automotive-grade defect control standards, the substrate is classified into defect levels. Any defect that exceeds the control threshold immediately triggers a non-conforming product interception command to ensure zero-defect outflow.

[0148] The wide-temperature-range failure risk prediction submodule is the core design of this embodiment for adapting to automotive-grade scenarios. Its core is a multi-sensor feature-wide-temperature-range impedance drift mapping model optimized for automotive-grade scenarios. The model takes a full lifecycle feature matrix as input and focuses on extracting core features of the copper paste, such as film thickness uniformity, internal density, room-temperature conductivity, thermal expansion coefficient, and wide-temperature-range impedance characteristics. Through an optimized deep neural network, it outputs the impedance drift rate of the copper paste circuit under a wide temperature range of -40℃ to 155℃, the aging failure rate under high-temperature and high-humidity environments, and the electromigration failure probability under high automotive loads. Simultaneously, it classifies risk levels according to the AEC-Q100 standard. For automotive-grade requirements, the model focuses on optimizing the identification accuracy for low-failure-probability scenarios, accurately identifying potential early failure risks. This provides a basis for process optimization and defective product interception, ensuring that only products meeting wide-temperature-range reliability requirements flow into downstream processes.

[0149] The automotive-grade interlayer matching evaluation submodule targets the heterogeneous integrated architecture of automotive-grade AI chips (Chiplet). Through an interlayer alignment matching model, it calculates the alignment deviation between the current layer copper paste and the interconnect structure of the upper and lower layers, predicting the risk of open circuits and impedance mismatch in interlayer interconnects. Simultaneously, considering the requirements of automotive vibration and temperature shock, it evaluates the fatigue resistance of the interlayer interconnect structure and predicts the risk of interconnect failure under long-term automotive conditions. When the alignment deviation or failure risk exceeds the threshold, it immediately triggers the interception of non-conforming products and outputs printing pre-compensation parameters to optimize the printing alignment accuracy of subsequent substrates, ensuring the stability and reliability of high-speed interconnects between chips.

[0150] The process closed-loop control module optimizes the cross-process collaborative closed-loop architecture to meet the high consistency and high reliability requirements of automotive-grade scenarios. Its core components include an automotive-grade wet film feedforward pre-control unit, a wide-temperature-range-oriented sintering process dynamic adaptation unit, and a zero-deviation-oriented printing parameter iterative optimization unit. The specific operation flow is as follows:

[0151] The automotive-grade wet film feedforward control unit addresses the high consistency requirements of automotive-grade mass production by collecting real-time data on copper paste viscosity, thermal expansion characteristics, and production line environment. Based on a wide temperature range adaptation benchmark, it pre-adjusts parameters such as printing speed, squeegee pressure, and screen gap of the printing press, controlling the consistency of the printed wet film from the source. This minimizes batch fluctuations in wet film linewidth and thickness. Furthermore, the unit features parameter self-verification, ensuring the reasonableness of all pre-adjusted parameters to prevent quality issues caused by parameter overshoot. All adjustment records are linked to the substrate batch, making them traceable and auditable.

[0152] The wide-temperature-range-oriented dynamic adaptation unit for sintering processes is the core execution unit of this embodiment. Addressing the automotive-grade wide-temperature-range high reliability requirements, the core wet-film-sintering multi-field coupling nonlinear dynamic adaptation formula has been extended and optimized. This unit receives wet-film detection data, wide-temperature-range failure risk prediction results, and interlayer compatibility assessment results. Through the optimized core formula, it calculates and generates sintering process parameters specific to each substrate, including four core parameters: peak temperature, holding time, heating rate, and cooling rate, and sends them to the sintering furnace in real time. The sintering furnace completes the sintering and curing of the copper paste according to the specific process curve. This not only compensates for wet-film deviations in the previous printing process, ensuring precise control of the 5µm linewidth, but also, by optimizing the sintering process curve, regulates the grain growth and microstructure of the copper paste, improving the density, thermal shock resistance, and wide-temperature-range impedance stability of the copper paste lines, thus optimizing its long-term reliability from a material perspective. The entire process does not interrupt the production line's cycle time, achieving personalized process optimization and reliability improvement for each substrate.

