MECH60561 Chap.8 Verification, Failure Modes and Uncertainty
Verification, Failure Modes and Uncertainty
Verification, Failure Modes and Uncertainty builds confidence from checks that can fail independently. Equilibrium, benchmark comparison and mesh refinement support verification, while comparison with physical behaviour supports validation. The chapter diagnoses locking, hourglass motion, rigid-body modes and singular local stress from their symptoms.
It then separates physical input uncertainty from numerical error and carries a stated material range through to a decision quantity without mislabelling sensitivity as probability. An evidence matrix ties each conclusion to the check that supports it and exposes where several apparently different checks rely on the same assumption.
Decision thresholds turn an input range into engineering action rather than decorative uncertainty.
What this chapter covers
- 01
Verification and validation
- 02
Equilibrium evidence
- 03
Benchmark comparison
- 04
Locking diagnosis
- 05
Hourglass behaviour
- 06
Rigid-body motion
- 07
Input sensitivity and decision thresholds
Worked example · free
Bound displacement from modulus uncertainty
- 2State the baseline displacement and modulus and the held-fixed assumptions.
- 4Scale displacement inversely at the lower and upper modulus bounds.
- 2Report a sensitivity range and avoid calling it a probability interval.
Key terms
- Validation
- Evidence that a model represents physical behaviour adequately for a stated use.
- Sensitivity
- The change in a model output caused by a defined change in an input or assumption.
- Singularity
- An idealised location where a field quantity may not approach a finite value under refinement.
Verification, Failure Modes and Uncertainty FAQ
Does solver convergence prove accuracy?
No. It shows that the numerical iteration met a criterion for the stated discrete equations. Accuracy also needs credible model assumptions, mesh evidence, equilibrium, benchmark comparison and validation appropriate to the requested engineering output.
How do I distinguish locking and hourglass motion?
Locking produces an artificially stiff response because the interpolation introduces a spurious constraint. Hourglass behaviour produces nonphysical low-energy deformation, often with under-integration. Their symptoms and remedies differ, so use benchmark displacement, energy and deformation pattern to diagnose them.
How should uncertainty be reported?
Name the uncertain input, its evidence-based range, what remains fixed and how the decision quantity changes. A parameter sweep is a sensitivity study, not a probability distribution. Compare the output range with the engineering threshold that controls the decision.
How strong should a verification claim be?
Match the wording to the evidence. Reaction balance supports global equilibrium, a benchmark supports behaviour under overlapping assumptions, and mesh stability applies to a named output at a named location. Validation needs relevant physical comparison. State what each check proves and what it cannot prove, then identify the next observation that would reduce the uncertainty controlling the decision.
Exam move
Organise verification as a ladder. Begin with units, total mass, applied resultant, reaction balance, symmetry and predicted deformation. Add a benchmark or simpler model whose assumptions overlap in a controlled way. Perform refinement on the named output and location, then compare with physical data where available.
Build a diagnostic table linking excessive stiffness to locking or over-constraint, checkerboard motion to hourglass behaviour, near-zero translation to insufficient restraint and rising peak stress to a singular idealisation. For each mechanism, state an observation, a discriminating test and a targeted correction.
Keep uncertainty categories separate: geometry, material, loads and supports are physical inputs, while mesh and solver settings are numerical. Start with transparent one-at-a-time sensitivity, then examine interactions for influential inputs. Compare the resulting output range with a design limit or decision boundary.
Use careful conclusions such as verified for equilibrium, stable for this output or not resolved by current evidence. No single successful check validates the whole model. Create an evidence matrix with rows for the main conclusion and columns for equilibrium, benchmark, discretisation, input sensitivity and physical comparison. Fill each cell with an observation rather than a green tick.
For example, reaction error below a stated tolerance supports global balance but says little about a local stress. A beam benchmark supports one idealised response but not a contact region. This format exposes where several checks depend on the same assumption and therefore are not truly independent. Practise writing diagnostic hypotheses before repairs.
If the response is too stiff, list over-constraint, unit error, section data and locking, then identify a test that separates them. If a deformation pattern has alternating zero-energy motion, inspect integration and stabilisation rather than increasing material stiffness. If a model translates nearly rigidly, trace supports and connectivity.
If local stress rises without bound, review the physical load or support idealisation. Only after the mechanism is supported should the model change. For uncertainty, start with a baseline table of geometry, material, load and support inputs. Give each range an evidence source or a clearly stated scenario.
Vary influential inputs in physically meaningful combinations, because support stiffness and contact state or modulus and density may not act independently. Distinguish sensitivity bounds from probability intervals. A probability statement needs distributions and dependencies that a simple parameter sweep does not supply. Compare the output range with the actual engineering threshold.
If every credible case leads to the same decision, the model may be robust despite numerical spread. If the decision changes, recommend better data, a more detailed model or a design with more margin. End every conclusion with scope: the output, location, load case and assumptions for which the evidence applies. That boundary prevents a verified global response from being used to certify an unexamined local failure mode.