A digital twin calibration for an automated material handling system in a semiconductor fab

Abstract

To address the complex, dynamic, and stochastic nature of an automated material handling system (AMHS) in a semiconductor fabrication facility (fab), practitioners have used a high-fidelity discrete-event simulation as its digital twin model for decision-making over several decades. Previous studies have focused on fast digital twin-based decision-making in AMHSs under the assumption that their digital twin models are credible enough to prescribe decisions. However, parameter uncertainty and intrinsic bias in an AMHS digital twin model can lead to an inaccurate representation of its field system. To address the challenge, this paper introduces the Bayesian calibration, which modularly estimates a digital twin outcome and its discrepancy using Gaussian process priors. A calibration framework for digital twin-based decision-making is also presented using an AMHS example. Our experimental results with various AMHS operating scenarios demonstrate that: (1) a sophisticated digital twin calibration is necessary, especially when AMHSs operate under heavy-workload scenarios; and (2) exploring model bias considerably decreases the prediction error of an AMHS digital twin within a limited number of field observations. Moreover, we discuss the applicability of the approach to digital twins in various fields.