Process Monitoring Methods For Wafer Fabrication Processes

As the core of the integrated circuit industry,wafer fabrication processes are the foundation of modern computer,communication,consumer electronics,and information industry,which are also related to every aspect of the national economy.As a technology to ensure the safety of fabrication processes and improve product quality,process monitoring is widely used in wafer fabrication processes.Data-driven multivariate statistical process monitoring has become one of the hot spots in both academic and industrial circles.As a typical batch process,the wafer fabrication process has some common characteris-tics,such as multi phases and uneven batch duration.At the same time,it also has some unique characteristics,such as lot-wafer structure,multi-machine with the same recipe,on-off actions between steps and between-batch state drift,etc.The traditional batch process monitoring meth-ods seldom focus on these unique characteristics of the wafer fabrication process.Based on the previous researches,this paper proposes a variety of improved process monitoring methods aiming at the characteristics of the wafer fabrication process,including:According to the characteristics of the wafer fabrication process,we proposed a series of data preprocessing operations to reduce the complexity of process data.These preprocessing operations include phase division,phase alignment,state drift compensation,variable selection and machine difference correction.All these operations are suitable for real-time fault detec-tions.Hence,combining the Multiway(PCA)principal component analysis model,we propose an improved real-time fault detection framework for the industrial wafer fabrication process.Specifically,these improvements are:(a)Considering the redundant process variable and machine difference problems in the in-dustrial wafer fabrication process,this paper proposes a variable selection method based on the within-batch mean-variance.This method divides all the process variables into the univariate part,which is suitable for univariate statistical process control charts,and the multivariate part,which is suitable for the multivariate monitoring model in the former section.Besides,a machine difference correction method based on the single machine standardization is proposed.With this method,a multivariate process monitoring model is enough to monitor a group of machines with the same process recipe.In this way,the number of fab-wide processes monitoring models can be effectively diminished,reducing modeling and maintenance costs(b)Considering the multi phases,uneven batch duration,and on-off actions in the wafer fabrication process,this paper proposes a differential weighted distance based two-step stationary-transition phase division method as well as a transition phase alignment method and realizes the real-time phase judgment in the online monitoring.Finally,a multi-phase multiway principal component analysis model is established on the divided and aligned datasets to solve the model mismatching problem caused by the uneven batch duration problem.In this way,the fault false alarm rate is reduced,and the transition phase’s fault detection capability is improved(c)Considering the between-batch state drift in the continuous batches,this paper proposes a simplified weighted least squares state drift estimation method,a state drift modeling method based on the seasonal autoregressive moving average model,and a moving win-dow multi-batch state drift forecast-compensation method.This compensation method improves the signal-to-noise ratio of the within batch dynamics in the whole dataset.In this way,the influence of state drift is eliminated.A multi-phase multiway principal com-ponent analysis model is established on the compensated dataset to obtain better fault detection capability.