LEDs are ubiquitous as light sources for a wide variety of use cases. During product development, they have to be tested under different environmental con- ditions. Since standard LED based lighting systems have a lifetime of several thousand hours, this usually does not include extensive lifetime prediction tests. The state-of-the-art are empirical and physical models based on the limited data that is available.
The aim of this thesis is to combine the existing approaches with techniques from machine learning. One particularly interesting method is kriging, as it allows to estimate both the lifetime and its uncertainty based on the measurement data. This enables new ways for adaptive design of experiments (DOE), which may help to streamline and speed up the testing procedures. Lastly, it is of great interest to investigate learning strategies that include hard physical constraints of the underlying process, such as positivity or monotonicity.