4. CAE model readiness

A design exploration campaign multiplies every modeling decision by the number of design points it evaluates. If the baseline CAE model is not trustworthy, the campaign does not produce insight — it produces a large, expensive set of consistently wrong answers. Model readiness is therefore not a preliminary formality; it is the gate that determines whether the DOE results can be interpreted at all. This section defines what "ready" means in practice: a mesh whose quality metrics are controlled and documented, solver settings that are stable and reproducible, and a correlation baseline that anchors the model to physical reality before any parameter is varied.

Mesh Quality: Metrics That Actually Matter

Mesh quality is not a single number. Different solvers and element types are sensitive to different defects, and a mesh that converges for a static linear analysis may fail immediately in an explicit or CFD context. The practical approach is to define a small set of pass/fail thresholds tied to the solver family, then verify them programmatically rather than by visual inspection.

For structural implicit analyses with second-order tetrahedral or hexahedral elements, the following metrics are commonly monitored:

These thresholds are starting points, not universal law. A crashworthiness model using LS-DYNA will prioritize minimum element size and time-step stability, while an underhood CFD model will prioritize skewness and orthogonality near walls. The key discipline is to record the thresholds in the campaign setup file so that every design point is meshed to the same standard. If the mesh is regenerated per design point — common in shape or topology optimization — the quality check must run automatically and reject any variant that violates the thresholds.

Solver Settings: Stability and Reproducibility

Solver settings must be frozen before the campaign launches. Changing a convergence tolerance or a time-step control mid-campaign invalidates comparison across design points. The readiness checklist should capture the following categories:

• Convergence criteria: force and displacement residuals, energy norms, and the maximum number of iterations. For implicit nonlinear analyses, a residual target of 1e-3 with energy-based convergence is a common baseline; for CFD, scaled residuals of 1e-4 with monitored lift/drag stabilization.
• Time integration: implicit vs. explicit, time-step size or automatic time-stepping bounds, and mass scaling limits if applicable.
• Contact definition: penalty stiffness, friction coefficients, and contact thickness. Contact instability is one of the most common causes of non-convergence in design sweeps.
• Material models: fully defined curves, strain-rate dependence, and failure criteria. A missing hardening table will cause the solver to extrapolate silently.
• Output requests: consistent field and history output across all design points so post-processing scripts do not break.

A reproducible setup is one where the solver input deck is generated from a parameterized template, not edited by hand. The template should expose only the design variables — thickness, material grade, rib position — and lock everything else. This is the single most effective safeguard against accidental variation.

Correlation Baselines: Anchoring the Model

Before a campaign explores new designs, the baseline model must reproduce a known physical result within an acceptable tolerance. This is the correlation baseline. It answers the question: "If I run the current model at the current design, does it match test or historical data?"

Typical correlation targets depend on the physics:

• Structural stiffness: first mode frequency within 5%, static deflection within 10%.
• Crash: peak force within 15%, intrusion within 10%, and pulse shape qualitatively matched.
• Thermal/CFD: outlet temperature within 5 K, pressure drop within 10%, and mass flow balance within 2%.

If the baseline does not meet these targets, the campaign should not launch. Instead, the discrepancy must be resolved — through mesh refinement, material calibration, or boundary condition correction — and the baseline re-validated. A common failure mode is to "tune" the model until it matches one test point, then launch a campaign that explores far from that point. The correlation baseline is only meaningful if the model is physically correct, not merely curve-fitted.

Readiness Gate: A Practical Checklist

The following checklist can be used as a formal gate before launching a design exploration campaign. Each item should be signed off with a recorded value or artifact.

Once these gates are passed, the model is ready. The campaign can then vary design parameters with confidence that differences between design points reflect design changes, not modeling noise.

The diagram above shows the three readiness tracks as parallel gates. Each track has its own pass criteria, and all three must be satisfied before the campaign launches. The green items at the bottom of each column represent the final verification artifacts — a convergence study, a locked template, and an overlay plot — that provide evidence the gate has been passed. This structure prevents the common mistake of treating model readiness as a single checkbox rather than a multi-disciplinary verification.

Common Failure Modes and How to Avoid Them

Three failure modes account for most campaign invalidations. First, mesh quality drift: the baseline mesh is clean, but the parameterized mesh generator produces distorted elements at extreme design points. Mitigation: run the quality check on the first and last design points of each parameter range before launching the full DOE. Second, solver setting leakage: a user edits the input deck to fix one design point and forgets to revert. Mitigation: generate all decks from a single template and hash-check the non-variable sections. Third, correlation overfitting: the baseline matches one test but fails at other conditions. Mitigation: validate against at least two independent load cases and document the error margins.

When these three areas are controlled, the design exploration campaign rests on a defensible foundation. The results can then be interpreted as design-driven differences rather than modeling artifacts, which is the entire purpose of running a campaign in the first place.