Survey of recent advanced statistical models for early life failure probability assessment in semiconductor manufacturing

In semiconductor manufacturing, early life failures have to be screened out before delivery. This is achieved by means of burn-in. With the aim to prove a target reliability level and release burn-in testing of the whole population, a burn-in study is performed, in which a large number of items is investigated for early life failures. However, from a statistical point of view, there is substantial potential for improvement with respect to the modeling of early life failure probabilities by considering further available information in addition to the performed burn-in studies. In this paper, we provide ideas on how advanced statistics can be applied to efficiently reduce the efforts of burn-in studies. These ideas involve scaling the failure probability with respect to the sizes of the different products, as well as taking advantage of synergies between different chip technologies within the estimation of the chips´ failure probability level.