4. Baseline selection and reference configuration

In a multi-iteration CAE study, every comparison you make is relative. A displacement of 4.2 mm means nothing until you know which configuration produced it and which configuration you are measuring it against. The baseline — sometimes called the reference configuration — is the single, explicitly frozen state of the model against which all other iterations are evaluated. If the baseline is ambiguous, every downstream delta, ranking, and design decision inherits that ambiguity. This section addresses how to choose a defensible baseline, how to configure it so it is reproducible, and how to detect the failure modes that silently corrupt comparisons.

What the Baseline Actually Is

A baseline is not simply "the first run" or "the model we had on Monday." It is a complete, versioned specification of a simulation state that includes at minimum: the geometry revision (CAD file name and revision hash), the mesh (element count, element order, mesh file or seeding parameters), material definitions with their source, boundary and load conditions, solver settings (contact formulation, nonlinear solution controls, convergence tolerances), and the software version with any relevant patches. Two runs that differ in any of these fields are not the same configuration, even if the geometry looks identical on screen.

In practice, teams distinguish three baseline types, and the choice among them should be deliberate:

• As-is baseline — the current production or legacy design, captured exactly as it exists. This is the correct anchor when the study question is "does the new design improve on what we ship today?"
• Nominal baseline — a cleaned, idealized configuration with nominal dimensions and nominal material properties. This is the correct anchor when the study question is "how sensitive is performance to design variables?" because it isolates the variable effect from manufacturing scatter.
• Certification or requirement baseline — a configuration that has already passed a validation test or satisfies a regulatory load case. This is the correct anchor when the study must demonstrate non-regression against an approved state.

Mixing these types within one study is the most common source of contradictory conclusions. If iteration 3 is compared to the as-is baseline and iteration 7 to the nominal baseline, the ranking between them is meaningless. Fix one baseline type per study and state it in the study charter.

Configuring the Reference So It Can Be Reproduced

The baseline must be reproducible by someone who was not present when it was created. A practical configuration record contains the fields below. The values shown are illustrative for a bracket stiffness study and are not measured results.

Store this record alongside the result files, not in a separate spreadsheet that drifts out of sync. A useful discipline is to embed the record as a header block in the solver input deck or as a metadata sidecar file committed to the same repository as the model. When a reviewer asks "what exactly did we compare against?", the answer should be a file path, not a recollection.

Choosing the Anchor: Decision Logic

The selection of a baseline is a decision with consequences, so it should follow explicit criteria rather than convenience. The figure below maps the decision path from study intent to baseline type, and shows the configuration fields that must be frozen once the type is chosen.

Worked Example: Establishing a Baseline for a Bracket Study

Consider a study comparing five rib layouts for a cast bracket. The team's question is "which layout reduces peak von Mises stress under LC-02 without increasing mass by more than 8%?" Because the question is about variable effects, the nominal baseline is appropriate. The team proceeds as follows.

1. Export the nominal geometry from CAD at revision bracket_A_rev03.step and record its hash. Do not use the "latest" CAD pointer, which can change without notice.
2. Generate the mesh with a documented seeding recipe (global size 1.8 mm, 2nd-order tets, local refinement 0.9 mm at the lug fillet) and save the mesh file. Record element count and a mesh-quality summary (minimum Jacobian, percentage of elements below 0.3).
3. Assign Al 6061-T6 properties from the approved material database and record the database entry ID.
4. Apply LC-02 exactly as defined in the load-case library: 2.5 kN vertical at the lug center, fixed translation at the four bolt-hole cylindrical faces. Record the load-case ID rather than re-describing the loads, so the definition stays single-sourced.
5. Run the baseline, archive the result file, and immediately re-run from the archived input deck. Compare peak stress and maximum displacement. If they agree within 1%, the baseline is locked.

Only after this lock do the five rib-layout iterations get built. Each iteration changes exactly one thing — the rib geometry — while inheriting every other field from the baseline record. This is what makes the comparison a controlled experiment rather than a collection of unrelated runs.

Failure Modes and How to Detect Them

Three failure modes account for most baseline problems in practice. Baseline drift occurs when the "baseline" is quietly re-run with a newer solver or a regenerated mesh and the archived results are not updated; the symptom is that iteration deltas shrink or grow inexplicably between study phases. Detect it by comparing the baseline re-run against the archived baseline on every study restart. Baseline contamination occurs when an iteration accidentally inherits a change intended for another iteration — for example, a material swap that was supposed to apply only to layout C. The symptom is an outlier iteration whose results move in a direction no geometry change could explain; detect it by diffing the input decks of each iteration against the baseline record. Baseline mismatch occurs when different iterations are compared to different baseline types, producing a ranking that cannot be defended; detect it by requiring every comparison table to cite the baseline ID in its header.

What a Locked Baseline Delivers

A properly configured baseline converts a pile of simulation results into a comparable dataset. It gives every delta a defined origin, makes the ranking auditable, and allows a reviewer to reproduce any comparison from the archived record alone. The cost is modest — a few hours of disciplined record-keeping at the start of the study — while the cost of an ambiguous baseline is a study whose conclusions cannot be trusted and must be redone. In the chapters that follow, every comparison metric and ranking method assumes this anchor is in place; the quality of the ranking can never exceed the quality of the baseline it is measured against.