I think there needs to be balance. There should be higher stakes in general the farther you dig, and more rewards. underground trading partners, exotic animals and plants that could be useful and or dangerous. Pockets of poisonous gas, Ruins of underground cities with powerful artifacts within or possibly ancient texts revealing powerful magics or forgotten recipes for unique alloys (as well as possible guardians), strange and bizarre metals and minerals unavailable near the surface, strange magical curses from close proximity to the underworld. etc etc. Perhaps dwarven caravans should only come with wagons when you reach the caverns and establish a road going off-map? The key is to make the underground both more difficult and grants even greater opportunity-not just in a way that makes the game too hard or too easy-both is the key.
This requires world generation, or at least site generation, to arrange things according to metagame concerns: how rewarding and/or challenging is this feature
for the player? Which is fine--roguelikes have been doing that for ages, putting shinier loot and tougher monsters at deeper levels of the dungeon. But around here, that suggestion will be met with a lot of snorting about how "gamey" and unrealistic and lacking in Procedural Generation Magic it is*.
Everyone needs to stop whining about that and get behind the idea of an explicit difficulty curve associated with depth. Yes, like the 2d version.
I note with minor amusement that we've just recapitulated the gamist/simulationist debate in RPG design. So think of it in those terms. If this were a tabletop RPG, the GM could draw a map of the world and randomly choose a location that has a portal to hell. Then you'd point to a spot on the map to build your fortress, and most likely not hit the portal to hell. But that's bad GMing. A
good GM will have you pick a location, and then decide that that's where the Hellmouth is.**
So I'd say what we need is a
procedural gamemaster: an algorithm that manages the world, or the corner of it where the player is operating, to present challenges and rewards. It doesn't have to be a Hellmouth--I'm all for lost civilizations, forgotten beasts, exotic minerals, and secret networks of dwarven roads. But the game can't just pregenerate a world and then expect me to prospect around for a site that's not boring.
* even though it would still be procedurally generated--just not as a simulation of some physical process.
** Or he'll announce that "Five years ago, a portal to hell opened up in Bronzehammer. Everything within a hundred miles is now a demon-haunted wasteland. Your mission, if you choose to accept it, is to close that portal." That's a different kind of story, and a good sequel to the game where you built Bronzehammer and discovered that, against all luck, you're on top of the Hellmouth. Which is to say that if the GM algorithm decides to put a portal to hell under your site, then once the portal opens it's part of the persistent world.