Glass via defect detection method based on PSO-XGBoost

Abstract

The present application belongs to the field of integrated circuit package testing, and particularly relates to a TGV defect detection method based on PSO-XGBoost. A defect model is established by taking the cavity, open circuit and composite defects in TGV as the research object, and the S parameters of different defects are extracted based on the established defect model; the TGV cavity defect radius, TGV open circuit defect length and TGV composite defect combination form are taken as the data classification basis, and the labels are set; a PSO-XGBoost classification model is established, the particle swarm optimization (PSO) algorithm is used to optimize the extreme gradient boosting (XGBoost), and the best hyperparameter is found based on the PSO-XGBoost algorithm; the S parameters of different defects are classified and processed by using the PSO-XGBoost model, the defect type is predicted, and the TGV defect detection accuracy is improved. The detection method avoids the secondary damage of TGV in the defect detection process, solves the problems that the traditional TGV detection method is difficult to accurately detect the micro defect and composite defect and has low detection efficiency, and has certain reference value for optimizing the design and manufacturing process of TGV, improving the TGV defect and improving the three-dimensional package yield.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional integrated circuit defect testing, specifically to a glass via defect detection method based on PSO-XGBoost (Particle Swarm Optimization, eXtreme Gradient Boosting). Background Technology

[0002] 3D integration is an important approach to achieving high-performance integrated circuits with more functions and reduced chip footprint. Currently, 3D integrated circuits mainly use through-silicon vias (TSVs) to achieve vertical stacking of multiple wafers, which has advantages such as small size, low power consumption, and support for heterogeneous integration. However, the semiconductor properties of silicon easily introduce parasitic capacitance, substrate coupling, and high-frequency losses, and the size and alignment accuracy requirements of TSVs themselves limit the development of higher-density integration. Through-glass vias (TGVs), with their excellent electrical insulation, low dielectric constant, and low insertion loss, are gradually becoming ideal vertical interconnect media in fields such as RF packaging and optoelectronic integration, which can further improve system integration and signal transmission performance.

[0003] Compared to TSV-based 3D integration technology, TGV technology can significantly improve signal insulation performance and integration density, attracting widespread attention in the industry. However, in the laser etching, copper electrodeposition, and subsequent thermal cycling processes of TGV, various defects are easily generated due to fluctuations in process parameters and the difference in thermal expansion coefficients between copper and glass. Uneven electrodeposition can easily form central voids, and void extension or crack propagation can directly lead to open circuit failures. The coexistence of multiple defects such as voids, cracks, and abnormal apertures can form composite defects. These defects can interrupt electrical paths, increase interconnect resistance, and reduce structural mechanical strength, seriously affecting the reliability of TGV interconnects and packaging yield. Therefore, establishing efficient and accurate detection methods for TGV defects is crucial to ensuring the performance of 3D packaging.

[0004] With the advent of the artificial intelligence era, machine learning (ML) has demonstrated significant advantages in complex feature recognition and fault classification. Currently, defect detection in TGV (Transmitted Telephone Vehicle) systems still relies primarily on traditional methods such as X-ray imaging and acoustic microscopy, which suffer from high costs, low efficiency, insufficient sensitivity to minute voids and early open-circuit defects, and difficulty in identifying the complex features of compound defects. Conventional machine learning methods depend on manual parameter tuning based on experience, making them prone to getting trapped in local optima and limiting their accuracy and generalization ability for detecting multiple types of defects. Applying a method combining intelligent optimization and ensemble learning to TGV defect detection would enrich defect detection methods, improve the accuracy of identifying voids, open circuits, and compound defects, and reduce yield losses during manufacturing. Against the backdrop of the rapid development of 3D ensemble technology, research on TGV defect detection based on intelligent optimization ensemble learning will become a highly valuable research direction in the TGV field. Summary of the Invention

[0005] The purpose of this invention is to provide a PSO-XGBoost-based TGV defect detection method. This method involves establishing three-dimensional simulation models of TGV void defects, open-circuit defects, and composite defects, extracting RF transmission characteristic parameters of TGVs under different defect types, and constructing a defect dataset. The Particle Swarm Optimization (PSO) algorithm is used to globally optimize the hyperparameters of the Extreme Gradient Boosting (XGBoost) classifier, forming a PSO-XGBoost defect classification model. Training this model enables accurate classification and identification of TGV void, open-circuit, and composite defects, solving the problems of low accuracy and poor efficiency in traditional detection methods, and significantly improving the accuracy and real-time performance of TGV defect detection.

[0006] To achieve the above objectives, this invention provides a TGV defect detection method based on PSO-XGBoost, comprising:

[0007] A three-dimensional physical model of a defect-free TGV was established in HFSS simulation software, and the S-parameters of the defect-free TGV were extracted based on the RF transmission characteristics.

