Modeling semiconductor performance and yield with empirical data using Monte Carlo methods

Date

2009-08

Journal Title

Journal ISSN

Volume Title

Publisher

Texas Tech University

Abstract

In this dissertation, a Monte Carlo semiconductor performance model based on empirical relationships is introduced. This novel approach results in a very low input dimension macromodel based on a small training sample and is shown to have equal or better precision and accuracy than a typical high dimension multivariate regression model. In order to compensate for input dimension, the regression error, which is typically neglected, is characterized and used as an input to a Monte Carlo model. This error modeling technique intentionally induces error into the model in an attempt to improve precision on long term forecasts. In addition, these techniques allow a sensitivity analysis and forecast to be made based on transistor targets only, meaning that no test lots are required to tune the process. The techniques described in this dissertation may also have other applications, because they can be applied to any situation that requires highly characterized outputs based on a small sample of inputs from a much larger population.

Description

Citation