Machine Learning Algorithms for Semiconductor Device Modeling

This chapter describes machine learning algorithms from the viewpoint of semiconductor device modeling. Machine learning has become one of the most favorable modeling techniques for device, circuit, and system-level modeling. Electrical characteristics of the device such as I-V and C-V characteristics represent an input-output relationship; also temperature, process variation, and power variations can be modeled by machine learning algorithms. Semiconductor device modeling is an area of extensive research. Technology computer aided design (TCAD) models are used to design and analyze semiconductor devices. They are also used as a first step in compact model development. Compact models are used for circuit design and functional verification. Accuracy of these models directly translates to design accuracy, high yield, and more profit. However, development of TCAD and compact models is a huge undertaking. It requires concrete understanding of underlying physics. It also requires a lot of simulation resources (computation time and memory), making it very complex and expensive. Additionally, models developed for one device cannot be ported, i.e. used for another device, if governing physical principles of two devices are different. This chapter is motivated by the need of acceleration of model development for semiconductor devices. Research efforts have been directed in semiconductor device modeling and machine learning fields separately; however, the use of machine learning in device modeling is yet to be explored by the research community. In this chapter, we describe various machine learning algorithms for semiconductor device modeling, challenges, and trade-offs associated with them. As a case study, we have proposed electrothermal modeling of GaN-based high electron mobility transistor (HEMT) devices. A data-driven approach has been implemented for a temperature range varying from 300 K to 600 K, based on one of the core methods of machine learning techniques, i.e, decision trees. The performance of the proposed models was validated through the simulated test examples. The attained outcomes show that the developed models predict the HEMT device characteristics accurately depending on the determined mean squared error between the actual and anticipated characteristics.