| The continual push of performance limits for upcoming technology generations has resulted in a flurry of activity to improve manufacturing practices in the semiconductor industry. While the advent of high density, low pressure plasma etch systems has enabled chip makers to meet current performance demands without sacrificing throughput, these benefits are accompanied by increased complexity requiring good process characterization, monitoring and control.;A comprehensive model of the plasma etch equipment-based on sensor data is constructed, which identifies modes of behavior corresponding to normal operation and specific failures. Plasma etching is viewed as a complex process exhibiting hybrid behavior---that is, the process contains both continuous and discrete dynamics. The continuous machine state, characterized by real-time tool signals under normal operating conditions, changes abruptly as a result of machine failures. However, the failures themselves are best classified into discrete groups corresponding to a particular type of faulty behavior. Thus, at a higher level, the state of the process can be described as nominal (i.e. no machine failures), or faulty, where the faulty state is further subdivided into categories corresponding to different causes or failure modes. At a lower level, the continuous dynamics evolve depending on the discrete state of the process. The description of the process is further complicated since, due to the nature of single wafer processing, these continuous dynamics are evolving over different time scales (a) on a second by second basis, within the processing time of a wafer, (b) from wafer to wafer, and (c) from lot to lot.;Time-series and linear modeling techniques are used to characterize the continuous behavior of the machine at three time-scales. The decomposition into different time-scales also facilitates the development of a robust procedure for fault detection using statistical process control techniques. To enhance the fault detection mechanism, models are developed which capture long term trends in the signals, visible on a lot to lot basis, which are mainly caused by changing machine dynamics due to machine aging. Methods for feature selection, extraction and classification are investigated to determine the limitations of current sensor data, and whether such data can effectively be used to identify discrete failure modes. Mixture models are built which provide likelihood estimates for assignment to a fault category based on sensor variables. These are combined in a graphical model encoding the relationships among the variables of interest. |
