The Noise, Vibration and Harshness development of vehicle components still heavily relies on full vehicle tests. To reduce costs, the number of simulations is increasing. Commonly used simulation methods (Multi-body-Simulation, Finite-Element-Method and Statistical-Energy-Analysis) are only valid within a limited frequency range and need high computational resources. The manually created models require validation through measured data. Holistic digitalization is therefore not achievable with today’s simulation methods. Machine learning as a different approach lets an algorithm compute the physical relation between input and output. Once this relation is found, the trained neural network can predict the output to any new given input. Time- and frequency-domain input data show different suitability. Network architecture and hyper-parameters are essential for the outcome. The network training is supported by Finite-Element computations.