Recent work has shown constrained Bayesian optimization to be a powerful technique for the optimization of industrial processes. We adapt this framework to the set-up and optimization of atmospheric plasma spraying processes. We propose and validate a Gaussian process modeling structure to predict coatings properties. We introduce a parallel acquisition procedure tailored on the process characteristics and propose an algorithm that adapts to real-time process measurements to improve reproducibility. We validate our optimization method numerically and experimentally, and demonstrate that it can efficiently find input parameters that produce the desired coating and minimize the process cost.