Applied time series analysis for forecasting process cycle times and process yields in the semiconductor manufacturing industry.

Abstract

Complementary metal oxide semiconductor (CMOS) integrated circuits (ICs) are the dominant technology in the semiconductor industry. Meeting delivery date is important to ensure customer satisfaction and maintain momentum between manufacturing divisions. Two major concerns are to produce the right quantity of ICs in the expected period of time. Yield and cycle time are two critical parameters used to assess process performance. The main objectives of this work were to obtain reliable forecasts of yields and cycle times and to monitor the process for detection of upsets. We explored the use of time series models such as ARIMA, transfer function and intervention models. A simultaneous outlier treatment and forecasting strategy was developed which combined the joint estimation of model parameters and outlier effects procedure with some new control charts. This method is particularly suited for highly correlated processes that are frequently subjected to large outliers. Choice of time basis was an important issue in this work (i.e., week of emergence from or entrance to the process). For ARIMA modeling, cycle time series were found to follow a highly correlated AR(1) process whereas yield series were just white noise. Five-step ahead ARIMA forecasts were often inaccurate due to the presence of frequent and large outliers. One step-ahead ARIMA forecasts were quite satisfactory. Transfer function models relating yield data to the process capacity data were built. Transfer function models for overall process cycle times were constructed using as inputs cycle times for various stages in the process. Five step-ahead forecasts were highly improved using these transfer functions. (Abstract shortened by UMI.)