基于滑动窗口集成即时学习算法的ALD工艺压强预测  

Study on ALD Process Pressure Prediction Based on Sliding Window Integrated Just-In-Time Learning Algorithm

作　　者：孙世娜 陈焰[1] 伍祖铭 刘振强 明帅强 夏洋 SUN Shi-na;CHEN Yan;WU Zu-ming;LIU Zhen-qiang;MING Shuai-qiang;XIA Yang(Faculty of Information Engineering and Automation,Kunming University of Science and Technology;Jiaxing Microelectronics Instrumentation and Equipment Engineering Center,Chinese Academy of Sciences;Jiaxing Kemin Electronic Equipment Technology Co.,Ltd.)

机构地区：[1]昆明理工大学信息工程与自动化学院 [2]中国科学院嘉兴微电子仪器与设备工程中心 [3]嘉兴科民电子设备技术有限公司

出　　处：《化工自动化及仪表》2024年第5期879-886,共8页Control and Instruments in Chemical Industry

基　　金：中科院关键技术团队项目(批准号:GJJSTD20200003)资助的课题。

摘　　要：原子层沉积(ALD)设备被广泛应用于半导体、新材料、光伏产业中制备高质量纳米膜。针对制膜过程中关闭真空计,导致压强数据缺失,无法判断反应源是否通入及设备是否正常运行的问题,采用软测量方法预测制膜过程中的压强。为提高软测量对不同工艺条件及配方的预测泛化能力,提出采用基于滑动窗口的集成即时学习法(BaggingMwLWPLS)对压强进行实时建模及预测,使用滑动窗口划分数据集,采用Bagging集成算法将局部加权偏最小二乘法(LWPLS)的预测值平均融合作为预测压强。结果表明:该方法能自适应预测不同工艺配方的实时压强,且预测快速、准确、不易受噪声干扰,实现设备压强状态和性能的实时监控。Atomic layer deposition(ALD)device enjoys wide application in semiconductors,new materials and photovoltaic industries to prepare high-quality nanofilms.Aiming at the vacuum gauge to be closed in the membrane-making process and the loss of pressure data and the difficulty in judge whether the reaction source is connected and whether the equipment is running normally,the soft sensor method was adopted to predict the pressure in preparing the membranes.For purpose of improving the prediction generalization ability of soft sensor for different process conditions and formulas,a sliding window-based ensemble just-in-time learning(BaggingMwLWPLS)algorithm was proposed to model and predict the pressure in real time,including having the sliding window employed to divide the data set and the Bagging ensemble algorithm adopted to average the predicted values of local weighted partial least squares(LWPLS)as the predicted pressure.The results show that,the method can adaptively predict the real-time pressure of different process formulations,and the prediction is fast,accurate,and not susceptible to noise interference so as to realize real-time monitoring of both equipment pressure status and performance.

关 键 词：工艺压强 滑动窗口 局部加权偏最小二乘 集成学习 自适应预测 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]