Fan-out wafer level packaging rewiring structure service life prediction method based on machine learning
Through the life prediction method of fanout wafer-level packaging devices based on machine learning, using neural network and Bayesian optimization method, the high life test cost caused by the complex structure of the rewiring layer in fanout wafer-level packaging is solved. It realizes efficient life prediction and reliability analysis, and improves production design efficiency.
Patent Information
- Application Number
- CN202411042104.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-29
AI Technical Summary
The structure and process of rewiring layer in fanout wafer-level packaging are complex, resulting in high life test time cost and high simulation difficulty, and it is difficult for the existing technology to effectively predict the device life.
Using a machine learning-based method, the model parameters-failure stress data set is generated, and the thermal mismatch stress is predicted through neural network regression model, and the model parameters are optimized in combination with Bayesian optimization method to establish the mapping relationship between failure stress and temperature cycle life to realize the fan Prediction of life of out-of-form wafer-level packaging devices.
It improves component selection, reliability allocation and expected efficiency, improves quality control level, and reduces the reliability iteration efficiency of electronic packaging of production design units.
Smart Images

Figure CN119047301A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of semiconductor devices, and in particular to a method for predicting the life of a rewiring structure of a fan-out wafer-level package based on machine learning. Background Art
[0002] Fan-out wafer-level packaging has the potential for heterogeneous integration and 3D stacking, and can be combined with a variety of advanced packaging technologies. It is the cornerstone of the future evolution of advanced packaging technology. Fan-out wafer-level packaging has a solid foundation for use, helping key components such as programmable logic arrays, central managers, and digital signal processing modules to develop towards high density, high performance, and high reliability. As the military, aviation, aerospace and other fields gradually develop towards small size, light weight, and high reliability, fan-out wafer-level packaging will play a key role in the packaging transformation of military and aerospace devices.
[0003] Delamination of the redistribution layer is a common failure mode in fan-out packaging. However, the redistribution layer structure and redistribution process are complex, the structure and materials are highly personalized, the redistribution layer simulation is difficult, and the time cost of directly conducting life tests is high. Summary of the invention
[0004] The present application provides a method for predicting the life of a fan-out wafer-level packaging rewiring structure based on machine learning. Using machine learning methods to predict the life of fan-out wafer-level packaging devices is conducive to improving the efficiency of component selection, reliability allocation and prediction, and provides strong technical support for improving the quality control level of fan-out wafer-level packaging devices.
[0005] In a first aspect, a method for predicting the life of a rewiring structure of a fan-out wafer-level package based on machine learning is provided, comprising:
[0006] Generate a data set of model parameter-failure stress data set, the data set of model parameter-failure stress data set includes multiple samples, and the data of each sample includes independent variable data and dependent variable data, including: randomly permuting and combining within the value range of packaging parameters of fan-out wafer-level packaging to obtain independent variable data of multiple samples; bring the independent variable data of multiple samples into the simulation model for stress simulation to obtain dependent variable data values of multiple samples;
[0007] Through machine learning regression prediction, the model parameters-failure stress dataset is learned, and a regression prediction model is built to predict thermal mismatch stress.
[0008] In combination with the first aspect, in certain implementations of the first aspect, the independent variable data include redistribution layer pattern elements, geometric parameters, and material linear elastic parameters; the redistribution layer pattern elements include parallel, intersection, and turning; the geometric parameters include geometric parameters of silicon wafers, solder balls, redistribution layers, and molding compounds; the material linear elastic parameters include parameters of silicon, epoxy resin, copper, and solder.
[0009] In combination with the first aspect, in certain implementations of the first aspect, the input data of the regression prediction model include: package length, package width, silicon ratio, RDL line spacing, RDL line width, Si thickness, Cu thickness, EMC thickness, PI thickness; the dependent variable data is the simulation value of shear stress at the copper layer-UBM layer interface.