[0008] Taking voids, open circuits and composite defects in the TGV fabrication process as the research object, three-dimensional simulation models of TGV defects of different sizes and locations were established in HFSS.

[0009] Based on the established defect simulation model, S-parameters corresponding to different types and degrees of defects are extracted.

[0010] The effects of cavity size, open circuit location, and combination of composite defects on TGV signal transmission were analyzed using the extracted S-parameters.

[0011] Obtain the radius of TGV void defects and set labels based on defect type;

[0012] Obtain the length of TGV open circuit defects and set labels based on defect type;

[0013] Obtain TGV composite defect combinations and set labels based on defect type;

[0014] A PSO-XGBoost-based TGV defect detection method is proposed, which optimizes the XGBoost classifier based on the PSO algorithm to find the optimal parameters and obtain the optimal parameters.

[0015] In this implementation, the acquisition of sample data based on the PSO-XGBoost algorithm to find the optimal parameters includes:

[0016] Establish an XGBoost classification model;

[0017] The sample data is read and input into the XGBoost classification model, and preprocessed.

[0018] The importance of all features was evaluated using the XGBoost method, and features with greater importance than the average importance were selected.

[0019] This helps reduce the number of features, improves model performance, and reduces the risk of overfitting;

[0020] Separate features and labels, and encode the category labels;

[0021] Convert the category label to an integer;

[0022] Normalize features: Scaling features to a given range of 0 to 1 helps the model converge better;

[0023] Determine the range and step size of the XGBoost hyperparameters to be searched;

[0024] Based on the determined parameter range and step size, global optimization is performed using the PSO algorithm to generate candidate hyperparameter combinations;

[0025] For each set of hyperparameters obtained by PSO optimization, use these parameters to train the XGBoost model and evaluate the model's performance on the validation set.

[0026] During the optimization process, the model performance evaluation results corresponding to each set of hyperparameters are recorded, and the set of hyperparameters with the best performance is found.

[0027] Train the XGBoost model using the best random seed and best hyperparameters, make predictions on the test set, and output a classification report and accuracy.

[0028] If the current accuracy is higher than the historical best accuracy, then update the optimal hyperparameter combination;

[0029] Finally, the XGBoost model was retrained using the found optimal combination of hyperparameters, and its performance was evaluated on the test set.

[0030] Output the best result.

[0031] The present invention provides a PSO-XGBoost-based TGV defect detection method. By training the S-parameters in the TGV model for void defects, open circuit defects, and composite defects, and using different S-parameters, the PSO-XGBoost model is used to classify the defects and predict the type of defects, thereby improving the accuracy of TGV defect detection. Attached Figure Description

[0032] Figure 1 This is a schematic flowchart of the TGV defect detection method based on PSO-XGBoost of the present invention.

[0033] Figure 2 A diagram of a defect-free GSG-TGV model created for HFSS simulation software;

[0034] Figure 3 Model diagrams of TGV void defects, open circuit defects, and composite defects created for HFSS simulation software;

[0035] Figure 4 S-parameter plots for 500 groups of void defects with different radii;

[0036] Figure 5 S-parameter plots for 500 groups of open-circuit defects of different lengths;

[0037] Figure 6 S-parameter plots for 500 different combinations of composite defects;

[0038] Figure 7 Optimize the XGBoost flowchart for the particle swarm optimization algorithm;

[0039] Figure 8 Confusion matrix diagram for defect detection in PSO-XGBoost model;

[0040] Figure 9 ROC curve for defect detection in the PSO-XGBoost model. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] like Figure 1 The diagram shown is a flowchart illustrating the TGV defect detection method based on PSO-XGBoost provided by this invention. Figure 1 As shown, the specific PSO-XGBoost-based TGV defect detection method may include the following steps:

[0043] A three-dimensional physical model of a defect-free GSG-TGV (ground-signal-ground) was established in HFSS simulation software, and the S-parameters of the defect-free GSG-TGV were extracted.

[0044] like Figure 2 As shown, the defect-free GSG-TGV physical model is a hyperbolic conical through-hole array structure. The overall layout adopts a ground-signal-ground (GSG) symmetrical layout, which includes two symmetrically arranged ground TGVs (G-TGV) and a centrally located signal TGV (S-TGV).

[0045] The through-hole carrier is borosilicate glass, and copper metal interconnect layers are disposed on the upper and lower surfaces of the glass substrate. The metal layer material is copper.

[0046] The specific parameters of the TGV are: the length (hole depth) of a single TGV L = 200 μm, the radius of the copper conductive pillar at the upper and lower ends R = 18 μm, and the radius of the copper pillar at the middle of the hole depth r = 12 μm;

[0047] The borosilicate glass substrate has a thickness of 200 μm, a length of 200 μm, and a width of 250 μm; the metal interconnect layer has a thickness of 5 μm; there is no gap between the copper pillars and the glass substrate to ensure the integrity of the vertical interconnect.

