Failure-Model Calibration & Test Data

Calibration is an evidence chain

A failure model is credible only to the extent that its parameters can be traced to appropriate material tests and then validated independently. The objective is not to force a simulation to reproduce one observed outcome. It is to identify a parameter set that represents the material across the stress states, strain rates and temperatures relevant to the structural assessment, with known uncertainty and a documented range of applicability.

Start with material condition

Record alloy or material specification, product form, heat treatment, thickness, orientation, manufacturing route, batch or heat where available, and baseline mechanical properties. Dynamic failure can be sensitive to these details. Combining data from nominally similar but differently processed material can introduce scatter that a calibration algorithm incorrectly interprets as constitutive behaviour.

Build a test matrix around the required states

• Smooth tension provides a basic tensile reference but does not span the full failure surface.
• Notched tension changes triaxiality and helps constrain pressure sensitivity of ductile failure.
• Shear-dominated specimens provide low-triaxiality information and help constrain Lode dependence.
• Compression data establishes flow behaviour in a different stress state and helps separate plasticity from fracture parameters.
• Dynamic tension, compression or shear tests establish rate dependence over the relevant range.
• Temperature-controlled tests are needed when initial or deformation-induced temperature materially changes failure.

Dynamic data quality matters

High-rate tests require careful wave analysis, specimen equilibrium checks and data reduction. A quoted 'strain rate' is usually an average or representative quantity, while the specimen experiences a time-varying field. Use the actual processed histories and document filtering or corrections. Poor dynamic data can make a complex failure law look well calibrated while the parameters are compensating for experimental artefacts.

Use specimen simulations in the calibration loop

The stress state in a notched or shear specimen is not perfectly uniform and evolves as deformation localises. Simulating the specimen with the same constitutive framework used in the structural model allows parameters to be identified against force-displacement, strain-field and fracture-location observables rather than relying only on nominal analytical stress-state estimates.

Separate flow and failure calibration

First establish elastic and plastic flow behaviour over strain, rate and temperature. Then calibrate failure initiation and damage evolution. Changing both simultaneously can produce non-unique solutions: a softer plastic law combined with a later failure threshold can reproduce the same global force history as a stronger law with earlier failure. Keeping the stages separate improves parameter identifiability.

• High-Strain-Rate Material Behaviour
• Material Failure Models for Impact FEA

Parameter identification

Parameter identification may use direct fitting, optimisation or Bayesian methods. Whatever method is chosen, define the objective functions before tuning and include multiple observables where possible. Force-displacement, local strain, failure location and post-test geometry provide more constraint than a single scalar endpoint. Weighting should reflect measurement quality and the intended use of the model.

Identifiability and over-parameterisation

A model can contain more adjustable parameters than the test programme can uniquely constrain. In that case several parameter sets may fit the calibration data but give different structural predictions. Examine parameter correlation, confidence intervals or posterior distributions rather than reporting only a best-fit vector. If parameters are poorly identifiable, simplify the model or acquire more informative tests.

Mesh size is part of the calibration

For local damage and element-deletion formulations, the calibrated response can depend on element size. Record the characteristic mesh used in specimen calibration and assess how transfer to the structural mesh affects dissipated energy and failure localisation. If the solver provides regularisation or element-size-dependent failure data, use it consistently and verify its behaviour rather than assuming it removes all mesh sensitivity.

Independent validation

Reserve at least one relevant experiment that is not used to identify the parameters. The validation case should exercise the combined constitutive and failure model at conditions representative of the intended structural use. Compare multiple quantities: timing, force or velocity history, deformation mode, failure location and residual geometry. Agreement in only one endpoint is weak evidence if the underlying response sequence is wrong.

Uncertainty and sensitivity

Material scatter, test measurement uncertainty and parameter uncertainty should be carried into the structural assessment when they affect the decision. Sensitivity analysis can identify which parameters deserve additional testing. If the conclusion is robust across credible parameter ranges, the model may be adequate even with imperfect calibration; if the conclusion changes, the uncertainty must be resolved or conservatively bounded.

Traceable calibration record

• Material provenance recorded — Product form, heat treatment, orientation and batch where available.
• Raw and processed test data retained — Including filtering and data-reduction assumptions.
• Calibration and validation sets separated — No hidden reuse of validation data during tuning.
• Solver implementation documented — Model form, units, reference rates, temperature conventions and erosion treatment.
• Mesh sensitivity assessed — Especially for softening and deletion-based damage.
• Parameter uncertainty recorded — Not just the nominal best-fit values.
• Application range stated — Strain, rate, temperature and stress-state range supported by evidence.

Avoiding over-calibration

A highly parameterised model can reproduce a large calibration data set yet still extrapolate poorly. Prefer the simplest formulation that captures the response features that materially affect the structural decision. Add model complexity only when the existing form shows a systematic, physically meaningful deficiency that the available test programme can actually constrain.

Calibration and validation are different activities. Once a result has been used to tune the parameters, it is no longer independent validation evidence.