Abstract
WIP and Move are two important production indices for a semiconductor manufacturing fab. The shop managers need to know the production condition and progress based on the variation of these indices and make proper judgment and corresponding decisions based on their experiences. However, this kind of execution provides only understanding of history or facts that happened before. In order to predict possible production variation; to prevent problems getting worse; and to facilitate the whole system performance, it is necessary to have an early warning mechanism that can predict possible future production deviations.Therefore this study focuses on how to predict the variation of WIP and Move to provide managers insights for material flow and control. The neural networks techniques are used to verify the proposed prediction model with data from one local semiconductor factory. The variables of the neural prediction model are chosen by reference to literature and consultation of experts. Through the process of net pruning to delete variables which have minor influence one by one to find the neural networks model with the best variables combination. The results showed the proposed prediction model can effectively predict the variation of production indices one day ahead, and had its practical contribution. The prediction error of WIP is about 7.33% and the prediction error of Move is about 7.63%.