DEV Community

Cover image for CityEngine vs GenAnywhere: Procedural Rules for a Campus That Never Stops Changing
Generate Build
Generate Build

Posted on Edited on

CityEngine vs GenAnywhere: Procedural Rules for a Campus That Never Stops Changing

A university central plant is a building that gets rebuilt every fifteen years without ever being demolished. Boilers come out, chillers go in, a pipe bridge gets added, a stack gets shortened. The facilities team asked for a model of the plant and the yard that could survive those changes, not a picture of how it looked in October. The campus planning office had already been burned once by a model that looked right and measured wrong. That pushed me toward CityEngine and a rule file instead of a generated scene.

Feed it a table, not a description

The campus GIS already had footprints for every building and a plant asset table with heights, floor counts, and construction dates. I imported both, joined them on the building ID, and set the scene to the campus coordinate system. The asset data did the work that a prompt cannot: every boiler house, cooling tower, and warehouse came in at its recorded height, in its recorded place, with its own attributes still attached. It also had the year each stack was last inspected, which mattered more than the geometry once we started phasing replacements.

A small CGA rule

The CGA rule for the plant buildings is short. Heights come from attributes, floors are split by a floor-to-floor value, and the roof type is a parameter:
Change the height attribute and the stack of floors changes with it. Change the roof pitch and the silhouette follows. That sounds trivial until the third scheme, when the facilities director wants every building raised two meters to match the new flood datum. One attribute, one regenerate, and the LOD is rebuilt from the same source.

Python for the boring part

CityEngine's Python scripting handled the repetition. I looped over a list of schemes, set the height attribute from a CSV, exported glTF and GLB for each one, and dumped a small JSON with floor areas and stack heights. The script also wrote a manifest with file names, dates, and the asset table version, so a reviewer could tell which export came from which snapshot. The export list also covers OBJ, which is what the old structural software on campus still reads. None of that is clever code. It is just the gap between a model you regenerate and one you remodel.

Where the pipe bridge breaks

The one thing the rules could not invent was the pipe bridge between the boiler hall and the chiller plant, because it is not a building. I modeled it by hand, as a swept profile along a path, and attached it to the same coordinate system. That mixture is normal in procedural modeling: rules for the repetitive stock, manual geometry for the one-offs. The model stays editable where it matters and accurate where it is weird. The rule file lives next to the asset table, and both are versioned in the same repository as the drawings.

The same site, described to GenAnywhere

I gave GenAnywhere the boundary of the plant yard. It returned a handsome industrial campus with a glass office block, a landscaped forecourt, and three matching stacks. The stacks were too short, the pipe bridge was gone, and the coal yard had become a parking lot with trees. It even put a green roof on the boiler hall, which would have been a nice idea if the roof could carry it. As a concept image for a facilities newsletter, it was convincing. As asset data, it had nothing to query.

The handoff that mattered

The facilities team wanted a browser viewer, so the glTF went into a three.js scene with a simple layer toggle: buildings, pipe bridges, underground lines. The viewer runs on the facilities intranet, and it saves nothing between sessions, which is exactly what they wanted. A GIS analyst opened the same GLB in ArcGIS and checked the footprint against the parcel layer. The OBJ went to the structural consultant. Everyone got the same geometry at the same coordinates, and the digital twin stayed a twin instead of turning into three unrelated models.
The plant will change again in 2029. The stack schedule, the pipe sizes, and the phase dates are all attributes now, and attributes are easier to audit than screenshots. When the next change lands, the rule file and the asset table are still the source, and nobody has to start from a picture.

Top comments (0)