Abstract <p>Structural mechanical simulations of short-fiber reinforced plastics are significant for evaluation of noise, vibration and harshness performance. In order to get an overall picture of the structural-dynamic behaviour of different conditioning states the dynamic elasticity constants to describe linear-viscoelasticity must be known. Whilst the dynamic modulus of elasticity can be determined experimentally as a function of frequency and fiber orientation for different temperatures and humidities, the challenge is significantly greater for dynamic, direction-dependent shear moduli and Poisson’s ratios. Based on a two-step homogenisation approach, a dataset was generated by matching homogenised stiffness predictions with experimentally determined directional elastic moduli to derive effective matrix properties and orthotropic elastic constants. Artificial neural networks were trained on the generated material dataset comprising three thermoplastic matrix materials to predict orthotropic elastic and damping properties as a function of fiber orientation, frequency, and environmental conditions, and were integrated into the finite element simulation as a surrogate model of the material behaviour. The prediction quality was evaluated by comparing simulated frequency response functions of bending specimens with experimental data. The results show sufficient correlation for 0° and 90° cutout orientations. For 30° and 45° orientations, the deviations were greater due to shear coupling. A method with constant Poisson’s ratios also showed sufficient prediction quality for samples with shear coupling. The two-step homogenisation approach tended to overestimate stiffness and underestimate damping.</p> Graphical abstract <p></p>