Simulation metamodel development using uniform design and neural networks for automated material handling systems in semiconductor wafer fabrication

Simulation is very time consuming, especially for complex and large scale manufacturing systems. The process of collecting adequate sample data places limitations on any analysis. This paper proposes to overcome the problem by developing a neural network simulation metamodel that requires only a comparably small training data set. In the training data set, the configuration of all input data is generated by uniform design and the corresponding output data are the result of simulation runs. A dispatching problem for a complex simulation model of an automated material handling system (AMHS) in semiconductor manufacturing is introduced as an example. In the example, there are 23 4-levels factors, resulting in a total of 423 possible configurations. However, by using the method proposed in this paper, only 28 configurations had to be simulated in order to collect the training data. The results show that the average prediction error was 3.12%. The proposed simulation metamodel is efficient and effective in solving a practical application.