Deep Learning-Based Multi-Horizon Forecasting for Automated Material Handling System Throughput in Semiconductor Fab

Abstract

Automated material handling systems (AMHS) aim to supply wafers to the right place at the right time in semiconductor fab. Appropriate management in response to the changing production environment is necessary to achieve this goal, and the implementation of optimal material handling based on this management can result in maximum output and increased operating profit. An optimal material handling environment can be maintained by monitoring the production system status and it can be more effective if the system status can be predicted in advance. This study presents the possibility of application and development of a deep learning-based (DL) framework for multi-horizon forecasting of throughput in the AMHS. For this purpose, we acquired the training data that can improve forecasting performance by interpolating anomalies present in the data through an anomaly detection model. The application of a DL model trained with the refined data resulted in outperformance compared to the statistical methods, and robust prediction performance was further confirmed through time series cross-validation. This study is important as it presents a forecasting method that can preemptively respond to the change in the production environments in a semiconductor fab by extracting the major factors that influenced the forecasting along with high performance.