1. Introduction2. Modeling and RTSPC2.1 Long Term Trends in Optical Emission Data2.2 Modeling the Effect of Window Clouding(2)(3)3. Analysis and Results3.1 Linearization of Optical Emission Data3.2 Improved Fault Detection4. A Framework for Fault Classification5. Conclusions and Future Work[1] S. F. Lee, E. D. Boskin, H. C. Liu, E. Wen, C. J. Spanos, “RTSPC: A Software Utility for Real...[2] D. C. Montgomery, Introduction to Statistical Quality Control, 2nd ed., John Wiley & Sons, 1991.11. IntroductionIn order to meet the increasing demands of the semi-conductor industry to improve yield while simulta-neously decreasing circuit geometries, recent efforts havefocused on characterizing and controlling variability incritical manufacturing processes such as plasma etching.Real-time tool signals from three sensors (SECSIImachine information, RF monitor and optical emissionspectroscopy) collected in-situ provide valuable informa-tion about the machine state. Effective monitoring ofthese signals serves two purposes: (1) it provides adescription of the machine and chamber states which canbe used to predict final wafer characteristics and (2) itprovides a means of detecting and identifying equipmentmalfunctions in real time without interrupting the pro-cess.Recent efforts have focused on using optical emis-sion data as a valuable source of information about theplasma state. However, measurements of this type exhibitatypical trends due to the confounding effect of windowclouding and machine aging. This behavior is cyclical inthe sense that the machine state can be “reset” by preven-tative maintenance (PM) events. This cycle of long termtrends can result in an increased false alarm rate duringfault detection. This paper describes models which char-acterize this behavior enabling the integration of opticalemission signals with other sensor data so that real-timestatistical process control (RTSPC [1]) techniques can beapplied to perform fault detection. By specificallyaccounting for long term trends, these models partiallydecouple the machine state from the state of the plasma;such decoupling reduces the false alarm rate due to pre-ventative maintenance events, thus resulting in a faultdetection mechanism which is robust over time.The detection of an out-of-control condition by thefault detection mechanism indicates the possible presenceof a fault. In order to confirm the hypothesis that a faulthas occurred and to identify an assignable cause, a meth-odology to classify faults into discrete categories is devel-oped. This paper presents the framework of a diagnosticsystem incorporating both qualitative information, pro-vided through the expert knowledge of human operators,and quantitative information derived from empiricalequipment models as well as historical and maintenancerecords. This framework provides a systematic method ofdrawing inferences from the available evidence andaccounts for uncertainty by retaining a measure of likeli-hood for each classification decision.Data collected from a Lam TCP 9600 plasma etcherwere used to construct the empirical models used for faultdetection and classification.2. Modeling and RTSPCTraditional statistical process control (SPC) tech-niques assume that the underlying process is stationary,i.e. that the mean and variance do not vary with time, andthat the observations are identically, independently, andnormally distributed (IIND) [2]. Presuming that thesetrends are present in data representing normal operatingbehavior, application of these techniques directly tomachine data that contain trends results in increased falsealarm and missed alarm rates [1]. To avoid theseincreased false and missed alarm rates, past work usedtime-series modeling techniques to filter out the timedependent trends; traditional or multivariate SPC meth-ods were then applied to the resulting residuals to moni-tor the machine behavior. This system, known as RTSPC,was shown in [1] to be effective in monitoring real-timeand wafer-to-wafer data. Our investigation was motivatedby the need to extend RTSPC to include long term vari-ability on a lot-to-lot basis.2.1 Long Term Trends in Optical Emission DataExamination and analysis of optical emission dataover long periods of time shows a different type of trendRobust Fault Detection and Fault Classification of Semiconductor Manufacturing EquipmentAnna M. Ison and Costas J. SpanosDepartment of EECS, University of California, Berkeley, CA 94720-1772office:(510)642-9584, fax:(510)642-2739, email:[email protected],WWW: http://bcam.eecs.berkeley.eduAbstractIn this paper we extend our current multivariate statistical process control system for faultdetection to deal with long term variability on a lot-to-lot basis. Long term trends in opticalemission data collected from a plasma etcher are characterized through data transformationsand linear modeling techniques. By filtering the known effects of machine aging, these mod-els facilitate the integration of optical emission data with other sensor signals, resulting in afault detection system which is robust over time. A methodology to classify the detectedfaults into discrete categories is also currently under development. We present the frameworkof a diagnostic system which incorporates various data types and accommodates uncertaintywhile providing a systematic method of drawing inferences from the available evidence.2than that typically handled by time series models. Asdepicted in Figure 1, the endpoint signal (a measure ofthe intensity of the plasma for a particular wavelength)exhibits an exponential decay. In this figure, the averagevalue of the endpoint taken over each lot is plotted withrespect to the wafercount. Because the wafercount isreset to zero after a preventative maintenance (PM)event, the plot shows the endpoint signal evolving overthe course of a maintenance cycle, where the chamberstate is initially clean but becomes progressively dirtieras more wafers are processed. The trend is clearly visi-ble and is repeatable as demonstrated by the five differ-ent maintenance cycles which are overlaid in this plot.The data shown in Figure 1 span a total period of eightmonths during which there were five PM events corre-sponding to chamber and window cleans.2.2 Modeling the Effect of Window CloudingTime series models are known to capture the depen-dencies among a sequence of data points, with theassumption that these readings are taken at regularlyspaced intervals. However,