Module 12 · Lesson 12.5
The Model Doctor
The method, applied: name the discriminating check before you name the cause.
Why this matters
Everything in this module so far has been a fault and its symptoms. This lesson is about the procedure — and the single most important thing about it is an ordering that feels wrong the first time.
Name the check before you name the cause. An engineer who reaches for a diagnosis first will find evidence for it, because a model has thousands of numbers and some of them will fit any story. Committing to a discriminating check first is what stops that.
By the end of this lesson you should be able to
- Apply the seven-part method to an unfamiliar model
- Choose a check that discriminates between candidate causes rather than confirming one
- Distinguish local from global symptoms as a first cut
- Say what a correction must be verified against
The seven parts
- 1.Symptoms — what you actually see, before you know what is wrong.
- 2.First check — the cheapest thing that discriminates between candidates.
- 3.Likely cause — the diagnosis.
- 4.Confirmatory test — the thing that proves it rather than assuming it.
- 5.Correction — what to change.
- 6.Verification — how you know the correction worked.
- 7.Lesson — what generalises.
Steps 2 and 4 are what make it a method rather than a checklist. A check that would come out the same whatever the cause is not a check.
Local or global: the first cut
Before anything else, ask whether the symptom is confined or spread. The two classes of cause almost never overlap:
Local — one bay, one member, one node. Candidates: connectivity, a wrong assignment, a local restraint or release, a mesh problem in that region.
Global — the whole model, uniformly. Candidates: units, material, load magnitude, analysis type, a systematic geometry error.
A unit error affects everything by the same factor. A lost member affects one bay. Establishing which you have takes one look at the results and eliminates half the possibilities.
The discriminating check
The skill is choosing a check whose two possible outcomes point at different causes.
Weak: 'check the section properties' — they might be right, and you have learned almost nothing.
Strong: 'compare a determinate moment with a hand calculation' — right means the error is in stiffness; wrong means it is in geometry or load. Either outcome halves the search.
More strong ones, worth having ready:
- Total load against total reaction — balanced means nothing is lost; unbalanced means connectivity or conditioning.
- Refine the mesh once — a shrinking increment means convergence; a constant ratio means a singularity.
- Replace pins with springs — a sawtooth that vanishes was caused by rigid supports.
- Remove a constraint from one floor — behaviour that changes there identifies the constraint.
- Turn averaging off — large jumps mean the mesh is not resolved.
Correction, then verification
A correction is not finished when the model runs. It is finished when a check that would have failed before now passes, and the effect the correction was meant to preserve is confirmed to still be there.
That second half gets forgotten. Replacing an over-stiff stub element with a rigid link fixes the conditioning — and if the stub was modelling a connection eccentricity, the eccentric moment must still be in the column afterwards. A correction that removes the fault and the feature has not corrected anything.
Try it
Model Doctor
80 original, fictional cases across 12 fault categories. Read the symptoms, commit to a diagnosis, then open the method.
Restraints and supports
6 in this category
- Difficulty
- foundation
- Cases written
- 80 of a target 80
The frame that will not solve
A single-bay, single-storey portal frame modelled with beam elements. Both column bases are given vertical restraint only. Gravity load is applied to the rafter.
Symptoms
- The solver stops with a message about a singular matrix or a zero pivot
- No results are produced at all
- Adding more load makes no difference to the message
What is your diagnosis?
Nothing is revealed until you commit. That is the discipline the module teaches: an engineer who reads the answer first will find evidence for it in any model.
About these cases
- Every case is invented. None is a retelling of a real project, and none is taken from any source.
- Where a real failure is relevant it is named once, in Module 13, with a pointer to the published investigation — not reconstructed.
- 80 cases are written against a target of 80. The categories still short are listed in the coverage matrix.
What this shows: Name the discriminating check before you name the cause — an engineer who reaches for the diagnosis first will find evidence for it.
Worked example
Diagnosing an unfamiliar model
Given
- A multi-storey frame where beam deflections under gravity are almost zero at every level
- Column axial loads are erratic — some far above their tributary area, others near zero
- Total reactions equal the applied load exactly
- No warnings
Find
The cause, following the method
Check yourself
A single bay of a floor deflects three times more than its neighbours under identical load, and the columns under it carry almost nothing. Total reactions are correct. What is the strongest first check?
Practice
A model has 96 elements. A count finds 187 moment releases. Roughly how many releases would you expect if they had been applied to every element end? Give the number.
Summary
- Name the discriminating check before naming the cause
- Local or global is the first cut, and it eliminates half the candidates
- A check whose outcomes point at the same cause is not a check
- Confirm the diagnosis rather than assuming it
- Verify that the correction fixed the fault and kept the feature
This is educational material. It uses simplified examples to teach principles, and must not be relied on for real design or safety-critical decisions. Module overview and checkpoint