Bayesian Prediction of SMART Power Semiconductor Lifetime with Bayesian Networks

Abstract

In this paper Bayesian networks are used to predict complex semiconductor lifetime data. The data of interest is a mixture of two lognormal distributed heteroscedastic components where data is right censored. To understand the complex behavior of data corresponding to each mixture component, interactions between geometric designs, material properties, and physical parameters of the semiconductor device under test are modeled by a Bayesian network. For the network’s structure and parameter learning the statistical toolboxes BNT and bayesf Version 2.0 for MATLAB have been extended. Due to censored observations MCMC simulations are necessary to determine the posterior density distribution and evaluate the network’s structure. For the model selection and evaluation goodness of fit criteria such as marginal likelihoods, Bayes factors, predictive density distributions, and sums of squared errors are used. The results indicate that the application of Bayesian networks to semiconductor data provides useful information about the behavior of devices as well as a reliable alternative to currently applied methods.