neural networks have the ability to serve as universal function
approximators, but they are not interpretable and don't generalize well outside
of their training region. Both of these issues are problematic when trying to
apply standard neural ordinary differential equations (neural OD