Modeling and prediction of smart power semiconductor lifetime data using a Gaussian process prior

In automotive industry end-of-life tests are necessary to verify that semiconductor products operate reliably. Due to limited test resources it is not possible to test all devices and thus, accelerated stress tests in combination with statistical models are commonly applied to achieve reliable forecasts. Challenging thereby is the highly complex data that shows mixture distributions and censoring. For the main purpose, the extrapolation to other test conditions or designs, neither frequently used acceleration models like Arrhenius, nor complex models like Bayesian Mixtures-of-Experts or Bayesian networks give accurate lifetime predictions, although, the latter two are precise in case of interpolation. To compensate the limitations of ordinary linear based regression models, we propose the application of a Gaussian process prior. The proposed model shows a high degree of flexibility by exploiting sums or products of appropriate covariance functions, e.g. linear or exponential, and serves as a reliable alternative to currently applied methods.