Data-Driven Non-Intrusive Modeling Framework for Multiphysics Digital Twin Models of Large-Scale Electronic Devices

This paper proposes a novel non-intrusive reduced-order modeling framework to implement the multiphysics digital twin modeling of large-scale electronic devices. In this framework, a non-intrusive model order reduction technique is used to project the full-order model onto a low-dimensional subspace to obtain a non-intrusive predictive reduced-order model, namely the digital twin model. To efficiently decompose the large-scale matrices, an improved singular value decomposition approach is adopted to achieve an acceleration of 2 to 3 orders of magnitude. Meanwhile, a dynamic model correction approach is proposed to improve the long-term prediction of digital twin models, by combining a proper orthogonal decomposition basis update approach with operator inference method. The proposed modeling framework is first used to build a thermomechanical digital twin model of a 2.5-D integrated packaging. Compared with the full-order model, the simulation efficiency of the digital twin model is improved by more than 3 orders of magnitude while keeping the relative mean square error (RMSE) of temperature and stress prediction less than 0.5% and 4%, respectively. Further, compared with traditional model order reduction techniques, digital twin models built using the proposed ones have higher prediction accuracy, especially in long-term prediction applications. At last, a thermal test module is designed, and its thermal digital model and online monitoring system prototype are built. The RMSE of predicted real-time tem-perature of the thermal digital twin model is smaller than 0.4% compared with the online monitoring results at the test points.