A neural-network-based local inverse mapping technique for building statistical DMOS models

This paper presents a methodology to circumvent the time consuming standard approach for statistical model development. The methodology is a two step process. The first part defines the relationship between electrical device parameters and model device parameters by means of training a neural network. The second stage uses the neural network to create worst-case model parameter sets. In order to select an appropriate set of worst-case electrical parameters, a multivariate statistical analysis is performed, such that correlations between device parameters are taken into account. The neural network approach also enables a Monte-Carlo model to be generated. The advantages of the proposed methodology are its speed improvement and accuracy.