3. Dummy model selection and validation

In frontal occupant restraint simulation, the choice of occupant model determines what physical quantities you can predict, how much confidence you can place in those predictions, and how much effort validation will cost. LS-DYNA offers three broad families: the Hybrid III 50th percentile crash dummy, the THOR (Test Device for Human Occupant Restraint) family, and Human Body Models (HBMs). Each occupies a different point on the spectrum between regulatory compliance and biofidelity, and each demands a different validation strategy.

Model Families and Their Positioning

The Hybrid III 50th percentile male is the regulatory workhorse. Its geometry and mass distribution are defined by FMVSS 208 and ECE R94, and its instrumentation set — head accelerometers, upper/lower neck load cells, chest accelerometers and deflection potentiometer, lumbar spine load cell, and femur/tibia load cells — maps directly onto injury criteria such as HIC15, Nij, chest deflection, and femur force. Because the model is stiff and its joints are simple revolute or spherical constraints, it is numerically robust and fast. Its weakness is biofidelity: the stiff thoracic spine, simplified ribcage, and limited shoulder articulation make it a poor predictor of real-world chest and clavicle loading, and it cannot represent submarining or out-of-position kinematics with high fidelity.

THOR was developed to close that gap. THOR-50M retains the 50th percentile mass and stature but adds a flexible thoracic spine, four-point chest deflection measurement, a more human-like pelvis and iliac wing geometry, and improved shoulder and clavicle load paths. It is instrumented to capture chest deflection at multiple locations, abdominal pressure, and pelvis rotation — quantities that matter for predicting rib fracture and submarining. The trade-off is complexity: more contact interfaces, more joint degrees of freedom, and a greater sensitivity to initial positioning. THOR is increasingly accepted in regulatory and consumer-test protocols, but its validation matrix is still narrower than Hybrid III's.

Human Body Models (e.g., THUMS, GHBMC) represent the occupant as deformable finite-element anatomy: bones with cortical/trabecular layers, organs, muscles, ligaments, and skin. They can predict local strain, stress, and fracture risk rather than lumped injury criteria. The cost is computational — a single frontal load case may run for hours to days — and the validation burden is much higher because each tissue material model must be correlated against cadaveric and volunteer data. HBMs are typically reserved for research, injury mechanism studies, and cases where Hybrid III and THOR cannot resolve the question.

Positioning: The Dominant Source of Error

No model is more accurate than its initial position. For frontal loading, the critical positioning parameters are:

• H-point relative to the seat reference point (SRP), typically within ±10 mm in X and Z.
• Pelvis angle — the tilt of the pelvis about the lateral axis, which controls lap-belt engagement and submarining risk.
• Torso angle — recline relative to vertical, affecting chest-to-belt contact and airbag interaction.
• Head position — both the head CG location and the neck angle, which set the initial distance to the airbag module.
• Limb posture — knee and ankle angles, and arm position for out-of-position cases.

In LS-DYNA, positioning is usually performed with the *INITIAL_FOOTPRINT and *INITIAL_VELOCITY cards, or by driving the model with a pre-simulation using *BOUNDARY_PRESCRIBED_MOTION to settle it into the seat. A common practical approach is a two-stage settle: first apply gravity and a small seat compression, then lock the model and apply the belt pre-tension. The settle must be checked for residual kinetic energy — if the model is still moving when the main event begins, the initial conditions are contaminated.

Joint Stiffness and Its Effect on Response

Joint stiffness in the dummy model is not a single number; it is a set of torque-angle curves for the neck, lumbar spine, shoulder, hip, knee, and ankle. These curves are defined in LS-DYNA through *MAT_ cards with load-curve references or through discrete elements such as *ELEMENT_DISCRETE with nonlinear springs. The Hybrid III neck, for example, uses a piecewise-linear moment-rotation curve that is much stiffer than a human neck in flexion. THOR uses a more compliant, multi-segment neck with separate flexion and extension curves.

Joint stiffness directly affects the timing and magnitude of chest and head loading. A stiffer neck transmits more moment to the head, raising HIC; a stiffer lumbar spine reduces chest deflection but increases pelvis rotation. When correlating a model, joint stiffness is often the first parameter to tune because it has a first-order effect on kinematics and is relatively well-characterized by component tests.

Correlation Metrics for Frontal Loading

Validation is not a single pass/fail test; it is a quantitative comparison between simulation and physical test across a matrix of signals. The standard approach uses the CORrelation and Analysis (CORA) method, which combines corridor, cross-correlation, and phase-shift metrics into a single score between 0 and 1. A score above 0.8 is generally considered good; above 0.9 is excellent. The signals typically evaluated in frontal loading are:

In practice, you compute CORA scores per signal, then a weighted average. If the weighted score falls below the acceptance threshold, you inspect the worst-performing signals first. A phase shift in chest deflection usually indicates a belt or airbag timing error; a magnitude error in femur force usually indicates a knee bolster stiffness or geometry mismatch.

Practical Validation Workflow

A reproducible validation sequence for a frontal sled or full-vehicle model looks like this:

1. Component-level check. Run the dummy model in isolation against published component test data (neck pendulum, chest impact, knee impact). Verify that the model reproduces the certification corridors before using it in a full system.
2. Positioning verification. Settle the model into the seat, then export the H-point, torso angle, and head position. Compare against the physical test setup. A mismatch greater than 10 mm in H-point or 2° in torso angle will dominate your correlation error.
3. Signal extraction. Use *DATABASE_ASCII or *DATABASE_BINARY_D3PLOT to output the same channels as the physical test. Filter both simulation and test data with the same SAE J211 filter class (typically CFC 600 for accelerations, CFC 180 for forces).
4. CORA evaluation. Compute scores per signal, then weighted average. Document the weighting scheme.
5. Iteration. Tune joint stiffness, belt routing, and airbag timing in that order. Re-run and re-score. Stop when the weighted score exceeds the project threshold or when further tuning degrades other signals.

The diagram above summarizes the selection and validation logic: choose the model family based on the injury mechanism you need to resolve, position it with the same care as the physical test, tune joint stiffness before touching belt or airbag parameters, and score against the full signal set rather than a single peak value. A model that passes component checks and positioning verification will usually correlate well; a model that fails those early gates will not be rescued by tuning the restraint system.