[0010] In combination with the first aspect, in certain implementations of the first aspect, the simulation model is a thermodynamic microelement model of the redistribution layer, and the temperature change ΔT and the thickness of each layer of the package and the side length of the package are used as model inputs, the thermal expansion coefficient a, elastic modulus E, and Poisson's ratio v of the packaging material are basic material parameters of the model, and the average value of the resultant force of the normal stress ε exerted on the copper wire microelement in the redistribution layer and the average value of the resultant force of the shear stress τ exerted on the microelement at the material interface are outputs of the simulation model.
[0011] In conjunction with the first aspect, in some implementations of the first aspect, the regression prediction model is a neural network regression model that satisfies:
[0012] The number of hidden layers of the neural network model is 2, the number of neurons in the first layer is 292, and the number of neurons in the second layer is 230;
[0013] The activation function of the neural network model is the ReLU function, and the mathematical expression of the ReLU activation function is:
[0014] ReLU(x)=max(0,x)
[0015] The loss function is L2 regularized, and the Dropout layer method is introduced, using the L2 regularized loss function:
[0016]
[0017] w is the weight coefficient matrix of the independent variable, x is the independent variable matrix, n is the number of training samples, y i is the dependent variable, i.e. the true response of the i-th sample, w T x i is the predicted response of the i-th sample, λ represents the regularization strength, and its value is 4E-08.
[0018] In combination with the first aspect, in some implementations of the first aspect, the regression prediction model is a Gaussian process regression model, and the Gaussian distribution covariance matrix of the Gaussian process regression model satisfies:
[0019]
[0020] in σ in the covariance matrix f =3E+08, l=489~490.
[0021] In combination with the first aspect, in certain implementations of the first aspect, the model parameters of the regression prediction model are optimized by a Bayesian optimization method.
[0022] In combination with the first aspect, in some implementations of the first aspect, the Bayesian optimization method uses a tree structure for parsing.
[0023] In combination with the first aspect, in some implementations of the first aspect, the method further includes:
[0024] Input the test samples into the trained regression prediction model for testing.
[0025] In combination with the first aspect, in certain implementations of the first aspect, the regression prediction model performs stress prediction for a package rewiring structure with the same preset temperature change and the same preset material.
[0026] Compared with the prior art, the solution provided by this application includes at least the following beneficial technical effects:
[0027] The present invention establishes a set of fan-out wafer-level packaging temperature cycle life methods based on the failure stress regression prediction machine learning model and the mapping relationship between failure stress and temperature cycle life. After knowing the parameter data of the advanced packaging device to be predicted, the method obtains the failure stress prediction value of the fan-out wafer-level packaging device to be predicted through the machine learning failure stress regression prediction model, and then maps the failure stress prediction value to the temperature cycle life prediction value according to the mapping relationship between failure stress and temperature cycle life. It is helpful for product user designers to perform reliability analysis in the component selection, product design initial reliability allocation, and sample life prediction stages, and helps to improve the reliability iteration efficiency of electronic packaging of component production design units. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 This is the overall flow chart of machine learning prediction of fan-out wafer-level package life.
[0029] Figure 2 This is a schematic diagram of a typical fan-out packaging structure.
[0030] Figure 3This is a diagram of the components of a typical fan-out wafer-level package simulation model.
[0031] Figure 4 It is a schematic diagram of the thermal mismatch stress on the hexahedral microelement under volume constraint.
[0032] Figure 5 It is a schematic diagram of the hexahedral microelement stress model.
[0033] Figure 6 It is a diagram of the neural network algorithm structure.
[0034] Figure 7 This is a schematic diagram of the introduction of the Dropout layer regularization method. DETAILED DESCRIPTION
[0035] The present application is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0036] Figure 1 A method for predicting the life of a fan-out wafer-level package rewiring structure based on machine learning is shown in an embodiment of the present application. The specific method steps are as follows.
[0037] Step 1: Generate a dataset of model parameters-failure stress dataset.
[0038] Specifically, the variation range of geometric parameters of main structures such as the plastic packaging material, silicon wafer, and redistribution layer of the fan-out wafer-level packaging is extracted, and a simulation model is established with reference to the wiring elements of the redistribution layer. Thermodynamic analysis is carried out based on the linear elastic parameters of metal materials such as silicon, epoxy plastic packaging material, polyimide, and copper and aluminum, and a model parameter-failure stress database under temperature cycling is established.
