Hierarchical Validation
Theory: errors can cancel
Overview
If analysis matches the full-product test, isn't the model correct?
It may be right for the matched quantity, but not necessarily for others. A model has many assumptions and constants; compensating one error by tuning another makes the tuned quantity match exactly, but the error remains and shows up when conditions or outputs change. Hierarchical validation prevents this: split into material, component, joint and system, and check each with independent tests. Then you know which assumption drives which result, and the system test becomes the final check. Let's see how compensation happens in a simple beam.
Cantilever with a rotational spring at the root
The "actual" beam: steel, 1 m long, 50 mm wide, 20 mm thick (EI = 6,667 N·m², 7.85 kg/m), with root rotational spring k_θ = 6.67×10⁴ N·m/rad (EI/(k_θL) = 0.1). Tip load P = 1 kN, distributed load q = 1 kN/m. Natural frequencies from 40 Euler–Bernoulli beam elements.
Validation as a pyramid
Hierarchical validation was systematised from the 1990s at US national laboratories, which had to show simulation credibility once nuclear testing ended. Oberkampf and colleagues drew a multi-tier pyramid from simple-physics tests up to full-system tests, with many tests at the base and few at the top. ASME V&V 10 (2006) codified this as guidance for solid mechanics.
Worked examples
Example: fixed-root model with Young's modulus tuned to the first frequency
| Quantity | Actual | Fixed, tuned E (142.5 GPa) | Fixed, E = 200 GPa |
|---|---|---|---|
| 1st frequency | 13.77 Hz | 13.77 Hz (match) | 16.31 Hz (+18.5%) |
| 2nd frequency | 89.77 Hz | 86.26 Hz (−3.9%) | 102.20 Hz (+13.8%) |
| 3rd frequency | 257.50 Hz | 241.54 Hz (−6.2%) | 286.16 Hz (+11.1%) |
| Tip deflection, 1 kN/m | 26.25 mm | 26.32 mm (+0.3%) | — |
| Tip deflection, 1 kN tip load | 65.00 mm | 70.18 mm (+8.0%) | 50.00 mm (−23.1%) |
| Root strain, 1 kN tip load | 1,500 μ | 2,105 μ (+40.4%) | 1,500 μ |
Both the first frequency and distributed-load deflection match, yet strain is 40% off.
The model compensated for root rotation by making the whole beam softer (lowering E). For the first mode and distributed-load deflection, root rotation and beam bending contribute in similar proportions, so both happen to match — two quantities agreeing feels like validation. Under a tip load, root rotation contributes differently, so deflection is 8% off, and higher frequencies drift too. Strain is worst: it is inversely proportional to E, so the lowered E shows up directly as 40% excess. Strain gauges would disagree, and strain-based fatigue assessment would be wrong. Determining E from material tests and the spring from a fixture test, hierarchically, prevents such compensation.
Running hierarchical validation
- Decompose the product into material, component, joint, subsystem and system tiers.
- Plan tests at each tier that check that tier's assumptions.
- Don't retune lower-tier constants at higher tiers.
- In system tests, compare several quantities not used in calibration (displacement, strain, frequency).
- When results disagree, trace back to which tier's assumption is responsible.
“The frame matched the vibration test”
An equipment frame's model had been tuned to its first measured frequency by slightly thinning welded plates. Later a crack appeared elsewhere in service. The real flexibility was not at the welds but at the bolted floor anchorage, so the tuned model's stress distribution was wrong and underestimated stress at the cracked location. Rebuilding the model with anchorage stiffness from a separate test raised stress there, matching reality.
Common mistakes
Mistakes and fixes
| Mistake | Effect | Fix |
|---|---|---|
| Tuning only to system tests | Errors cancel | Validate per tier |
| Validating on tuned quantities | Apparent agreement | Compare other quantities |
| Non-physical tuning | Fails as conditions change | Locate the real cause |
| Assuming rigid joints | Wrong stiffness and stress | Test joints separately |
| Ignoring test uncertainty | Over-tuning | Compare within uncertainty |
I'd like to learn related topics.
Related pages include ASME V&V 10, validation metrics, code verification, solution verification and ASME V&V 40.
Related Topics
Experience the theory firsthand with the interactive simulator for this field
All Simulators