Wheel suspension models relying on differential algebraic equation solvers for simulations are computationally resourceful and often not hard real-time capable. In this work, a neural network approach is developed that is based on the NARX model. A multibody vehicle simulation in ADAMS CAR and a high-precision object-oriented SIMULINK-model are utilized to obtain adequate datasets for NN training and testing purposes. Different settings of the model are tried. The final NARX model, containing 5 hidden nodes, is trained in open- and closed-loop configuration by BAYESIAN Regularization. The generalization ability is proven in several different, driving-situation-based test instances. It is shown that the NN can model wheel carrier acceleration and spring / damper reaction force effectively while requiring significantly lower computation time than a DAE-based simulation.