CN-122021229-A - TSV-Cu microstructure fatigue life prediction method based on condition generation countermeasure network

CN122021229ACN 122021229 ACN122021229 ACN 122021229ACN-122021229-A

Abstract

The invention relates to fatigue life prediction of a TSV-Cu of a microelectronic packaging vertical interconnection structure, which belongs to the technical field of life prediction and comprises the following steps of S1, establishing thermal cycle and vibration coupling signal time domain-frequency domain conversion, calibrating a small sample training set, S2, introducing an improved multi-physical field intensity coupling module into an input layer of a condition generation countermeasure network model, enriching coupling characteristics of the small sample training set so as to improve the perceived sensitivity and prediction precision of the condition generation countermeasure network prediction model to coupling load, S3, resetting the convolution layer number and convolution kernel size of the condition generation countermeasure network, improving the global characteristic mapping capability of the model, realizing secondary enhancement of the model prediction precision, S4, aiming at the defect that the condition generation countermeasure network adopts first enhancement and then screening on data, introducing physical priori constraint and corresponding loss evaluation index into a generator and a discriminator of the model, further improving the robustness and generalization capability of the model, and realizing fatigue life prediction. Compared with the existing machine learning algorithm, the method can realize TSV-Cu fatigue life prediction under the condition of high-fidelity small sample data, and the cycle and cost of the fatigue test are compressed to a certain extent.

Inventors

HUANG ZHIYONG
QIAN HONGJIANG
ZHU QINGYUN
Shen Zeshuai
Feng tianan
ZHENG LIANGQI

Assignees

四川大学

Dates

Publication Date: 20260512
Application Date: 20241112

Claims (8)

1. A fatigue life prediction method for a microelectronic package vertical interconnect structure TSV-Cu for a service condition generation countermeasure network, comprising the steps of: s1, converting thermal cycle and vibration coupling load from time domain signals to frequency domain signals by using a fast Fourier transform method, wherein the frequency domain signals converted by each sample correspond to fatigue life test results, and the calibration of a small sample data set is completed; s2, improving a multi-physical field intensity coupling model, introducing an acceleration rate vector of a fourth degree of freedom to a traditional second-order transient dynamics equation, and constructing a third-order transient dynamics equation so that the coupling model is more in line with a real complex working condition; S3, constructing a condition generation countermeasure network model according to the improved multi-physical field coupling model; s4, model super-parameter optimization mainly relates to adjustment of the number of convolution layers and the size of a convolution kernel; s5, constructing a related loss objective function, and realizing TSV-Cu microstructure fatigue life prediction.
2. The method for predicting fatigue life of a TSV-Cu in a microelectronic package vertical interconnect structure for a use condition generation countermeasure network according to claim 1, wherein in S1, the fast fourier transform uses an angular frequency ω as an independent variable function, and a relation of converting a time domain signal f (t) into an f (ω) frequency domain signal is: Where ω=2pi f is the angular frequency and f is the frequency.
3. A method for fatigue life prediction of TSV-Cu for microelectronic package vertical interconnects that use conditions to generate an countermeasure network according to claim 1, characterized in that in S2, the modified third-order transient dynamics equation is: wherein T, M, C, K is a time matrix, a mass matrix, a damping matrix and a rigidity matrix respectively, u and Q are displacement vectors and temperature vectors, and F (t) and Q represent the sum of a thermovibration coupling dynamics equation and a heat flow vector.
4. The method for predicting the fatigue life of the TSV-Cu of the microelectronic package vertical interconnection structure for generating the countermeasure network according to the use condition of claim 1, wherein the input of the countermeasure network is a frequency domain signal after the fast Fourier transform is completed, the frequency domain signal mainly comprises two load information of heat vibration, wherein four characteristic matrixes including time, quality, damping and rigidity are used as the input, and the output of the model is the fatigue life.
5. The method for predicting fatigue life of TSV-Cu for microelectronic package vertical interconnects using a condition generation countermeasure network according to claim 1, wherein in S3, the target formula of the condition generation countermeasure network model may be: V(G,D)=E[logD( x )]+E[log(1-D( x ))] wherein G, D is the generator and arbiter functions, respectively, x is the sampling point, and E is the desired value. In theory, the best effect of the arbiter is when the generator is fixed, dmax.
6. The method for predicting fatigue life of a TSV-Cu in a microelectronic package vertical interconnect structure for a use condition generation countermeasure network according to claim 1, wherein in S4, super-parametric optimization of a model can be realized by Adam optimizer, optimal learning rate parameters can be automatically adjusted, and training early stop steps are enabled to prevent model overfitting. For the selection of fixed parameters, a Relu function is selected as an activation function by the model, and the number of preset training rounds is 2000. The whole model framework adopts a generator and a discriminator, and is matched with the improved third-order transient dynamics equation provided by the invention to be used as a pre-coupling load signal processing method.
7. A method for fatigue life prediction of a microelectronic package vertical interconnect structure TSV-Cu for a service condition generation countermeasure network according to claim 1, wherein in S5, the related objective loss function, i.e., a priori equation, can be constructed as: θ=min[P(D)-P(G)] where θ is the loss function error value, and P (D) and P (G) are the predicted values of the discriminator and the generator.
8. A method of fatigue life prediction for a microelectronic package vertical interconnect structure TSV-Cu using condition generation against a network according to claim 1 wherein in case the model is optimal based on small sample training, part of the fatigue life data not used as condition generation against network training is used as a test set to achieve fatigue life prediction.

