Full-automatic probe position inspection system for probe card

Abstract

The invention provides a full-automatic probe position inspection system for a probe card, and the system comprises a morphology collection module which is used for carrying out the three-dimensional morphology scanning of a probe group of the probe card, and obtaining the three-dimensional point cloud data of the probe group; the probe position analysis module is connected to the morphology acquisition module, comprises a machine learning model and is used for processing the three-dimensional point cloud data and generating probe position deviation data and probe health state scores; the compensation execution module is connected to the probe position analysis module and is used for scoring based on the probe position deviation data and the probe health state; calculating an optimal contact position, and driving the carrying platform to move to the position; and the verification module is used for verifying the accuracy of the needle position after the contact is executed at the optimal contact position. Passive detection, predictive maintenance and process closed-loop optimization are crossed, and the test yield, the equipment service life and the manufacturing process level are comprehensively improved through data driving.

Description

Technical Field

[0001] This application relates to the field of semiconductor technology, and in particular to a fully automated pin inspection system for probe cards. Background Technology

[0002] As a critical interface in wafer testing, the precise position and health status of probes on the probe card directly determine the accuracy and reliability of the test results. In existing technologies, probe position inspection largely relies on manual operation and two-dimensional visual inspection. This method has inherent drawbacks such as high subjectivity, low efficiency, and difficulty in acquiring three-dimensional spatial orientation information.

[0003] With the continuous miniaturization of chip manufacturing processes and the development of 3D packaging technology, the number of probe tips on probe cards has increased dramatically, and the arrangement density has been continuously improved. This has placed almost stringent requirements on the coplanarity of probe positions, alignment accuracy, and long-term contact stability. Traditional methods not only cannot efficiently and accurately complete the detection of large-scale probe groups, but also lack the ability to predict and assess potential risks of probes.

[0004] Furthermore, existing needle position compensation strategies are often based on simple geometric averages, resulting in insufficient optimization of overall contact yield. Moreover, existing solutions cannot make predictive judgments on the health trend of probes, nor can they correlate massive amounts of testing data with probe card manufacturing processes. Therefore, they cannot provide feedback and optimize processes from the source to prevent defects. Summary of the Invention

[0005] Based on the above problems, this invention proposes a fully automated pin inspection system for probe cards, which can leap from passive detection to predictive maintenance and process closed-loop optimization, and comprehensively improve test yield, equipment life and manufacturing process level through data-driven approaches.

[0006] This invention provides a fully automated pin position inspection system for probe cards, comprising: The topography acquisition module is used to perform three-dimensional topography scanning on the probe group of the probe card and acquire three-dimensional point cloud data of the probe group. The needle position analysis module is connected to the morphology acquisition module and includes a machine learning model for processing the three-dimensional point cloud data and generating needle position deviation data and probe health status scores. The compensation execution module, connected to the needle position analysis module, is used to calculate the optimal contact position based on the needle position deviation data and probe health status score, and drive the stage to move to that position. The verification module is used to verify the accuracy of the needle position after contact is performed at the optimal contact position.

[0007] Preferably, the machine learning model is a physical information neural network, which is configured as follows: The original 3D point cloud data is processed to generate a high-level feature representation corresponding to each probe; the high-level feature representation includes a numerical description for quantifying the probe's spatial pose, structural integrity, and population consistency. Historical performance data from the probe card status archive; Output the needle position deviation data and probe health status score; The historical performance data includes historical contact resistance values, historical pin position deviation data, and the corresponding number of test cycles in a time series.

[0008] Preferably, the probe health status score is achieved by the physical information neural network through the following multi-task learning, including: A mechanical health assessment task is used to evaluate the deviation of the probe from the standard morphology based on the high-level feature representation and output a first score; The electrical health assessment task is used to evaluate the stability and degradation trend of contact resistance based on fused historical performance data and output a second score; The trend prediction task is to predict the probability of morphological degradation after a specific number of cycles in the future based on the changes in morphological features over a time series, and output a third score. The physical information neural network obtains the final probe health status score by weighting and fusing the first score, the second score, and the third score through a shared feature extraction layer and their respective task-specific layers.

[0009] Preferably, the needle position analysis module is further configured as follows: Based on the probe health status score, a confidence weight is assigned to the spatial deviation vector of each probe in the needle position deviation data, and this weight is positively correlated with the probe health status score. For probes with a health status score below a preset threshold, the compensation amount in the needle position deviation data is marked as low confidence.

[0010] Preferably, the compensation execution module calculates the optimal contact position through an optimization algorithm. The input of the optimization algorithm is the needle position deviation data after the confidence weight correction, and its objective function is configured to maximize the estimated overall contact yield.

[0011] Preferably, the optimization algorithm used by the compensation execution module is a confidence-weighted multi-objective adaptive optimization algorithm, which is configured as follows: A six-degree-of-freedom optimization model is established, which includes spatial position compensation and attitude adjustment. Its decision variables include the global translation compensation of the platform in the X, Y, and Z directions and the rotation compensation around the X and Y axes. Construct a multi-objective function that includes the following sub-objectives: The primary sub-objective is to maximize the weighted contact yield prediction function, which is calculated by combining the magnitude of the spatial deviation vector of each probe, the horizontal offset from the center of the pad, and the corresponding confidence weight. Secondary sub-targets, minimizing the total contact force, are estimated by the total deformation of the probe group in the Z direction; Constrain sub-targets to ensure that the contact pressure uniformity index of all probes is below a preset threshold.

[0012] Preferably, the compensation execution module is configured as follows: Receive process information of the wafer under test, wherein the process information includes at least one of the following: the feature size of the wafer, the thickness of the uppermost intermetallic dielectric layer, and the maximum pin density; Based on the process information, the weights of different optimization objectives in the multi-objective optimization function are dynamically allocated, including: When the feature size is smaller than the first size, and / or the thickness of the uppermost intermetallic dielectric layer is smaller than the second size, the weight of the sub-objective of minimizing the total contact force in the multi-objective optimization function is increased; in the weighted contact yield prediction function, the weighting factor of the magnitude of the spatial deviation vector is decreased. When the maximum pin density is higher than the first density, the weighting factor related to the horizontal offset of the pad center in the sub-objective of the maximum weighted contact yield prediction function is increased.