[0039] Figure 2 The schematic diagram of a fan-out wafer-level packaging rewiring structure. The typical rewiring layer structure consists of multiple layers of metal wiring wrapped in organic materials, which includes aluminum micro-bumps connecting the silicon wafer and the rewiring, copper rewiring for internal interconnection, and the Under Bump Metallization (UBM) layer connecting the solder balls and the rewiring. Since PI has suitable indicators such as thermal expansion coefficient and elongation, and excellent electrical performance, these structures are usually wrapped with PI to play an insulating and supporting role.
[0040] According to the failure mode and structural characteristics of the fan-out wafer-level packaging rewiring structure, Figure 3The parameters and elements shown in the figure are used to establish a model of a typical fan-out wafer-level packaging device. The model contains common redistribution layer pattern elements (parallel, intersection, and turning), covers the key structures (silicon wafers, solder balls, redistribution layers, and plastic packaging materials) and their geometric parameters that affect the life of fan-out wafer-level packaging devices, and selects the material parameters of fan-out wafer-level packaging device materials according to the failure stress type of the redistribution structure to generate a data set of model parameters-failure stress data sets, as shown in Table 1.
[0041] Table 1 Structure of training dataset for machine learning thermal mismatch stress prediction model
[0042]
[0043] The data set of the model parameter-failure stress data set includes multiple samples. The data of each sample includes independent variable data and dependent variable data. The independent variable data is the specification parameters of the packaging model, which may include pattern elements, geometric parameters, and material linear elastic parameters of the redistribution layer. The dependent variable data is failure stress data. The failure stress data can be obtained by simulating the independent variable data by inputting the independent variable data into a finite element model of the packaging structure (such as a thermodynamic microelement model of the redistribution layer). Specifically, firstly, random permutations and combinations are performed within the range of package parameter values to obtain the independent variables of the data. The randomly generated independent variable group is brought into the simulation model to obtain the copper layer-UBM layer interface shear stress simulation value of each set of packaging parameters to obtain the dependent variable.
[0044] In one embodiment, in the thermodynamic microelement model of the redistribution layer, for an isotropic elastic deformation material, when the internal microelement is subject to body constraints, the thermal mismatch stress in each direction is as follows: Figure 4 As shown. The magnitude of thermal mismatch stress is:
[0045]
[0046] Where a is the thermal expansion coefficient of the material, ΔT is the temperature difference of the material, E is the elastic modulus of the material, and v is the Poisson's ratio of the material.
[0047] However, due to the different thermal expansion coefficients and elastic moduli of different materials, the deformation amount is also different when the temperature changes, so shear stress will also be generated at the interface of different materials. Figure 5 is the model of normal stress and shear stress on the microelement. The normal strain-stress equation of the microelement in the x, y, and z directions is:
[0048]
[0049]
[0050]
[0051] The shear stress-strain equation of the microelement is:
[0052]
[0053]
[0054]
[0055]
[0056]
[0057]
[0058] Among them, the normal stress between microelements in the redistribution layer is the main failure stress that causes the redistribution layer to break, and the shear stress at the interface between copper and other materials is the main failure stress that causes the redistribution layer to delaminate. According to the mathematical principle of thermal mismatch stress and the failure mechanism of redistribution interconnection, the temperature change ΔT and geometric parameters such as the thickness of each layer of the package and the side length of the package are used as model inputs, the thermal expansion coefficient a, elastic modulus E, and Poisson's ratio v of the package material are the basic material parameters of the model, and the average value of the resultant normal stress ε on the copper wire microelement in the redistribution layer and the average value of the resultant shear stress τ on the microelement at the material interface are the simulation model outputs.
[0059] Step 2: Use machine learning regression to predict the learning model parameters-failure stress dataset, and build a machine learning algorithm to predict thermal mismatch stress.
