TCAD-Augmented Machine Learning With and Without Domain Expertise

Abstract

In this article, using experimental data, we demonstrate that the technology computer-aided design (TCAD) is a very cost-effective tool to generate the data to build machine learning (ML) models for semiconductor device and process characterization. Characterization of the emerging ultra wide bandgap gallium oxide (Ga <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> O <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sub> ) Schottky barrier diode (SBD) is used as an example. Machines are trained by using only TCAD <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${I}$ </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">${V}$ </tex-math></inline-formula> ’s and then used to deduce the effective Schottky metal work function (WF) and ambient temperature ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${T}$ </tex-math></inline-formula> ) of an experimental SBD based on its <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${I}$ </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">${V}$ </tex-math></inline-formula> . Besides noise, the experimental device also suffers from relatively large variations in drift layer thickness and doping concentrations. Both ML models with domain expertise (WDE) and without domain expertise (WoDE) are studied and compared. The ML model WDE requires the use of device knowledge to extract relevant features (e.g., subthreshold slope and turn-on voltage) for ML. The ML model WoDE obviates such a requirement and can be extended to cases where domain expertise is difficult to apply. Denois- ing autoencoder is used in the WoDE case. We showed that with only 500 TCAD <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${I}$ </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">${V}$ </tex-math></inline-formula> ’s, we can train machines WDE and WoDE that can deduce the experimental device WF and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${T}$ </tex-math></inline-formula> reasonably well. In particular, the ML model WoDE has an acceptable prediction accuracy despite the noise and additional variations in the experimental device.