A university lab built an AI that models a skyscraper's structure in 5 minutes. Structural engineers still have to check its work.
University of Miami researchers built CivilBot, a tool that turns a plain-language structural description into a working SAP2000 model up to 30 times faster than doing it by hand. For GCs running design-assist, faster structural iteration cuts both ways.
A structural engineer can now describe a building in plain English and get a working analysis model back in minutes instead of days. University of Miami researchers built the tool, called CivilBot, and it's a real preview of what AI does to the front end of structural design — and a reminder of exactly where a GC still needs to slow down and check.
What did the researchers build?
Minghui Cheng, an assistant professor jointly appointed to University of Miami's civil engineering department and school of architecture, built CivilBot with graduate student Ziheng Geng, Hunan University associate professor Ran Cao, and University of Illinois Chicago computer science professor Lu Cheng. An engineer feeds it a plain-language description of a structure — beam lengths, supports, loads — and CivilBot generates the code needed to build the corresponding model in structural analysis software such as SAP2000. That's the step that normally eats a huge share of an engineer's modeling time: not deciding how a building should behave structurally, but the manual, repetitive work of clicking that decision into existence inside the software.
The team reports the tool cuts that modeling process from days to hours, on the order of 20 to 30 times faster than doing it by hand. As a demonstration, they had CivilBot generate the code for all the structural conditions of downtown Miami's Freedom Tower — a real, complex, landmark building — in about five minutes. The tool is on its third version and currently handles 2D models, with 3D capability still in development.
What does this mean for a GC running design-assist or early BIM coordination?
Structural drawings and models are a hard dependency for a lot of preconstruction work — early trade coordination, clash detection, long-lead procurement decisions that hinge on structural framing. When the engineer of record's modeling step compresses from days to hours, two things follow, and they pull in opposite directions:
| What changes | Why it matters to a GC |
|---|---|
| Structural model iterations arrive faster | Design-assist teams can get feedback on framing changes in near-real-time instead of waiting through a modeling cycle |
| Early deliverables are less likely to be "final" | A model produced in five minutes is easier to revise again in five minutes — don't treat a fast handoff as a frozen one |
| More design churn is possible before documents freeze | BIM coordination teams need to confirm which model version is current before running clash detection against it |
| AI-generated model code can carry hallucinated errors | The engineer of record's review step doesn't go away — if anything it matters more, since a wrong support condition or load path can look identical to a correct one in a model viewer |
That last point isn't hypothetical. Separate academic research on this exact class of problem — using large language models to automate structural modeling across software platforms — is specifically focused on reducing hallucination rates in multi-step model generation, because plausible-looking but wrong output is a known failure mode for this kind of tool. CivilBot's own limitation to 2D models, with 3D still being validated, is the team building in caution rather than shipping a black box.
The takeaway
Faster structural modeling is a real, useful development — for the engineer of record's own workflow, it's a legitimate multiple-of-productivity gain. But "faster model" isn't the same as "reviewed model," and a GC or design-assist team pulling structural data into a coordination model needs to keep asking that question explicitly as tools like this spread: is what I'm looking at an engineer-reviewed deliverable, or a first pass that took five minutes to produce? CAB flagged the same review-before-trust gap when covering structural health monitoring technology after the Surfside NIST findings — the tooling gets better every quarter, but the engineer's sign-off is still the thing standing between a model and a building.
Forward this to the person on your team who's still arguing AI is overhyped. Subscribe at constructionaibrief.com.
Next time a structural deliverable comes back faster than expected, ask what got reviewed versus what got generated — before it goes into your coordination model.
- What is CivilBot?
- CivilBot is an AI tool built by University of Miami professor Minghui Cheng, graduate student Ziheng Geng, and collaborators at Hunan University and the University of Illinois Chicago. An engineer describes a structure in plain language — beam lengths, supports, loads — and CivilBot generates the code to build a working structural analysis model in software like SAP2000.
- How much faster is CivilBot than modeling by hand?
- Researchers report it turns a modeling process that normally takes days into one that takes hours — up to roughly 20 to 30 times faster on the repetitive model-setup work. As a demonstration, the team had CivilBot generate the full structural-condition code for downtown Miami's Freedom Tower in about five minutes.
- Does CivilBot design the building or just build the model?
- It builds the analysis model, not the design. The engineer still supplies the structural description and loads, still has to verify the model matches intent, and still runs and interprets the analysis. CivilBot removes the manual data-entry step between a design idea and a working model — it doesn't remove the engineer of record's judgment or stamp.
- Can AI-generated structural models be trusted without review?
- No. CivilBot is currently limited to 2D models, with 3D still in development, and separate academic research on this exact problem — automating structural modeling with large language models — is focused specifically on reducing hallucination errors in multi-step model generation, because LLM-generated engineering code can produce plausible-looking mistakes. Every output still needs an engineer's review before it's used for design or permitting.
- What should a GC take from this if they're not the structural engineer?
- If your engineer of record starts turning around structural models in hours instead of days, don't assume the earlier deliverable is a locked deliverable. Ask what's been reviewed and stamped versus what's a fast first pass, and build that distinction into your design-assist and BIM coordination schedule instead of treating every model handoff the same way.