[0153] The zero-deviation-oriented printing parameter iterative optimization unit addresses the high consistency requirements of automotive-grade mass production. It receives final inspection data and AI analysis results, analyzes systematic deviations in the printing process through high-precision iterative algorithms, and corrects the printing press's process parameters in reverse to continuously reduce printing deviations and improve batch consistency. All parameter iterations are executed according to the change management process, first undergoing small-batch verification before full-scale application, ensuring that parameter changes do not affect product quality stability. At the same time, all iteration processes are fully documented, including the reason for the change, verification data, and approval records, meeting automotive-grade change control requirements.

[0154] The edge-cloud collaborative data processing module has been specifically optimized for automotive-grade full lifecycle data traceability requirements. It is divided into an edge computing layer and a cloud platform layer, and the specific operation process is as follows:

[0155] The edge computing layer is deployed locally on the production line, undertaking all real-time detection, inference, and control tasks to ensure normal operation even when the production line is offline. At the same time, the edge layer encrypts, tags, and backs up all production data, storing it in a local redundant array according to automotive-grade requirements to ensure that the data is not lost or tampered with. The edge layer securely uploads the encrypted production data to the cloud platform according to a preset cycle, and at the same time receives model optimization parameters and algorithm coefficients from the cloud to complete the iterative updates of the local system.

[0156] The cloud platform is deployed on the enterprise's private cloud, and its core functions include automotive-grade full lifecycle data management, model iteration and optimization, multi-production line collaborative control, quality traceability, and root cause analysis. The cloud platform receives all data uploaded from the edge layer, binds it with the substrate DPM code for encrypted storage, and stores it for a period of no less than 15 years. The data has the characteristics of being tamper-proof, traceable, and auditable, fully meeting the traceability requirements of the automotive electronics industry. Based on the accumulated mass production data, the AI ​​model and core algorithms are iteratively optimized regularly. The optimized parameters are distributed to the edge layer after a strict verification and approval process. At the same time, the cloud platform connects with the enterprise's MES, PLM, and quality management systems to achieve full-process integration of production data, providing complete data support for root cause analysis of quality problems, process optimization, and product certification. In addition, the cloud platform supports collaborative control of multiple factories and production lines, realizing the unification of quality standards and experience sharing across different production bases.

[0157] In this embodiment, considering the characteristics of automotive-grade AI chips with wide temperature range high reliability and 5µm ultra-fine lines, the core wet film-sintering multi-field coupling nonlinear dynamic adaptation formula of this invention has been deeply extended. A wide temperature range reliability enhancement correction factor and a precise control term for ultra-fine line sintering shrinkage have been added to the original formula. Simultaneously, the coordinated control of heating and cooling rates has been optimized. The complete extended formula is as follows:

[0158] Compared to the original formula, this embodiment adds two major core correction factors and optimizes the parameter dimensions of the formula, as detailed below:

[0159] I. Wide Temperature Range Reliability Enhancement Correction Factor

[0160] This factor is specifically designed for automotive-grade wide-temperature-range reliability requirements. It strongly links sintering process optimization with the wide-temperature-range stability of copper paste circuitry. By adjusting sintering process parameters, it optimizes the microstructure of the copper paste, improving its resistance to thermal shock, electromigration, and wide-temperature-range impedance stability. The calculation formula is as follows: ;

[0161] The physical meaning and calculation logic of each parameter in the formula are as follows:

[0162] The wide-temperature-range impedance deviation coefficient is calculated using the following formula: ;

[0163] in, These represent the maximum and minimum impedance values ​​of the copper paste circuit within the temperature range of -40℃ to 155℃. This is the rated impedance value at room temperature. This coefficient reflects the impedance stability of the copper paste circuit over a wide temperature range. The larger the deviation coefficient, the more drastic the impedance changes with temperature, and the worse the stability over a wide temperature range.

[0164] The coefficient of thermal expansion mismatch is calculated using the following formula: ;

[0165] in, The coefficient of thermal expansion of copper paste. The coefficient of thermal expansion of the substrate material reflects the degree of thermal expansion mismatch between the copper paste and the substrate. The larger the mismatch coefficient, the more likely interface delamination and interconnect failure will occur during temperature cycling.

[0166] These are the wide-temperature-range impedance deviation sensitivity coefficient and the thermal expansion mismatch sensitivity coefficient, respectively. They are calibrated through automotive-grade temperature cycling tests and thermal shock tests, with dimensionless values ​​ranging from 0.3 to 1.2 and 0.2 to 1.0, respectively. These coefficients determine the degree to which the wide-temperature-range characteristics can be adjusted in the sintering process. When the wide-temperature-range stability deviation is large, the optimization of the sintering process is strengthened to improve the wide-temperature-range reliability of the copper paste.

