I'm looking for some algorithmic advice.  I don't expect code, but if anyone can suggest a link or search words to help narrow it down that'd be great!  

Given a 3D grid of cells (int [][][]), where 0==unoccupied and 1==occupied, convert it into rectangular polygonal solids ("boxes") such that the maximum number of adjacent occupied cells convert into the largest possible boxes.  Essentially aggregating the implied cubes into larger polygon-representation boxes.

Another way of saying it: given a voxel space, determine the minimal set of maximal axis-aligned bounding boxes that fully contain only the occupied voxels.  (which suggests an octree as potential intermediate representation, but total brain-fart on how I'd actually use it once built)

For example, take a 2D case:  on a 5x5 grid, the entire top row is occupied, but only the first 4 rows of the first column are occupied (8 occupied cells total).  Ideally, that would produce a 5x1 unit box along the top, and a 1x3 unit box along the left.  Then consider the broader cases, like if the first two rows were entirely filled (14 total, solution:  5x2, 1x2). Then extend to 3D...

Even obtaining a "pretty good" solution would probably suffice.  (I suspect the optimal solution is hard for arbitrary data)  I have an ugly hack that sort of works, via brute force nested loop manual span scanning, but it's time to look for a better way!  Thanks for any ideas.

Right, could be "hard" as in "NP-Hard".  That's why I'd settle for "pretty good".

This is a mashup, feeding the voxel-type sketch's data into a polygon-type sketch. The polygon sketch will only be interesting with the larger aggregated volumes, but the voxel sketch needs all intended cells filled.

I'll give your "surface" suggestion some thought, it might help, but will be tricky.  There's two approaches, right?: have the voxel-sketch "track" the larger boxes, or have the polygon-sketch "discover" the larger boxes.  Voxel-sketch jumps all around, with no notion of "travelling to adjacent neighbor", so that makes that tricky.

Might be able to concurrently maintain two voxels, one for internal use, one for export with surfaces only.  Or could post-process with some sort of "hollowing-out" edge detector, hmm..

Anyway, much thanks for the suggestion, and sparking further ideas.