CAD/BIM Tips & Tricks
Your CAD Model Looks Great, So Why Is the Engineer Nervous?
11 September 2026
“Can you change that diameter?”
“Look what AI did.”
The engineer looks at the screen.
It’s beautiful. The geometry is clean and the proportions look right. There are no odd surfaces or mysterious rectangles floating in from another dimension.
“Wow,” the engineer says. Then he leans in closer and points at the model.
“Can you change that diameter?”
And just like that, the real test has begun.
Pretty Isn’t the Same as Parametric
Ask an AI housekeeper to clean everything it can see and you may come home to a spotless house. Look under the bed, however, and the illusion could end rather abruptly. In engineering and design, what remains hidden inside a model matters far more than a little forgotten dust.
AI is getting good at producing CAD models that look professionally made. But anyone who works in CAD knows that appearance is only the beginning. The real value lies in how the model behaves when something changes.
In a parametric model, dimensions, constraints and feature relationships tell the design what to do when something changes. Increase a hole diameter, move a feature or alter a component’s length and the model should rebuild accordingly. Features that depend on the edited geometry should follow. Everything else should stay put.
Until its behavior has been tested, that impressive model is still a 3D picture with ambitions.
Without dependable behavior, you haven’t created an intelligent engineering model. Until its behavior has been tested, that impressive model is still a 3D picture with ambitions.
Enter CADEngBench
On 10 August 2026, Harmanjot Singh, Abhra Dubey and Jorge Alejandro Amador Herrera posted the CADEngBench paper to arXiv. The paper examines parametric design, assembly reasoning and physics simulation. Its central question was refreshingly direct: Does AI-generated CAD actually work?
The researchers weren’t awarding points for a handsome rendering. They were looking for evidence that the model behaved like usable engineering CAD. They wanted to know whether the model could meet engineering requirements, survive changes, support edits, behave credibly in a simulation, and connect properly with other parts. It couldn’t just pose for the picture. It had to perform.
CADEngBench-P tests 300 parametric parts twice. AI first generates each part from an engineering brief. It then edits existing CAD without damaging anything that should remain untouched, producing 600 tasks in total.
The question is no longer simply whether AI can draw a part. It’s whether an engineer can trust the result.
CADEngBench-A turns to assemblies. Using 150 pairs of components, it tests whether AI can determine how the parts should connect, identify the exact faces or edges that meet and predict how the resulting joint should move.
So, the question is no longer simply whether AI can draw a part. It’s whether an engineer can trust the result.
Is It Usable CAD?
Before anyone worries about physics, joints or clever edits, the model has to clear a more basic hurdle: Does it work as CAD?
The AI-generated CadQuery program must run, create a valid solid, export it as a STEP file and allow that file to be imported again. STEP is a vendor-neutral format for moving precise 3D geometry between CAD programs. It preserves a part’s shape and structure, though generally not the native file’s complete parametric history or constraints.
The benchmark also checks the model’s boundary representation, or B-Rep, which mathematically describes a solid through its faces, edges and vertices. It tells the CAD system, “Yes, this is a proper object,” rather than surfaces doing a convincing impression of one.
An invalid B-Rep can cause later work to fail even when the model looks fine. Boolean operations, meshing, manufacturing preparation and edits all rely on geometry the software can interpret.
Passing this stage doesn’t automatically mean the part meets its brief. CADEngBench also checks dimensions, features, interfaces and task-specific requirements. Where relevant, CADEngBench applies limited design-for-manufacturing (DFM) checks covering wall thickness and hole geometry.
The research paper describes these as general geometry checks, not manufacturing certification. That’s important. A benchmark examining a part on screen can’t account for every material, process, tolerance or shop-floor reality.
A BIM model can look complete while concealing a coordination clash, for example, between structural and MEP systems.
What Happens When a Parameter Changes?
Now the researchers start changing values to see whether the model stages a small digital rebellion.
The evaluator adjusts one parameter through several legitimate settings. Each time, the model must remain valid, the parameter must control the intended feature, and protected properties must remain unchanged.
This reveals problems a finished model can hide. A dimension may display the correct value while controlling the wrong feature. Two dimensions may be linked when they should be independent. A model may rebuild perfectly at its default size, then fail when someone enters another valid value.
CADEngBench also tests functional editing. The AI receives existing CAD and a requested change. It must make that change without altering geometry that should remain untouched.
Consider the difference between “move this hole” and “move this hole without damaging the mounting face, changing the wall thickness or disturbing other dependent features.” That small addition turns a simple geometry instruction into a more realistic engineering task.