[0060] The data type of the input data obtained by the regression prediction model from the package modeling simulation is continuous, and the scales and ranges of the respective variables are quite different. The data weight of the model will be biased towards data with larger scales, making the model weights distorted. In one implementation, it is also necessary to perform input data standardization preprocessing on the data in the model parameter-failure stress data set. Standardization of input data refers to converting a variable of a certain dimension into a normal distribution with a mean of 0 and a variance of 1.
[0061] In one implementation, a neural network regression model may be used to predict thermal mismatch stress.
[0062] like Figure 6 As shown in the figure, the neural network regression model includes an input layer, a hidden layer, and an output layer. The complete neural network regression algorithm also requires the activation function used by each neuron. The loss function regularization method of the output result is selected and adjusted. The data must be preprocessed before importing it into the model. After the model is established, the 9 specification parameters of the package structure to be predicted are entered into the input layer as the characteristic parameters of the model. After a complex prediction process, the output layer will output the predicted failure thermal mismatch stress.
[0063] In some embodiments, the number of hidden layers of the neural network model is 2, the number of neurons in the first layer is 292, and the number of neurons in the second layer is 230.
[0064] In some embodiments, the activation function of the neural network model is a ReLU function. The mathematical expression of the ReLU activation function is:
[0065] ReLU(x)=max(0,x)
[0066] In some embodiments, due to the randomness of the packaging simulation model in the meshing process, even the same model will not produce completely consistent results in multiple simulations. In the calculation of the shear stress and normal stress of the microelement, certain noise will be generated. L2 regularization of the loss function and the introduction of the Dropout layer method can improve the model overfitting problem and enhance the model robustness.
[0067] The loss function using L2 regularization is:
[0068]
[0069] w is the weight coefficient matrix of the independent variable, x is the independent variable matrix, n is the number of training samples, y i is the dependent variable, i.e. the true response of the i-th sample, w T x i is the predicted response of the i-th sample, λ represents the regularization strength, and its value is 4E-08.
[0070] like Figure 7 As shown in the figure, the Dropout layer prevents overfitting by randomly killing some hidden layer neurons. During each training, neurons are randomly removed, which can prevent one neuron from being overly dependent on another neuron, thereby cutting off the transmission of erroneous information such as noise and error, and preventing overfitting.
[0071] In one implementation, a Gaussian process regression model may be used to predict thermal mismatch stress.
[0072] The Gaussian distribution covariance matrix of the Gaussian process regression model satisfies:
[0073]
[0074] in
[0075] σ in the covariance matrix f =3E+08, l=489.67.
[0076] In some embodiments, the above neural network regression model or Gaussian process regression model can be optimized using Bayesian optimization method. The principle of Bayesian method is to establish an expected improvement equation (Expected Improvement, EI) to determine the probability of finding a better point:
[0077]
[0078] Based on the tree structure, the EI formula is transformed into
[0079]
[0080] In the formula, y* is a threshold set by humans, and y*-y is the expected improvement of the next point compared to y*. p(y|x) is the possible range of optimized function values when the search point has been selected, so the meaning of the formula can be interpreted as the difference between a certain search point x and the threshold y. * The next step is to search for point x n+1 The basis for choosing is to make this expectation as large as possible:
[0081]
[0082] The structure of the EI formula makes it tend to favor search strategies with smaller mathematical expectations and larger variances in probability distributions, allowing the Bayesian optimization method to make a better choice between mining strategies and exploration strategies.
[0083] The model optimization results are shown in Table 2.
[0084] Table 2 Horizontal comparison of machine learning regression algorithms
[0085]
[0086] For the test of the above-mentioned regression prediction model, the embodiment of the present application designs three layout types, and five test chips with different structures form a daisy chain test chip sample matrix, as shown in Table 3.
[0087] Table 3 Daisy chain test chip sample matrix
[0088] Group A B C D E Layout Type Type 1 Type 2 Type 2 Type 3 Type 3 Silicon wafer thickness (um) 200 200 400 600 400 Plastic compound thickness (um) 400 600 600 800 800 Copper wiring layer thickness (um) 5 8 8 5 5 PI thickness (um) 10 12 12 10 10 Line width(um) 35 55 55 125 125
[0089] The temperature cycle life test was carried out on the samples, and the average number of failure cycles for each group of samples is shown in Table 4.
