With decades of research and development, confocal microscopes have been the work horse of scientific and industrial 3D measurement. However, due to its requirement for axial scanning, its range of application is limited by its slow measurement speed. Chromatic confocal measurement systems have been developed to eliminate the need for mechanical scanning. Nevertheless, they are still bottle-necked by the transfer and processing of densely sampled spectral data. In this article, Bayesian experimental design is applied to the chromatic confocal measurement scheme,
allowing for more efficient spectral sampling. Recurrent neural network (RNN) is trained to approximate full Bayesian experimental design with much less computation. Simulations have demonstrated that experimental
design approximated by RNN provides better results than an equidistant sampling scheme and performance close to full Bayesian experimental design.