[0167] II. Precise Control of Sintering Shrinkage in Ultra-Fine Circuits

[0168] This factor is designed to precisely control the sintering shrinkage of 5µm ultra-fine lines. It is used to precisely regulate the copper paste shrinkage rate during the sintering process, ensuring that the linewidth deviation after sintering is controlled within ±0.5µm, meeting the process requirements of 5µm ultra-fine lines. The calculation formula is as follows: ;

[0169] The physical meaning and calculation logic of each parameter in the formula are as follows:

[0170] The relative deviation of the wet film linewidth, that is, the relative deviation between the measured linewidth of the wet film and the design target linewidth, is the core input for sintering shrinkage control.

[0171] The standard deviation of wet film linewidth uniformity reflects the degree of dispersion of linewidth in different regions of the substrate, and is used to achieve differentiated sintering control in different regions of the substrate.

[0172] The linewidth shrinkage sensitivity coefficient, calibrated through orthogonal experiments on sintering 5µm ultra-fine lines, has a dimensionless value range of 0.8-1.5. This coefficient determines the adjustment effect of linewidth deviation on the sintering process, ensuring that the linewidth after sintering accurately meets the design requirements.

[0173] Meanwhile, this embodiment optimizes the output vector of the formula, expanding it from a 3×1 column vector to a 4×1 column vector, and adds a dynamic adjustment amount for the sintering cooling rate. By optimizing the cooling rate to control the grain cooling process of the copper paste, its microstructure and internal stress are further optimized, improving its thermal shock resistance and wide-temperature stability. The corresponding process coupling coefficient matrix... The synchronous optimization is converted to a 4×4 matrix to ensure dimensional matching with the input parameters, thereby achieving coordinated and precise control of the four major sintering process parameters.

[0174] In this embodiment, when the system is in mass production on the automotive-grade AI chip packaging production line, it strictly follows the requirements of the automotive-grade quality management system and executes the following complete steps to achieve zero-defect control and wide-temperature-range reliability optimization of a single substrate throughout the entire process. The specific steps are as follows:

[0175] The substrate enters the production line through the feeding machine. The barcode scanner reads the unique DPM code on the substrate and enters the substrate's design information, batch information, and raw material information into the system to establish a full life cycle data traceability file. All subsequent test data, process data, and quality data are bound to and stored with this DPM code.

[0176] The system's wet film feedforward control unit collects real-time data on the viscosity, coefficient of thermal expansion, and other characteristics of the incoming copper paste, as well as production line environmental data. Based on automotive-grade wide temperature range adaptation standards, it pre-adjusts the printing machine's process parameters and verifies the rationality of the pre-adjusted parameters to ensure the stability and consistency of the printing process. All parameter adjustment records are simultaneously entered into the substrate's traceability file.

[0177] The screen printing machine completes high-precision screen printing of 5μm linewidth copper paste lines on the substrate according to the pre-adjusted process parameters. After printing, the substrate is sent to the wet film non-destructive testing station via an anti-static conveyor belt.

[0178] The substrate enters the wet film non-destructive testing station. The production line encoder synchronously triggers all sensing units to complete the nanoscale online detection of all parameters of the wet film, including the geometric parameters of the 5µm linewidth, the nanoscale film thickness distribution, the copper paste viscosity, the coefficient of thermal expansion, environmental data, etc. After all the test data is verified for integrity, it is bound to the substrate DPM code and transmitted to the edge computing end in real time, and the traceability file is entered at the same time.

[0179] The dynamic adaptation unit for sintering process receives wet film detection data and, combined with the pre-judgment of wide-temperature-range failure risk, calculates and generates sintering process parameters specific to the substrate through an extended core formula. These parameters include peak temperature, holding time, heating rate, and cooling rate. After all process parameters have been verified for rationality, they are sent to the sintering furnace in real time. The sintering furnace completes parameter pre-adjustment and waits for the substrate to enter. The entire parameter calculation and sending process is recorded, making it traceable and auditable.