[0090] Table 4 Daisy chain test sample temperature cycle life
[0091] Test sample set A B C D E Shear stress simulation value (Mpa) 4.59 9.94 5.58 0.94 3.83 Average temperature cycle life (times) 100.31 31.45 81.89 136.11 121.30
[0092] The Pearson correlation coefficient between the simulated value of the shear stress at the copper-PI interface of the daisy chain test chip and the temperature cycle life of the sample is 0.97, which is significantly correlated. Based on the correlation between the two, a mapping from failure stress to temperature cycle life is established.
[0093] Although the present invention is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope defined by the claims of the present invention.
Claims
1. A method for predicting the life of a rewiring structure of a fan-out wafer-level package based on machine learning, characterized in that: include: Generate a data set of model parameter-failure stress data set, the data set of model parameter-failure stress data set includes multiple samples, and the data of each sample includes independent variable data and dependent variable data, including: randomly permuting and combining within the value range of packaging parameters of fan-out wafer-level packaging to obtain independent variable data of multiple samples; bring the independent variable data of multiple samples into the simulation model for stress simulation to obtain dependent variable data values of multiple samples; Through machine learning regression prediction, the model parameters-failure stress dataset is learned, and a regression prediction model is built to predict thermal mismatch stress.
2. The method according to claim 1, characterized in that The independent variable data include redistribution layer pattern elements, geometric parameters, and material linear elastic parameters; the redistribution layer pattern elements include parallel, intersection, and turning; the geometric parameters include the geometric parameters of silicon wafers, solder balls, redistribution layers, and plastic packaging materials; the material linear elastic parameters include the parameters of silicon, epoxy resin, copper, and solder.
3. The method according to claim 1, characterized in that The simulation model is a thermodynamic microelement model of the redistribution layer. The temperature change ΔT and the thickness of each layer of the package body and the side length of the package body are used as model inputs. The thermal expansion coefficient a, elastic modulus E and Poisson's ratio v of the packaging material are the basic material parameters of the model. The average value of the resultant force of the normal stress ε on the copper wire microelement in the redistribution layer and the average value of the resultant force of the shear stress τ on the microelement at the material interface are the simulation model outputs.
4. The method according to claim 1, characterized in that: The input data of the regression prediction model include: package length, package width, silicon ratio, RDL line spacing, RDL line width, Si thickness, Cu thickness, EMC thickness, PI thickness; the dependent variable data is the simulation value of shear stress at the copper layer-UBM layer interface.
5. The method according to claim 1, characterized in that: The regression prediction model is a neural network regression model that satisfies: The number of hidden layers of the neural network model is 2, the number of neurons in the first layer is 292, and the number of neurons in the second layer is 230; The activation function of the neural network model is the ReLU function, and the mathematical expression of the ReLU activation function is: ReLU(x)=max(0,x) The loss function is L2 regularized, and the Dropout layer method is introduced, using the L2 regularized loss function: w is the weight coefficient matrix of the independent variable, x is the independent variable matrix, n is the number of training samples, y i is the dependent variable, i.e. the true response of the i-th sample, w T x i is the predicted response of the i-th sample, λ represents the regularization strength, and its value is 4E-08.
6. The method according to claim 1, characterized in that The regression prediction model is a Gaussian process regression model, and the Gaussian distribution covariance matrix of the Gaussian process regression model satisfies: in σ in the covariance matrix f =3E+08, l=489~490.
7. The method according to claim 1, characterized in that The model parameters of the regression prediction model are optimized by Bayesian optimization method.
8. The method according to claim 7, characterized in that The Bayesian optimization method uses a tree structure for parsing.
9. The method according to claim 1, characterized in that: The method further comprises: Input the test samples into the trained regression prediction model for testing.
10. The method according to claim 1, characterized in that The regression prediction model performs stress prediction on a package rewiring structure with the same preset temperature change and the same preset material.
