Skip to content
Queensferry

Computational Engineering

Model Doctor

Deliberately flawed models, and the method for finding out what is wrong with them.

Every case follows the same seven parts, because that structure is the method the course teaches: symptoms, the first check that discriminates, the likely cause, the confirmatory test, the correction, the verification that it worked, and the lesson that generalises.

Note the order. The first check comes before the diagnosis, deliberately. An engineer who reaches for the diagnosis first will find evidence for it; naming the discriminating check first is what stops that.

Every case here 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 (14 foundation, 37 practitioner, 29 advanced) against a target of 80. The remaining cases and the interactive that scores them are listed in the coverage matrix.

Restraints and supports (6)

The frame that will not solve

foundation

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
Work it out first, then open the diagnosis
1. First check
Count the restrained directions before looking at anything else. In two dimensions you need at least three independent restraints; in three dimensions, six.
2. Likely cause
No horizontal restraint. The frame is free to translate as a rigid body.
3. Confirmatory test
Run a modal or stability analysis. A rigid-body mode at zero frequency, showing the whole frame sliding, names the direction exactly.
4. Correction
Restrain the bases horizontally, as the real foundation does. If the real base genuinely slides, model a spring rather than nothing — a real foundation has friction, and zero is not a stiffness the mathematics will accept.
5. Independent verification
The analysis runs, and the horizontal reactions sum to the applied horizontal load. Under pure gravity on a symmetric portal they should be equal and opposite.
6. The lesson
The model needs restraint in directions the structure has no load in. The requirement comes from the arithmetic, not from the statics.

The three-dimensional truss made of bars

practitioner

A roof truss built entirely from bar elements in a three-dimensional model. Supports are pinned at both ends. The loading is vertical and in the plane of the truss.

Symptoms

  • The solver reports instabilities at most of the truss nodes
  • Where it runs, the out-of-plane displacements are enormous and meaningless
  • The same truss modelled in two dimensions solves without complaint
Work it out first, then open the diagnosis
1. First check
Ask which element in the model resists a node moving out of the truss plane. If the answer is 'none', you have found it without running anything.
2. Likely cause
Bar elements in a three-dimensional model with no out-of-plane restraint at the interior nodes.
3. Confirmatory test
A stability or modal analysis returns as many near-zero modes as there are interior nodes, each one moving a single node sideways.
4. Correction
Either restrain the model to its plane, or use beam elements for the chords so they carry out-of-plane bending — which is closer to the real structure anyway, because the chords are continuous.
5. Independent verification
Zero-stiffness modes disappear. Compare chord forces with the two-dimensional model: they should barely change, which confirms the fix restored stability without altering the behaviour.
6. The lesson
Element choice and model dimension are one decision, not two. Bars are safe in a plane and dangerous in space.

The diaphragm that holds the floor up

advanced

A multi-storey frame where the floor slabs are replaced by rigid diaphragm constraints to save elements. The constraint was defined without restricting it to the horizontal directions.

Symptoms

  • Beam deflections under gravity are almost zero at every level
  • Column axial loads are strange: some columns pick up far more than their tributary area, others almost nothing
  • Total reactions still equal the applied load
Work it out first, then open the diagnosis
1. First check
Plot the deflected shape of one floor under gravity alone. A floor that stays flat while carrying load is being held, not stiffened.
2. Likely cause
The rigid constraint includes the vertical translation and the horizontal rotations.
3. Confirmatory test
Remove the constraint from a single floor and re-run. If that floor's beams suddenly deflect sensibly and its columns pick up their tributary loads, the constraint is the cause.
4. Correction
Restrict the diaphragm to the two horizontal translations and, where the software allows it, the rotation about the vertical axis. Never the vertical translation or the horizontal rotations.
5. Independent verification
Beam deflections match a hand calculation for the span and load. Column axial loads match a tributary-area take-down to within the redistribution you would expect from frame action.
6. The lesson
A diaphragm models in-plane stiffness. A floor slab is stiff in its plane and negligible out of it, and a constraint that does not respect that difference invents a structure you did not design.

The zig-zag base moments

advanced

A cross-section through a rectangular water-retaining structure, modelled with beam elements for the wall and base. Every node along the base is given a pinned restraint because the base sits on soil.

Symptoms

  • Wall moments look sensible
  • The base moment diagram zig-zags between adjacent nodes rather than varying smoothly
  • Reactions alternate in sign along the base
Work it out first, then open the diagnosis
1. First check
Look at the reactions rather than the moments. Alternating signs along a uniformly loaded base is not something soil can do.
2. Likely cause
Rigid point supports substituted for a continuous elastic medium.
3. Confirmatory test
Replace the pins with springs of any plausible stiffness. If the sawtooth vanishes and the moment becomes smooth, the supports were the cause.
4. Correction
Use spring supports with a modulus of subgrade reaction. A defensible first estimate: assume the foundation settles about 10 mm under full unfactored load, which makes the total spring stiffness a hundred times the load.
5. Independent verification
The moment diagram is smooth. Run the analysis at both ends of a plausible soil-stiffness range and check the design is not sensitive to the value you cannot know precisely.
6. The lesson
Infinite stiffness is a modelling choice, and usually a poor one. Wherever a real support has finite stiffness and the structure is indeterminate, a spring is more honest than a pin.

Half a building, and the half that walked away

practitioner

A symmetric building is modelled as one half, with the cut face on the plane of symmetry, to save analysis time. Vertical restraint is applied at the column bases. Wind is applied perpendicular to the cut.

Symptoms

  • The model solves
  • The sway under wind is roughly twice what a hand estimate suggests
  • The reactions on the cut face are all zero
Work it out first, then open the diagnosis
1. First check
Read the reactions on the symmetry plane. A correctly restrained symmetry plane carries reaction; zero everywhere means no restraint was applied.
2. Likely cause
The symmetry plane was treated as a free edge rather than as a boundary condition.
3. Confirmatory test
Build the full structure for one load case and compare the sway. If the half-model is roughly twice as flexible, the symmetry restraints are the difference.
4. Correction
Restrain displacement normal to the plane and the two rotations about axes lying in it, on every node of the cut face.
5. Independent verification
Sway matches the full model to within the meshing difference, and the reactions on the symmetry plane sum to half the applied lateral load.
6. The lesson
A symmetry plane is a support condition, not a place where the model happens to stop. And symmetric geometry is not enough: the loading, the mass and the restraints must all be symmetric before a half-model is valid at all.

The foundation spring that was never a spring

advanced

A tall core is supported on a raft modelled with vertical springs beneath it. The spring stiffness field was left at the software's default. Wind is applied to the building.

Symptoms

  • Base rotation of the core is almost exactly zero
  • The building's fundamental period is shorter than a hand estimate
  • Bending moment at the base of the core is at the top of the expected range
Work it out first, then open the diagnosis
1. First check
Read the base rotation. Any real soil rotates. A rotation of zero, or of the order of 10⁻¹⁰ radians, is a rigid support wearing a spring's name.
2. Likely cause
A default spring stiffness several orders of magnitude above anything a soil provides.
3. Confirmatory test
Divide the base moment by the base rotation and compare with the stiffness that was supposedly entered. If the implied stiffness is far above it, the spring is not doing what was asked.
4. Correction
Derive the spring stiffness from a soil parameter with a source — a modulus of subgrade reaction from the geotechnical report — and record where it came from.
5. Independent verification
Base rotation is non-zero and consistent with the stated stiffness. Repeat the analysis at half and twice the stiffness: if the design changes, soil stiffness is a governing assumption and belongs in the model record.
6. The lesson
A default is a number nobody derived. Where a stiffness governs, it needs a source, and the check is whether the design survives the plausible range rather than whether the number looks reasonable.

Releases and connectivity (6)

The beam released at every mesh division

foundation

A continuous steel beam over three spans, subdivided into eight elements per span so that results can be plotted smoothly. Moment releases were applied to the ends of every element to model the pinned connections at the supports.

Symptoms

  • The analysis fails, or returns displacements of the order of kilometres
  • The beam appears to have become a chain
  • Removing the releases makes it run
Work it out first, then open the diagnosis
1. First check
Count the releases. If the number of moment releases is close to twice the number of elements, they have been applied to elements rather than to members.
2. Likely cause
Releases applied element by element instead of member by member.
3. Confirmatory test
Display the releases graphically. Every internal node showing a hinge symbol on both sides is the picture of the fault.
4. Correction
Release only the two ends of each physical member, where the real connection is. Everywhere else the member is continuous steel and the model should say so.
5. Independent verification
The three-span moment diagram shows hogging over the internal supports. Compare the peak sagging moment with the standard coefficient for a three-span beam.
6. The lesson
Elements are a discretisation; members are the structure. Releases describe connections, and connections happen to members.

The columns with no axial load

foundation

A braced steel frame with simple beam-to-column connections. To model the pinned connections, pinned *restraints* were placed at the ends of every beam.

Symptoms

  • The columns carry almost no axial load
  • Total reactions are far larger than the applied load would suggest at the base
  • The beams behave correctly in isolation
Work it out first, then open the diagnosis
1. First check
Sum the reactions. If the total is right but it is appearing in the wrong places, the load path has been short-circuited rather than lost.
2. Likely cause
Pinned restraints applied at beam ends, where moment releases were intended.
3. Confirmatory test
Display the restraint symbols. A support symbol at every beam end, mid-air, is unambiguous.
4. Correction
Delete the restraints and apply moment releases about both axes at the ends of the beams that have simple connections.
5. Independent verification
Column axial loads accumulate down the building and match a tributary-area take-down. Reactions appear only at the foundations.
6. The lesson
Restraints join the model to the world. Releases describe what happens inside it. Both are sometimes called 'pinned', and that is the whole trap.

