Out-of-Training-Range Synthetic FinFET and Inverter Data Generation Using a Modified Generative Adversarial Network

Abstract

In this letter, a novel variation of Generative Adversarial Network (GAN) is proposed and used to predict device and circuit characteristics based on design parameters. Unlike regular GAN which takes white noise as inputs, this modified GAN uses device or circuit parameters as inputs. Unlike regular Physics-informed GAN (PI-GAN) which incorporates differential equations in the training process, this modified GAN learns physics through the inputs and has one extra step of supervised learning. FinFET is used as a device example and Technology Computer-Aided-Design (TCAD) is used to generate its current-voltage ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${I}_{D}{V}_{G}$ </tex-math></inline-formula> , <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${I}_{D}{V}_{D}{)}$ </tex-math></inline-formula> and capacitance-voltage ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${C}_{G}{V}_{G}$ </tex-math></inline-formula> ) curves as the training data by varying the gate length ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text{L}_{\text {G}}{)}$ </tex-math></inline-formula> , fin top width ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text{W}_{\text {TOP}}{)}$ </tex-math></inline-formula> , and gate metal workfunction (WF). A CMOS inverter with source contact defects is used as a circuit example and a SPICE simulator is used to generate its Voltage Transfer Characteristics (VTC) by varying the source contact resistances. We show that 1) the GAN model is able to generate both the device and circuit electrical characteristics based on the input parameters, 2) it can predict the characteristics of the device and circuit out of the training range (in a testing volume 3.7x to 4.6x larger than the training volume), and 3) it is further verified on experimentally measured data in the inverter case that it does not overfit and has learned the underlying physics.