Well, the article does mention that Part IV adds erosion. Note also that this particular source is a rare example of working based on a mesh rather than a grid (which complicates the logic - in particular, when do you split/merge nodes? - but should be cheaper at scale).

People can try something fully physics-based (or rather, physics-inspired) even for earlier stages, but there are problems:

* You still need some kind of nondeterministic input so you don't always generate the same world.

* You must do the whole world at once, rather than being able to generate each area independently.

* This requires the computation to run for a long time, and needs to feed back in on itself (think of "lake overflows a natural dam and carves a valley, then the tectonics lift it and change the low point anyway").

* It's very easy for your code to result in "boring" outputs, such as "all flat" or "infinitely deep valleys".