Hey,

I’m working on a system that processes IFC (building) models and tries to extract structured information from them (areas, elements, categories like balconies, facades, etc.).

The geometric part is mostly deterministic and working fine.

The problem I’m facing is at the semantic layer.

For example:

• What exactly defines a “balcony” in a model?
• Which elements belong to it? (slabs, railings, proxies, etc.)
• How do you distinguish facade vs exterior wall vs decorative elements?

Right now I’m trying to do this with rules based on:

• geometry (position, orientation, bounding boxes)
• relationships (adjacency, containment)
• element types (IfcSlab, IfcRailing, etc.)

But it breaks down in real-world models because:

• models are inconsistent
• element naming is unreliable
• geometry alone is ambiguous

So I’m stuck between:

1. Making increasingly complex rule-based systems
2. Introducing some kind of probabilistic / AI-assisted interpretation layer

My goals:

• keep the system explainable and deterministic where possible
• avoid overengineering
• still handle messy real-world IFC files

Questions:

• How would you approach defining these kinds of “semantic zones”?
• Is a hybrid (rules + AI suggestion layer) a reasonable approach?
• Are there known patterns for this kind of problem?

Would really appreciate guidance from anyone who has worked with IFC/BIM or similar structured-but-messy data.

Thanks