[0013] Preferably, the needle position analysis module is further configured as follows: Based on the spatial distribution of the needle position deviation data described in the probe card, identify group deviations exhibiting specific patterns; The identified group deviation patterns are matched with a pre-stored database of manufacturing process defects for probe cards; When the matching degree exceeds the threshold, one or more suspected process defect types are associated and labeled in the needle position deviation data; Based on the suspected process defect type of the associated annotation, the calculation method of the health status score of the corresponding probe is adaptively adjusted; The types of process defects include uneven probe plating, sintering deformation of the ceramic substrate, or stress warping of the wiring layer.

[0014] Preferably, the system includes an optimization module, which is configured to: A central process knowledge base is constructed based on multi-probe card data labeled with suspected process defect types. The probe card data includes probe card pin position deviation data, probe health status score, and associated probe card metadata. The probe card metadata includes at least the probe card model, production batch number, and cumulative test cycle count. This knowledge base is used to establish a benchmark model of defect indicators and test cycle counts, and multi-dimensional correlation analysis is conducted based on this model to identify process anomalies. Based on the analysis results, a quantitative package of process parameter optimization suggestions is generated.

[0015] Preferably, the construction of a central process knowledge base and the establishment of a benchmark model include: Probe cards marked with at least one suspected process defect type are identified as defect probes. For each probe card, a defect severity index is calculated, which is a weighted function of the confidence level of all suspected process defect types and the inherent severity level. Based on historical data, benchmark function models were established to show the changes in the proportion of defect probes and the average defect severity index with the number of test cycles, and the standard deviation of each expected value was determined.

[0016] The multi-dimensional correlation analysis includes: When the actual proportion of defect probes or the actual average defect severity index of the same batch deviates from its expected value by more than r times the standard deviation, a manufacturing process anomaly alarm is generated. Track the standardized residuals of multiple consecutive batches of the same model in chronological order of production time; when the residuals of N consecutive batches all exceed the preset positive threshold, or when the slope value of the residual sequence fitted by linear regression is greater than zero and passes the significance test, it is determined to be a systematic process drift and an early warning is generated. Within the same range of test cycles, the defect indicators of different models are compared, and high-risk model-process defect combinations are identified by combining design features.

[0017] The beneficial effects of this invention include: By coordinating the entire process of morphology acquisition, analysis, compensation, and verification through a control module, fully automating pin inspection is achieved, significantly improving detection efficiency and consistency, and reducing reliance on manual experience. Employing a machine learning model (physical information neural network) integrating physical information for pin analysis not only quantifies pin deviation with high precision but also integrates historical data to comprehensively evaluate the mechanical health, electrical health, and future degradation trends of the probes, achieving a leap from detection to diagnosis and prediction. The compensation execution module calculates the optimal contact position based on pin deviation data corrected for confidence weights using a multi-objective adaptive optimization algorithm. This algorithm dynamically adjusts the optimization target weights according to the process information of the wafer under test, achieving precision and scenario-adaptive compensation strategies, effectively improving overall contact yield and protecting the delicate wafer structure. By linking microscopic pin position deviations with macroscopic manufacturing processes, the system can automatically identify collective deviation patterns caused by manufacturing defects. Through multi-dimensional correlation analysis, it can accurately locate the root cause of process anomalies and ultimately generate a quantitative process parameter optimization suggestion package. This transforms the system from a simple quality inspection tool into a powerful process optimization driving engine, which can continuously promote the improvement of probe card manufacturing processes and improve product quality and consistency from the source. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a fully automated pin position inspection system for probe cards provided in an embodiment of this application. Detailed Implementation

[0019] The present application will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0020] See Figure 1 This application provides a fully automated pin position inspection system for probe cards, comprising: The topography acquisition module is used to perform three-dimensional topography scanning on the probe group of the probe card and acquire three-dimensional point cloud data of the probe group; the data accurately records the coordinates (X,Y,Z) and attitude information of each probe tip in three-dimensional space; The needle position analysis module is connected to the morphology acquisition module and includes a machine learning model for processing the three-dimensional point cloud data and generating needle position deviation data and probe health status scores. The compensation execution module, connected to the needle position analysis module, is used to calculate the optimal contact position based on the needle position deviation data and probe health status score, and drive the stage to move to that position. The verification module is used to verify the accuracy of the needle position after contact is performed at the optimal contact position; The control module is used to coordinate the execution of the fully automated needle position inspection process by various modules.

[0021] In one possible implementation, the topography acquisition module includes a structured light scanner or a laser profilometer, which scans below the probe card via a moving mechanism to acquire three-dimensional point cloud data of the probe group.

[0022] The effects of the above technical solution are as follows: Traditional inspection methods only provide geometric information about positional alignment. This system, through machine learning models and multi-task learning, not only provides accurate deviation data but also performs comprehensive health scoring and prediction of probe mechanical damage and electrical performance degradation trends. This enables proactive early warning and predictive maintenance, effectively avoiding sudden test failures and reducing the risk of wafer scrap. The introduction of the concept of reliability weights makes compensation calculations more scientific, preventing "bad data" from low-health probes from interfering with overall alignment. A multi-objective optimization algorithm that dynamically adjusts weights based on process information can simultaneously pursue high yield while protecting the ultra-fine chip structure (by minimizing contact force), making it particularly suitable for testing advanced processes and 3D packaged chips. The system not only completes one-time probe alignment but also forms a continuous optimization closed loop through feedback from the verification module and correlation analysis of test yield data by the control module. It can trace collective deviations to manufacturing process defects and directly generate test program optimization suggestions, feeding back physical world findings to digital world test strategies, comprehensively improving the intelligence level and output quality of the testing process.

[0023] All data generated by the system (morphology, deviation, health score, historical records) is fully saved and analyzed, providing solid data support for probe card life management, precise repair or replacement, helping enterprises shift from traditional periodic preventive maintenance to more economical and efficient predictive maintenance, and significantly reducing the total cost of ownership associated with probe cards.

