3. Crash-specific mesh defect taxonomy
Crashworthiness simulation is unusually unforgiving of mesh defects. Unlike a static stiffness run, an explicit LS-DYNA frontal, side, or rear impact model is driven by contact, large deformation, and strain-rate-dependent material response, so a defect that is harmless in a linear solve can dominate the energy balance of an impact event. This section defines a crash-specific defect taxonomy, traces each class back to its root cause in model preparation, and assigns severity levels that depend on the load case being analyzed.
Why a Crash-Specific Taxonomy Is Needed
General FEA mesh quality metrics — aspect ratio, Jacobian, warpage, skew — were developed for implicit structural analysis. In crash, three additional factors change what "bad" means:
- Contact is the primary load path. A folded or inverted shell element on a bumper or B-pillar face can penetrate a rigid barrier or a neighboring part, producing non-physical interpenetration and spurious contact forces.
- Explicit time step is governed by the smallest element. A single 0.8 mm sliver in a 5 mm nominal mesh can reduce the stable time step by an order of magnitude, inflating runtime without improving accuracy.
- Failure and erosion depend on element size. Strain-based failure criteria and element deletion are calibrated to a characteristic length; a locally refined or coarsened patch shifts the effective failure strain and changes the crash mode.
The taxonomy below therefore classifies defects by their mechanical consequence in an impact event, not by geometric appearance alone.
Defect Classes and Root Causes
| Class | Definition | Typical root cause in model preparation | Crash consequence |
|---|---|---|---|
| D1 — Sliver / degenerate shell | Element with aspect ratio > 20 or minimum edge < 0.5 mm on a 5 mm nominal mesh | Unmerged CAD edges, fillet defeaturing, imprint from a neighboring part | Time-step collapse; local hourglassing; premature deletion |
| D2 — Inverted / negative Jacobian | Shell normal reversed or solid element with negative volume | Mid-surface offset errors, mirrored geometry, hex meshing of a twisted extrusion | Immediate negative-volume termination in explicit solver |
| D3 — Free edge / T-junction | Unconnected nodes along a load-bearing seam | Part-by-part meshing without equivalencing; mismatched node density at a weld flange | Artificial hinge; spot weld carries no load; mode shape wrong |
| D4 — Penetration / initial overlap | Slave nodes initially inside the master surface | Clearance ignored during assembly; thickness offset not applied | Large spurious contact force at t = 0; energy spike; solver instability |
| D5 — Non-physical thickness | Shell thickness inconsistent with the referenced section or gauge | Section property not updated after a gauge change; mixed units | Wrong bending stiffness; incorrect intrusion and energy absorption |
| D6 — Mass scaling artifact | Added mass exceeding 1–2 % of total model mass | Aggressive mass scaling to recover time step lost to D1 | Momentum and kinetic energy distorted; barrier force under-predicted |
| D7 — Material / section mismatch | Element assigned a material or section not matching the physical part | Property inheritance error during part import; duplicated IDs | Wrong failure mode; non-physical rupture or no rupture at all |
Severity Classification by Load Case
Severity is not a property of the defect alone; it is a property of the defect in the context of a load case. The same 3 mm sliver may be tolerable in a rear impact model and critical in a frontal offset model. The following classification uses three levels:
- S1 — Critical. The defect changes the predicted crash mode, energy absorption, or occupant intrusion by more than the correlation tolerance (typically 10 % on barrier force or 5 mm on intrusion). The model must not be used for decision-making.
- S2 — Significant. The defect introduces a local artifact that is visible in the response but does not change the global mode. Acceptable for trend studies, not for final correlation.
- S3 — Minor. The defect affects runtime or local stress noise but not the global response. Acceptable with documentation.
Applying this to the three primary load cases:
| Defect | Frontal (offset deformable barrier) | Side (moving deformable barrier) | Rear (fixed barrier / pendulum) |
|---|---|---|---|
| D1 Sliver | S1 on front rail and shotgun; S2 elsewhere | S1 on B-pillar and rocker; S2 on door skin | S2 on rear rail; S3 on floor pan |
| D2 Inverted normal | S1 anywhere on the load path | S1 anywhere on the load path | S1 anywhere on the load path |
| D3 Free edge | S1 at rail-to-shotgun and rail-to-apron seams | S1 at B-pillar-to-rocker and door-to-B-pillar seams | S2 at rear rail-to-floor seams |
| D4 Penetration | S1 at bumper-to-barrier and rail-to-bumper interfaces | S1 at door-to-B-pillar and B-pillar-to-seat interfaces | S2 at rear bumper-to-barrier interface |
| D5 Thickness error | S1 on front rail, S2 on hood | S1 on B-pillar and rocker | S2 on rear rail |
| D6 Mass scaling | S1 if > 2 %; S2 if 1–2 % | S1 if > 2 %; S2 if 1–2 % | S2 if > 2 %; S3 if 1–2 % |
| D7 Material mismatch | S1 on any energy-absorbing member | S1 on any energy-absorbing member | S1 on any energy-absorbing member |
From Taxonomy to Detection
The taxonomy is only useful if it maps to a detection procedure. In practice, the following checks are run in sequence before any crash solve is submitted:
- Geometry checks: minimum edge length, aspect ratio, Jacobian, warpage, and duplicate nodes. These catch D1 and D2.
- Connectivity checks: free edges, T-junctions, and unmerged nodes along defined weld and adhesive lines. These catch D3.
- Assembly checks: initial penetration depth between all contact pairs, reported as a histogram. This catches D4.
- Property checks: shell thickness versus the referenced section, material ID versus the bill of materials, and total mass versus the target. This catches D5 and D7.
- Solver checks: added mass from mass scaling, initial time step, and energy ratio at t = 0. This catches D6.
The output of these checks is a defect register: a list of element IDs, defect class, severity for the load case under analysis, and the disposition (fix, accept with justification, or flag for correlation review). The register is the input to the ML detection pipeline developed in later chapters, where each defect class becomes a labeled example for training a classifier that can screen new models automatically.
Practical Notes on Severity Assignment
Two rules of thumb help when assigning severity in a real program. First, severity follows the load path. A defect on a member that carries the primary crush force — front rail in frontal, B-pillar and rocker in side, rear rail in rear — is almost always S1, because that member controls the energy absorption and the intrusion. A defect on a non-structural closure panel is usually S2 or S3. Second, severity follows the correlation tolerance. If the program's acceptance criterion is 10 % on barrier force and 5 mm on intrusion, then any defect that can plausibly move those numbers by more than half the tolerance should be treated as S1 until proven otherwise.
It is also worth noting that defects interact. A sliver (D1) often forces the analyst to apply mass scaling (D6), and the combination can be worse than either alone. A free edge (D3) near a penetration (D4) can produce a contact force spike that masks the real structural response. The defect register should therefore record combinations, not just individual defects, and the ML classifier should be trained on co-occurring labels rather than single-class labels.
The taxonomy and severity matrix above form the labeling schema for the detection models developed in the following chapters. Each defect class becomes a target label, each severity level becomes a weighting in the loss function, and each load case becomes a separate evaluation split. With this schema in place, the remaining work is to build the feature extraction and classification pipeline that can populate the register automatically from a raw crash mesh.