Abstract
An efficient and effective forecasting of production cycle times (CT) is a critical success factor in semiconductor manufacturing systems (SMS): inaccurate CT forecasts can have a negative impact on production scheduling, causing late deliveries, as well as on the amount of inventories and work-in-progress, which rapidly lose value over time because of the high risk of obsolescence. Therefore, since the 80s, several quantitative techniques have been developed to face this problem. Furthermore, Artificial Intelligence (AI) techniques are gaining importance, despite their potential is still not fully exploited even in the most advanced manufacturing systems. However, a synthetic overview of the techniques to forecast CT in SMS is still missing in the literature. As a result, it is difficult for decision makers to orient themselves and choose, among the many existing ones, the best model for their specific situation, comparing the different performance in terms of accuracy, data required, speed and easiness to use. This paper aims at presenting an overview of the quantitative techniques developed to forecast production CT in SMS. Firstly, a description of the methodology with which the literature review has been carried out is provided. Secondly, a taxonomy of forecasting techniques is proposed. Subsequently, a synthetic description of analytical, simulation, time-series and causal methods is presented. Within statistical techniques, a special focus is deserved to AI ones, since their popularity has dramatically increased in the last years. In particular, the most recent applications of artificial neural networks (ANN) in SMS – namely, hybrid methods and Long-Short-Term-Memory recursive neural networks – are described. Finally, a table with a qualitative comparison between the different methods is proposed.