[0024] In one possible implementation, the machine learning model is a physical information neural network, which is configured as follows: The original 3D point cloud data is processed to generate a high-level feature representation corresponding to each probe; the high-level feature representation includes a numerical description for quantifying the probe's spatial pose, structural integrity, and population consistency. Historical performance data from the probe card status archive; Output the needle position deviation data and probe health status score.

[0025] The historical performance data includes historical contact resistance values, historical pin position deviation data, and the corresponding number of test cycles in a time series.

[0026] The numerical description used to quantify spatial attitude includes the yaw angle and pitch angle of the probe body relative to the theoretical position, as well as the spatial coordinate deviation of the probe tip. The numerical description used to quantify structural integrity includes feature values ​​calculated based on point cloud data that reflect the degree of probe bending, twisting, or wear. The numerical description used to quantify population consistency includes the height and orientation of individual probes and the standard deviation of the overall mean or mode of the probe swarm. The number of test cycles refers to the cumulative number of times the probe card goes through the complete process from loading the wafer, performing contact, completing electrical testing, to separation and unloading.

[0027] In one possible implementation, the machine learning model is a multi-head neural network, comprising: A branch of a point cloud convolutional neural network is used to extract spatial features from the original 3D point cloud. A branch based on a recurrent neural network is used to extract time-series degradation features from the historical performance data; A fusion layer is used to combine the features of the above branches and generate the final output.

[0028] In one possible implementation, the probe health status score is achieved by the physical information neural network through the following multi-task learning, including: A mechanical health assessment task is used to evaluate the deviation of a probe from a standard morphology based on the high-level feature representation and output a first score; the high-level features include a numerical description for quantifying the deviation of the probe from the standard morphology and a numerical description for quantifying population consistency; the first score is the result of a weighted fusion of the above two types of numerical descriptions. The electrical health assessment task is used to evaluate the stability and degradation trend of contact resistance based on fused historical performance data and output a second score; The trend prediction task is to predict the probability of morphological degradation after a specific number of cycles in the future based on the changes in morphological features over a time series, and output a third score. The physical information neural network obtains the final probe health status score by weighting and fusing the first score, the second score, and the third score through a shared feature extraction layer and their respective task-specific layers.

[0029] The effects of the above technical solution are as follows: Traditional methods can only provide isolated conclusions such as positional deviation or increased resistance. This system, through multimodal data fusion, can reveal the intrinsic relationship between physical morphology and electrical performance. For example, it can diagnose abnormal resistance fluctuations as being caused by micrometer-level bending of the probe, resulting in a reduced contact area, thus providing a root cause analysis of the fault.

[0030] Through trend prediction tasks, the system can not only report the current status of probes but also provide early warnings of future failure risks. This allows maintenance personnel to schedule maintenance or replacement in advance, before probe performance deteriorates completely and leads to a large number of wafer test failures. This shifts from reactive, reactive maintenance to proactive, predictive maintenance, greatly reducing the risk of production interruptions and wafer scrapping losses.

[0031] The multi-task learning framework enables the three evaluation tasks to share knowledge and mutually regularize, meaning that when evaluating mechanical health, the model is also constrained by electrical health data, and vice versa. This mechanism effectively prevents the model from overfitting to a single data source, making its evaluation results more robust and reliable. The final weighted fusion score provides a more comprehensive health metric than any single dimension. The resulting probe health status score is not an isolated number, but a key reliability weight; based on this, the system can prioritize relying on the bias data of healthy probes in subsequent compensation execution, reducing the impact of abnormal data from unhealthy probes, thereby making the global pose compensation calculation more accurate and intelligent, directly improving the test yield.

[0032] In one possible implementation, the trend prediction task is configured to simultaneously output prediction scores at multiple time scales, the time scales being divided based on the fatigue failure stages of the probe material: Short-term morphological stability score, corresponding to the initial wear and plastic deformation stabilization stage of the probe; Mid-term morphological degradation early warning score corresponds to the fatigue crack initiation and stable propagation stages of the probe; Long-term probe life end probability score, corresponding to the stage of accelerated fatigue crack propagation until fracture of the probe; The third score is a weighted composite of the short-term, medium-term, and long-term scores.

[0033] The weighted composite weight of the short-term, medium-term, and long-term scores is jointly determined by the cumulative cycle count factor and the real-time health status factor, including: Based on the ratio of the cumulative number of test cycles C_current of the probe card to its theoretical total lifespan C_total, R_c=C_current / C_total, a set of basic weights (W_s_base,W_m_base,W_l_base) corresponding to the short-term, medium-term and long-term are obtained. Based on the first score (mechanical health score) S1, a weight correction coefficient K is obtained; when S1 is lower than a preset health threshold, K increases the proportion of the short-term weight W_s_base while decreasing the proportion of the long-term weight W_l_base; the weight correction coefficient K is negatively correlated with the first score. The final weight is the result of adjusting the basic weight by the weight correction coefficient K.

[0034] The cumulative cycle count factor is obtained by querying a predefined weight mapping table.

[0035] For example: A set of basic weights (W_s_base, W_m_base, W_l_base) is obtained based on the ratio R_c, which is achieved through a predefined piecewise linear interpolation rule: For example: when 0 ≤ R_c < 0.3, it is defined as the initial stage of the lifetime: W_s_base = 1 - R_c / 0.3, W_m_base = R_c / 0.3, W_l_base = 0; When 0.3 ≤ R_c < 0.7, it is defined as mid-life: W_s_base = 0, W_m_base = 1 - (R_c - 0.3) / 0.4, W_l_base = (R_c - 0.3) / 0.4; When 0.7≤R_c≤1, it is defined as late lifespan: W_s_base=0, W_m_base=0, W_l_base=1; K = 1 - S1, where S1 is the normalized mechanical health score (0 indicates complete damage, 1 indicates brand new). The final weight is obtained through the following steps: Use coefficient K to amplify the short-term basic weights while suppressing the medium- and long-term weights: W_s_temp = W_s_base * (1 + α * K); α is the magnification factor, for example, 1.5 W_m_temp = W_m_base * (1 - β * K); β is the inhibition factor, for example, 0.5 W_l_temp = W_l_base * (1 - β * K) The sum of the scaled temporary weights is normalized to 1 to obtain the final weights.