[0048] The essence of TGV is a set of vertical transmission lines that run through the glass substrate from top to bottom, which are used to realize the electrical interconnection and signal shielding of the upper and lower metal layers. The data used for training, testing and classification in this paper are all based on the relevant parameters extracted from this TGV model.

[0049] The core focus of TGV transmission performance analysis is the transmission state of the intermediate TGV signal. Common TGV defects mainly include signal TGV void defects, open-circuit defects, and composite defects. For example... Figure 3As shown, from left to right, the models are signal TGV void defects, open circuit defects, and composite defects.

[0050] Void defects mainly refer to the incomplete copper filling inside the signal TGV during the copper electrodeposition filling process of the TGV, due to fluctuations in electroplating process parameters, uneven electrolyte flow, etc., resulting in hollow cavities. During the subsequent packaging process, due to the effects of thermal cycling and mechanical stress, void defects may expand, directly affecting the electrical transmission performance and structural strength of the signal TGV, leading to increased interconnect resistance and aggravated signal transmission attenuation.

[0051] The open circuit defect is mainly caused by problems such as over-etching and metal layer breakage during the laser etching process, which causes the signal TGV copper pillar to be completely cut off, forming a state of no connection between the upper and lower parts. In the open circuit state, the low frequency electrical signal cannot be vertically transmitted through the intermediate signal TGV, and only the high frequency signal can be weakly transmitted between the upper and lower metal layers through coupling.

[0052] Composite defects mainly refer to the coexistence of multiple defects such as voids and open circuits. Their impact on TGV signal transmission is more complex. They are mostly caused by accumulated process errors and have a more complex effect on TGV signal transmission performance. The detection difficulty is much higher than that of single defects. They can lead to severe signal attenuation, transmission distortion, and even device failure. At the same time, they may destroy the shielding integrity of the GSG configuration and introduce additional crosstalk.

[0053] Taking the voids, open circuits, and composite defects that appear inside the TGV as the research object, a three-dimensional simulation model of the void defects, open circuit defects, and composite defects of the TGV was established in HFSS simulation software.

[0054] The impact of defects of different degrees on TGV signal transmission performance was analyzed using S-parameters.

[0055] like Figure 4 As shown, based on the established void defect model, S-parameters of different defects were extracted. The excitation frequency range was set to 1–20 GHz with a step size of 0.1 GHz. Simulations were performed on 500 groups of different void defects with radii ranging from 0.1 to 10.1 μm and a step size of 0.02 μm. The S-parameters of the 500 groups of void defects with different radii are shown below. Figure 4 As shown;

[0056] like Figure 5 As shown, based on the established open-circuit defect model, S-parameters of different defects were extracted. The excitation frequency range was set to 1–20 GHz with a step size of 0.1 GHz. Simulations were performed on 500 groups of different open-circuit defects with lengths ranging from 0.1 to 5.1 μm and a step size of 0.01 μm. The S-parameters of the 500 groups of open-circuit defects with different lengths are shown below. Figure 5 As shown;

[0057] like Figure 6 As shown, based on the established composite defect model, S-parameters of different defects were extracted. The excitation frequency range was set to 1–20 GHz with a step size of 0.1 GHz. Simulations were performed on 500 sets of composite defects with different combinations of cavity radius and open circuit length. The S-parameters of the 500 sets of composite defects with different combinations are shown below. Figure 6 As shown.

[0058] The chart-type data was converted into numerical data, the S-parameter data was extracted, and the data was classified according to different defect types. Labels were set, with 500 data sets in each category, so that the PSO-XGBoost machine learning algorithm could be used for data processing later.

[0059] In this embodiment of the invention, PSO-XGBoost combines the global optimization capability of the particle swarm optimization algorithm with the strong fitting advantage of the extreme gradient boosting algorithm. XGBoost, as a multi-decision tree ensemble model, features high fitting accuracy, strong generalization ability, and resistance to overfitting; the PSO algorithm achieves global optimization in the hyperparameter space through swarm intelligence, offering advantages such as fast convergence and simple parameter settings. The combination of these two algorithms adaptively optimizes the core hyperparameters of XGBoost, solving the problems of easily getting trapped in local optima and low efficiency in manual parameter tuning, and significantly improving the model's classification performance.

[0060] like Figure 7 As shown, the specific process for building and training the PSO-XGBoost model is as follows:

[0061] Build an XGBoost classification model; read sample data, input it into the model, and complete preprocessing;

[0062] XGBoost is used to evaluate the importance of all features, and core features with greater importance than the average importance are selected to reduce feature redundancy and the risk of overfitting.

