Machine learning and new AI algorithms inspire the scientific community to explore and develop new approaches for discovery of scientific laws and governing equations for complex physical and nonlinear dynamical systems. The question on how well deep learning approaches can create sense from a given set of input data is difficult to answer. Considering the unperturbed two-body problem, this paper investigates the approximation and prediction capabilities of three types of neural networks: Feed-Forward, Residual and Deep Residual. Used in a purely recurrent model, this three architectures are able to produce highly satisfying performances, very close to numerical integration tolerances. Furthermore, the effect of the mathematical representation (i.e. coordinate system) on the learning process is also investigated. From numerical results, it can be inferred that NN were able to better learn inherent dynamics characteristics in spherical coordinates without any apriori information than in Cartesian coordinate system. It is shown that a simple NN architecture is able to learn the symmetry of the central force and reproduce the conservation of the constants of the motion.