基于迁移学习的LiPON制备工艺模拟优化  

Transfer-Learning-Based Virtual Process Optimization for LiPON

作　　者：吴军君 王涛 王英楷[1,2] 王星辉 WU Jun-jun;WANG Tao;WANG Ying-kai;WANG Xing-hui(College of Physics and Information Engineering,Fuzhou University,Fuzhou,Fujian 350000,China;Institute of Micro-Nano Device and Solar Cells,Fuzhou University,Fuzhou,Fujian 350000,China)

机构地区：[1]福州大学物理与信息工程学院,福建福州350000 [2]福州大学微纳器件与太阳能电池研究所,福建福州350000

出　　处：《电子学报》2023年第3期687-693,共7页Acta Electronica Sinica

基　　金：国家自然科学基金(No.83417013);福建省自然科学基金(No.2019J06008,No.2018J01535)。

摘　　要：不同工艺参数对磁控溅射制备固态电解质薄膜LiPON的物理化学特性有巨大影响,使用机器学习对过程建模,能加强内部原理理解,优化参数提升薄膜性能.迁移学习通过挖掘历史数据集中的信息,提升模型精确度与泛化能力,从而更好地找到良好的工艺参数.本文以文献中磁控溅射制备LiPON的数据集为例,探究靶基距离、溅射功率、溅射气压对LiPON薄膜的离子电导率的影响.对比普通机器学习,迁移学习模型在多项误差指标上提升30%以上.通过模型遍历参数空间,搜寻最佳工艺组合,预测LiPON薄膜的离子电导率为2.04μS/cm,优于文献中的最优值,方差分析与实际样本证明了该方法具有可行性.Different process parameters have a huge impact on the physical and chemical properties of the LiPON thin films synthesized by magnetron sputtering.It has great significance to model the synthesis process for strengthening the understanding of internal principles and improving the properties of the thin films.Transfer learning can improve model ac⁃curacy and generalization ability by mining information in historical data sets,so as to better find good process parameters.This paper takes the datasets of LiPON synthesized by magnetron sputtering in literatures as examples to explore the influ⁃ence of target-substrate distance,sputtering power,and sputtering pressure on the ion-conductivity of LiPON films.Compar⁃ing with ordinary machine learning,the transfer learning model improves by more than 30%in multiple error metrics.The built model recommended the optimal parameters combination after traversing parameters space,and the predicted ion-con⁃ductivity of LiPON film is 2.04μS/cm,which is better than the maximum value in the literature.The mapped contour graph of process parameters and performance recommended for a process parameter range,and the performance of film is good and stable within the range.The analysis of variance and actual samples prove that the method is practical.

关 键 词：LIPON 迁移学习 机器学习 工艺优化 方差分析 

分 类 号：TM912[电气工程—电力电子与电力传动] TP181[自动化与计算机技术—控制理论与控制工程]