[0036] The effects of the above technical solution are as follows: Traditional single-life prediction models are too coarse to reflect the actual fatigue process of the probe. This solution divides the prediction into three stages that correspond perfectly to materials science: short-term (initial wear), medium-term (crack propagation), and long-term (fracture risk). This transforms the prediction result from a vague point in time into a clear, phased failure evolution roadmap. This allows maintenance personnel to take precise measures matching the risk level, such as observation in the short-term stage, planned maintenance in the medium-term stage, and emergency replacement in the long-term stage, greatly improving the scientific rigor and effectiveness of maintenance strategies.

[0037] Through a dynamic weight adjustment mechanism, the system intelligently focuses on the most relevant risks. In the early stages of the probe's lifespan, it prioritizes short-term stability, helping to detect manufacturing defects or installation problems early. In the middle stages, it focuses on monitoring the risk of mid-term crack initiation. In the later stages, it prioritizes warnings about the probability of fracture to prevent catastrophic failures. This age-dependent focus adjustment avoids false alarms such as excessive warnings about "long-term fracture risks" when the probe is still healthy, and also prevents missed warnings such as focusing only on "short-term wear" when the probe is nearing failure. This significantly improves the signal-to-noise ratio and operability of the warning information.

[0038] When a probe experiences a sudden drop in its mechanical health score (S1) due to an accidental impact or overload, the system automatically increases the weight of the short-term score in the final decision by immediately increasing the K value. This means that the system can instantly identify and respond to this abnormal, accelerated degradation, quickly switching its focus from long-term risks to emergency warnings of immediate failure risks, thus achieving rapid response to emergencies. This robustness is not found in systems with fixed weights or those based solely on the number of iterations.

[0039] The output scores across multiple time scales are valuable decision-making data. For example, a probe might have acceptable short-term stability but a high risk of degradation in the medium term. Maintenance personnel can use this to determine that the probe does not require immediate replacement but must be thoroughly reviewed or replaced after specific future test cycles (e.g., after 50,000 cycles). This granular insight allows for more refined maintenance planning, minimizing unnecessary downtime and production interruptions while ensuring equipment reliability. Accurate long-term end-of-life probability prediction enables companies to more accurately forecast probe card replacement needs in the coming months, thus achieving precise spare parts inventory management.

[0040] In one possible implementation, the training process of the physical information neural network ensures the physical rationality of its output by introducing a physical constraint loss function.

[0041] The training process includes a physical constraint loss function, which, in addition to the regular data fitting error, calculates whether the needle position deviation output by the model conforms to the mechanical model of the probe cantilever beam, to ensure that the output results are physically reasonable.

[0042] In one possible implementation, the training process of the machine learning model includes a composite loss function consisting of two parts: Data fitting loss: used to minimize the general error between the model-predicted needle position deviation data, health status scores, and the actual labeled data.

[0043] Physical constraint loss: used to ensure that the needle position deviation in the model output conforms to the mechanical model of the probe cantilever beam, thereby giving the output results solid physical rationality.

[0044] The specific construction method of the physical constraint loss function is as follows: Based on the probe's geometric and material properties (including but not limited to elastic modulus, moment of inertia of section, and cantilever length), a cantilever beam mechanical model is established; this model describes the mathematical relationship between the stress at the probe root and the deformation at the probe tip under a given load.

[0045] The spatial attitude (mainly pitch angle) output by the model is used as the input of the mechanical model to back-calculate or estimate the tip contact force required to cause the attitude change, and then calculate the theoretical deformation of the probe under this force.

[0046] The physical constraint loss term is constructed by calculating the difference between the theoretical deformation and the deviation of the needle tip spatial coordinates (mainly the Z-direction deviation) directly output by the model. This loss term penalizes "attitude-deformation" combinations that do not hold true in the mechanical model, thereby forcing the model to learn characteristics that conform to physical laws.

[0047] This mechanism forces the output of the neural network to conform to the basic cantilever beam mechanics principle. This means that there must be a reasonable mechanical relationship between the needle tip coordinate deviation predicted by the model and the probe attitude. This effectively avoids absurd predictions that are physically impossible and may be generated by a purely data-driven model (for example, a huge pitch angle corresponding to a very small Z-direction displacement). It ensures that each needle position deviation report is not only statistically accurate but also physically reliable, providing a solid guarantee for subsequent high-risk compensation decisions.

[0048] In one possible implementation, the needle position analysis module is further configured as follows: Based on the probe health status score, a confidence weight is assigned to the spatial deviation vector of each probe in the needle position deviation data, and this weight is positively correlated with the probe health status score. For probes with a health status score below a preset threshold, the compensation amount in the needle position deviation data is marked as low confidence.

[0049] Its workflow is as follows: Input reception: The module simultaneously receives two outputs from the machine learning model: Needle position deviation data: contains the deviation vector (ΔX, ΔY, ΔZ) of each probe in three-dimensional space.

[0050] Probe health status score: A quantitative score that integrates mechanical, electrical, and trend predictions, reflecting the reliability of the probe's current state.

[0051] Credibility Weight Mapping: A predefined mapping function is used to convert the health status score of each probe into a credibility weight. This weight is positively correlated with the score, i.e.: High health score → high reliability weight: Probes in good condition are considered to have highly reliable spatial bias measurements, which will play an important role in subsequent calculations.

[0052] Low health score → low credibility weight: The reliability of the bias data of probes in poor condition is questionable, and their influence in subsequent calculations will be weakened.

[0053] For probes with health scores below a preset threshold (e.g., scores below 30), the system will mark their compensation data as low confidence.

[0054] Ultimately, the module outputs a set of needle position deviation datasets adjusted for confidence weights.

