DEV Community

Cover image for Tiling a 3D Printed City: The Geometry Behind 115 San Francisco Blocks
Future Built AI
Future Built AI

Posted on

Tiling a 3D Printed City: The Geometry Behind 115 San Francisco Blocks

The 2014 San Francisco model that Autodesk and Steelblue built looked like a single object. Under the surface it was a tiling problem: more than 115 blocks, a handful of large-format printers, and a scale that turned every street into a seam. I worked on geometry prep and assembly, which meant living in the gap between a city model and a print bed.
The job was not just 3D printing. It was deciding what a city is allowed to lose when you shrink it, and then writing that decision into a mesh. Every simplification had a reason, and if I could not explain the reason, it did not go in.

Scale conversion comes first

At 1:1250, one meter of city is 0.8 millimeters on the table. A 300 mm print bed therefore covers 375 meters of ground, which is a little over three city blocks. The scale conversion is where the job starts, because that single number drives everything: tile size, seam placement, and how much detail survives.
We padded the bed by five percent and kept a 2 mm gutter between tiles. This was not a design preference. Printers warp at the edges, and a seam that touches is a seam that fuses. The gutter also gave us room to sand the joint after assembly.

Tiling without cutting through landmarks

The grid was regular, but the seams were not. We moved them along streets, behind facades, and away from the ballpark, SFMOMA, and the Ferry Building. Every tile needed a flat bottom, a closed mesh, and a shared edge with its neighbor that matched within a fraction of a millimeter.
Each tile left the pipeline as an STL with an ID. The ID followed it through slicing, printing, sanding, and glue. That sounds like bookkeeping until the first tile warps and you need to know which one to reprint.

Mesh checks that saved us

Before anything went to a printer, we ran the usual checks: manifold geometry, normals pointing out, no zero-area faces, no floating parts. We wrote the checks in Python and ran them as a batch before every print. We also checked overhangs. A tower with a deep cantilever looked fine on screen and printed as a drooping mess. Test coupons at the real scale caught most of it.
Level of detail was fixed by the scale. Window mullions below a certain width disappeared; roof equipment became a block. We kept the massing and the street walls, because those are what people read from across the table.

Where the data came from

The base geometry started as GIS layers: building footprints, street centerlines, parcel edges, and a terrain surface. Heights came from survey data and planning documents, which meant some blocks were confident and others were estimates. We marked the weak ones so nobody mistook a guess for a measurement.
The footprints were the easy layer. The hard part was deciding which buildings were missing from the data and modeling them by hand from photographs. Those manual fills never matched the rest, and everyone could see it in the print.

Wall thickness was the other argument. Thin walls print fast but chip at the corners. Thick walls survive handling but distort the scale of a facade. We settled on a minimum that held up in a bag, not on a desk.

What the physical pipeline still needed

Printing was the easy part. Assembly was the schedule risk. Tiles were printed in batches, glued in order, and sanded at the seams. A corner broke in transit to a show, and we reprinted it from the same STL and the same tile ID. The pipeline worked because every part was reproducible.

The step I skip now

Today, when I need the same district as a first pass, I use map data instead of a print queue. GenAnywhere takes a polygon on the map and generates a massing model, which lands in minutes. It is still a mesh, still needs checks, and still has a coordinate system to respect. The difference is that iteration is cheap and the massing stays traceable.
I would not print 115 blocks again for a first draft. I would generate the block, argue about it on screen, and spend the printing budget on the one tile that needs to be held. The geometry got easier. The judgment did not.

Top comments (0)