The beam that is not attached to anything

practitioner

A floor model assembled by copying a bay and moving it into position. The copied beams appear to meet the columns exactly on screen.

Symptoms

  • One bay deflects far more than its neighbours under the same load
  • The columns under that bay carry no load
  • The geometry looks perfect at every zoom level
Work it out first, then open the diagnosis
1. First check
Run the software's coincident-node check, or count nodes against what the geometry needs. It takes seconds and it is the first thing to do after any copy, move or import.
2. Likely cause
Duplicate nodes at the same coordinates, connecting nothing.
3. Confirmatory test
Display node numbers in that bay: two labels on top of one another. Or apply a small load and watch the two halves move apart.
4. Correction
Merge the coincident nodes to a suitable tolerance, then re-check.
5. Independent verification
The bay deflects like its neighbours and the columns under it carry a sensible share. The node count matches the geometry.
6. The lesson
The graphics show geometry, not connectivity. After any copy, move or import, check connectivity explicitly — the picture cannot tell you.

The truss chord carrying nothing

advanced

A floor containing a transfer truss, modelled in a building frame that uses a rigid diaphragm constraint at each floor. The truss chords lie in the floor plane and were included in the diaphragm.

Symptoms

  • The truss chords report almost zero axial force
  • The diaphragm constraint reports very large forces
  • The truss diagonals still carry sensible forces
Work it out first, then open the diagnosis
1. First check
Ask what is holding the chord's end nodes. If a constraint links them, the chord's axial behaviour has already been decided by something other than the chord.
2. Likely cause
Truss chords included in a rigid floor diaphragm.
3. Confirmatory test
Exclude the chord nodes from the diaphragm and re-run. Chord forces appear immediately and the constraint force falls.
4. Correction
Exclude the truss from the diaphragm, or analyse the truss in a separate model where its own stiffness governs. Both are legitimate; say which you did and why.
5. Independent verification
Chord force agrees with moment over truss depth from a beam idealisation of the truss.
6. The lesson
A rigid constraint is a silent, infinitely stiff member. Anything it links has had its behaviour decided for it, and the model will not warn you.

Released at both ends of the same joint

practitioner

A continuous beam runs over an internal column. Moment releases are applied to the ends of both spans at that column, intending to model a pinned splice.

Symptoms

  • The solver reports an unconstrained rotational freedom, or returns an enormous rotation at that node
  • The bending moment diagram shows nothing at all at the column
  • Deflection either side is far larger than expected
Work it out first, then open the diagnosis
1. First check
At every released node, count the members that still connect rotationally. If the answer is zero, the freedom is unsupported.
2. Likely cause
Both beam ends released at a node whose only other member is either absent or also released.
3. Confirmatory test
List the model's unconstrained freedoms. A good solver reports them by node and direction; the node in question will be named.
4. Correction
Release one side only. The joint is then a pin between the two spans, which is what a splice is, and the rotation belongs to the unreleased member.
5. Independent verification
The unconstrained-freedom list is empty, the moment at the splice is exactly zero, and the two spans behave as simply supported at that end.
6. The lesson
A release removes stiffness. Remove all of it at a node and the freedom has nothing left — which is a modelling error rather than a structural one, and the model is right to complain.

The edge beam that was asked to twist

advanced

A one-way slab is modelled as a grillage spanning onto an edge beam. The slab strips connect rigidly to the edge beam at every node. The edge beam is a shallow rectangular section.

Symptoms

  • Large torsional moments appear along the edge beam
  • The torsion is roughly constant along its length, changing sign at midspan
  • The edge beam fails a torsion check by a wide margin
Work it out first, then open the diagnosis
1. First check
Ask whether the torsion is needed for equilibrium. If removing the edge beam's torsional stiffness leaves the slab still able to span, it is compatibility torsion and it can be relieved.
2. Likely cause
Compatibility torsion attracted by a rigid connection between slab strips and a member with low torsional stiffness.
3. Confirmatory test
Re-run with a torsional release between the slab strips and the edge beam. If the slab moments redistribute slightly and everything still spans, the torsion was compatibility rather than equilibrium.
4. Correction
Release the torsional connection, or reduce the edge beam's torsional constant to a cracked value, and detail the beam for the residual. Both are decisions and both belong in the model record.
5. Independent verification
The slab still carries its load, the edge beam's torsion falls to something detailable, and the slab's span moments have risen by an amount that makes sense.
6. The lesson
Equilibrium torsion must be carried; compatibility torsion will be shed if the structure can crack and shed it. A linear model cannot tell them apart, and the engineer must.

Units, scale and signs (7)

The building a thousand times too big

foundation

A frame imported from a drawing package. The drawing is in millimetres; the analysis model expects metres. Sections and materials were assigned in the analysis package.

Symptoms

  • The model view auto-fits and looks entirely normal
  • Beams and columns render as hairlines
  • Deflections come out in kilometres
Work it out first, then open the diagnosis
1. First check
Switch on the solid section display. If the members stay as lines, the model is too big or the sections too small; if the view ends up inside a beam, the opposite.
2. Likely cause
Length unit mismatch on import.
3. Confirmatory test
Measure one bay in the model and compare with the drawing. A factor of exactly 1000 is not a coincidence.
4. Correction
Re-import with the correct unit, or scale the geometry by 0.001 and re-check every dimension.
5. Independent verification
Span, storey height and section depth all read correctly. Self-weight matches an estimate from volume times unit weight.
6. The lesson
The most common serious error in structural modelling is a factor of a thousand, and the view auto-fits so it never looks wrong. Check a dimension, not a picture.

The structure made of jelly

foundation

A user-defined material for a steel with the modulus typed as 210000, into a field whose unit is not what the user assumed.

Symptoms

  • The geometry is correct and the loads are correct
  • Deflections are a million times larger than expected
  • Bending moments are exactly right
Work it out first, then open the diagnosis
1. First check
Compare the moments with a hand calculation and the deflections with one. If the moments agree and the deflections do not, the error is in stiffness alone — modulus or second moment of area.
2. Likely cause
Young's modulus entered in the wrong pressure unit.
3. Confirmatory test
Read the material back with its unit displayed. Steel is 210 GPa; anything else is wrong.
4. Correction
Correct the modulus and re-run.
5. Independent verification
Span-over-deflection lands in the hundreds, not the fractions. The first natural frequency of a floor beam lands between about 4 and 10 Hz.
6. The lesson
Determinate moments do not depend on stiffness. That makes moment-versus-deflection the single most useful discriminator for narrowing a numerical error to its source.

The concrete that weighs a hundred times too much

practitioner

A user-defined concrete material with 2400 typed into the density field, in a program that expects a unit weight in kN/m³.

Symptoms

  • Self-weight deflections are about a hundred times too large
  • Applied-load results are correct
  • The first natural frequency is about ten times too low
Work it out first, then open the diagnosis
1. First check
Compute the total self-weight from volume times unit weight by hand and compare it with the model's reported total. One multiplication settles it.
2. Likely cause
Density confused with unit weight.
3. Confirmatory test
Divide the model's reported self-weight by the volume you can measure from the geometry. If the answer is about 2400 in a field labelled kN/m³, it is a mass in disguise.
4. Correction
Enter 24 kN/m³, or set the program's mass units correctly and enter 2400 kg/m³.
5. Independent verification
Total self-weight matches the hand calculation. The frequency returns to a plausible band for the span.
6. The lesson
Density and unit weight are different physical quantities that use the same word in conversation. The factor between them is g, and it is about 102 when the units are kg/m³ and kN/m³.

The load that lifts

foundation

A general finite element package where loads are defined in global axes with z upwards. A uniformly distributed floor load was entered as a positive z value.

Symptoms

  • The floor deflects upwards
  • Reactions are negative
  • Moments have the right magnitude and the wrong sign
Work it out first, then open the diagnosis
1. First check
Switch on the load graphics and look at which way the arrows point. It takes two seconds.
2. Likely cause
Sign convention on the gravity direction.
3. Confirmatory test
The deflected shape hogs where it should sag.
4. Correction
Negate the load, or use the program's gravity/self-weight function rather than typing a vector.
5. Independent verification
Reactions are positive upwards and sum to the applied load. The deflected shape sags.
6. The lesson
Every sign convention is defensible and none is universal. Check the arrows, and check moment conventions too — some packages use the left-hand rule.

The tank wall in compression

advanced

A circular reinforced-concrete tank modelled with shell elements. The hydrostatic pressure was applied normal to the wall, using the element local axis without checking which way it pointed.

Symptoms

  • The analysis runs cleanly with no warnings
  • Hoop forces are compressive everywhere
  • Vertical bending looks plausible
Work it out first, then open the diagnosis
1. First check
Look at the deflected shape, magnified. Water pushes a tank wall outwards. If the model's wall moves inwards, the load is inwards.
2. Likely cause
Pressure applied along the negative element normal.
3. Confirmatory test
Take a horizontal slice and check the hoop force against p·r by hand. The magnitude will match; only the sign is wrong.
4. Correction
Reverse the pressure, or set the element normals consistently outwards and reapply.
5. Independent verification
Hoop force is tensile and equal to pressure times radius at the level checked. A hand calculation and the model agree in sign and magnitude.
6. The lesson
A clean run is not a correct run. The sign of a result is a physical statement, and it needs checking against what you know the structure does — not against whether the solver complained.