[0055] The effects of the above technical solution are as follows: Traditional methods treat all probes equally, and compensation calculations are equivalent to taking the geometric mean of all deviations. This approach, however, implements an intelligent weighted average, giving healthy probes greater weight and suppressing noise interference from unhealthy probes. This makes the calculated globally optimal contact position more likely to trust probes with reliable conditions, thus significantly improving the accuracy of the compensation position and the final contact yield.

[0056] The system can automatically identify and isolate unreliable data sources; even if some probes produce extremely large and abnormal deviation data due to severe wear, contamination, or damage, these outliers will not have a disruptive impact on the overall compensation decision because of their low health score and small confidence weight. This effectively prevents the risk of entire wafer testing failing due to the failure of a single or a few probes, and enhances the system's stability under non-ideal operating conditions.

[0057] Low-reliability marking is a powerful maintenance signal that directly tells maintenance personnel that the probe is not only misaligned, but more importantly, it may be faulty and its reported location information is unreliable. This upgrades maintenance decisions from identifying which probe has a large location deviation to identifying which probe is unreliable, guiding maintenance personnel to prioritize checking or replacing these marked probes. This enables accurate and efficient maintenance, saving time and costs.

[0058] In the testing process, the chip cells corresponding to these low-confidence probes can be pre-labeled. When these cells fail to test, the system can attribute it to a problem with the probe itself, rather than a chip defect, thus avoiding misjudgments and improving the accuracy of yield analysis. Simultaneously, retesting procedures can be designed for these cells to further improve overall output quality.

[0059] In one possible implementation, the compensation execution module calculates the optimal contact position through an optimization algorithm. The input of the optimization algorithm is the needle position deviation data after the confidence weight correction, and its objective function is configured to maximize the estimated overall contact yield.

[0060] The compensation execution module employs a confidence-weighted multi-objective adaptive optimization algorithm, which is configured as follows: A six-degree-of-freedom optimization model is established, which includes spatial position compensation and attitude adjustment. Its decision variables include the global translation compensation of the platform in the X, Y, and Z directions and the rotation compensation around the X and Y axes. Construct a multi-objective function that includes the following sub-objectives: The primary sub-objective is to maximize the weighted contact yield prediction function, which calculates the weighted contact yield by combining the magnitude of the spatial deviation vector of each probe, the horizontal offset from the center of the pad, and the corresponding confidence weight. Secondary sub-targets, minimizing the total contact force, are estimated by the total deformation of the probe group in the Z direction; Constrain sub-targets to ensure that the contact pressure uniformity index of all probes is below a preset threshold.

[0061] In one possible implementation, the weighted contact yield is calculated in the following manner: For each probe in the probe group, calculate its yield contribution estimate based on the individual probe position deviation compensation vector, and multiply this estimate by its corresponding confidence weight and sum them up to obtain the overall weighted contact yield estimate.

[0062] For each probe in the probe group, a basic yield contribution factor is calculated based on the magnitude of its spatial deviation vector and its horizontal offset from the center of the pad. The weighted yield contribution value of each probe is obtained by multiplying its basic yield contribution factor by its corresponding confidence weight. The weighted contact yield is obtained by summing the weighted yield contribution values ​​of all probes.

[0063] In one possible implementation, the weighted contact yield is obtained by the following formula: To estimate the overall contact yield, the range is (0,1), with the yield being higher the closer to 1. M is the total number of probes on the probe card. Let be the confidence weight of the i-th probe; Let i be the spatial deviation vector of the i-th probe; , , These are the deviation vectors. Components along the X, Y, and Z coordinate axes; The maximum allowable deviation is set according to the wafer pad size; This is the horizontal offset between the i-th probe and the center of the pad; The maximum allowed horizontal offset is set based on the pad radius.

[0064] The effects of the above technical solution are as follows: With the ultimate goal of maximizing the estimated overall test yield, the system comprehensively considers various factors (location, health status) that have a real impact on yield, ensuring that the compensation strategy always serves the final production target. Through reliability weighting, it automatically performs intelligent trade-offs, prioritizing the accurate contact of healthy probes while sacrificing those deemed "unreliable" or with low health. This means that even if some probes on the probe card are faulty or degraded, the system can still ensure the normal operation of the majority of healthy probes through optimization strategies, thereby preventing localized problems from spreading and causing the entire wafer to fail to test, greatly improving the system's availability and stability.

[0065] The multi-objective optimization mechanism avoids the side effects of solely focusing on yield. By minimizing the uniformity of total contact force and constraint pressure, it actively reduces the mechanical wear of the probes while pursuing high yield, extending their service life and effectively preventing physical damage to the wafer chips. This is crucial for advanced process chips with increasingly smaller feature sizes, achieving a balance between performance and safety. The six-degree-of-freedom model enables the system to perform not only simple translation compensation but also precise tilt adjustments, which is essential for correcting microscopic warpage of the probe card or the wafer itself and achieving optimal group coplanarity.

[0066] In one possible implementation, the compensation execution module is configured as follows: Receive process information of the wafer under test, wherein the process information includes at least one of the following: the feature size of the wafer, the thickness of the uppermost intermetallic dielectric layer, and the maximum pin density; Based on the process information, a multi-objective optimization function for calculating the optimal contact position is dynamically constructed or selected; The weights of different optimization objectives in the multi-objective optimization function are dynamically allocated based on the process information, including: When the feature size is smaller than the first size (e.g., 5 nanometers) and / or the thickness of the uppermost intermetallic dielectric layer is smaller than the second size (e.g., 100 nanometers), in the multi-objective optimization function, the weight of minimizing the total contact force objective is increased; in the weighted contact yield prediction function, the weight factor related to the magnitude of the spatial deviation vector is reduced; at the same time, the setting value of the maximum allowable contact force threshold in the optimization model is reduced.

[0067] When the maximum pin density is higher than the first density (e.g., 10,000 pins per square millimeter), the maximum weighted contact yield prediction function is kept as the highest weighted sub-objective; in the weighted contact yield prediction function, the weighting factor related to the horizontal offset is increased or the setting of the maximum allowable horizontal offset is decreased; at the same time, in the constraints, the requirement for contact pressure uniformity is tightened to avoid edge probe pressure concentration due to tilted contact.