Description

TSV-Cu microstructure fatigue life prediction method based on condition generation countermeasure network Technical Field The invention belongs to the technical field of microelectronic fatigue life prediction, and particularly relates to a TSV-Cu microstructure fatigue life prediction method based on a condition-based generation countermeasure network. Background With the continuous improvement of design level and manufacturing process, the integration level and complexity of electronic products are also higher and higher, and these requirements are put on fault test and diagnosis of electronic systems. However, due to the high cost and high process of the chip interconnection structure TSV-Cu, and the characteristics of long test period and the like, the fatigue data of the TSV-Cu microstructure is usually in a small sample state. Therefore, how to realize high-precision fatigue life prediction of the TSV-Cu microstructure through small sample data becomes a serious difficulty in research. Disclosure of Invention The finding exactly solves the problem of the serious difficulty. Calibration of the small sample dataset of the TSV-Cu microstructure by a fast fourier transform method. The multi-physical field intensity coupling model is improved, an acceleration rate vector of a fourth degree of freedom is introduced to a traditional second-order transient dynamics equation, and a third-order transient dynamics equation is constructed, so that the coupling model is more in line with a real complex working condition. Constructing a condition generation countermeasure network model by an improved multi-physical field coupling model, performing super-parameter optimization on the model, and finally constructing a related loss objective function, thereby realizing the high-precision fatigue life prediction book of the TSV-Cu microstructure under the condition of a small sample The invention adopts the technical scheme that: 1) Converting thermal cycle and vibration coupling load from time domain signal to frequency domain signal by using fast Fourier transform method, wherein the frequency domain signal after each sample conversion corresponds to fatigue life test result, and the calibration of small sample data set is completed; Further, the thermal cycle and vibration load can be obtained by testing on an existing thermal-vibration fatigue test platform. Further, the fatigue test data of the TSV-Cu microstructure are small sample data, and the size of the collected sample set is 10-15. Further, the fast fourier transform uses an angular frequency ω as an argument function, and the relation of the time domain signal f (t) to the f (ω) frequency domain signal is: Where ω=2pi f is the angular frequency and f is the frequency. 2) Improving a multi-physical field strength coupling model; Further, the improved third-order transient dynamics equation is: wherein T, M, C, K is a time matrix, a mass matrix, a damping matrix and a rigidity matrix respectively, u and Q are displacement vectors and temperature vectors, and F (t) and Q represent the sum of a thermovibration coupling dynamics equation and a heat flow vector. Further, the improved kinetic equation will be used as a physical constraint for the subsequent generation of the antagonism network. 3) The method comprises the steps of constructing and generating an countermeasure network, wherein the input end of the network is a frequency domain signal after the fast Fourier transform is completed, and the frequency domain signal mainly comprises time, mass, damping and rigidity matrix information and is decoupling information of thermal-vibration load. The target output of the model is fatigue life; 4) The target formula for the condition generating countermeasure network model may be: V(G,D)=E[logD(x)]+E[log(1-D(x))] wherein G, D is the generator and arbiter functions, respectively, x is the sampling point, and E is the desired value. In theory, the best effect of the arbiter is when the generator is fixed, dmax. 5) The super-parameter optimization of the model can be realized by an Adam optimizer, the optimal learning rate parameter can be automatically adjusted, and the training stage stop is started to prevent the model from being overfitted. Further, for the selection of fixed parameters, the model selects Relu functions as the activation functions, and the number of preset training rounds is 2000. Furthermore, the whole model framework adopts a generator and a discriminator, and is matched with the improved third-order transient dynamics equation provided by the invention to be used as a pre-coupling load signal processing method. 6) The associated objective loss function, a priori equation, may be constructed as: θ=min|[P(D)-P(G)]| where θ is the loss function error value, and P (D) and P (G) are the predicted values of the discriminator and the generator. 7) Drawing stress-strain distribution conditions according to fatigue life obtained by generating antag