The user-defined section that is 10¹² out

practitioner

A plate girder defined by typing section properties directly, taken from a spreadsheet that works in millimetres, into a program working in metres.

Symptoms

  • The girder deflects far more than the rolled sections around it
  • It attracts almost no moment in the continuous frame
  • Its area and mass look plausible
Work it out first, then open the diagnosis
1. First check
Compute EI/L³ for the girder and for a neighbouring rolled beam. If the girder is orders of magnitude softer, the section is wrong, not the model.
2. Likely cause
Section property entered in the wrong length unit.
3. Confirmatory test
Recompute the second moment of area from the plate dimensions in the model's own units and compare.
4. Correction
Enter properties in the program's units, or better, define the section parametrically from its plate sizes and let the program compute the properties.
5. Independent verification
The girder attracts moment in proportion to its stiffness relative to the adjacent members, and its span deflection matches a hand calculation.
6. The lesson
Section properties carry different powers of length. A unit slip is 10³ in a dimension, 10⁶ in an area, 10⁹ in a section modulus and 10¹² in a second moment of area — and typing properties by hand is the way to meet all four.

Section properties from a catalogue in centimetres

foundation

Beam sections are entered by hand from a manufacturer's table that lists the second moment of area in cm⁴. The model works in metres and kilonewtons.

Symptoms

  • Deflections are about a hundred million times too small
  • The structure appears almost perfectly rigid
  • Member forces in a statically determinate part of the model are correct
Work it out first, then open the diagnosis
1. First check
Take one member and hand-calculate its deflection under a known load. An error of many orders of magnitude is a unit error, and the exponent tells you which one.
2. Likely cause
Second moments of area entered in cm⁴ into a model working in m⁴.
3. Confirmatory test
Compare one entered value with the same section's I in m⁴. A UB 457 × 191 × 82 has I ≈ 3.7 × 10⁻⁴ m⁴ and 37 100 cm⁴ — the same section, eight orders apart.
4. Correction
Multiply every catalogue value by 10⁻⁸, or set the model's units and re-import.
5. Independent verification
The hand deflection and the model deflection agree. Determinate member forces are unchanged, which confirms nothing else was disturbed.
6. The lesson
A stiffness error leaves statically determinate forces exactly right. That is what makes it survive a check of the member forces, and why the deflection is the thing to check first.

Geometry, offsets and axes (7)

The 5 mm element that ruined the answer

practitioner

A steel frame where the beam-to-column eccentricity was modelled by inserting a very short beam element between the column centreline and the connection point.

Symptoms

  • Results are noisy: similar members report noticeably different forces
  • Total loads and reactions differ by a small but non-zero amount
  • Any reported condition number is very large
Work it out first, then open the diagnosis
1. First check
Sort the elements by length. Anything two or more orders of magnitude shorter than the rest is a candidate before you look anywhere else.
2. Likely cause
Very short, very stiff elements used to model an offset.
3. Confirmatory test
Report the ratio of the largest to the smallest element stiffness, or the condition number if the program gives one. Remove the stubs and watch the residual fall.
4. Correction
Use a rigid link or the element's built-in end-offset property. A link constrains the nodes to move together without contributing a huge stiffness, which is exactly the behaviour wanted.
5. Independent verification
Total loads and reactions cancel to within rounding. Similar members report similar forces. The column still receives the eccentric moment — check it explicitly, because the correction must not lose the effect it was modelling.
6. The lesson
Infinite stiffness is safe as a constraint and dangerous as an element. When you need something rigid, say 'rigid'; do not say 'very stiff' and hope.

The simply supported beam with a horizontal reaction

practitioner

A floor model built to match the architect's model exactly: every node at top-of-steel level, every beam offset down to its centroid so the drawing and the analysis look identical.

Symptoms

  • Simply supported non-composite beams report large axial forces
  • Horizontal reactions appear where there is no horizontal load
  • Beam moments are lower than a hand calculation gives
Work it out first, then open the diagnosis
1. First check
Look for a horizontal reaction under vertical-only load. On a simply supported beam there should be none.
2. Likely cause
Elements offset from the nodes that restrain them, with horizontal restraint at both ends.
3. Confirmatory test
Release one support horizontally. If the axial force vanishes and the moment rises to the hand-calculated value, the arching was the cause.
4. Correction
Model non-composite beams on their own centroids and accept that the analysis model will not look like the drawing. Reserve offsets for cases where the eccentricity is doing real structural work — composite action, connection eccentricity.
5. Independent verification
Moment matches wL²/8. Axial force is zero. Vertical reactions are unchanged.
6. The lesson
The CAD model represents the structure; the analysis model represents its behaviour. Making the second look like the first is a common instinct and a reliable source of error.

The purlins bending the wrong way

practitioner

A roof with cold-formed purlins whose sections are strong about one axis and weak about the other. The purlins were drawn in two directions, some left-to-right and some right-to-left.

Symptoms

  • Alternate purlins report very different deflections under identical load
  • The pattern follows the order in which they were drawn, not the geometry
  • Section and material are identical throughout
Work it out first, then open the diagnosis
1. First check
Display the element local axes, or the solid section rendering. The rotated ones are visible immediately.
2. Likely cause
Element orientation inherited from drawing direction.
3. Confirmatory test
Reverse one offending element and re-run. Its deflection should join the others.
4. Correction
Set the section rotation angle consistently, or re-draw the elements in a consistent direction. Most packages have a command to align local axes across a selection.
5. Independent verification
All purlins in the same situation report the same deflection. The reported major-axis direction matches the roof slope.
6. The lesson
An element knows its orientation from its node order, which is a modelling artefact and not an engineering decision. Any section that is not doubly symmetric will punish you for ignoring it.

The moment that vanished at the corner

advanced

A grid of ground beams spanning between piles, carrying masonry walls. The beams meet at right angles at the corners of the grid.

Symptoms

  • Bending moments in the two beams meeting at a corner do not balance
  • The mismatch is largest where the wall load is heaviest
  • Nothing in the output is flagged
Work it out first, then open the diagnosis
1. First check
At the corner node, list every moment component in every member and sum them, including torsion. Equilibrium is exact; if your sum is not zero you have left a component out of the sum.
2. Likely cause
Torsional moment generated at a rigid corner and not carried through to the design.
3. Confirmatory test
Plot the torsion diagram on the perpendicular beam. The magnitude will be exactly the missing amount.
4. Correction
Design for the torsion, detail the corner to release it, or change the layout so it does not arise. All three are legitimate; ignoring it is not.
5. Independent verification
All three moment components sum to zero at the node. Reinforcement or connection design accounts for the torsion.
6. The lesson
Equilibrium is exact in the solution. Any apparent imbalance is a component you have not looked at, and torsion is the one most often forgotten.

The column bending about the wrong axis

practitioner

Columns in a steel frame are given a UC section. Some were drawn bottom-to-top and some top-to-bottom, and the section orientation was set once for all of them.

Symptoms

  • Sway in one direction is far larger than in the other, in a frame that is nominally symmetric
  • Some columns carry much more moment than their neighbours
  • The affected columns are exactly those drawn in one direction
Work it out first, then open the diagnosis
1. First check
Display the local axes, or colour the model by major-axis direction. The pattern will be immediately visible and will not match the structure.
2. Likely cause
Member local axes flipped by drawing direction, taking the section orientation with them.
3. Confirmatory test
Compare the major-axis second moment of area reported for two columns that should be identical. If they are swapped, this is it.
4. Correction
Set the orientation per member relative to a global reference rather than the local axis, or redraw the offending members consistently.
5. Independent verification
Sway is equal in the two directions of a symmetric frame, and every column of the same size reports the same major-axis stiffness.
6. The lesson
Local axes are a property of how the member was drawn. Anything defined relative to them — orientation, releases, loads, results sign — inherits that, and drawing direction is not something anyone records.

The brace that misses the work point

advanced

A braced bay is modelled with the brace running to the beam-column intersection. The detail as designed connects the brace to a gusset some distance along the beam.

Symptoms

  • The model shows the brace carrying pure axial force
  • The beam's moment diagram is smooth through the connection
  • The connection design comes back with a moment nobody expected
Work it out first, then open the diagnosis
1. First check
Overlay the model's centrelines on the connection detail. If the three lines do not meet at one point in the detail, the model and the steelwork disagree about the joint.
2. Likely cause
Brace centreline modelled through the work point when the detail is eccentric.
3. Confirmatory test
Compute the moment the eccentricity implies: the brace force times the perpendicular distance from the work point. If it matches what the connection design reported, that is the source.
4. Correction
Model the eccentricity with a rigid offset, or accept the centreline model and design the connection and adjacent members explicitly for the eccentric moment. Either is defensible; leaving it undiscussed is not.
5. Independent verification
The moment the model shows at the joint matches the moment the connection is being designed for, and the beam near the connection is checked for it.
6. The lesson
A centreline model assumes the centrelines meet. When the fabricator's detail says otherwise, the difference is a moment, and it is real whether or not the model contains it.

The arch that is shorter than it should be

practitioner

A shallow arch is modelled as a series of straight members between points on the true curve. Eight segments are used across the span.