[0068] When extremely small feature sizes and / or extremely thin dielectric layers are identified, the system determines that the wafer itself is structurally very fragile and cannot withstand large mechanical stresses. Therefore, it automatically increases the weight of the sub-objective of minimizing total contact force. In the weighted contact yield prediction function, it reduces the weight factor related to the magnitude of the spatial deviation vector, because excessive Z-axis displacement is the main cause of contact force overload. This means that the optimization algorithm will be more inclined to find poses that can significantly reduce the overall contact force even if a small amount of theoretical contact yield is sacrificed, thus placing the protection of the wafer from damage as the highest priority. At the same time, it reduces the setting value of the maximum allowable contact force threshold in the optimization model.

[0069] When an extremely high pin density is detected, the system determines that the alignment tolerance between the probe and the pad is very small. Even a slight horizontal offset can lead to a short circuit or poor contact with adjacent pads. Therefore, the system automatically increases the weighting factor related to the horizontal offset in the weighted contact yield prediction function, making the optimization algorithm more sensitive to horizontal deviations. It strives to avoid any large-scale horizontal misalignment that could cause a short circuit, prioritizing the improvement of alignment accuracy.

[0070] In one possible implementation, the needle position analysis module is further configured as follows: Based on the spatial distribution of the needle position deviation data described in the probe card, identify group deviations exhibiting specific patterns; The identified group deviation patterns are matched with a pre-stored database of manufacturing process defects for probe cards; When the matching degree exceeds the threshold, one or more suspected process defect types are associated and labeled in the needle position deviation data; Based on the suspected process defect type of the associated annotation, the calculation method of the health status score of the corresponding probe is adaptively adjusted; The types of process defects include uneven probe plating, sintering deformation of the ceramic substrate, or stress warping of the wiring layer.

[0071] The identification of group biases exhibiting specific patterns specifically includes: Spatial clustering and trend surface fitting are performed on the needle position deviation data; Extract the geometric features of the fitted trend surface, including but not limited to: the radius and direction of the overall curvature, the existence of a linear slope gradient, and local bulges or depressions in specific regions (such as edges or centers); The geometric features are matched with predefined group bias patterns.

[0072] When needle position deviation data is labeled as a specific suspected process defect type, the needle position analysis module is configured as follows: Based on the inherent characteristics of this process defect, the calculation method for the health status score of the affected probes is adjusted. For example, for deviations caused by sintering deformation of ceramic substrates, due to their high stability but uncompensable characteristics, the trend prediction score of the corresponding probe will be increased, while the mechanical health score will be decreased.

[0073] The diagnostic information provided by the system enables more refined operational decisions: for uneven plating, only cleaning or focused attention on specific probes is needed. For stress warping in the wiring layer, local repair or adjustment of test parameters is required.

[0074] For fundamental and irreversible defects such as sintering deformation of ceramic substrates, it is recommended to directly replace the probe card, avoiding wasting maintenance resources and time on unrepairable components. This root-cause-based decision-making achieves optimal resource allocation and effective cost control.

[0075] When the system confirms that a deviation stems from an uncompensable process defect, this information can be directly fed back to the testing procedure. For example, for areas with persistent poor contact due to substrate deformation, the testing procedure can mark them as systematically low-yield areas caused by hardware limitations. In the final data analysis, the impact of inherent hardware faults on the chip's yield can be removed, resulting in more accurate and purer chip performance data and improving the accuracy of product analysis.

[0076] In one possible implementation, the system further includes an optimization module, which is configured to: A central process knowledge base is constructed based on multi-probe card data labeled with suspected process defect types. The probe card data includes probe card pin position deviation data, probe health status score, and associated probe card metadata. The probe card metadata includes at least the probe card model, production batch number, and cumulative test cycle count. This knowledge base is used to establish a benchmark model of defect indicators and test cycle counts, and multi-dimensional correlation analysis is performed to identify process anomalies. Based on the analysis results, a quantitative process parameter optimization suggestion package is generated.

[0077] In one possible implementation, the construction of a central process knowledge base and the establishment of a benchmark model include: Probe cards marked with at least one suspected process defect type are identified as defect probes. For each probe card, a defect severity index is calculated, which is a weighted function of the confidence level of all suspected process defect types and the inherent severity level. Based on historical data, benchmark function models are established to show the changes in the proportion of defect probes and the average defect severity index with the number of test cycles, and the standard deviation σ of each expected value is determined. The benchmark function model can be in the form of an exponential function, a logarithmic function, or a quadratic function. The system selects the most suitable model for fitting based on the distribution characteristics of historical data.

[0078] In one possible manner, the defect severity index is obtained by the following formula: in, The score is the defect severity index, which can range from 0 to 10. The higher the score, the more severe the defect. P represents the number of process defect types marked on the probe card. The inherent severity level of defect type j is preset based on the degree of defect impact; for example, sintering deformation of the ceramic substrate. =3; uneven plating of the probe, LE=1; This is the maximum inherent severity level among all defect types; This represents the area percentage of the j-th type of defect, i.e., defect area / total probe card area; Let be the matching confidence level for the j-th type of defect, ranging from (0,1); The normalization coefficient is in the range of (1, 10), preferably 10.

[0079] The multi-dimensional correlation analysis includes: When the actual proportion of defect probes or the actual average defect severity index of the same batch deviates from its expected value by more than r times the standard deviation σ, a manufacturing process anomaly alarm is generated. Track the standardized residuals of multiple consecutive batches of the same model in chronological order of production time; when the residuals of g consecutive batches all exceed the preset positive threshold, or when the slope of the residual sequence fitted by linear regression is greater than zero and passes the significance test (p value < 0.05), it is determined to be a systematic process drift and an early warning is generated. Within the same range of test cycles, the defect indicators of different models are compared, and high-risk model-process defect combinations are identified by combining design features.