[0180] The substrate enters the sintering furnace, which completes the sintering and curing of the copper paste according to the process curve generated specifically for the substrate. This not only compensates for the wet film deviation of the previous printing process and ensures precise control of the 5µm linewidth, but also improves the wide temperature range stability and thermal shock resistance of the copper paste circuit by optimizing the sintering process curve. The real-time process data of the sintering process is simultaneously recorded into the substrate's traceability file.

[0181] After sintering, the substrate enters the final inspection station after sintering, and all sensing units are triggered simultaneously to complete the full-dimensional inspection of the copper paste after sintering, including surface morphology, nanoscale linewidth, sheet resistance characteristics, internal density, and wide temperature range impedance characteristics. After all the test data is verified, it is bound to the substrate DPM code and transmitted to the AI ​​analysis and prediction module, and at the same time, the traceability file is entered.

[0182] The AI ​​analysis and prediction module integrates full data from wet film and final inspection to generate a copper paste full lifecycle feature matrix. Through defect identification models, wide-temperature-range failure risk prediction models, and interlayer matching evaluation models, it outputs defect identification results, wide-temperature-range failure risk prediction results, and interlayer matching evaluation results. The system performs quality grading of the substrate according to automotive-grade zero-defect control standards. Any substrate with defects or risks exceeding the standard immediately triggers a safety interlock and is automatically diverted to the non-conforming product area to ensure that zero-defect products flow out. All analysis results and diversion decisions are simultaneously recorded in the traceability file.

[0183] The printing parameter iteration optimization unit receives final inspection data and AI analysis results, analyzes systematic deviations in the printing process, and generates printing parameter optimization schemes through high-precision iterative algorithms. The optimization schemes are applied to the subsequent substrate printing process after small-batch verification and approval in accordance with automotive-grade change management procedures, continuously improving the consistency of printing batches. All application, verification, and approval records for parameter iterations are fully retained, traceable, and auditable.

[0184] The entire process data of the substrate is encrypted and uploaded to the cloud platform through the edge layer, and bound to the substrate DPM code for long-term storage with a storage period of no less than 15 years, realizing full life cycle traceability of the product. Based on the accumulated mass production data, the cloud platform regularly iterates and optimizes the AI ​​model and core algorithm. After the optimized parameters are verified and approved, they are sent to the edge layer to realize the continuous evolution of the system. At the same time, all data is synchronized to the enterprise MES and quality management system to meet the requirements of IATF16949 system.

[0185] Qualified substrates that have completed the entire process of testing and control are unloaded by the unloading machine and enter the next packaging process. Unqualified substrates are isolated, labeled, reviewed and disposed of in accordance with automotive-grade non-conforming product control standards. All disposal processes are fully recorded to ensure that non-conforming products do not flow into downstream processes.

[0186] After mass production on the 5µm linewidth packaging line for automotive-grade AI chips, the system in this embodiment fully meets the extreme requirements of automotive-grade wide temperature range, high reliability, and zero defects. This helps the product successfully pass AEC-Q100 Grade 1 certification and IATF16949 system audit, bringing comprehensive improvements in quality, reliability, and operations, as detailed below:

[0187] In terms of zero-defect quality control, the system achieves full-process zero-defect control meeting automotive-grade requirements. For minute defects in 5µm ultra-fine lines, it achieves 100% full inspection and zero missed inspections, with a defect identification accuracy rate of over 99.9%. This completely solves the problem of later failures caused by missed inspections of minute defects in traditional solutions. Through cross-process feedforward compensation, it achieves precise control of sintering shrinkage in 5µm ultra-fine lines, with the linewidth deviation after sintering controlled within ±0.5µm. Film thickness uniformity is improved by 90%, and the dimensional accuracy and batch consistency of copper paste lines meet the ultimate requirements of automotive-grade mass production. Through pre-assessment and pre-compensation of interlayer matching, the interlayer interconnection failure rate between chiplets is reduced by over 98%, ensuring the stability of high-speed signal transmission. During mass production, the PPM value of appearance defects is reduced from 500 PPM in traditional solutions to below 5 PPM, and the PPM value of electrical performance defects is reduced from 300 PPM to below 3 PPM, fully meeting the zero-defect control target for automotive-grade products.