Symptoms

  • The arch's rise measured through the model is slightly less than specified
  • The thrust at the springings is a few per cent higher than a hand calculation gives
  • Refining to sixteen segments changes the thrust by about a quarter of the previous change
Work it out first, then open the diagnosis
1. First check
Measure the modelled rise at midspan and compare with the specified rise. Chords always cut inside a convex curve.
2. Likely cause
Straight-segment discretisation of a curved member, with too few segments.
3. Confirmatory test
Halve the segment length and watch the thrust. A converging sequence settles; the size of each step tells you how far from converged you are.
4. Correction
Use enough segments that the rise error is negligible, or place the nodes so the chords are secant to the curve — offsetting them so the chord midpoint sits on the true line — which converges much faster.
5. Independent verification
The modelled rise matches the specified rise to within a tolerance you have stated, and further refinement changes the thrust by less than that tolerance.
6. The lesson
Discretising geometry is a modelling decision with an error, and the error has a sign: chords always shorten a convex curve. Unlike a formulation error, this one does converge — so a refinement study settles it.

Loads and combinations (7)

The grillage that weighs twice what it should

foundation

A flat slab modelled as a grillage of beam elements running in both directions, with self-weight generated automatically from the element sections.

Symptoms

  • Total self-weight is twice the hand calculation
  • Deflections are about twice what is expected
  • Applied-load results are correct
Work it out first, then open the diagnosis
1. First check
Compute the slab volume from plan area times thickness, multiply by unit weight, and compare with the model's reported total. One line.
2. Likely cause
A grillage duplicates the slab, and so duplicates its self-weight.
3. Confirmatory test
Sum the volume of the grillage members. It will be twice the slab volume.
4. Correction
Halve the density used for self-weight, or apply the self-weight as an area load in one direction only. Say which you did, because the next person will check.
5. Independent verification
Total self-weight matches the hand calculation. Deflection under self-weight matches a plate solution to within the grillage's own approximation.
6. The lesson
A grillage is an abstraction of a slab, not a picture of one. Anything derived from the model's *volume* — self-weight, mass, quantity take-off — needs correcting for the duplication.

The envelope that misses the worst case

advanced

A foundation designed from an envelope of reactions across all load combinations. The envelope reports the maximum and minimum of each component independently.

Symptoms

  • Every individual component looks covered
  • The resultant force from a single real combination exceeds the resultant of the envelope extremes
  • Nothing is flagged anywhere
Work it out first, then open the diagnosis
1. First check
Compute the resultant for every individual combination and compare the largest with the resultant of the envelope. If they differ, the envelope is not covering you.
2. Likely cause
Component-wise enveloping of a vector quantity.
3. Confirmatory test
Find the combination that produces the largest resultant and check it against the envelope by hand.
4. Correction
Design for the coexisting values within each real combination, or envelope the resultant itself rather than its components.
5. Independent verification
The governing combination is named in the calculation, not just a set of extreme numbers. A spot check of that combination reproduces the design forces.
6. The lesson
An envelope is a convenience. The directions we resolve forces into are ours, not the structure's, and the critical case rarely lines up with them.

The nonlinear results that were added together

advanced

A basement held down against flotation by ground anchors, modelled with tension-only elements. The uplift case and the gravity case were run separately and the results combined in the post-processor.

Symptoms

  • The gravity-only case fails to solve
  • The uplift-only case solves and looks reasonable
  • The combined results show anchors in compression
Work it out first, then open the diagnosis
1. First check
List every nonlinear feature in the model — tension-only, compression-only, gaps, contact, plastic materials. If the list is not empty, superposition is off the table regardless of which solver was selected.
2. Likely cause
Results from a directionally nonlinear model combined after analysis.
3. Confirmatory test
Run the factored combination as a single analysis task. The anchors will be in tension or slack, never in compression.
4. Correction
Build factored load combinations inside the analysis task and analyse each one. The only post-processing operation that survives is the envelope.
5. Independent verification
No tension-only element reports compression. The combined case satisfies equilibrium on its own, not as a sum.
6. The lesson
A model can be nonlinear even when a linear solver was chosen. It is the elements and materials that decide, not the menu.

The seismic mass taken from the ultimate load

advanced

A seismic analysis of an office building where the floor mass was derived from the full ultimate-limit-state imposed load.

Symptoms

  • Natural periods are longer than comparable buildings
  • Base shear is high, and the design looks conservative
  • Inter-storey drifts are large
Work it out first, then open the diagnosis
1. First check
Divide the model's total seismic mass by the floor area. Compare with what an occupied office floor actually weighs.
2. Likely cause
Ultimate-limit-state imposed load used as seismic mass.
3. Confirmatory test
Re-run with the quasi-permanent imposed load and compare the period and base shear.
4. Correction
Use the mass likely to be present during the event, as the relevant code defines it, and state the assumption in the model record.
5. Independent verification
Total mass matches a take-off of the permanent load plus the appropriate imposed fraction. The period lands in a plausible band for the height and system.
6. The lesson
In a dynamic analysis, mass is a load case decision, not a material property. And 'more mass is conservative' is a hypothesis, not a rule — check where it puts you on the spectrum.

The finishes that were counted twice

foundation

A floor model has a superimposed dead load applied as an area load on the slab. The beams also carry a line load representing the same finishes, added when the beam model was built separately.

Symptoms

  • The total reaction exceeds the hand-calculated applied load by about a fifth
  • Beam moments are consistently higher than a hand check
  • Slab moments look correct
Work it out first, then open the diagnosis
1. First check
Sum the applied load by hand — area times pressure — and compare with the sum of the vertical reactions. They must agree exactly.
2. Likely cause
The same load applied through two paths after two models were merged.
3. Confirmatory test
Switch off the beam line loads and re-run. If the reaction total now matches the hand sum, the duplicate is confirmed.
4. Correction
Apply each load once, on the element that actually receives it, and record which that is.
5. Independent verification
Total reaction equals total applied load to rounding. Beam moments now agree with the hand check.
6. The lesson
The reaction-against-applied-load check takes a minute and catches every load error of this kind. It is the first thing to run on any model you did not build.

The continuous beam that was never patterned

practitioner

A four-span continuous beam is analysed with the full imposed load on every span, in a single load case, and designed from that result.

Symptoms

  • Support moments look reasonable
  • Span moments in the middle spans are noticeably lower than the end spans
  • A hand check of the middle span with alternate spans loaded gives a much larger sagging moment
Work it out first, then open the diagnosis
1. First check
Count the load cases. A continuous member needs the imposed load patterned, and one case cannot be an envelope.
2. Likely cause
Imposed load applied only in the all-spans-loaded arrangement.
3. Confirmatory test
Add the alternate-span cases and compare the span moment envelope. If the sagging moment rises materially, the single case was not governing.
4. Correction
Build the pattern cases, combine them into an envelope, and design from the envelope rather than from any one case.
5. Independent verification
The envelope's support moments come from the all-loaded case and its span moments from the patterned cases, which is the pattern to expect.
6. The lesson
An envelope is not a load case. Continuity means the worst arrangement differs for different effects, and a model that was only ever run one way cannot tell you that.

Wind applied to the surface it was calculated for

advanced

A pitched-roof building has wind pressure calculated per square metre of roof surface, and applied in the model as a load per square metre of plan area.

Symptoms

  • The total wind uplift is lower than a hand check on the roof
  • The discrepancy is a consistent factor across every roof panel
  • The factor is close to the cosine of the roof pitch
Work it out first, then open the diagnosis
1. First check
Sum the applied wind load in the model and compare with pressure times the true sloping area. A ratio equal to a trigonometric function of the geometry names the error.
2. Likely cause
A surface pressure applied to a projected area.
3. Confirmatory test
Divide the applied total by the hand total. If it equals cos(pitch) to two figures, that is the whole explanation.
4. Correction
Apply the pressure to the sloping area, or convert it to an equivalent projected-area value deliberately and record that you did.
5. Independent verification
Applied total matches pressure times sloping area, and the uplift reaction matches a hand check on the whole roof.
6. The lesson
A pressure needs the area it acts on stated. Plan and sloping areas differ by a factor everyone knows and few check, and the model will not ask which one you meant.

Meshing (7)

The opening the mesh could not see

practitioner

A shear wall with a door opening, meshed with a uniform element size chosen for the wall as a whole. Two elements span the width of the pier beside the opening.

Symptoms

  • Stresses beside the opening look moderate
  • Halving the element size increases them substantially
  • Halving again increases them again, though by less
Work it out first, then open the diagnosis
1. First check
Run the same model at three mesh sizes, each half the last, and tabulate the quantity you care about. One table settles what one model cannot.
2. Likely cause
Insufficient mesh density where the stress gradient is steep.
3. Confirmatory test
The change between successive refinements shrinks. If it does not, you are at a singularity instead.
4. Correction
Refine locally around the opening, keeping the transition gradual, and leave the rest of the wall coarse.
5. Independent verification
The quantity of interest changes by less than a few per cent on the next halving. State the mesh size you settled on and why.
6. The lesson
'The answer changed when I refined it' tells you the previous mesh was inadequate. It says nothing about whether the new one is.

The stress that would not converge

advanced

A flat slab modelled with shell elements, supported on columns modelled as single one-dimensional elements. The peak slab moment at the column is being refined towards.