[0080] The proposed optimization suggestions for the generation process parameters include: Based on the correlation analysis results, the most likely manufacturing steps and process parameters are identified; Generate quantitative recommendations that include target process parameters, suggested adjustment directions, and confidence levels.

[0081] This solution transcends the traditional use of pin inspection merely for screening qualified products, transforming inspection data into a driving force for iterative optimization of manufacturing processes, thus achieving the ultimate goal of quality management. By establishing a dynamic benchmark model linked to the number of test cycles, it effectively distinguishes between manufacturing defects and natural wear, significantly reducing false alarms and missed alarms. The multi-dimensional analysis model, particularly its early warning of systemic process drift, can identify potential problems before they substantially impact yield, shifting from reactive response to proactive prevention. The comprehensive evaluation method using a defect severity index solves the complex problem of determining multiple coexisting defects, making decision-making more objective. The use of rigorous standards such as statistical significance testing (p-value) to replace human experience-based judgment ensures that alarms and warnings are generated with evidence, greatly enhancing the scientific rigor and authority of decision-making.

[0082] By improving yield, reducing scrap, preventing batch accidents, and optimizing process parameters to extend probe card life, production costs are significantly reduced in many ways, and production efficiency is greatly improved due to the automation and intelligence of the process.

[0083] In one possible implementation, the verification module is further configured as follows: Perform one or more micro-contact tests at the optimal contact position and collect real-time contact resistance value arrays and contact pressure distribution data for each test; Based on the contact resistance value array and contact pressure distribution data, the following verification indicators are calculated: Contact yield: The proportion of probes whose contact resistance falls within the preset acceptable range to the total number of probes; Contact uniformity index: the standard deviation or coefficient of variation of the contact resistance values ​​of all probes; Pressure uniformity: the entropy or kurtosis of contact pressure distribution data; When the contact yield is lower than the preset standard, or the contact uniformity index / pressure balance exceeds the allowable range, the needle position inspection is determined to have failed the verification, and a re-inspection process coordinated by the control module is triggered.

[0084] The verification module is also configured to: Each verified contact resistance value array is associated with and stored with the corresponding probe health status score to form historical performance data in the probe card status file, which is used to update and optimize the machine learning model in the needle position analysis module.

[0085] The effects of the above technical solution are as follows: The verification module performs micro-contact tests and collects real-time electrical (contact resistance) and physical (contact pressure) data, providing objective and quantitative criteria for evaluating the accuracy of the optimal contact position. When verification fails, the system automatically triggers a re-check process, forming a highly efficient internal self-calibration closed loop. This completely changes the limitations of traditional methods that rely on manual interpretation or a single data source, greatly improving the reliability and success rate of the final test results.

[0086] This system not only focuses on the geometric position of the probes (ensuring this through the topography acquisition and probe position analysis modules), but also directly assesses the electrical quality (contact yield, uniformity) and physical contact state (pressure equalization) of the contact through the verification module. This leap from mere resemblance to perfect alignment ensures that the compensated position is not only spatially accurate, but also achieves stable, uniform, and low-resistance contact, fundamentally guaranteeing the yield of wafer testing.

[0087] The verification module associates and stores each successful test data point with the probe's health status and feeds it back to the front-end machine learning model. This mechanism makes the entire system a continuously learning organism. As the amount of data accumulates, the model in the needle location analysis module can more accurately understand the complex relationship between morphological features, health status, and final electrical performance, thereby continuously improving its prediction accuracy and the intelligence level of compensation decisions, achieving the system's self-evolution.

[0088] In one possible implementation, the compensation execution module further includes a real-time compensation unit, used for: During the test, the probe position drift is monitored in real time based on the changes in contact resistance data from the verification module. The stage position is dynamically adjusted to compensate for the thermal expansion effect of the probe card caused by temperature changes inside the testing machine.

[0089] In one possible implementation, the real-time compensation unit is further configured as follows: Establish a mapping relationship between thermal expansion compensation and the internal temperature of the testing machine; At the start of subsequent tests, an initial compensation offset is preloaded based on the real-time monitored initial temperature of the machine, in order to shorten the dynamic fine-tuning response time and reduce the number of corrections during the test process.

[0090] The effects of the above technical solution are as follows: This unit solves the long-standing industry problem of probe drift caused by thermal expansion of the probe card due to internal temperature variations in the testing machine. Instead of treating the optimal contact position as a fixed point, it tracks and compensates for it in real time as a dynamically changing quantity throughout the testing process, ensuring that the contact between the probe and the pad remains optimal from the start to the end of the test.

[0091] In testing tasks involving a large number of test vectors or requiring long-term operation, an increase in internal temperature is inevitable. Traditional systems often experience a drop in test yield due to this. This real-time compensation unit effectively suppresses yield degradation caused by thermal effects through dynamic fine-tuning, ensuring the consistency and reliability of test results. It is particularly suitable for demanding testing scenarios such as automotive electronics and high-reliability chips.

[0092] This unit works closely with the verification module, utilizing contact resistance—the most direct electrical performance indicator—as a feedback signal to form a higher-level, end-to-end active control closed loop. This allows the system to not only optimize positioning during the inspection phase but also proactively maintain optimal contact during actual testing, elevating the system's intelligence and automation to a new level. Through real-time monitoring and compensation, the system's tolerance to external environmental disturbances (such as workshop temperature fluctuations and the testing machine's own thermal cycling) is significantly enhanced. This reduces the stringent requirements for the testing room's ambient temperature, improves the system's adaptability and robustness under different operating conditions, and reduces operating costs from another perspective.

[0093] In one possible implementation, the system further includes a vision-assisted positioning module comprising first and second camera units for taking pictures of the probe and wafer before or after the three-dimensional topography scan for preliminary coarse positioning.

[0094] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method in any embodiment of this invention or the function of the device in embodiment 1 of this invention.

[0095] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the invention without departing from the principles and spirit of the invention, and all such changes should fall within the protection scope of the claims of the present invention.

