Pretty sure what you will end up with in the best case is a numerical integrator of the solar system as represented by a neural net. I also expect to have pretty crappy stability, but its an interesting question what sort of time step method it would approximate.