A fuzzy-neural system with error feedback to adjust classification for forecasting wafer lot flow time: A simulation study

Abstract

Estimating lot flow (cycle) time is a critical task for a wafer fabrication plant (wafer fab). Many recent studies have shown that pre-classifying wafer lots before estimating the flow times is beneficial to estimation accuracy. In this aspect, various classification approaches, e.g. k-means (kM), fuzzy c-means (FCM), and self-organization map (SOM), have been applied. After pre-classification, to estimate the flow times for lots belonging to different categories, different approaches (that are in fact the same approaches but with different parameter settings) are applied. However, these applications of classification approaches considered only the data of wafer lots, but ignored whether the classification approaches combined with the subsequent estimation techniques were suitable for the data. To tackle this problem, instead of trying many possible classification and forecasting approaches to find out the most suitable combination, a FCM and back propagation network (BPN) combination is chosen in the current study. In the proposed methodology, the classification results by FCM will be adjusted with forecasting error fed back from the BPN. In this way, if the FCM-BPN combination is not good enough for the data, then a forecasting error will be generated and fed back to the FCM classifier to adjust the classification results. After some replications, the FCM-BPN combination will become more suitable for the data. To evaluate the effectiveness, production simulation is applied in the present study to generate test data. According to experimental results, the forecasting accuracy of the proposed methodology is significantly better than those of many existing approaches. The effects of adjusting classification results with prediction error are also revealed.