Claims

1. A fully automated pin position inspection system for probe cards, characterized in that, include: The topography acquisition module is used to perform three-dimensional topography scanning on the probe group of the probe card and acquire three-dimensional point cloud data of the probe group. The needle position analysis module is connected to the morphology acquisition module and includes a machine learning model for processing the three-dimensional point cloud data and generating needle position deviation data and probe health status scores. The compensation execution module, connected to the needle position analysis module, is used to calculate the optimal contact position based on the needle position deviation data and probe health status score, and drive the stage to move to that position. The verification module is used to verify the accuracy of the needle position after contact is performed at the optimal contact position.

2. The system according to claim 1, characterized in that, The machine learning model is a physical information neural network, which is configured as follows: The original 3D point cloud data is processed to generate a high-level feature representation corresponding to each probe; the high-level feature representation includes a numerical description for quantifying the probe's spatial pose, structural integrity, and population consistency. Historical performance data from the probe card status archive; Output the needle position deviation data and probe health status score; The historical performance data includes historical contact resistance values, historical pin position deviation data, and the corresponding number of test cycles in a time series.

3. The system according to claim 2, characterized in that, The probe health status score is achieved by the physical information neural network through the following multi-task learning, including: A mechanical health assessment task is used to evaluate the deviation of the probe from the standard morphology based on the high-level feature representation and output a first score; The electrical health assessment task is used to evaluate the stability and degradation trend of contact resistance based on fused historical performance data and output a second score; The trend prediction task is to predict the probability of morphological degradation after a specific number of cycles in the future based on the changes in morphological features over a time series, and output a third score. The physical information neural network obtains the final probe health status score by weighting and fusing the first score, the second score, and the third score through a shared feature extraction layer and their respective task-specific layers.

4. The system according to claim 2, characterized in that, The needle position analysis module is further configured as follows: Based on the probe health status score, a confidence weight is assigned to the spatial deviation vector of each probe in the needle position deviation data, and this weight is positively correlated with the probe health status score. For probes with a health status score below a preset threshold, the compensation amount in the needle position deviation data is marked as low confidence.

5. The system according to claim 4, characterized in that, The compensation execution module calculates the optimal contact position through an optimization algorithm. The input of the optimization algorithm is the needle position deviation data after the confidence weight correction, and its objective function is configured to maximize the estimated overall contact yield.

6. The system according to claim 5, characterized in that, The compensation execution module employs a confidence-weighted multi-objective adaptive optimization algorithm, which is configured as follows: A six-degree-of-freedom optimization model is established, which includes spatial position compensation and attitude adjustment. Its decision variables include the global translation compensation of the platform in the X, Y, and Z directions and the rotation compensation around the X and Y axes. Construct a multi-objective function that includes the following sub-objectives: The primary sub-objective is to maximize the weighted contact yield prediction function, which is calculated by combining the magnitude of the spatial deviation vector of each probe, the horizontal offset from the center of the pad, and the corresponding confidence weight. Secondary sub-targets, minimizing the total contact force, are estimated by the total deformation of the probe group in the Z direction; Constrain sub-targets to ensure that the contact pressure uniformity index of all probes is below a preset threshold.

7. The system according to claim 6, characterized in that, The compensation execution module is configured as follows: Receive process information of the wafer under test, wherein the process information includes at least one of the following: the feature size of the wafer, the thickness of the uppermost intermetallic dielectric layer, and the maximum pin density; Based on the process information, the weights of different optimization objectives in the multi-objective optimization function are dynamically allocated, including: When the feature size is smaller than the first size, and / or the thickness of the uppermost intermetallic dielectric layer is smaller than the second size, the weight of the sub-objective of minimizing the total contact force in the multi-objective optimization function is increased; in the weighted contact yield prediction function, the weighting factor of the magnitude of the spatial deviation vector is decreased. When the maximum pin density is higher than the first density, the weighting factor related to the horizontal offset of the pad center in the sub-objective of the maximum weighted contact yield prediction function is increased.

8. The system according to claim 3, characterized in that, The needle position analysis module is further configured as follows: Based on the spatial distribution of the needle position deviation data described in the probe card, identify group deviations exhibiting specific patterns; The identified group deviation patterns are matched with a pre-stored database of manufacturing process defects for probe cards; When the matching degree exceeds the threshold, one or more suspected process defect types are associated and labeled in the needle position deviation data; Based on the suspected process defect type of the associated annotation, the calculation method of the health status score of the corresponding probe is adaptively adjusted; The types of process defects include uneven probe plating, sintering deformation of the ceramic substrate, or stress warping of the wiring layer.

9. The system according to claim 8, characterized in that, The system includes an optimization module, which is configured to: A central process knowledge base is constructed based on multi-probe card data labeled with suspected process defect types. The probe card data includes probe card pin position deviation data, probe health status score, and associated probe card metadata. The probe card metadata includes at least the probe card model, production batch number, and cumulative test cycle count. This knowledge base is used to establish a benchmark model of defect indicators and test cycle counts, and multi-dimensional correlation analysis is conducted based on this model to identify process anomalies. Based on the analysis results, a quantitative package of process parameter optimization suggestions is generated.

10. The system according to claim 9, characterized in that, The construction of a central process knowledge base and the establishment of a benchmark model include: Probe cards marked with at least one suspected process defect type are identified as defect probes. For each probe card, a defect severity index is calculated, which is a weighted function of the confidence level of all suspected process defect types and the inherent severity level. Based on historical data, benchmark function models were established to show the changes in the proportion of defect probes and the average defect severity index with the number of test cycles, and the standard deviation of each expected value was determined. The multi-dimensional correlation analysis includes: When the actual proportion of defect probes or the actual average defect severity index of the same batch deviates from its expected value by more than r times the standard deviation, a manufacturing process anomaly alarm is generated. Track the standardized residuals of multiple consecutive batches of the same model in chronological order of production time; when the residuals of N consecutive batches all exceed the preset positive threshold, or when the slope value of the residual sequence fitted by linear regression is greater than zero and passes the significance test, it is determined to be a systematic process drift and an early warning is generated. Within the same range of test cycles, the defect indicators of different models are compared, and high-risk model-process defect combinations are identified by combining design features.

Images