[0063] Separate features and labels and encode them with integers; normalize the features and scale the feature values ​​to the range of 0 to 1 to improve the model convergence efficiency.

[0064] Determine the optimization range and step size of XGBoost hyperparameters, and generate candidate hyperparameter combinations through global optimization using the PSO algorithm;

[0065] Train the model for each set of hyperparameters and evaluate its performance on the validation set; record and select the optimal hyperparameters.

[0066] The model is trained using the optimal random seed and optimal hyperparameters, and it predicts on the test set and outputs a classification report and accuracy.

[0067] If the current accuracy is higher than the historical best, update the optimal hyperparameter combination; finally, retrain the model with the optimal parameters and evaluate its performance.

[0068] Output the best detection result.

[0069] The confusion matrix of the classification results is shown below. Figure 8 As shown in the figure, the ROC curve is as follows: Figure 9 As shown, the overall defect classification accuracy of TGV reached 98.33% according to calculations.

[0070] The accuracy rates for identifying open circuit defects, void defects, and composite defects reached 96.52%, 100%, and 98.48%, respectively. ROC curve analysis showed that the AUC values ​​of the three types of defects were all higher than 0.99, indicating that the model has excellent classification and identification capabilities for different types of TGV defects.

[0071] The TGV defect detection method based on PSO-XGBoost proposed in this invention achieves accurate classification of open circuits, voids and composite defects in TGVs, with an overall accuracy of 98.33%. This method solves the problems of low accuracy, poor efficiency and high cost of traditional detection methods for small defects and composite defects. It has important reference value for optimizing TGV design and manufacturing and improving the yield of 3D packaging, and can be widely used in TGV defect detection scenarios.

Claims

1. The purpose of this invention is to provide a TGV defect detection method based on PSO-XGBoost, comprising: A TGV defect model is established, and TGV defect S-parameters are extracted. By using supervised machine learning methods, the S-parameters in the TGV model for void defects, open circuit defects, and compound defects are trained. Based on different S-parameters, the PSO-XGBoost model is used to classify the defects and predict the types of defects that occur, thereby improving the accuracy of TGV defect detection.

2. The TGV defect detection method based on PSO-XGBoost according to claim 1, characterized in that, A three-dimensional physical model of a defect-free GSG-TGV (ground-signal-ground) was established in HFSS simulation software, and the S-parameters of the defect-free GSG-TGV were extracted.

3. The TGV defect detection method based on PSO-XGBoost according to claim 1, characterized in that, Taking the voids, open circuits, and composite defects that appear inside the TGV as the research object, a three-dimensional simulation model of the void defects, open circuit defects, and composite defects of the TGV was established in HFSS simulation software. Based on the established defect model, S-parameters for different defects are extracted.

4. The TGV defect detection method based on PSO-XGBoost according to claim 3, characterized in that, Labels were set based on the data classification criteria of TGV void defect radius, TGV open circuit defect length, and TGV composite defect combination form. The optimal hyperparameters are obtained by finding the best hyperparameters using the PSO-XGBoost algorithm.

5. The TGV defect detection method based on PSO-XGBoost according to claim 4, characterized in that, Establish an XGBoost classification model; The sample data is read and input into the XGBoost classification model, and preprocessed. The XGBoost algorithm is used to evaluate the importance of all features and select features with greater importance than the average importance. This helps reduce the number of features, improves the model's performance, and reduces the risk of overfitting.

6. The TGV defect detection method based on PSO-XGBoost according to claim 5, characterized in that, Separate features and labels, and encode the category labels; Convert the category label to an integer; Normalize the features by scaling them to a given range of 0 to 1.

7. The TGV defect detection method based on PSO-XGBoost according to claim 6, characterized in that, Determine the range and step size of the XGBoost hyperparameters to be searched; Based on the determined hyperparameter range and step size, global optimization is performed using the PSO algorithm to generate candidate hyperparameter combinations. For each set of hyperparameters obtained by PSO optimization, these parameters are used to train the model, and the model's performance is evaluated on the validation set.

8. The TGV defect detection method based on PSO-XGBoost according to claim 7, characterized in that, During the optimization process, the model performance evaluation results corresponding to each set of hyperparameters are recorded, and the set of hyperparameters with the best performance is found. The model is trained using the best random seed and the best hyperparameters, and predictions are made on the test set. A classification report and accuracy are then output.

9. The TGV defect detection method based on PSO-XGBoost according to claim 8, characterized in that, If the current accuracy is higher than the historical best accuracy, then update the optimal hyperparameter combination; Finally, the model is retrained using the found optimal combination of hyperparameters, and its performance is evaluated on the test set to output the best result.

Images