Physics-Informed Neural Network for Predicting Out-of-Training-Range TCAD Solution with Minimized Domain Expertise

In this paper, a Si nanowire transistor is used to demonstrate the possibility of using a physics-informed neural network to predict out-of-training-range TCAD solutions without accessing internal solvers and with minimal domain expertise. The machine can predict a 10 times larger range than the training data and also predict the inversion region behavior with only subthreshold region training data. The physics-informed module is trained without human-coded differential equations making this extendable to more sophisticated systems.