You may also want to experiment with different initial states in addition to random. E.g. have the states intermixed as scanlines, or start with concentric circles of multiple layers etc. I'd expect those might help to better understand the automata's behavior.