In this study, we aim at designing rotorcraft trajectories whose associated noise footprint is minimized. The values for the objective function are provided by an industrial black-box. Therefore, a method that is adapted to Black-box Optimization (MADS) is used to solve this problem. We propose a neural network surrogate defined thanks to our knowledge of the problem to replace the expensive black-box. Numerical results using the proposed surrogate within the MADS algorithm are promising in terms of both computational time savings and accuracy of the computed solution.