Session: K9-10: NANOSCALE THERMAL TRANSPORT MODELING AND MACHINE LEARNING II

Paper Number: 138384

138384 - Data-Driven Prediction of Thermal Field in Field-Effect Transistors Using Deep Neural Networks 

Abstract: 

The state-of-the-art unit in the logic devices of microprocessors is fin field-effect transistor (FinFET). The fin in FinFET serves as the channel for current flow between the source and drain terminals, leading to self-heating due to interactions between electrons and phonons. As the fin size of the FinFET decreases to smaller than the mean free path of silicon, accurate prediction of the thermal field demands the solution of the phonon Boltzmann transport equations (BTE). However, numerically solving the phonon BTE for a large array of FinFET transistors is computationally intractable. In this work, we employ a numerical model to simulate the steady-state temperature profile for FinFET transistors. The predicted peak temperature rise was 170K higher than the predictions by the conventional Fourier’s law of heat conduction. Based on the numerical solution, we then develop a data-driven model using deep neural networks as an efficient surrogate for the fast prediction of the temperature field. The inputs to the neural network were the coordinates of cell centers of the mesh and the output was temperature. The neural network is trained and tested on BTE solutions and the important hyperparameters are optimized. The efficiency and accuracy of the neural network model are quantified.

Presenting Author: Jyoti Panda University of Notre Dame

Presenting Author Biography: Dr. J. P. Panda is a researcher in heat transfer, data-driven modeling, and computational fluid dynamics (CFD) at the University of Notre Dame, USA. He received his PhD in CFD (Turbulence Modeling) from the Indian Institute of Technology Kharagpur. Dr. Panda has several years of experience in developing CFD models and applying them to various practical problems in engineering and science. He is currently working on a research project at University of Notre Dame, USA, focusing on the development of data-driven models for heat transfer in nano-transistors using phonon Boltzmann transport equation solutions and explainable artificial intelligence (AI).

Authors:

Jyoti Panda University of Notre Dame
Jiahang Zhou University of Notre Dame
Pan Du University of Notre Dame
Bo Zhang University of Notre Dame
Wenjie Shang University of Notre Dame
Jianxun Wang 46556
Tengfei Luo University of Notre Dame