【作者】 李航；

【导师】 贾护军； 邢朝洋；

【作者基本信息】 西安电子科技大学 ， 电子信息硕士（专业学位）， 2023， 硕士

【摘要】 随着集成电路的特征尺寸不断接近工艺极限,摩尔定律受到的制约日益增加。三维集成技术作为后摩尔时代延续摩尔定律的关键技术,能够使全局互连长度降低并改善集成电路中的延时与功耗,满足基于三维集成技术的信息处理微系统模块对于小型化、高性能及低成本的要求。然而随着信息处理微系统模块尺寸不断缩小,集成度不断提高,使其功率密度增加,产生的热量若不能及时散发出去将会产生热积聚现象,进而严重影响信息处理微系统模块的可靠性,因此对信息处理微系统模块进行电热耦合分析具有重要的意义。本文首先建立双TSV模型并仿真,确定了影响TSV温度的三个结构参数,并通过正交实验设计得出不同的结构参数组合,同时建立双层TSV阵列对不同的结构参数进行仿真得到对应的温度数据,并建立了BP神经网络模型对双层TSV中的最高温度进行预测,预测值与仿真结果的绝对误差绝大多数在3℃以下,相对误差绝大多数在3%以下,这说明绝大多数预测值与有限元仿真结果接近,所建立的BP神经网络模型可以用于TSV阵列中温度的预测。为提升BP神经网络模型的性能,本文使用遗传算法对BP神经网络的初始权值与阈值进行优化,使用GA-BP神经网络模型预测得到的结果与仿真结果的绝对误差绝大多数在2℃以下,相对误差绝大多数在1%以下,证明经过遗传算法优化使得BP神经网络的性能有所提升。为进一步提升BP神经网络的性能,使用遗传算法对BP神经网络中的迭代次数、学习率与隐含层神经元数进行优化,最终确定:学习率为0.0109、最大迭代次数为6464、第一层隐含层神经元数为5、第二层隐含层神经元数为6。使用优化了上述参数的GA-BP神经网络预测得到的结果与仿真结果的绝对误差绝大多数在0.5℃以下,相对误差绝大多数在0.5%以下,预测结果与仿真结果有着极高的拟合度,这说明经过该优化过程后使得BP神经网络的性能得到了极大提升。本文针对传统TSV结构参数设计方法效率过低的问题,提出了一种快速、准确的TSV阵列结构参数智能优化设计方法,相比于使用传统方法,设计效率得到了大幅提升,并且得到的温度优化结果与有限元仿真结果相比误差较小,其中T_TSV的平均误差为0.87%,T_Bump的平均误差为0.25%,T_RDL1的平均误差为0.64%,T_RDL2的平均误差为0.63%,T_max的平均误差为1.45%,说明仿真结果符合预期,体现了该智能优化方法的准确性及可行性,对三维集成微系统的研究与设计具有很大的启发意义。更多还原

【Abstract】 As the characteristic size of integrated circuits approaches the technological limit,the restriction of Moore’s law increases day by day.As a key technology to continue Moore’s Law in the post-Moore era,3D integration technology can reduce the length of global interconnection,improve the delay and power consumption in integrated circuits,and meet the requirements of information processing microsystem module based on 3D integration technology for miniaturization,high performance and low cost.However,as the size of the information processing microsystem module continues to shrink and the integration degree continues to improve,its power density increases.If the generated heat cannot be dissipated in time,heat accumulation will occur,which will seriously affect the reliability of the information processing microsystem module.Therefore,it is of great significance to conduct electrothermal coupling analysis on the information processing microsystem module.In this thesis,a double TSV model has been established and simulated,and three structural parameters affecting the temperature of the TSV have been determined,and different structural parameter combinations have been obtained through orthogonal experimental design.At the same time,a double layer TSV array has been established to simulate different structural parameters to obtain the corresponding temperature data,and a BP neural network model has been established to predict the highest temperature in the double layer TSV.The absolute error between the predicted value and the simulation result is mostly below 3℃,and the relative error is mostly below 3%,which indicates that most of the predicted value is close to the finite element simulation result,and the established BP neural network model can be used to predict the temperature in TSV array.In order to improve the performance of the BP neural network model,the genetic algorithm in this thesis has been used to optimize the initial weight and threshold of the BP neural network.The absolute error between the predicted results and the simulation results using the GA-BP neural network model is mostly below 2℃,and the relative error is mostly below 1%.It is proved that the performance of BP neural network is improved by genetic algorithm optimization.In order to further improve the performance of BP neural network,genetic algorithm has been used to optimize the number of iterations,learning rate and number of hidden layer neurons in BP neural network.Finally,it is determined that the learning rate is 0.0109,the maximum number of iterations is 6464,the number of neurons in the first hidden layer is 5,and the number of neurons in the second hidden layer is 6.The absolute error of GA-BP neural network with optimized above parameters and simulation results are mostly below 0.5℃,and the relative error is mostly below 0.5%.The predicted results and simulation results have a very high degree of fitting,which indicates that the performance of BP neural network has been greatly improved after the optimization process.Aiming at the low efficiency of traditional TSV structural parameter design methods,a fast and accurate intelligent optimization design method for TSV array structural parameters has been proposed in this thesis.Compared with traditional design methods,the efficiency has been greatly improved,and the error of temperature optimization results is smaller than that of finite element simulation results.The average error of T_TSVis 0.87%,the average error of T_Bumpis 0.25%,the average error of T_RDL1is 0.64%,the average error of T_RDL2is0.63%,and the average error of T_maxis 1.45%,indicating that the simulation results are in line with expectations,which reflects the accuracy and feasibility of the intelligent optimization method.It is of great enlightening significance to the research and design of3D integrated Microsystems.更多还原