Symptoms

  • Each halving of the element size increases the peak moment by roughly the same factor
  • The increase shows no sign of settling
  • Away from the columns everything converges nicely
Work it out first, then open the diagnosis
1. First check
Look at the ratio between successive refinements rather than the values. A constant ratio is the signature of a singularity; a shrinking one is convergence.
2. Likely cause
A dimensionless support applied to a dimensionless-thickness element.
3. Confirmatory test
Refine twice more. A converging quantity settles; this one keeps climbing at the same rate.
4. Correction
Stop chasing the peak. Take the design action from a quantity that converges: the integrated moment over the column strip, or the punching shear from the change in column axial force — which comes from equilibrium and has no mesh in it at all. Optionally, thicken the elements over the column zone or connect the slab to the column perimeter with links.
5. Independent verification
The integrated quantity is stable across mesh sizes. The punching check uses the reaction, not the peak stress, and the calculation says so.
6. The lesson
Not every quantity converges, and refining until the answer stops changing is a rule with an exception you must be able to recognise. Where a peak is singular, the resultant is not.

The hanging node

practitioner

A slab meshed finely in one bay and coarsely in the next, with the two regions meeting along a line where one coarse element abuts two fine ones.

Symptoms

  • A line of odd results follows exactly the boundary between the two mesh densities
  • Displacements are discontinuous across that line
  • The rest of the model is fine
Work it out first, then open the diagnosis
1. First check
Plot the displacement contour and look for a discontinuity. Displacement is continuous within a properly connected mesh, always.
2. Likely cause
Hanging nodes at a mesh-density transition.
3. Confirmatory test
Zoom in on the transition and display node connectivity, or apply a load and watch the mesh open along the line.
4. Correction
Transition gradually so element edges match one to one, or use the software's conforming-transition feature. Change size over several elements, not in one step.
5. Independent verification
Displacement contours are continuous everywhere. The odd line disappears.
6. The lesson
Elements connect through shared nodes and nothing else. A node that merely lies on an edge is not attached to it, however convincing the picture looks.

The mesh the auto-mesher made in a hurry

practitioner

An irregular slab meshed automatically with a coarse target size, producing several elements with aspect ratios above ten and corners near 170°.

Symptoms

  • Stresses vary erratically between adjacent elements
  • Nodal averaging hides the variation entirely
  • Displacements look plausible
Work it out first, then open the diagnosis
1. First check
Switch nodal averaging OFF and look at the unaveraged element stresses. Large jumps between neighbours are the mesh telling you it is not resolved.
2. Likely cause
Distorted elements, with averaging concealing the symptom.
3. Confirmatory test
Run the mesh-quality report: aspect ratio, skew, and the smallest Jacobian.
4. Correction
Reduce the target size so the mesher has room to make better shapes, split obtuse elements, and merge acute ones. Fix the geometry where slivers come from the outline rather than the mesher.
5. Independent verification
Unaveraged stresses vary smoothly between neighbours. Quality metrics are within limits and the converged answer is unchanged by further refinement.
6. The lesson
Averaging is a presentation choice that removes your best diagnostic. Turn it off when checking, and back on when reporting.

Two meshes that touch but do not connect

practitioner

A shell-meshed wall meets a shell-meshed floor slab along a line. The two were meshed independently at different sizes, and the nodes along the joint do not coincide.

Symptoms

  • The wall and slab deflect independently, with a visible step in the deflected shape at the joint
  • Almost no force is transferred between them
  • The model solves without complaint
Work it out first, then open the diagnosis
1. First check
Plot the deflected shape and look for a discontinuity. Continuous structures deflect continuously; a step means the two sides are not joined.
2. Likely cause
Independently meshed surfaces with non-coincident nodes along a shared line.
3. Confirmatory test
Count the nodes on the joint line belonging to each mesh. If the two counts differ, or the coordinates do not match, they are not connected.
4. Correction
Mesh the two surfaces together so they share nodes along the line, or connect them with constraints along the joint. Sharing nodes is simpler and less likely to be got wrong.
5. Independent verification
The deflected shape is continuous across the joint, and the force transferred along it balances what the slab is shedding to the wall.
6. The lesson
Elements connect through shared nodes and nothing else. Two surfaces drawn touching are not joined, and no solver will mention it, because a disconnected model is a perfectly valid model of something.

The stress that jumps at the mesh transition

advanced

A shell model has a fine mesh around an opening and a much coarser mesh away from it. The transition is made over a single row of elements.

Symptoms

  • Stress contours show a discontinuity along the transition line
  • The jump is well away from any change in geometry or load
  • Refining the coarse region reduces the jump
Work it out first, then open the diagnosis
1. First check
Compare the element size either side of the line. A ratio much beyond about two in one step is where transition artefacts appear.
2. Likely cause
An abrupt mesh size transition, giving different discretisation error either side.
3. Confirmatory test
Grade the transition over several rows and re-run. If the discontinuity disappears, it was the transition.
4. Correction
Grade the mesh — change size by no more than about a factor of two per row — so neither side is much better resolved than its neighbour.
5. Independent verification
Stress contours are continuous across the former transition, and the peak stress at the opening is unchanged, which confirms the fix affected the artefact rather than the result.
6. The lesson
A feature that moves when the mesh moves is a property of the mesh. It is worth learning to recognise, because it looks exactly like a stress concentration on a contour plot.

The slab meshed at the wrong level

practitioner

A composite floor is modelled with shell elements for the slab and beam elements for the downstand beams. The shells are placed at the top of the slab and the beams on their own centrelines.

Symptoms

  • The floor is much more flexible than a composite hand calculation predicts
  • The beams carry almost all the moment and the slab almost none
  • Refining the shell mesh does not help
Work it out first, then open the diagnosis
1. First check
Ask where the model thinks each centroid is. If the shell and the beam share a level, there is no lever arm and no composite action is possible.
2. Likely cause
Shell and beam elements at the same level, with no offset representing the true separation of their centroids.
3. Confirmatory test
Compute the composite second moment of area by hand using the transformed section and compare with the stiffness the model implies. The model's will be close to the bare beam's.
4. Correction
Offset the beam elements below the shell by the distance between the two centroids, or replace both with a single beam element using transformed section properties.
5. Independent verification
The floor's stiffness matches the transformed-section hand calculation, and the slab now carries compression while the beam carries tension.
6. The lesson
Composite action is a lever arm. A model that puts both parts at the same level has removed the mechanism entirely, and it will report a perfectly consistent answer for the structure it actually contains.

Materials and section properties (7)

Two checkers, two answers, one bridge

advanced

A continuous concrete bridge deck analysed independently by two engineers. One used the short-term concrete modulus, the other the long-term value.

Symptoms

  • Total load and reactions agree exactly
  • Support and span moments differ by more than twenty per cent
  • Both models are internally consistent
Work it out first, then open the diagnosis
1. First check
Compare the two models' assumption records before comparing their results. Any assumption that differs is a candidate, and the modulus is the one that moves an indeterminate structure most.
2. Likely cause
Different concrete modulus assumptions, both defensible, never reconciled.
3. Confirmatory test
Re-run one model with the other's modulus. The moments should converge.
4. Correction
Agree which modulus is appropriate to which check — short-term for transient effects, long-term for sustained — and record the decision. Where both matter, run both and envelope them.
5. Independent verification
Both models produce the same moments from the same assumptions. The remaining differences are traceable to differences that were agreed.
6. The lesson
An independent check that disagrees has done its job. The purpose of the check is to surface the assumption, and both engineers being right about different questions is the most common way it happens.

The floor that would not vibrate

foundation

A composite floor where the slab was assigned the steel material by mistake, having been copied from a steel plate element.

Symptoms

  • The floor's natural frequency is far above expectation
  • Static deflections are much smaller than expected
  • Total self-weight is about three times too high
Work it out first, then open the diagnosis
1. First check
List the materials actually assigned to each element group. It takes one screen and rules out the whole class.
2. Likely cause
Wrong material assigned to the slab elements.
3. Confirmatory test
Divide the reported self-weight by the slab volume. The answer will be about 77 kN/m³ rather than 24.
4. Correction
Assign the correct concrete material, with the dynamic modulus if this is a vibration model.
5. Independent verification
Self-weight matches the take-off. The first frequency lands in the band a floor of that span should occupy.
6. The lesson
When two independent symptoms move at once — mass and stiffness — the cause is usually one input that affects both. Materials are the commonest.

The cracked stiffness applied to half the frame

advanced

A concrete frame where cracked-section stiffness was applied to the beams to reflect service behaviour, but the columns and walls kept their gross-section properties.

Symptoms

  • Sway deflections are larger than a comparable building
  • Beam moments are lower and column moments higher than a first pass suggested
  • Every equilibrium check passes
Work it out first, then open the diagnosis
1. First check
List the stiffness modifier applied to every member group. Any group without one, in a model where others have one, is a deliberate decision or an omission — and you must be able to say which.
2. Likely cause
Inconsistent cracked-section assumptions across member types.
3. Confirmatory test
Re-run with gross properties everywhere, and again with cracked properties everywhere. If the answers bracket the inconsistent model, the inconsistency is what is driving it.
4. Correction
Apply a consistent, stated set of modifiers appropriate to the limit state being checked, and record them. Serviceability and ultimate generally want different sets.
5. Independent verification
Drift falls within the bracket set by the two consistent runs. The modifiers are recorded in the model record with a reason.
6. The lesson
In an indeterminate structure, stiffness is a *relative* quantity. A modifier applied to one member type is a statement about every other, whether or not you meant it to be.

