基于有限元分析和机器学习的跌落所致封装结构力学行为预测

Prediction of Mechanical Behavior of Package Structure Subjected to Drop Impact Based on Finite Element Analysis and Machine Learning

目前电子封装行业中,在进行封装结构力学可靠性研究时需要开展大量的有限元仿真分析,存在仿真模型建模流程复杂且计算过程漫长的常见问题。鉴于此方面的技术瓶颈,首先使用ABAQUS有限元软件对封装结构跌落过程动力响应进行数值模拟并获取特征候选值,建立了4×3的以抗跌落可靠性评估的关键特征变量为输入特征值和以应力和等效塑性应变为输出特征值的数据组。在有限元分析结果基础上,应用相关性驱动神经网络的机器学习方法对数据组进行训练,进而得到相应预测模型。最后,通过与数值模拟结果进行对比,验证了所提出神经网络预测结果与有限元模拟结果具有很好的吻合度。结果表明,基于有限元仿真模型的机器学习方法,在预测封装结构复杂工况下力学性能可靠性方面具有巨大的潜力。

In the electronic packaging industry,the evaluation of mechanical reliability of the packaging structures is usually studied by means of finite element simulations,which is however considerably time-consuming due to modeling creation and iterative analysis.To eliminate these technical limitations,the dynamic response of the packaging structure during the drop processwas firstly simulated by ABAQUS in the present study,and the values of feature candidates were obtained to construct a 4×3 data set.In fact,the main feature variables of the drop resistance evaluation is taken as the input features,while the stress and equivalent plastic strain are taken as the output feature values.Consequently,the data set was trained by the correlation-driven neural network to obtain the corresponding machine-learning based prediction model.Finally,compared with the finite element simulation,the predicted results of machine-learning based neural network are similar to those of finite element simulation results.The results of this paper show that the machine learning method based on finite element simulations has great potential in predicting the reliability of the mechanical performance of the packaging structure under complicated working conditions.