[0188] Regarding the improvement in reliability over a wide temperature range, the system's pioneering online prediction of wide temperature range failure risks enables full reliability inspection of every product across the wide temperature range. This completely changes the lagging mode of traditional solutions that can only verify reliability through offline sampling tests. It can accurately identify potential early failure risks during the production process, achieving a 100% defective product interception rate. This fundamentally prevents substandard products from entering automotive scenarios. Through the coordinated optimization of sintering process and wide temperature range reliability prediction, the system can optimize the microstructure of copper paste at the micro level, improving its resistance to thermal shock, electromigration, and impedance stability over a wide temperature range. The final automotive-grade AI chip shipped exhibits an impedance drift rate of ≤3% after 1000 temperature cycles from -40℃ to 155℃, far below the industry standard requirement of 10%. In a 2000-hour aging test at 155℃ high temperature and humidity, the product failure rate is 0, fully meeting the AEC-Q100 Grade 1 certification requirements. This significantly improves the operational safety and reliability of the autonomous driving domain controller and reduces the after-sales failure rate and safety risks of vehicles.

[0189] In terms of automotive-grade mass production operation and compliance, the system's fully automated closed-loop control achieves fully automated management and control of the copper paste production process. Downtime caused by process adjustments has been reduced by more than 95%, and the overall OEE of the production line has increased from 70% in the traditional solution to 92%, significantly improving capacity utilization. At the same time, the system replaces the traditional manual offline inspection and manual process adjustment mode, which not only significantly reduces labor costs but also avoids quality risks caused by human error, ensuring the stability and consistency of the mass production process. The system's full lifecycle data traceability system ensures that the data of each product from raw materials to shipment is tamper-proof, traceable, and auditable. The storage period meets the full lifecycle requirements of automotive products and fully complies with the IATF16949 system and automotive industry traceability regulations, providing complete data support for product certification, customer complaint handling, and quality control. In addition, the system's strict change management mechanism ensures that all process changes are under control, meeting the mass production stability requirements of automotive-grade products. Product changeover time has been shortened by more than 85%, enabling rapid adaptation to the production needs of different models of automotive-grade AI chips.

[0190] In terms of adaptability to automotive scenarios, the system's dedicated optimization for harsh automotive conditions gives the copper paste circuitry excellent resistance to vibration, temperature shock, and damp heat aging. The final automotive-grade AI chip shipped has a long-term operating failure rate of less than 0.01% under all automotive operating conditions, down from 0.5% in traditional solutions. This fully meets the functional safety requirements of L2-L4 level autonomous driving. At the same time, the copper paste circuitry's wide-temperature impedance stability has been significantly improved, ensuring the AI ​​chip's stable computing performance under extreme low and high temperatures. It will not experience a decrease in computing power or signal distortion due to temperature changes, thus guaranteeing the stable operation of the autonomous driving system in various harsh environments and providing solid underlying hardware support for the safe deployment of intelligent connected vehicles.

[0191] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An online inspection system for conductive copper paste with AI computing power, applied to the production control of multilayer rewiring layer (RDL) conductive copper paste with line width / spacing ≤ 8µm, comprising a multi-sensor detection module, an AI analysis and prediction module, a process closed-loop control module, and an edge-cloud collaborative data processing module, characterized in that: The multi-sensor detection module includes a non-destructive testing station for wet film after printing and a final inspection station after sintering, which are set sequentially along the production line process. They are used to acquire multi-dimensional physical parameters of copper paste in the uncured wet film state and multi-dimensional performance and structural parameters of the circuit after sintering and curing, respectively. The AI ​​analysis and prediction module is deployed at the edge computing end and is used to integrate all parameters obtained from the wet film non-destructive testing station and the final inspection station after sintering, and synchronously output the following two types of core control results, including but not limited to: first, the full-type defect identification results of the copper paste circuit after sintering; second, the long-term operational reliability risk prediction results of the copper paste circuit under the rated working conditions of the AI ​​chip. The process closed-loop control module adopts a cross-process collaborative closed-loop architecture and has a built-in dynamic adaptation unit for sintering process. The dynamic adaptation unit for sintering process has a built-in wet film-sintering multi-field coupling nonlinear dynamic adaptation formula, which is used to calculate and generate exclusive sintering process setting parameters for the current single substrate based on the real-time parameters of the wet film of the single substrate obtained by the wet film non-destructive testing station, combined with the failure risk prediction results and interlayer matching evaluation results output by the AI ​​analysis and prediction module. The parameter deviation of the previous printing process is actively compensated by the subsequent sintering process. The edge-cloud collaborative data processing module includes an edge computing layer and a cloud platform layer. The edge computing layer is used to realize real-time data processing, local inference of AI models, and real-time control of process closed loop. The cloud platform layer is used to construct a digital twin of copper paste quality, accumulate full-link process and inspection data, and iteratively optimize the prediction model of the AI ​​analysis and prediction module and the process adaptation model of the process closed loop control module.