Concrete stiffness at the wrong state

practitioner

A reinforced concrete frame is analysed for serviceability deflection using gross uncracked section properties throughout.

Symptoms

  • Computed deflections are comfortably within limits
  • A hand check allowing for cracking gives roughly twice the deflection
  • The frame's fundamental period is shorter than measured on similar buildings
Work it out first, then open the diagnosis
1. First check
Ask what section state the model used, and whether that state is consistent with the check being made. Ultimate and serviceability want different answers.
2. Likely cause
Gross uncracked section properties used for a serviceability deflection check.
3. Confirmatory test
Re-run with cracked properties for members whose service moment exceeds the cracking moment. If the deflection roughly doubles, that is the effect.
4. Correction
Use stiffness appropriate to the check: cracked for serviceability deflection, and state which members were treated as cracked and why.
5. Independent verification
The model's deflection agrees with a hand calculation using the same cracked properties, and the assumption is recorded.
6. The lesson
A material model is chosen for the question being asked. The same frame needs different stiffness for deflection, for load distribution and for dynamics, and using one set for all three is a decision that should be conscious.

Two moduli for the same timber

advanced

A glulam roof structure is analysed for deflection using the fifth-percentile modulus of elasticity taken from a strength-class table.

Symptoms

  • Deflections are noticeably larger than the designer expected
  • The structure is being upsized on deflection despite comfortable strength utilisation
  • A colleague's check of the same structure gives a smaller deflection
Work it out first, then open the diagnosis
1. First check
Read the modulus the model used and compare it with the strength class table. If it is the fifth-percentile value, ask which check it is being used for.
2. Likely cause
The fifth-percentile modulus applied to a deflection calculation that calls for the mean.
3. Confirmatory test
Re-run with the mean modulus. The deflection ratio should equal the ratio of the two moduli exactly, since deflection is inversely proportional to E.
4. Correction
Use the modulus the check calls for, and record which was used where. Timber's two values exist because they answer different questions.
5. Independent verification
The ratio between the two runs equals the ratio of the moduli, confirming nothing else changed.
6. The lesson
Material properties are not single numbers. Where a code offers more than one value, choosing between them is an engineering decision, and a model that silently uses the first one in the list has made it for you.

A material change that did nothing

foundation

A steel frame is upgraded from S275 to S355 to solve a utilisation problem. The model is re-run with the new grade assigned.

Symptoms

  • Every deflection is identical to the previous run, to the last digit
  • Member forces are identical
  • Only the utilisation ratios have changed
Work it out first, then open the diagnosis
1. First check
Ask which side of the check the changed parameter belongs to. Stiffness drives the analysis; strength drives the resistance.
2. Likely cause
None — the model is behaving correctly, and the surprise is a misunderstanding of what grade affects.
3. Confirmatory test
Compare the modulus reported for the two grades. Both will be 210 GPa.
4. Correction
None needed. Record why the analysis was unchanged, so the next person does not repeat the investigation.
5. Independent verification
Utilisations have fallen in the ratio of the two yield strengths for members governed by strength, and are unchanged for members governed by deflection — which is the tell for which check governs each member.
6. The lesson
Not every surprise is an error. A case that turns out to be correct behaviour is worth as much as one that is not, and 'the model did nothing' is sometimes exactly right.

The self-weight that never arrived

foundation

A steel frame model has self-weight switched on. A custom material was defined for one group of members, with the density field left blank.

Symptoms

  • Total vertical reaction is a few per cent below the hand-calculated total
  • The shortfall matches the weight of one group of members
  • Those members show no self-weight load at all
Work it out first, then open the diagnosis
1. First check
Sum the vertical reactions and compare with hand-calculated self-weight plus applied load. Any shortfall is a load that is not there.
2. Likely cause
A custom material with no density, giving zero self-weight for the members using it.
3. Confirmatory test
List materials and their densities. A blank or zero entry names the group immediately.
4. Correction
Enter the density, and check every custom material for missing fields.
5. Independent verification
The reaction total matches the hand total to rounding, and the previously missing group now reports self-weight in proportion to its volume.
6. The lesson
A blank field is a value, and it is usually zero. The fields that matter are the ones you did not fill in, which is why a load total against a hand check belongs on every model.

Second-order effects and buckling (6)

The beam with 800 MPa of elastic stress

practitioner

A slender beam under a heavy central point load, analysed linearly. The deflection magnification factor the program applied to draw the deflected shape was only 2.5.

Symptoms

  • The deflected shape and moment diagram look entirely sensible
  • Peak stress is around 800 MPa
  • There is no axial force at all
Work it out first, then open the diagnosis
1. First check
Look at the deflection magnification the program chose. A small factor means large real deflections, and it is on the screen before you look at any number.
2. Likely cause
Linear analysis of a structure whose deflections are far outside the small-displacement assumption.
3. Confirmatory test
Re-run nonlinearly. Deflections fall, axial force appears, and the deflected shape becomes V-shaped rather than smoothly curved — the signature of a hinge with catenary action.
4. Correction
Use a nonlinear analysis with material and geometric nonlinearity, and design the beam so it does not need one.
5. Independent verification
Stresses lie within the material's range. Axial force and deflection are consistent with catenary equilibrium at the deflected geometry.
6. The lesson
Linear analysis will happily report an elastic stress far above anything the material can carry. It has no way to object, because objecting is not among its assumptions.

The buckling mode read as a deflected shape

practitioner

A linear buckling analysis of a braced frame. The first mode is being reported as the frame's deflection under load.

Symptoms

  • Reported displacements are of the order of unity in an unspecified unit
  • The shape is smooth and plausible
  • There is no load case attached to the displacements
Work it out first, then open the diagnosis
1. First check
Ask what analysis produced the numbers. A buckling analysis produces a load factor and a set of shapes; if a displacement magnitude appears, it is meaningless.
2. Likely cause
Eigenvector magnitude mistaken for a displacement.
3. Confirmatory test
Re-run with a different normalisation, or scale the loads. The load factor stays put; the 'displacements' change by whatever factor the normalisation chose.
4. Correction
Use the buckling analysis for what it gives: the load factor, which mode it belongs to, and where in the structure that mode is active. Use a static analysis for displacements.
5. Independent verification
The load factor is reported with the mode it belongs to and the members participating in it. Displacements come from a static or second-order analysis.
6. The lesson
A buckling analysis answers 'at what load' and 'in what shape'. It does not answer 'how much does it move', and it never has.

The load factor read as a factor of safety

advanced

A steel frame whose lowest buckling load factor is 4.2. The designer records this as 'a factor of safety of 4.2 against buckling' and proceeds.

Symptoms

  • Nothing in the output objects
  • Member utilisations from the linear analysis are all below one
  • No second-order analysis has been run
Work it out first, then open the diagnosis
1. First check
Compute α/(α−1). If it is not close to 1, second-order effects are in the answer whether or not they are in the analysis.
2. Likely cause
Buckling load factor treated as a margin rather than as an amplifier.
3. Confirmatory test
Run a second-order analysis with the factored combination and compare moments. They should rise by roughly the amplification factor in the sway-sensitive parts.
4. Correction
Include second-order effects: a P-delta or nonlinear analysis on factored combinations, with imperfections scaled from the buckling mode where the code requires them.
5. Independent verification
Second-order moments exceed first-order moments by roughly α/(α−1) in the parts of the frame participating in the mode, and by very little elsewhere — because amplification is local to where buckling is starting.
6. The lesson
α_cr describes how much the loads you already have are being magnified. Every structure has second-order effects; the factor tells you how much of them you can afford to ignore.

A perfectly plumb frame

practitioner

A multi-storey frame is analysed for gravity load only. No equivalent horizontal forces are applied, and the frame is modelled perfectly vertical.

Symptoms

  • Sway under gravity load alone is exactly zero on a symmetric frame
  • Column moments are lower than a check including initial imperfections
  • The buckling load factor is comfortably high
Work it out first, then open the diagnosis
1. First check
Look at the sway under gravity load alone. If it is exactly zero, the model contains no imperfection at all.
2. Likely cause
Initial imperfections not represented, either geometrically or as equivalent horizontal forces.
3. Confirmatory test
Apply the equivalent horizontal forces and re-run. Column moments will rise; the amount tells you how sensitive the frame is to imperfection.
4. Correction
Apply equivalent horizontal forces derived from the code's initial sway imperfection, in each direction, combined with the gravity cases.
5. Independent verification
Sway under gravity is non-zero and in the direction of the assumed imperfection, and column moments include the second-order contribution.
6. The lesson
Exactly zero is almost always a modelling artefact. A structure that is perfectly symmetric, perfectly plumb and perfectly loaded exists only in the model.

Second-order results that were added together

advanced

A frame is analysed with P-delta effects included, for gravity and for wind as separate cases. The design combinations are formed by factoring and adding the two second-order results.

Symptoms

  • Combination results are lower than a direct second-order analysis of the combination
  • The difference grows with the gravity load
  • The buckling load factor is around four
Work it out first, then open the diagnosis
1. First check
Ask whether the combination was analysed, or assembled. Any nonlinear analysis must be run on the combined loading.
2. Likely cause
Superposition applied to a nonlinear analysis.
3. Confirmatory test
Analyse one combination directly with P-delta and compare with the assembled version. The direct analysis will give more.
4. Correction
Build each design combination as a load case and run the second-order analysis on it. There will be more analyses; there is no way round it.
5. Independent verification
Directly analysed combinations replace the assembled ones, and the amplification of each is consistent with its own axial load rather than with an average.
6. The lesson
Superposition is a property of linear analysis. The moment geometric stiffness enters, every combination is its own problem — and the software will let you add the results anyway.

