3. HBM validation fundamentals

Human Body Models (HBMs) are the computational surrogate for the human occupant in AI-assisted crash simulation. Validating an HBM means demonstrating that, for a defined set of loading conditions, its predicted kinematic and injury responses fall within an acceptable band around physical reference data. That reference data comes from two complementary sources: post-mortem human subjects (PMHS, i.e., cadaver tests) and living volunteers. This chapter establishes the vocabulary and the quantitative machinery you will use throughout the rest of the course: which HBM families exist and what each is built for, how validation protocols are structured, and how correlation is actually scored.

HBM families and what each is designed to represent

HBMs are not interchangeable. They differ in geometry source, mesh formulation, material laws, and the injury criteria they can output. The practical taxonomy is:

• Rigid-body / multibody models (e.g., MADYMO-style facet or ellipsoid models). Fast, stable, and well suited to restraint-system design sweeps and occupant packaging. They carry contact stiffness and joint properties but cannot resolve local tissue strain, so they are weak for tissue-level injury prediction.
• Finite-element HBMs (e.g., the GHBMC and THUMS families). Solid/shell meshes with organ-level geometry, nonlinear materials, and strain-based injury criteria. These are the models used when the question is "where and how severely does tissue deform," and they are the primary target of AI-surrogate acceleration.
• Statistical / morphable models. A base mesh plus a statistical shape model (or a parametric scaling scheme) that produces a population of occupants — stature, BMI, age, sex. Validation must then be demonstrated not only for the nominal 50th-percentile model but across the morphable space.

Two further distinctions matter for validation. First, posture and boundary conditions: an HBM validated in a driver posture with a lap-shoulder belt is not automatically validated for a reclined or out-of-position occupant. Second, anthropometry: a model validated at 50th-percentile male does not transfer to a small female or a child without separate validation evidence. Every validation claim in this course is therefore scoped as model + anthropometry + posture + loading condition.

Reference data: cadavers versus volunteers

PMHS tests provide the only ethical route to injurious and near-injurious loading, and they supply the tissue-level and skeletal-fracture data that injury criteria are calibrated against. Their limitations are real and must be carried into the validation logic: embalming and tissue-condition changes, absent active muscle tone, non-physiological boundary conditions in some rigs, small sample sizes, and substantial inter-specimen variability in geometry and bone quality. Volunteer tests, by contrast, provide physiological muscle tone, repeatable and well-instrumented kinematics, and clean boundary conditions — but only at sub-injury severity. The two datasets therefore play different roles: volunteers constrain the kinematic response (excursions, timing, belt interaction), while cadavers constrain the injury response (force–deflection, fracture timing, strain thresholds).

How a validation protocol is structured

A defensible protocol is built from discrete load cases, each with a documented input, a documented reference response, and a documented acceptance criterion. The typical construction sequence is:

1. Select load cases that span the model's intended use: frontal sled at multiple severities, lateral sled or pendulum, belt-only and airbag interactions, and — where relevant — out-of-position or reclined configurations.
2. Digitize the reference: extract the measured time histories (head and chest acceleration, chest deflection, belt force, pelvis acceleration, knee excursion) and the reported test conditions (pulse shape, peak, duration, restraint geometry).
3. Reproduce the boundary conditions in the model: the sled pulse applied to the buck, belt routing and pretensioner/foad characteristics, contact definitions, and the initial posture.
4. Run and compare using objective correlation metrics, not visual overlay alone.
5. Record the outcome per load case, including failures, so the validation matrix is auditable.

For a reproducible example, consider a frontal sled case at 40 km/h with a 50th-percentile male FE HBM, three-point belt with pretensioner and load limiter. The input pulse is a trapezoidal acceleration with a 25 g plateau over 40 ms and a 90 ms total duration. Reference channels are head resultant acceleration, upper-chest resultant acceleration, and chest deflection. The model is run for 150 ms with a 1 ms output interval; the belt is modeled with a force-limiting characteristic of 4 kN. The observable outputs are the peak values and their timings, plus the full time histories for correlation scoring. If the model's chest deflection peaks at 38 mm at 62 ms while the reference peaks at 44 mm at 58 mm, that is a timing-and-magnitude discrepancy worth investigating before any correlation score is computed — check belt routing, pretensioner timing, and chest material stiffness in that order.

Correlation metrics

Objective comparison between a simulated curve and a reference curve is normally done with a composite metric that combines magnitude, phase, and shape agreement. The most widely used is the CORrelation and Analysis (CORA) method, which returns a score in [0, 1] from three sub-scores: corridor (magnitude), cross-correlation (phase), and a shape/trend term. A score of 1.0 is perfect agreement; in practice, published HBM validation studies treat roughly 0.7–0.8 as "good" and below about 0.6 as "marginal or poor," though the threshold must be stated explicitly in your protocol rather than assumed.

Two other metrics are common and worth knowing. The ISO/TS 18571 objective rating method combines a corridor score, a phase score, and a magnitude score into a single rating, and is often used where a standardized, regulator-recognizable metric is required. The Analysis of Correlation and Kinematics (AnCoR) approach and simpler normalized-error measures (e.g., peak error as a percentage of reference peak, and peak-time error in ms) are useful as transparent sanity checks alongside CORA. A practical rule: report at least one composite metric plus the raw peak magnitude and peak-time errors, because composite scores can hide a large timing error behind a good shape score.

Reading the validation matrix

The output of a validation campaign is a matrix: load cases on one axis, response channels on the other, each cell holding a metric value and a pass/fail flag against the pre-declared threshold. This matrix is the artifact that AI-surrogate work depends on — a surrogate trained to reproduce an HBM is only as trustworthy as the HBM's validated domain, and the matrix defines that domain precisely. When you later see a surrogate's error reported, the correct question is always: error relative to what, over which cells of this matrix?

The matrix above is illustrative, not measured: it shows the reporting convention you should adopt. Each cell carries a composite score and a status against a threshold declared before the runs. Marginal and failing cells are not hidden — they are the cells that tell you where the model's validated domain ends, and they are exactly the cells an AI surrogate must be tested against most carefully.

Practical validation checklist

With this foundation — model families, the two reference streams, protocol construction, and per-channel correlation scoring — you can read any HBM validation study critically and, more importantly, build the validation matrix that will govern how you train and trust the AI surrogates in the chapters that follow.