For a given set of generic rigid bodies where one defines a point of rotation for each body, single or multiple interface points on each body and the associated loads acting on these interfaces, one can formulate the dynamics of the assembled system in a decoupled manner. The decoupled representation of the dynamics would be decoupled both from the active constraints, and also on a component level, where the dynamics of each body would be decoupled from one another. Based on this theoretical framework, a neural network (NN) structure which would model a single component of interest is proposed. The component can be fully described through its mass matrix and a vector containing the generalized body forces and the body's quadratic velocity vector, both of which the NN should learn. This representation of the dynamics serves only as justification for the proposed NN structure, the solution strategy used to obtain the system dynamics and the generalized interface loads can be freely chosen. The proposed NN is tasked with learning the relationship between the component's motion and its derivatives, and the acting interface loads on that component in a manner conforming to the aforementioned description of the system dynamics. The NN constructed based on the novel description of the system dynamics introduced here significantly reduces the complexity of modeling multibody systems. That is due to the possibility to process the dynamics of the system one component at a time. Furthermore, the acquired component models can be seamlessly embedded into existing codes for multibody simulation, and a NN was successfully trained to capture the dynamics of a rigid multibody system accurately. NN design techniques and considerations from previous works can be further incorporated into the outlined approach to improve its accuracy.