Three-dimensional integrated circuit glass through hole defect detection method based on WOA-Light GBM

Abstract

The invention belongs to the field of integrated circuit testing, and particularly relates to a glass through hole defect detection method based on WOA-Light GBM. Aiming at the problems that the existing detection means is low in efficiency and high in cost and the precision depends on human experience, three-dimensional models of a TGV structure under different defect conditions are established through electromagnetic simulation software, the three-dimensional models comprise defect models of uneven through hole contours and multi-stage side wall roughness characteristics, and corresponding S parameters are extracted as judgment bases; and constructing a feature data set. On this basis, a LightGBM classification model is constructed, a whale optimization algorithm is adopted to carry out global search and optimization on key hyper-parameters, an optimal parameter combination is obtained, and automatic classification and identification of the defect TGV are realized. Therefore, according to the glass through hole defect detection method based on the WOA-LightGBM, high-precision detection of various types of defects can be achieved in a lossless mode, the algorithm training efficiency is high, the method is suitable for the large-scale production detection process, and flexibility is provided for manufacturing of high-standard and high-reliability glass through holes.

Description

Technical Field

[0001] The present invention relates to the field of three-dimensional integrated circuit defect testing, and in particular to a through-glass hole defect testing method based on a lightweight gradient boosting machine (WOA-LightGBM) using a whale optimization algorithm. Background Art

[0002] Three-dimensional integration is a technology with great potential, which can improve the performance and functionality of integrated circuits without increasing the chip footprint. At present, mainstream three-dimensional integrated circuits usually use through silicon via (TSV) technology to achieve vertical stacking of multiple chips. This method has the advantages of compact size, low energy consumption and support for heterogeneous integration. However, traditional TSVs are affected by substrate coupling, crosstalk and increased insertion loss under RF conditions. Therefore, in the RF / microwave field, glass is a better insertion material. Through glass via (TGV) has gradually become an important choice in high-frequency and high-reliability circuit applications due to the excellent electrical insulation properties, high transparency and low dielectric constant of glass materials. TGV has broad application prospects in three-dimensional packaging fields such as RF components, optoelectronic integrated circuits and MEMS devices.

[0003] Current TGVs are typically manufactured using a laser-induced wet etching process. This method can cause non-uniformity in the through-holes. Variations in the laser's focal length and power, as well as the effect of the laser pulse width, can lead to non-uniformity in the irradiated area. The chemical reactions of the wet etching process can shape the TGVs into cylindrical, conical, and hyperbolic profiles. During the wet etching process, irregularities in the chemical solution, deposits, and bubbles can form pits on the TGV sidewalls, resulting in sidewall roughness. Chemical reagents with varying compositions can lead to inconsistent etching rates, which in turn affects sidewall roughness. In high-density vertical interconnect structures, both the TGV profile and sidewall roughness can affect signal transmission. Therefore, in practical applications, TGV defect detection is an essential component to improving manufacturing accuracy.

[0004] With the continuous advancement of artificial intelligence (AI), machine learning (ML), one of its core branches, is being widely applied in key areas such as industrial inspection and intelligent manufacturing. Currently, the detection of through-glass via (TGV) structural defects still relies primarily on traditional methods such as X-ray imaging and acoustic microscopy. These methods suffer from high cost, low efficiency, and reliance on expert experience for accuracy, making them unable to meet the quality control requirements of future high-density packaging and high reliability. However, machine learning can uncover more complex microscopic defect signatures. If applied to TGV defect detection, machine learning models can extract subtle differences from large amounts of data that are difficult for humans to discern. This is crucial for detecting hidden defects that are difficult to detect using traditional methods. This not only improves detection sensitivity but also enables the early detection of potential problems, preventing serious defects from occurring later in the production process, significantly enhancing detection accuracy and efficiency. In this context, research on TGV defect detection using machine learning methods offers broad application prospects and exploration value for the TGV research field. Summary of the Invention

[0005] In view of the problems that traditional methods are not accurate enough and cannot meet the high standards of TGV manufacturing, the present invention provides a glass through-hole defect detection method based on WOA-LightGBM to solve the problem of TGV defects in the existing glass through-hole manufacturing process. The specific solution is to establish a three-dimensional simulation model of TGV non-uniform through-hole profile and different sidewall roughness in HFSS simulation software, extract the S parameters under different defects based on the defect model, construct a defect data set, use a supervised LightGBM machine learning method based on the whale optimization algorithm, train the S parameter data set of defective TGV, use the WOA-LightGBM model to classify and process it, and predict the type of defect. This invention can not only achieve high-precision detection of various types of defects without loss, but also has high algorithm training efficiency, and is suitable for large-scale production detection processes.

