Comment by Peteragain

The linked original paper is readable and answered many questions. They simply embrace the idea that some "crazy shape" will work, and then do "machine learning" in a simulation to find it.