2. The online detection system for conductive copper paste for AI computing power according to claim 1, characterized in that: The wet film-sintering multi-field coupling nonlinear dynamic adaptation formula is as follows: ； in, This is the dedicated sintering process parameter matrix for the i-th substrate. This is the sintering process reference parameter matrix. This is the process nonlinear coupling coefficient matrix. Here is the dynamic weight vector, and ⊙ is the Hadamard product operator. Let be the relative deviation vector of the wet film parameters of the i-th substrate. For reliability sensitivity coefficient, The high-temperature electromigration failure probability of the i-th substrate is output by the AI ​​analysis and prediction module. To preset the failure probability threshold, Let be the interlayer interconnect matching deviation vector of the i-th substrate, and the interlayer interconnect matching deviation vector This is a 1×3 dimensionless vector, whose elements are, in order, the deviation of the current layer of copper paste pattern position, the deviation of sintering shrinkage rate, and the deviation of linewidth uniformity. This is the vector of interlayer alignment compensation coefficients.

3. The online detection system for conductive copper paste for AI computing power according to claim 2, characterized in that: The exclusive sintering process parameter matrix With sintering process reference parameter matrix All are 3×1 column vectors, with vector elements representing the peak sintering temperature, sintering holding time, and sintering heating rate, respectively.

4. The online detection system for conductive copper paste for AI computing power according to claim 2, characterized in that: The relative deviation vector of wet film parameters It is a 4×1 column vector, and the vector elements are, in order, the relative deviation of wet film thickness, the relative deviation of wet film linewidth, the relative deviation of copper paste viscosity, and the comprehensive deviation of production line environment. All of them are dimensionless relative deviation values, which are calculated from the real-time data collected by the wet film non-destructive testing station.

5. The online detection system for conductive copper paste for AI computing power according to claim 2, characterized in that: The process nonlinear coupling coefficient matrix It is a 3×4 dimensionless matrix, and each element in the matrix is ​​calibrated by sintering orthogonal experiments and response surface methodology using AI computing power to package special conductive copper paste.

6. The online detection system for conductive copper paste for AI computing power according to claim 2, characterized in that: The dynamic weight vector Given a 4×1 dimensionless vector, the formula for calculating each element of the vector is: ； in, As the benchmark weight, This is the bias sensitivity coefficient. It is the nth element of the relative deviation vector of wet film parameters.

7. The online detection system for conductive copper paste for AI computing power according to claim 2, characterized in that: The reliability sensitivity coefficient The dimensionless value ranges from 0.2 to 1.5, determined by aging tests of copper paste circuits under rated high-temperature conditions for AI computing chips. The preset failure probability threshold The dimensionless value ranges from 0.01 to 0.1, determined by the packaging reliability level requirements of AI computing chips.

8. The online detection system for conductive copper paste for AI computing power according to claim 2, characterized in that: The interlayer alignment compensation coefficient vector It is a 3×1 dimensionless vector, calibrated through alignment tests of multi-layer RDL stacked packages, used to achieve pre-optimization of interlayer interconnect matching in the sintering process.

9. The online detection system for conductive copper paste for AI computing power according to claim 1, characterized in that: The full-type defect identification results of the AI ​​analysis and prediction module are generated through the built-in lightweight multimodal defect identification model. The model adopts a hybrid architecture of YOLOv8 backbone network and Transformer context enhancement branch. The backbone network has built-in CBAM attention mechanism and multi-scale feature fusion branch.

10. The online detection system for conductive copper paste for AI computing power according to claim 1, characterized in that: The cloud platform layer of the edge-cloud collaborative data processing module is used to bind the wet film inspection data, final inspection full-dimensional data, failure risk prediction results, and exclusive sintering process parameters of each substrate with the unique identification code of the substrate, and synchronize them to the enterprise manufacturing execution system (MES) to generate a full life cycle quality data package for each product, realizing forward traceability and reverse traceability.

Images