[0006] The present invention is achieved through the following technical solutions:

[0007] A physical model of a defect-free TGV was established in the HFSS full-wave electromagnetic simulation software, and the S parameters of the defect-free TGV were extracted;

[0008] Establish a defect-free TGV equivalent circuit model in ADS simulation software and extract the S parameters of the defect-free TGV;

[0009] The relative error analysis of the S parameters obtained by HFSS simulation software and ADS simulation software was performed to verify the effectiveness of the model;

[0010] Taking the common defect types in the TGV manufacturing process as the research object, a three-dimensional simulation model of TGV uneven through-hole shape and different sidewall roughness defects was established in HFSS simulation software.

[0011] Based on the established defect model, the S parameters under different defect types are extracted;

[0012] The S-parameters are used to analyze the effects of different via shapes and sidewall roughness on TGV signal transmission performance.

[0013] Design different TGV through-hole taper sizes, and use the taper angle size as the basis for data classification and set labels;

[0014] Design different TGV through-hole sidewall roughness levels, and use the roughness size as the basis for data classification and set labels;

[0015] A through-glass hole defect detection method based on WOA-LightGBM, which uses the gradient boosting decision tree algorithm to find the optimal parameters for optimization and obtain the optimal parameters;

[0016] In this embodiment, the implementation process of the TGV defect type detection and classification algorithm based on the LightGBM classification model includes:

[0017] 1. Establish a LightGBM classification model;

[0018] 2. Read sample data for TGV defect type detection from the data source. Data preprocessing includes data cleaning, handling missing values, and handling outliers.

[0019] 3. Evaluate the importance of all features. Determine the importance of defect features for TGV defect classification. Select features with greater than average importance based on their importance values ​​to reduce redundant features.

[0020] 4. Separate features and labels. Divide the data into features and labels. Features are the input data used to train the model, while labels are the target outputs predicted by the model. Encode the category labels and convert them into numeric types.

[0021] 5. Normalize the input features and scale the feature values ​​to a given range;

[0022] 6. Set the whale optimization method and set parameters such as the initial population of WOA, the maximum number of iterations, and the search step size.

[0023] 7. Determine the hyperparameter range that needs to be optimized for the LightGBM model, including learning rate, number of leaf nodes, maximum depth, minimum number of samples per leaf node, subsample ratio, etc.

[0024] 8. Use WOA to optimize hyperparameters: Train the LightGBM model based on the current hyperparameter configuration and evaluate its performance on the validation set. Use metrics such as accuracy, precision recall, and average as optimization targets and calculate the fitness value for each set of hyperparameter configurations.

[0025] 9. Use the entire training set to train the LightGBM model using the best hyperparameter combination obtained through WOA optimization. After training, perform classification prediction on the test set. If the accuracy of the current model is higher than the historical best accuracy, update the best parameter combination.

[0026] 10. Train the final model using the optimal hyperparameter combination and perform classification predictions on the test set. Output a classification report, confusion matrix, and evaluation metrics such as accuracy, recall, precision, and average, ultimately generating a classification result for the TGV defect type.

[0027] The present invention provides a through-glass hole defect detection method based on WOA-LightGBM. By training the S parameters in the TGV model of various defects existing in the process manufacturing, the WOA-LightGBM model is used to classify the defects according to different S parameters, and the type of defects is predicted, thereby improving the accuracy of TGV defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 FIG1 is an implementation structure diagram of the through-glass hole defect detection method based on WOA-LightGBM of the present invention;

[0029] Figure 2 Diagram of a fault-free TGV model built for HFSS simulation software;

[0030] Figure 3 The S parameter curve of TGV simulating different sidewall roughness

[0031] Figure 4 Optimize the LightGBM model flow chart for the Whale Optimization Algorithm; DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0033] like Figure 1 As shown in FIG, the present invention provides a flow chart of a method for detecting defects in glass through holes based on WOA-LightGBM. Figure 1 Specifically, the WOA-LightGBM-based through-glass hole defect detection method may include the following steps:

[0034] A physical model of a defect-free TGV under ideal conditions was established in the HFSS simulation software, and the S parameters of the defect-free TGV were extracted;

[0035] like Figure 2 As shown in the figure, under ideal conditions, the through-hole shape of the defect-free TGV is cylindrical, its conductive metal material is copper, and the material of the glass interposer is borosilicate glass (BF33);

[0036] The specific physical geometric parameters of TGV modeling are as follows: the height H of TGV TGV =220um, radius R of the copper column TGV =10um, the height of the glass interposer is H glass =220um, and the bottom length and width of the glass interposer are both 300um; an equivalent circuit model of a defect-free TGV is established in ADS simulation software, and the S parameters of the defect-free TGV are also extracted;

[0037] The S parameters of the defect-free TGV in HFSS simulation software and ADS simulation software were extracted for error analysis to verify the effectiveness of the proposed model.

