机器学习赋能电子封装材料可靠性优化

Machine Learning Enabling the Reliability Optimization of Electronic Packing Materials

刘之豪

11930201

硕士

材料工程

吴景深

2021-05-23

2021-06-15

南方科技大学

深圳

近年来，电子封装技术不断向小型化、集成化的方向发展，不同材料尺寸的封装产品在热载荷下引起的翘曲问题严重影响产品的可靠性。传统的封装可靠性分析与优化方法主要采用试验设计(Design of Experiment, DOE)和有限元分析。然而，电子封装结构复杂性不断增加，导致设计变量维度爆炸，传统分析方法难以保证及时性、高效性与经济性。针对当前现状，本文通过ABAQUS有限元仿真方法，模拟热循环载荷下不同封装设计参数对应的翘曲值，结合机器学习方法提出电子封装结构的翘曲预测模型，以及设计参数的优化模型，并通过数值实验方法检验了模型效果。本文主要研究内容包括：分析电子封装翘曲失效机理，确定封装结构、尺寸和材料种类。基于ABAQUS二次开发工具建立封装的有限元模型。借助拉丁超立方抽样方法得到不同封装设计参数组合，并借助ABAQUS仿真方法计算得到翘曲值，最终形成封装翘曲模拟数据集：五维设计变量与对应的一维翘曲值标签。封装设计参数的翘曲预测问题。分析数据集本身状态，采用线性回归、神经网络等方法拟合该翘曲模拟数据集映射的函数，并使用网格搜索对模型进行调参优化。以均方误差作为评价指标，对比神经网络和线性回归在测试集上的效果。结果表明，对于本研究的封装数据集，神经网络的拟合表现更好。以最小化翘曲值为目标的封装设计参数优化问题。基于建立离线式代理模型，使用粒子群优化算法寻找全局最优解，适合封装数据充足的场景，在样本充分的条件下找到92\%近似程度的全局最优解；基于高斯过程-贝叶斯优化算法，建立在线式代理优化模型，通过采集函数引导探索新样本，适合封装样本数据缺乏的场景。算法最在11次探索的情况下，找到85\%近似程度的全局最优解。本文研究成果对于封装设计参数的翘曲值预测，以及封装参数优化方面有参考价值，可提高电子封装的可靠性分析的效率，降低设计探索次数，节省电子封装设计成本。

In recent years, electronic packaging technology has continued to develop in the direction of miniaturization and integration. Because of warping caused by thermal load, the difference in the size and performance of various materials in the package will seriously affect the reliability of the package. Design of experiment (DOE) and finite element analysis(FEA) are applied in traditional packaging reliability analysis. However, the increasing complexity of electronic package structures has caused an explosion in the dimensions of design variables. Traditional analysis methods are difficult to ensure timeliness, efficiency, and economy. Considering the situation, ABAQUS is applied to simulate the warpage values corresponding to different package design parameters under thermal cycle loads.Combining machine learning methods, a warpage prediction and optimization model for an electronic package are proposed. Moreover, numerical experiment method is carried out to verify the effectness of the model.The main research contents of this paper include: Analyze the warpage failure mechanism of the electronic package, select the package structure, size and material type. ABAQUS secondary development tool is applied to establish the finite element model of the package. Furthermore, different package design parameter combinations are obtained with the Latin hypercube sampling method, and the warpage value is calculated by the ABAQUS simulation method. A package warpage simulation dataset is formed: five-dimensional design variables and corresponding one-dimensional warpage value label. The problem of warpage prediction with respect to package design parameters is studied. The statistical characteristics of data is analyzed. Linear regression and neural network algorithms are applied to fit the warpage simulation dataset mapping function. Grid search combined is used to optimize the model parameters. The mean square error is used as the evaluation index to compare the effects of neural network and linear regression on the test dataset. The results show that for the packaged dataset of this study, the neural network fits better.Optimization of packaging design parameters with the objective of minimizing warpage value. For the scene with sufficient data, the off-line agent optimization model is established, and the particle swarm optimization algorithm is applied to find the global optimal solution. The optimal solution of 92\% approximation degree is obtained. In the case of scarce samples, the Gaussian process-Bayesian optimization algorithm, an online surrogate optimization model is established to guide the exploration of new samples through the collection function. The algorithm can find the global optimal solution with 85\% approximation in the case of 11 explorations.The research results of this paper are of reference value to the reliability parameter optimization of packaging. The efficiency of reliability analysis of electronic package can be increased. The number of design explorations will be reduced. Therefore, the overall cost of electronic package reliability design will be lower.

电子封装 翘曲 试验设计 代理模型 优化

packaging warpage DOE surrogate optimization

中文

独立培养

南方科技大学

刘之豪. 机器学习赋能电子封装材料可靠性优化[D]. 深圳. 南方科技大学,2021.