An effective length that came from a menu

practitioner

Columns in an unbraced frame are checked using an effective length factor of 1.0, taken as the software's default, without a stability analysis.

Symptoms

  • Column utilisations are comfortable
  • A frame buckling analysis gives a load factor of about 3
  • Back-calculating the effective length from that factor gives a value well above 1.0
Work it out first, then open the diagnosis
1. First check
Ask where the effective length came from. If the answer is 'the default', it has not been established for this frame.
2. Likely cause
An effective length factor assumed rather than derived, for a frame whose sway behaviour does not match the assumption.
3. Confirmatory test
Run a linear buckling analysis and back-calculate the effective length from the critical load. Compare with what the design check used.
4. Correction
Use the effective length the frame's own stability gives, or brace the frame so the assumed value becomes true. Record which was done.
5. Independent verification
The design check's effective length matches the one implied by the buckling analysis, and the utilisations reflect it.
6. The lesson
Effective length is a property of the whole frame, not of the member. A default value is an assumption about the structure around the column, and it is one the model can actually settle.

Dynamic models (7)

The floor that vibrates in the model and not in reality

practitioner

A footfall analysis of an office floor, modelling structural self-weight only. Finishes, services, partitions and the imposed load likely to be present were left out.

Symptoms

  • Predicted frequency is higher than measured on the completed floor
  • Predicted response factors are worse than complaints suggest
  • Static deflections agree with hand calculation
Work it out first, then open the diagnosis
1. First check
Divide the model's total mass by the floor area and compare with what an occupied office floor actually weighs, superimposed loads and partitions included.
2. Likely cause
Non-structural mass omitted from a vibration model.
3. Confirmatory test
Add a plausible superimposed mass and re-run. Frequency should fall towards the measured value.
4. Correction
Include the mass likely to be present in service — not the ultimate imposed load, and not zero. Run a sensitivity study across the plausible range, because the answer moves with it.
5. Independent verification
Frequency lands in a plausible band. The sensitivity study shows how much the response depends on the mass assumption, and the report states the range.
6. The lesson
In a vibration model, mass is not a leftover from the static case. It is a design input with its own range, and the worst case is often the *lightest* one.

The base shear that went missing

advanced

A response-spectrum analysis of an irregular building where the first six modes were extracted, capturing 61 % of the mass in one direction.

Symptoms

  • The analysis runs cleanly
  • Base shear is well below an equivalent static estimate
  • Mode shapes look sensible
Work it out first, then open the diagnosis
1. First check
Read the effective-modal-mass table before reading any result. It is the first number a response-spectrum analysis should be judged on.
2. Likely cause
Insufficient modes extracted.
3. Confirmatory test
Extract more modes and watch both the cumulative participation and the base shear rise together.
4. Correction
Extract enough modes to reach the participation the relevant code requires, or add a missing-mass correction for the residual. State which you did.
5. Independent verification
Cumulative participation meets the target in every direction, including torsion. Base shear is within a sensible margin of an equivalent static estimate.
6. The lesson
A modal analysis reports only what it modelled. Participating mass is the measure of how much of the structure that was, and it is the check that comes before every other.

The time history that missed the peak

practitioner

A linear time-history analysis run with a step chosen to keep the run short. The step is about a fifth of the fundamental period.

Symptoms

  • The response trace looks smooth and plausible
  • Peak response is lower than a response-spectrum analysis of the same structure
  • Halving the step raises the peak substantially
Work it out first, then open the diagnosis
1. First check
Halve the step and re-run. If the answer changes, the original step was inadequate — the same convergence discipline as a mesh, applied in time.
2. Likely cause
Integration time step too large for the frequencies that matter.
3. Confirmatory test
Halve again. When two successive halvings stop changing the answer, the step is adequate.
4. Correction
Choose a step small enough for the highest mode contributing to the response — a common working rule is a tenth of that mode's period, or finer.
5. Independent verification
The peak is stable under further halving. The step used is recorded, with the mode that governed the choice.
6. The lesson
Time-step convergence is mesh convergence in another variable. The same rule applies: halve it until the answer stops moving, and state where you stopped.

Vibrating under factored load

practitioner

A floor vibration analysis takes its mass from the ultimate limit state combination, because that load case was already built.

Symptoms

  • The computed fundamental frequency is lower than a hand estimate
  • The floor is reported as a vibration problem
  • Re-running with a smaller mass moves the frequency comfortably clear
Work it out first, then open the diagnosis
1. First check
Ask which load case the mass came from. Vibration mass is the mass actually there — self-weight, finishes and a fraction of the imposed load, unfactored.
2. Likely cause
Modal mass derived from a factored ultimate load combination.
3. Confirmatory test
Recompute the frequency with the unfactored mass. Frequency goes as one over the square root of mass, so the ratio is predictable and confirms the diagnosis.
4. Correction
Build a mass case from unfactored permanent load plus the fraction of imposed load expected to be present during the vibration event, and say which fraction and why.
5. Independent verification
The frequency ratio between the two runs equals the square root of the mass ratio, and the result agrees with the 18/√δ estimate.
6. The lesson
Design combinations are for checking resistance. Dynamics happens in the real building, under the load that is actually on the floor, and reusing a convenient case is a decision with a direction.

Not enough modes to carry the mass

advanced

A response spectrum analysis of an irregular building extracts the first six modes, which is what the software offered by default.

Symptoms

  • Participating mass is around 60 % in one horizontal direction
  • Base shear from the spectrum analysis is well below an equivalent static estimate
  • Adding modes raises the base shear steadily
Work it out first, then open the diagnosis
1. First check
Read the participating mass before reading any result. It is the first number a modal analysis should be judged on.
2. Likely cause
Too few modes extracted to represent the mass in the direction of interest.
3. Confirmatory test
Extract more modes and watch the participating mass and the base shear. Both should rise and then settle.
4. Correction
Extract modes until the participating mass reaches the required fraction in every direction, and add a missing-mass correction for the remainder.
5. Independent verification
Participating mass exceeds the required threshold in all directions, the base shear has settled, and it is consistent with an equivalent static estimate.
6. The lesson
A modal analysis represents only the modes it extracted. Participating mass is the measure of how much of the building that is, and a default mode count knows nothing about your structure.

A diaphragm that was not rigid

advanced

A long, narrow building with a timber floor deck is analysed with rigid diaphragm constraints at each level, to distribute lateral load to the walls by relative stiffness.

Symptoms

  • Lateral load is distributed to the walls strictly in proportion to their stiffness
  • The end walls carry much less than a flexible-diaphragm hand calculation gives
  • The computed fundamental period is shorter than expected
Work it out first, then open the diagnosis
1. First check
Compare the diaphragm's in-plane stiffness with the walls'. The rigid assumption needs the deck to be much stiffer than the vertical elements, and a long narrow plan makes that harder, not easier.
2. Likely cause
Rigid diaphragm constraints applied to a deck that is flexible relative to the walls.
3. Confirmatory test
Model the deck with shell elements at its real in-plane stiffness and compare the wall force distribution. If the end walls attract materially more, the diaphragm was not rigid.
4. Correction
Model the deck's in-plane stiffness explicitly, or distribute by tributary area, and check the walls for the more onerous of the two.
5. Independent verification
The distribution lies between the rigid and tributary extremes, and the deck's own in-plane forces are checked.
6. The lesson
A constraint is infinitely stiff in the direction it constrains. Applying one is a statement about the structure, and 'rigid diaphragm' is a default in most software and an assumption in every building.

Damping that was assumed rather than justified

practitioner

A footbridge's pedestrian comfort check uses 2 % critical damping, taken from a general table, for a bare steel structure with welded connections and no finishes yet installed.

Symptoms

  • Predicted accelerations are just inside the comfort limit
  • The check passes with little margin
  • A sensitivity run at 0.5 % damping fails the limit by a factor of four
Work it out first, then open the diagnosis
1. First check
Ask where the damping value came from, and run the check at the bottom of its plausible range before doing anything else.
2. Likely cause
A damping ratio taken from a general table rather than justified for this structure in its as-built state.
3. Confirmatory test
Sweep damping across the plausible range and plot the acceleration. If the verdict changes within the range, damping governs the design.
4. Correction
Design for the low end of the justifiable range, or commit to measuring the completed structure and installing tuned damping if it falls short. Both are decisions; assuming the middle of a table is not.
5. Independent verification
The check passes at the lower bound of the damping range, or a contingency is designed and recorded.
6. The lesson
When one assumption controls the answer, the design must survive its plausible range rather than its central value. Damping is almost always that assumption in a comfort check.

Model transfer and interoperability (7)

The column that did not survive the import

practitioner

A frame exported from an information model and imported into an analysis package. One column was drawn in the source model as a component type the exchange format does not carry.

