Scientific Machine Learning Enables Multiphysics Digital Twins of Large-Scale Electronic Chips
Xiao Li, Qiwei Zhan, Bozhao Sun, Haoqiang Feng, Yonghu Zeng, Huabing Wang, Xiaofan Yang, Wen‐Yan Yin
Abstract
We propose a scientific machine learning (SciML) algorithm toward 3-D dynamic digital twins, which represent multiphysics coupling effects in large-scale electronic chips. The SciML is a burgeoning topic, and here we refer to an organic fusion of scientific computing, model order reduction, and machine learning (ML) method. The proposed model order reduction compresses multiphysics information by a data-driven non-intrusive technique, thus getting rid of the access to backend source code; the proposed ML intrinsically infers the partial differential equation operators encoding the physical process. Numerical experiments showcase that the proposed digital twins have superior properties in real-time intelligent computing and generalization capability in predictive modeling. Nevertheless, it should be mentioned that the presented work is still in an early stage of intelligent digital twins.