[0038] Today's TGVs are typically manufactured using a laser-induced wet etching (LIWE) process. This method can cause non-uniformity in the through-holes. Changes in the laser's focal length and power, as well as the influence of the laser pulse width, can lead to non-uniformity in the irradiated area. Furthermore, the chemical reaction of the wet etching can shape the TGV into cylindrical, conical, and hyperbolic profiles.

[0039] During the wet etching process, irregularities in the chemical solution, deposits, and bubbles can form pits on the TGV sidewalls, resulting in sidewall roughness. Chemical reagents with different compositions can lead to inconsistent etching rates, which in turn affects sidewall roughness. Furthermore, because laser etching creates an uneven via profile, a larger taper angle concentrates the laser energy at the top of the via, while the etching effect at the bottom is weaker. This uneven etching results in varying degrees of sidewall roughness. At high frequencies, when the skin depth approaches the sidewall roughness, the TGV may experience signal scattering and reflection, leading to signal attenuation. In high-density vertical interconnect structures, both the TGV profile and sidewall roughness affect signal transmission.

[0040] Current research on the signal transmission performance of TGVs mainly focuses on single-layer 2.5D / 3D vertical interconnect structures. This paper proposes to establish a TGV physical model with through-holes of different profiles and different sidewall roughness to study the impact of non-uniform TGVs on the microwave performance of integrated circuits.

[0041] In the simulation, hemispheres with a radius of Rq are periodically embedded in the outer surface of the copper pillar to simulate the sidewall roughness. The hemisphere model is an ideal simplification that greatly simplifies the complex sidewall shape. By assuming that the sidewall roughness consists of a series of hemispherical pits, the complex surface topography can be simplified to a few basic geometric shapes, facilitating analysis and calculation. Sidewall roughness is typically characterized at the microscale, and the hemisphere model can be used to approximate these microscopic pit-like shapes. For most applications, the hemisphere model provides sufficient accuracy for describing the TGV sidewall roughness.

[0042] like Figure 3 Shown are the S-parameter curves of TGVs simulating through-holes with different sidewall roughness.

[0043] Taking the uneven TGV through-hole profile and the different sidewall roughness of the through-hole as the research objects, a three-dimensional simulation model of the uneven TGV through-hole profile and different sidewall roughness was established in the HFSS simulation software.

[0044] The S-parameters are used to analyze the impact of through-holes with different tapers and sidewall roughness on TGV signal transmission performance.

[0045] Based on the TGV model with different taper profiles, the S parameters of TGVs with different defects were extracted. The defect of TGV non-uniform through-hole is characterized by the size of the taper angle θ. The simulation settings are set to simulate TGV non-uniform through-hole defects with different taper angles.

[0046] Based on the specific TGV models with different sidewall roughness, the S parameters of different defects were extracted. The defect representation of the TGV sidewall roughness is a hemisphere model with a radius of Rq embedded on the outer surface of the copper pillar. The simulation settings are set to simulate the defect TGV with different sidewall roughness.

[0047] After the simulation is completed, the S parameter data is exported into a numerical table file through Execl. The data is classified based on different defect types, and three defect type labels are set.

[0048] In an embodiment of the present invention, LightGBM is a machine learning method based on the gradient boosting decision tree (GBDT). GBDT itself is an integrated learning method that improves the performance of the model by constructing a series of decision trees. Each tree takes into account the error of the previous tree when it is constructed, and continuously adjusts the model through the gradient descent method to minimize the loss function. LightGBM greatly outperforms traditional GBDT in training speed and memory usage by introducing technologies such as histogram optimization, leaf-first growth, and category feature support. It has obvious advantages in improving tree training efficiency, reducing memory usage, and improving prediction accuracy. It is one of the most effective algorithms for handling complex machine learning tasks.

[0049] like Figure 4 As shown, the specific process is:

[0050] 1. Establish a LightGBM classification model;

[0051] 2. Read sample data for TGV defect type detection from the data source. Data preprocessing includes data cleaning, handling missing values, and handling outliers.

[0052] 3. Evaluate the importance of all features. Determine the importance of defect features for TGV defect classification. Select features with greater than average importance based on their importance values ​​to reduce redundant features.

[0053] 4. Separate features and labels. Divide the data into features and labels. Features are the input data used to train the model, while labels are the target outputs predicted by the model. Encode the category labels and convert them into numeric types.

[0054] 5. Normalize the input features and scale the feature values ​​to a given range;

[0055] 6. Set the whale optimization method and set parameters such as the initial population of WOA, the maximum number of iterations, and the search step size.

[0056] 7. Determine the hyperparameter range that needs to be optimized for the LightGBM model, including learning rate, number of leaf nodes, maximum depth, minimum number of samples per leaf node, subsample ratio, etc.

