Sound simulations by physical modelling are interesting to transcribe the physics underlying the functioning of a musical instrument. These simulations make it possible to listen to a virtual instrument with a mode of operation representative of the musician-instrument interaction. The work consists of studying the contribution of machine learning (ML) methods in the understanding of the relationships between the shape of a trumpet and the sound simulated. The physical model used is based on an acoustical modeling of the resonator, a mechanical model of the excitator, and an aeroelastic coupling between the excitator and the resonator. From different samples of the input impedance of the resonator, time domain simulations are generated to constitute a training set of sounds. Supervised learning is next trained to the data, with the impedance as input and sound descriptors as outputs, using classical ML methods (neural networks). The ML model is finally used to optimize the sound descriptors levels, according to the input impedance. To illustrate the approach, different "targets" for the sound features are considered (brightness, intonation), and a validation is conducted with the simulations. The approach is a first stage toward a "customization" of an instrument according to different perceptual dimensions.