A Proposal, Not a Part: The Pre-Solver Audit for Generative CAD
Generative CAD has reached the point where a prompt can return a bracket, a housing, or a heat-sink volume that looks ready for review. That is precisely the danger. The output is visually coherent, dimensionally plausible, and often topologically interesting, which makes it easy for a team to treat the model as a draft part rather than a proposal. In a CAE workflow, that shortcut is expensive. A solver does not care whether geometry looks reasonable; it cares whether connectivity, dimensions, tolerances, materials, and feature history are explicit and consistent. The audit described here is a gate between generative output and simulation input. It is deliberately human-owned, traceable, and boring in the right places.
The hazard is plausibility, not obvious failure
Most bad simulation inputs announce themselves. A missing body, a reversed normal, or a unit mismatch will usually surface before the run completes. Generative CAD is different. Its failures tend to be plausible: a fillet that blends into a neighboring rib, a wall thickness that averages to the right value but varies beyond intent, a boss that intersects a shell without a clean Boolean history. The geometry may pass a visual check and still misrepresent the design intent. If the model reaches the solver in that state, the analysis can converge on a shape that no one deliberately designed and no one can reproduce from a drawing.
The audit gate should therefore begin with a status change. Generative output is not a released model. It is a proposal with unknown provenance until an engineer accepts it. That acceptance should be recorded in the same way a supplier part approval is recorded: who reviewed it, what assumptions were verified, what was changed, and what remains uncertain.
Define the gate before the prompts arrive
A useful gate has a short scope and a clear exit condition. The model leaves the gate only when it can be handed to a CAE analyst without verbal translation. That means the CAD file, the audit record, and the assumptions travel together. A generative model that cannot meet that condition is not rejected forever; it is returned for clarification or converted into a conventional model with a named owner.
The gate should be applied before meshing, not after. Once an analyst has spent a day repairing geometry, the organization is psychologically invested in the result. The audit is cheaper when it happens while the model is still a proposal.
Topology and connectivity
Start with body count and connectivity. Generative tools often produce unified surfaces, zero-thickness regions, or bodies that touch but do not share topology. Each condition changes how a solver treats load paths. The reviewer should confirm that the body count matches the intended physical assembly, that contacts are intentional, and that no unintended gaps or overlaps exist. Where the generative output has merged features that should remain separate, the audit should note whether the merge is acceptable for the analysis or must be reversed.
A simple connectivity map is enough: body name, intended neighbor, connection type, and reviewer initials. Attach it to the model. The map is not bureaucracy; it is the difference between a solver contact definition and a guess.
Dimensions, tolerances, and design intent
Generative CAD is good at satisfying a volume or a stiffness target. It is not automatically good at respecting manufacturing tolerances. A wall may be dimensionally correct at the nominal value while sitting on a tolerance stack that makes it unproducible. The audit should compare critical dimensions against the design requirements and flag any dimension that exists only because the generator placed it there.
Tolerances deserve separate attention. If the model omits tolerance callouts, the CAE analyst may assume nominal geometry and miss the sensitivity that matters. The reviewer should identify which dimensions are functional, which are manufacturing-driven, and which are still open. For US teams, that often means checking against ASME Y14.5-style callouts where the drawing exists, and explicitly marking any dimension that has no released tolerance basis.
Material assignments and mass properties
A generative model can carry a material name without carrying a defensible material specification. The audit should verify that every body has an assigned material, that the assignment matches the intended manufacturing process, and that density and elastic properties are consistent with the source data. Mass properties are a quick sanity check: compare the model mass against the target and against any predecessor part. A large discrepancy is not automatically an error, but it must be explained before the solver inherits it.
The reviewer should also record whether the material assignment is provisional. A provisional material is acceptable for early trade studies, but it should be visible in the handoff packet so the analyst can treat the results accordingly.
Feature history and provenance
The feature tree is where generative models often become opaque. Some tools produce editable histories; others return a final shape with limited construction logic. The audit should determine whether the feature history is sufficient to answer a basic question: how was this shape made? If the answer is only a prompt and a seed, the model is not ready for a design review that depends on reproducibility.
Provenance does not require a perfect history. It requires an honest record. Note the tool, version, prompt, constraints, seed if available, and any post-generation edits. If the model was repaired manually, say so and name the person. A clean provenance record lets a CAE analyst distinguish between a design assumption and an artifact of generation.
The handoff packet
The gate exits with a small packet: the native CAD file, a neutral format if required, the connectivity map, the dimension and tolerance review, the material assignment table, the mass-property check, and the provenance note. The packet should also list open items. An open item is not a failure; it is an assumption that the analyst must own.
The packet is the real deliverable of the audit. It converts a generative proposal into a simulation input with a known boundary. Without it, the solver becomes the place where unresolved design questions go to hide.
Human ownership remains the control
Generative CAD can compress the drafting phase, but it cannot absorb the engineering accountability that sits between geometry and simulation. The pre-solver audit is a deliberate slowdown at the point where speed is most tempting. It keeps the organization from confusing a plausible shape with a released design, and it gives the CAE analyst a model that can be defended in a review rather than merely meshed. In 2026, that distinction is still the difference between a useful generative tool and an expensive source of confident nonsense.
