Atomistic Thermal Modeling of GAA FETs Using Machine Learning Potentials

Ultra-thin Si fins-based Field-Effect Transistors (FETs) implemented in CMOS devices are subjected to self-heating, which reduces their performance and reliability. Large scale atomistic simulation of transistors makes it possible to model thermal transport and examine thermal performance of nanometer scale FETs. The Gate-all-around (GAA) FinFET introduces stack of nano-materials into the device architecture to enhance the performance of the FinFET, but this also leads to strong self-heating effects. The thermal transport modeling of these FETs including stack of nano-materials becomes significantly challenging due to a lack of well parameterized force fields for the materials used in FETs, e.g., the High-K metal gate (HKMG) stack used to wrap around a silicon nanosheet. HKMG stacks also use amorphous oxides as well as gate metal, for which the thermal transport mechanisms needs to be explored using the extended phonon model. We develop machine learning potentials (MLPs) for stack of nano-materials used in FETs and perform Molecular Dynamics (MD) simulations of GAA FinFETs using these MLPs. The developed MLPs present a significant advantage compared to ab initio methods as they can accommodate a high number of atoms, enabling modeling of nanometer scale devices. We analyze the thermal transport from the Si-nanosheet to the HKMG stack using non-equilibrium MD (NEMD) simulations. We use the NEMD simulations to estimate the thermal properties such as conductivity and thermal boundary resistance of the nanosheet and HKMG stack layers. An analysis of the thermal transport mechanisms is performed using spectral decomposition of the heat current to identify contributing mechanisms from the extended phonon models (locons, diffusons, and propagons) within the amorphous oxides and gate metal.
"This presentation is in memory of Dr. Natalio Mingo."