Then Comes the Physics
For 164 eligible parts, CADEngBench compares the generated design with a reference using matched linear-static analysis in CalculiX, a free, open-source finite element analysis software package used to simulate how components respond to loads.
Both versions are analyzed using the same material properties, supports, loads and analysis conditions. The benchmark then compares responses such as stress, displacement and stress concentration.
Finite element analysis isn’t a magic truth machine. Its usefulness depends on sound assumptions, suitable boundary conditions and sensible interpretation. What CADEngBench provides is a fair comparison: the generated part and the reference part sit the same virtual exam.
Virtual physics doesn’t give a hoot about a glamorous rendering.
That exam can reveal what a visual comparison misses. Two parts may look almost identical while small differences in thickness, holes, fillets or feature relationships change how they respond to a load. Virtual physics doesn’t give a hoot about a glamorous rendering.
Will the Parts Work Together?
Of course, individual parts rarely spend their lives alone. Buildings, infrastructure and manufactured products depend on components that fit and move together correctly.
Finite element analysis can reveal stress concentrations that a visual comparison of two CAD models may miss.
CADEngBench-A examines 150 body pairs across joint types including rigid, revolute, slider and ball relationships. Each AI system receives both bodies, rendered views and indexed faces and edges. It must identify the joint and where the components connect. The benchmark then tests whether that joint permits the right movement while preventing the wrong movement.
The results show why “close enough” isn’t enough. Across eight AI models, tested on the 120 assembly pairs used for the reported comparisons, systems often found the general connection area but struggled to recover the recorded joint and exact connecting faces or edges.
In plain English, AI could often say, “Those parts probably meet around here.” Explaining exactly how they meet was another matter.
The Verdict: Don’t Fire Your Engineers or Designers
So, how did AI perform? Is it ready to take over from the world’s engineers and designers? Based on this evidence, the answer is clear: No.
The AI systems did substantially better when editing supplied CAD than when generating CAD from scratch. Even then, apparently simple changes could expose their limitations. Adding a single independent body was relatively easy. Modifying a construction history involving joins, cuts or multiple bodies was another story.
The generation results were even more sobering. Across the eight systems, 1,030 of 2,400 generated programs passed the first stage for executability and solid validity. Only 432 also met every applicable engineering and DFM requirement. In other words, fewer than one in five attempts cleared all those checks. Producing geometry that opened and looked plausible was no guarantee that the model had followed its engineering brief.
Producing geometry that opened and looked plausible was no guarantee that the model had followed its engineering brief.
I think we’re a long way from AI being a suitable substitute for the judgement and mental agility of humans who are experts in their craft.
This doesn’t mean AI can’t create CAD. It means the statement “AI created the model” tells us very little about whether that model is ready to use. You still need to ask:
- Is it valid?
- Is it parametrically sound?
- Does it preserve design intent?
- Can it survive a controlled edit?
- Does it behave like the reference under the same simulated load?
- Will it connect correctly in an assembly?
A pretty screenshot answers none of those questions.
More CAD Means More to Manage
There’s a broader lesson for architects, engineers, DOTs and AEC or MEP teams. If AI makes CAD and BIM content faster to produce, organizations may soon have much more of it.
That could help teams test more ideas before committing time and money. But abundance also brings more files to check, versions to compare, standards to enforce and references to maintain.
Someone still has to decide which version is authoritative. Someone must find the subtle standards violation, identify what changed between submissions and repair the reference that vanished after a folder was renamed. AI hasn’t yet learned to sigh despairingly when the references in 4,000 files need updating on a Friday afternoon.
As AI generation gets easier, verification and data management become more important. When a model takes hours to create, people tend to inspect it closely. When AI can produce hundreds of variations quickly, review can’t depend on patience and memory alone.
Where Axiom Fits
Axiom has spent decades helping CAD professionals with the practical work that surrounds design data. Its tools help teams inspect, compare, correct and maintain files at scale. For example:
- SpecChecker™ performs automated quality-assurance checks on MicroStation files.
- DgnCompare™ helps users identify differences between DGN files.
- Global File Changer™ applies repeated changes across multiple design files.
These tools don’t decide whether an AI-generated design is structurally sound. That remains the responsibility of qualified professionals using suitable standards, analysis and judgment. What they can do is reduce the manual work surrounding quality control, file maintenance and large-scale change.
AI may make CAD creation faster than ever. It may also make checking, comparing and maintaining the resulting data more consequential than before.
The real test of a model isn’t whether it looks good. It’s what happens when someone leans in and asks, “Great. Now can you change just this one thing?”
Axiom's President
Oscar Albornoz
Feel free to share this article on your social media:
