Scanning an airport terminal is a data problem before it is a modeling problem. The capture covered roughly 2.1 million square feet of building: terminal, several parking structures, and the roads between them. The deliverable was not a mesh for viewing. It was a 3D model detailed enough to support facility operations and MEP management: HVAC, plumbing and drainage, electrical systems, and the equipment attached to each. The digital twin had to hold geometry and asset data in the same place.
I worked on the capture and modeling pipeline for CS TECH Ai. This is the technical account, in the order the work happened.
Control network and registration
Every scan station tied to a control network set with targets on structural columns and permanent features. We treated the network as the contract. If registration drifted outside tolerance, we rescanned instead of rationalizing the residual. Residuals were reviewed per station and per floor, never as a single project average that hides the worst corner.
The check that catches drift
A point cloud that looks aligned can still be wrong. We measured known distances between control points in the registered cloud, ran section cuts through slabs and ceilings, and compared floor elevations across a level. Laser scanning errors hide in long corridors and at the top of tall atria, so those areas got extra stations and a second operator.
Classify before you model
Automatic classification handles floors, ceilings, and walls well. It is less reliable on a cable tray running beside a duct at the same height. Our workflow was semi-automatic: machine classes for the bulk, manual passes for the overhead services band. A short Python script trimmed that band out of the full file so the modeling team worked with a fraction of the points.
Above a public ceiling, ducts, pipes, and trays share a few feet of depth, and every one of them has an owner. That band is where the model earns its keep.
Modeling by system
HVAC
Ducts were modeled from centerlines with dimensions read from the cloud. Terminals, VAV boxes, and air handlers received equipment records, not just geometry, because a duct with no connected equipment is only half useful.
Plumbing and drainage
Gravity sets the rules, so pipe slope was checked in the model rather than assumed. Risers were traced floor to floor before any horizontal run was drawn, which kept the drainage network connected in something closer to reality.
Electrical
Panels, feeders, conduit, and cable trays. Where panel schedules existed, they were linked as attributes. Where they did not, the panel was flagged as incomplete instead of invented, and the flag carried through to delivery.
Every element carried a stable ID and an asset record: system, type, size, location, and equipment information. The schema was agreed with the airport before modeling started and followed BIM naming rules where they existed. That agreement is the only reason the output survived review.
Checks that found real problems
We ran consistency checks instead of trusting the eye. Duct sizing against the scanned outline. Pipe continuity at penetrations. Equipment counts per mechanical room against the photo log. The checks found a missing riser, two trays that changed elevation without a fitting, and a rooftop unit on the wrong side of a parapet.
Interoperability
The model shipped as an open BIM reference with linked asset data, so the operator could read it in more than one platform. Geometry lived in one file structure and the asset records in another, joined by ID. That split kept the data portable when the next software version arrived, which it always does.
The analytics layer
Energy analysis uses the same geometry and equipment data. Because the model carries counts and sizes, a simulation starts from installed conditions rather than design intent. Maintenance planning uses the asset records to schedule work and find shutoffs. Security and emergency planning uses room and system locations to test response routes against the building as it exists.
Where fast massing fits
For option studies I lean on GenAnywhere: I draw the area on the map, let it generate a massing model, and iterate. It carries no control network and no systems, so it stays upstream of the scan. Its job is to tell us which option deserves a survey, not to compete with one.
What I would standardize next time
Ship the control report with the model. Agree the classification taxonomy before the first scan. Use asset IDs that match the maintenance system, not the model tree. Run a section-cut pass on every floor before sign-off.
None of it is exotic. It is the difference between a model that gets opened and one that gets archived.



