Bayesian Experimental Design for LEDs using Gaussian Processes

We present a novel design of experiments approach for LEDs based on Gaussian processes. The method aims to decrease the measurement effort associated with characterizing LEDs based on their spectral power distribution (SPD) and derived quantities of interest (QOIs), such as the luminous flux or color coordinates. It is both easy to implement and based on open source software. We showcase the approach on an example taken from automotive applications. For the considered example, we are able to decrease the total number of measurements by over 75%, while the model is able to predict the SPD and derived QOIs with relative errors of less than 5%.