3. Finite element occupant models

Physics-informed machine learning for occupant safety does not replace the explicit finite element (FE) solver; it sits on top of it. Surrogate models are trained on solver outputs, hybrid workflows call the solver inside the loop, and any credible claim about a learned model's fidelity is ultimately anchored to a validated LS-DYNA or PAM-CRASH occupant model. This chapter therefore treats the FE occupant model as the reference "ground truth generator" and examines what a practitioner must understand about its architecture, material definitions, contact treatment, and cost profile in order to use it correctly as a data source and as a physical constraint.

Model architecture: what an occupant model actually contains

A full-vehicle crash model and a stand-alone occupant model are different objects with different purposes. The occupant model — typically a Hybrid III 50th percentile male, THOR, or a human body model (HBM) such as THUMS or GHBMC — is a deformable multibody-to-FE hybrid: rigid bodies for skeletal segments, deformable solids for flesh and internal organs, discrete beams and springs for joints and ligaments, and a belt/airbag subsystem that interacts with the dummy through contacts. In LS-DYNA the model is assembled from keyword blocks: *PART and *SECTION define the structural discretization, *MAT_* cards define constitutive behavior, *CONTACT_* cards define interaction, *CONSTRAINED_* cards tie parts together, and *DATABASE_* cards control what is written out. PAM-CRASH uses an equivalent structure through its material (MAT), contact (CONT), and link/kinematic condition cards.

The practical consequence for ML work is that the model is not a monolithic black box. A surrogate that predicts chest deflection is really learning a mapping from a belt-routing geometry, an airbag pressure curve, and a pulse — the inputs — to a response computed by a specific set of parts, materials, and contacts. If any of those are changed between training and deployment, the surrogate's validity domain changes with them.

Material models that dominate occupant response

Occupant models are dominated by a small number of material families, and each has parameters that materially affect the output signal:

  • Rigid materials (*MAT_RIGID, *MAT_020) for skeletal segments. Cheap, but the mass and inertia assigned here drive the whole-body kinematics, so they must match the dummy certification data.
  • Foam and low-density materials (*MAT_LOW_DENSITY_FOAM, *MAT_SOIL_AND_FOAM) for padding, armrests, and the dummy's own flesh. The stress–strain curve is tabulated and rate-sensitive; the loading and unloading branches differ, which is exactly the hysteresis a surrogate must reproduce.
  • Fabric and membrane materials (*MAT_FABRIC) for airbags, with permeability and rate dependence controlling inflation and venting.
  • Viscoelastic and hyperelastic models for soft tissue in HBMs, where the relaxation time constants govern how the model responds to loading rate.
  • Piecewise linear plasticity (*MAT_PIECEWISE_LINEAR_PLASTICITY) for belt webbing and structural brackets, with strain-rate scaling via *DEFINE_CURVE tables.

A common and expensive mistake is to treat these as interchangeable. A surrogate trained on a model whose foam curve was later updated will silently extrapolate. The defensible practice is to version the material deck alongside the training data and to record, for every simulation, the hash of the keyword file that produced it.

Contacts: where most of the physics and most of the cost live

Occupant models are contact-dominated. Belt-to-chest, airbag-to-face, knee-to-knee-bolster, and self-contact of the dummy's own limbs all require explicit contact definitions. In LS-DYNA the workhorses are *CONTACT_AUTOMATIC_SURFACE_TO_SURFACE for belt and airbag interaction and *CONTACT_AUTOMATIC_SINGLE_SURFACE for self-contact; PAM-CRASH uses penalty-based contact with comparable card structure. Three parameters dominate both accuracy and runtime:

  • Penalty stiffness (SLSFAC / PENOPT): too soft and the belt penetrates the chest, producing non-physical deflection; too stiff and the explicit time step collapses.
  • Friction (FS, FD): belt slip against clothing changes chest loading substantially; the coefficient is a calibration parameter, not a physical constant.
  • Contact thickness and segment size: the belt is a shell with a defined thickness; if the shell is coarser than the local curvature of the chest, contact is missed entirely.
Contact penetration is the single most common cause of a physically implausible occupant result. Before trusting any simulation used for training data, check the contact energy and the penetration depth in the glstat and sleout outputs. A negative sliding energy or a penetration exceeding the shell thickness invalidates the run.

The computational cost profile

High-fidelity occupant simulation is expensive for a structural reason: the explicit time step is set by the smallest element, and occupant models contain very small elements in the flesh, face, and belt. A typical Hybrid III model runs at a stable time step of roughly 0.5–1.0 µs; a 150 ms event therefore requires on the order of 150,000–300,000 cycles. With 200k–1M elements, a single run on 32–64 cores takes hours, and a design-of-experiments sweep of a few hundred runs becomes a multi-day cluster campaign. This cost is precisely the motivation for surrogates — but it also means the training set is small relative to what a purely data-driven model would want, which is why physics-informed constraints matter.

Occupant FE Model: Architecture, Physics Coupling, and Cost Chain Model Architecture Rigid skeletal segments *MAT_RIGID / *MAT_020 Deformable flesh / organs foam, viscoelastic, hyperelastic Belt & airbag subsystem *MAT_FABRIC, *MAT_ELASTIC Joints & constraints *CONSTRAINED_*, discrete beams Contact definitions belt–chest, airbag–face, self Output: acceleration, deflection, force Physics Coupling Crash pulse (input) *BOUNDARY_PRESCRIBED_MOTION Belt force–displacement retractor + pretensioner curve Airbag inflation *AIRBAG_* pressure / mass flow Contact penalty forces SLSFAC, friction FS/FD Injury metrics: HIC, chest defl., Nij Computational Cost Chain Smallest element size flesh, face, belt ≈ 1–3 mm Stable time step Δt ≈ 0.5–1.0 µs Cycle count 150 ms / Δt ≈ 1.5–3 × 10⁵ Wall-clock per run hours on 32–64 cores DoE campaign hundreds of runs → days Cost is driven by the smallest element; accuracy is driven by contact and material calibration.

Working with the model as a data source

When the FE model is used to generate training data, three practical controls make the difference between a usable dataset and an expensive one. First, fix the output sampling: request *DATABASE_ASCII at a defined interval (e.g., every 0.5 ms) and record the channel names explicitly, because a surrogate trained on a mislabeled channel is worse than no surrogate. Second, run a convergence check on the quantities of interest — refine the chest flesh mesh once and confirm that peak chest deflection changes by less than a few percent; if it does not, the coarse model is not a valid reference. Third, log the contact energy and the added-mass ratio from glstat for every run and reject any run where the added mass exceeds roughly 1% of the total model mass, since that indicates the time step was scaled and the dynamics are no longer physical.

A reproducible pattern for surrogate training: generate a Latin hypercube over the design variables (belt anchor position, pretensioner force, airbag vent size, pulse scaling), run each point through the validated occupant model, and store the keyword-file hash, the sampled inputs, and the extracted injury metrics together. The hash is what lets you prove later that the surrogate and the FE model are describing the same physics.

The takeaway for the rest of this course is that the FE occupant model is both the source of truth and the source of cost. Its architecture determines which inputs are meaningful, its material and contact definitions determine which physics a surrogate must respect, and its time-step-driven cost determines how much data is realistically available. Physics-informed learning is, in large part, the discipline of encoding the second and third of these into the model so that the first remains trustworthy.

← Previous 📖 Contents Next →