Numerical study on the applicability of learning methods to equipment models for fab-level simulation

Abstract

Recently, research about modeling of semiconductor production process has been actively developed. There are four equipment models which are affine models, exit recursion models, flow line models and detailed models. They know the physics of the production process and predict cycle time, lot residency time and throughput time of the semiconductor production. However, there are certain settings in which these equipment models show bad performance. In addition to that, to endure the randomness of the system, some learning-based models have been considered. In this paper, neural network, random forest, and XGBoost models are newly revised and applied to predict cycle time, lot residency time and throughput time of the semiconductor production. Then, the performance of learning-based models and the equipment models are compared based on mean value, mean absolute error, and percent error.