Physics-Informed Deep Learning Models in Semiconductor Device Simulation: Recent Progress

Abstract

As semiconductor devices become smaller, their structures and physical movements become more complicated, and the simulation of technology computer-aided design (tcad) requires more computing power. Physics-driven deep learning solves this problem by introducing governing equations, boundary conditions and physical restrictions into neural networks, which can go beyond the limits of methods using traditional data with low explanatory power and poor generalization ability. In this paper, it is helpful to use the physics-informed neural networks (pinns), neural operators, physics-informed compact models and AI to optimize semiconductor equipment. These methods can not only combine physical principles, but also speed up some simulation and design work. However, due to the training cost, physical accuracy, out-of-range reliability of data in training, it is difficult to combine with the existing TCAD workflow. Therefore, it can be used in various situations. In general, the review results show that the better mixing mechanism of physics-driven model can supplement or accelerate some TCAD work, rather than replace TCAD. The extent to which it can be used depends on the type of equipment, the range of offset and the distribution of data used for training.