[0057] 8. Use WOA to optimize hyperparameters: Train the LightGBM model based on the current hyperparameter configuration and evaluate its performance on the validation set. Use metrics such as accuracy, precision recall, and average as optimization targets and calculate the fitness value for each set of hyperparameter configurations.

[0058] 9. Use the entire training set to train the LightGBM model using the best hyperparameter combination obtained through WOA optimization. After training, perform classification prediction on the test set. If the accuracy of the current model is higher than the historical best accuracy, update the best parameter combination.

[0059] 10. The final model is trained using the optimal hyperparameter combination and classified on the test set. A classification report, confusion matrix, and evaluation metrics such as accuracy, recall, precision, and average are output, ultimately generating a classification result for TGV defect types. The calculated classification accuracy reaches 98.91%.

[0060] The glass through-hole defect detection method based on WOA-LightGBM proposed in the present invention has an accuracy rate of 98.91% for detecting defect types occurring in TGV. This detection method avoids direct contact with the TGV during the defect detection process to cause damage, and solves the current problem of vacancies in the field of TGV process manufacturing defect detection and low detection accuracy. It has certain reference value for optimizing the process manufacturing of TGV.

Claims

1. The purpose of the present invention is to provide a through-glass via defect detection method based on WOA-LightGBM to solve the problem of vacancies in the existing three-dimensional integrated circuit testing field. The present invention is achieved through the following technical solutions: A defect physical model of TGV was established in the full-wave electromagnetic simulation software HFSS, and the S parameters of the defect-free TGV were extracted; An equivalent circuit model of a fault-free TGV was established in ADS simulation software, and the S parameters of the defect-free TGV were extracted; The S parameters obtained by simulation using HFSS simulation software and ADS simulation software are subjected to relative error analysis to verify the effectiveness of the model.

2. The through-glass hole defect detection method based on WOA-LightGBM according to claim 1, characterized in that: Taking the common defect types in the TGV manufacturing process as the research object, a three-dimensional simulation model of TGV uneven through-hole shape and different sidewall roughness defects was established in HFSS simulation software. Based on the established defect model, the S parameters of TGV under different defect types are extracted; S parameters are used to analyze the effects of different through-hole shapes and different sidewall roughness on TGV signal transmission performance.

3. The through-glass hole defect detection method based on WOA-LightGBM according to claim 2, characterized in that: Design the taper size of TGV through-holes, and use the taper angle as the basis for data classification and set labels; Design the roughness of the TGV through-hole sidewall, and use the roughness as the basis for data classification and set labels; Based on the gradient boosting decision tree algorithm, the optimal parameters are found for optimization to obtain the optimal parameters.

4. The through-glass hole defect detection method based on WOA-LightGBM according to claim 3 is characterized by: Build a LightGBM classification model; Read sample data for TGV defect type detection from the data source. Data preprocessing includes data cleaning, handling missing values, and handling outliers. Evaluate the importance of all features. Determine the importance of defect features for TGV defect classification. Select features with greater than average importance based on their importance values ​​to reduce redundant features.

5. The through-glass hole defect detection method based on WOA-LightGBM according to claim 4 is characterized by: Separate features and labels, dividing the data into features and labels. Features are the input data used to train the model, while labels are the target outputs predicted by the model. Encode the category labels and convert them to numeric types. Normalize the input features and scale the feature values ​​to a given range.

6. The through-glass hole defect detection method based on WOA-LightGBM according to claim 5 is characterized by: Set the whale optimization method and set WOA parameters such as the initial population, maximum number of iterations, and search step size. Determine the hyperparameter range that needs to be optimized for the LightGBM model, including learning rate, number of leaf nodes, maximum depth, minimum number of samples per leaf node, subsample ratio, etc.

7. The through-glass hole defect detection method based on WOA-LightGBM according to claim 6 is characterized by: Optimize hyperparameters using WOA: Train the LightGBM model based on the current hyperparameter configuration and evaluate its performance on the validation set. Use metrics such as accuracy, precision recall, and average as optimization targets and calculate the fitness value for each set of hyperparameter configurations.

8. The through-glass hole defect detection method based on WOA-LightGBM according to claim 7 is characterized by: The best hyperparameter combination obtained through WOA optimization is used to train the LightGBM model using the entire training set. After training, classification prediction is performed on the test set. If the accuracy of the current model is higher than the historical best accuracy, the best parameter combination is updated.

9. The through-glass hole defect detection method based on WOA-LightGBM according to claim 8 is characterized by: The final model is trained using the optimal hyperparameter combination and classified on the test set. A classification report, confusion matrix, and evaluation metrics such as precision, recall, precision-recall, and average are output, ultimately generating a classification result for the TGV defect type.

Images