If I understood correctly, on page 2 they say they generated their training and test cases using random initial positions and velocities for the 3-body problem. But this is exactly the sort of case that a non-sophisticated solver should be able to handle easily as well. Brutus is special in the sense that it tries hard to give some sort of a guarantee about the accuracy of its solution to a chaotic problem. But the neural network solver provides no such guarantee, and has not been trained or evaluated on "hard" examples. I'm not sure the comparison is fair or particularly meaningful. I wonder if exactly the same sort of "positive" result could be achieved with a standard symplectic integrator, it's not at all obvious from the paper that they have ruled that out.