Symptoms

  • The imported model solves cleanly
  • One bay deflects more than its neighbours
  • Element count is one lower than the source model, and nobody counted
Work it out first, then open the diagnosis
1. First check
Count. Nodes, elements, total length by section, total mass — compare each against the source model before looking at a single result.
2. Likely cause
Information lost in a model exchange.
3. Confirmatory test
Overlay the two models and difference them, or compare the take-off by section type.
4. Correction
Re-export with the member represented as a type the format carries, or add it by hand and record that you did.
5. Independent verification
Counts and quantities match the source. Deflections are consistent bay to bay.
6. The lesson
Every exchange loses something. The question is not whether, but what — and the only reliable answer comes from a quantity comparison, not from looking at the picture.

The pinned frame that came back rigid

advanced

A steel frame exchanged between two analysis packages. The source model used moment releases at every simple connection; the exchange format carried the geometry and sections but not the releases.

Symptoms

  • Column moments are much larger than in the source model
  • Beam sagging moments are smaller
  • Total load and reactions agree exactly
Work it out first, then open the diagnosis
1. First check
Count the releases in the imported model. Zero, in a frame with simple connections, is the whole diagnosis.
2. Likely cause
Connection information lost in the exchange.
3. Confirmatory test
Display the release symbols. Compare beam end moments with the source model: they should be zero and are not.
4. Correction
Reapply the releases, ideally by script so it is repeatable, and record the step in the transfer procedure so the next exchange does not repeat it.
5. Independent verification
Beam end moments are zero at simple connections. Beam and column moments match the source model within rounding.
6. The lesson
Geometry survives an exchange far better than behaviour does. Releases, offsets, constraints and load cases are the things to check, precisely because they are invisible in the picture.

The analytical model that inherited the physical geometry

practitioner

A frame converted directly from a physical information model, where every beam sits at its true level with its top flange aligned and every column runs between floor faces.

Symptoms

  • Very short elements appear inside every connection zone
  • Results are noisy
  • Column effective lengths derived automatically are shorter than they should be
Work it out first, then open the diagnosis
1. First check
Sort elements by length. A crop of elements of the order of a section depth, all inside connection zones, names the cause immediately.
2. Likely cause
Physical geometry used directly as the analytical model.
3. Confirmatory test
Look at where the short elements sit. If they are all at connections, the conversion is the source.
4. Correction
Use the authoring tool's analytical model rather than its physical one, and rationalise it: nodes at intersections, members on centrelines, offsets expressed as offsets rather than as stub elements.
5. Independent verification
No element is much shorter than the rest. Column lengths are storey heights. Total loads and reactions cancel cleanly.
6. The lesson
A physical model and an analytical model answer different questions, and neither is a rationalisation of the other. Automating the conversion does not remove the engineering decision; it hides it.

Pinned in one model, fixed in the next

practitioner

An analysis model with pinned beam-to-column connections is exported to a second package for a check. The exchange format has no concept of an end release.

Symptoms

  • The receiving model solves and gives sensible-looking results
  • Beam midspan moments are about a third lower than the sending model's
  • Column moments are higher
Work it out first, then open the diagnosis
1. First check
Read the beam end moments in the receiving model. A pinned connection carries none, so any non-zero end moment says the release did not arrive.
2. Likely cause
End releases lost in an exchange whose format does not carry them.
3. Confirmatory test
Sum the applied load in both models — it will match — then compare a single beam's end moment. The redistribution is the signature.
4. Correction
Re-apply the releases in the receiving model, and add them to the handover note as something the route does not carry.
5. Independent verification
Beam end moments are zero, midspan moments match the sending model, and the column moments come back down.
6. The lesson
Geometry survives an exchange and behaviour does not. The check is on the receiving model's behaviour, not on whether the import reported an error.

The import that changed the units

foundation

A model built in kilonewtons and metres is imported into a package configured in newtons and millimetres. The import reports success.

Symptoms

  • The geometry looks correct on screen
  • Deflections are reported as enormous numbers
  • Member forces are a thousand times the expected values
Work it out first, then open the diagnosis
1. First check
Measure a known dimension in the receiving model — a bay width, a storey height. If it reads 6 000 where it should read 6, the units are the whole story.
2. Likely cause
Unit system not carried by the exchange, leaving values reinterpreted on arrival.
3. Confirmatory test
Check three quantities of different dimension — a length, a force and a modulus. The factors will be consistent powers of the length and force conversions.
4. Correction
Set the receiving package's units to match the file before importing, or convert the file. Then re-check the same three quantities.
5. Independent verification
A known dimension reads correctly, and a single member's deflection under a known load matches a hand calculation.
6. The lesson
A number without its unit is not data. An exchange that reports success has confirmed that the file parsed, and nothing at all about what the values mean.

Forces from a model that no longer exists

advanced

Connection design proceeds from a schedule of member forces exported from the analysis model six weeks earlier. The frame has since been re-analysed after a grid change.

Symptoms

  • Connection forces do not match the current model when spot-checked
  • Some connections are over-designed and some under
  • Nobody can say which analysis run the schedule came from
Work it out first, then open the diagnosis
1. First check
Ask which model version the schedule came from. If the answer is a date rather than a version, the link is already broken.
2. Likely cause
Analysis results exchanged as data, then outliving the model that produced them.
3. Confirmatory test
Spot-check three connections against the current model. Disagreement in both directions — some higher, some lower — indicates a different geometry rather than a different load factor.
4. Correction
Re-export from the current model at an identified version, re-run the connection design, and record the version on the schedule.
5. Independent verification
Every connection force traces to a named model version, and spot-checks agree exactly.
6. The lesson
Never exchange results. Re-run them. A received force has no model behind it, cannot be checked, and will still be in circulation long after the model has changed.

A model a long way from the origin

practitioner

A structural model is built on the project's site survey coordinates, several hundred kilometres from the coordinate origin, and imported into an analysis package.

Symptoms

  • The model solves but the results contain visible numerical noise
  • Members that should carry identical forces differ in the third significant figure
  • Moving the model to the origin removes the noise entirely
Work it out first, then open the diagnosis
1. First check
Read the model's coordinate range. Anything above about 10⁵ in the units in use is worth moving, and it costs nothing to check.
2. Likely cause
Site coordinates carried into analysis, consuming floating-point precision.
3. Confirmatory test
Translate the model to the origin and re-run. If the noise disappears, the coordinates were the cause.
4. Correction
Analyse near the origin and keep the survey transformation as a separate step for setting out. The two purposes want different coordinate systems.
5. Independent verification
Members that should be identical now report identical forces, and the equilibrium residual falls by orders of magnitude.
6. The lesson
Analysis wants numbers near unity and setting-out wants site coordinates. Carrying one into the other is a transfer decision, and precision is quietly spent by geometry nobody thought was an input.

Conditioning and numerical health (6)

The reactions that nearly balance

practitioner

A large model with several thousand elements, a mixture of very stiff links and slender bracing, and no reported errors.

Symptoms

  • Total applied load and total reaction differ by about 2 %
  • No warnings were issued
  • The deflected shape looks plausible
Work it out first, then open the diagnosis
1. First check
Total loads against total reactions, direction by direction. It is the cheapest check in structural computing and it should be run on every model, every time.
2. Likely cause
Ill-conditioning from an extreme stiffness range, lost connectivity, or both.
3. Confirmatory test
Report the stiffness range and the condition number if available. Remove or replace the stiffest elements and watch the residual fall.
4. Correction
Replace very stiff elements with rigid links or constraints, remove near-zero stiffnesses, and fix any coincident nodes. Then re-run the balance check.
5. Independent verification
The residual falls to a rounding-level fraction of the applied load. Repeat the check after every subsequent model change.
6. The lesson
Total loads against total reactions costs nothing and catches a large fraction of everything that can go wrong. Run it first, and run it again after every change.

The model in national grid coordinates

advanced

A bridge model imported with its real-world survey coordinates, placing it several hundred kilometres from the model origin.

Symptoms

  • Small local results are noisy
  • Repeating the run after a trivial edit changes results in the last digits
  • The same model, moved to the origin, is clean
Work it out first, then open the diagnosis
1. First check
Look at the coordinate values. If they have six or more digits before any that vary between nodes, precision is being spent on position.
2. Likely cause
Model located far from the origin, consuming floating-point precision.
3. Confirmatory test
Translate the model to the origin and re-run. Clean results confirm it.
4. Correction
Work in a local coordinate system near the origin and keep the survey transformation as a separate, documented step.
5. Independent verification
Results are repeatable to full precision and unchanged by trivial edits.
6. The lesson
Floating-point precision is a finite budget. Coordinates, stiffness range and element aspect ratio all draw on the same account.

The stub that broke the matrix

practitioner

A frame model contains a 5 mm long member, created accidentally when two nodes were placed almost on top of each other during a geometry import.

Symptoms

  • The solver warns about conditioning, or produces a very large condition number
  • Results are noisy and change when the model is renumbered
  • The equilibrium residual is orders of magnitude larger than usual
Work it out first, then open the diagnosis
1. First check
Sort the members by length and look at the shortest. A ratio beyond about 10³ between the shortest and the typical member is where conditioning starts to bite.
2. Likely cause
A near-zero-length member creating an extremely large stiffness term.
3. Confirmatory test
Delete the stub, merging its two nodes, and re-run. The condition number and the residual should both fall dramatically.
4. Correction
Merge coincident nodes with a tolerance appropriate to the model, and re-check the shortest member afterwards.
5. Independent verification
The condition number is back in its usual range, the residual is at rounding level, and renumbering the model no longer changes the answer.
6. The lesson
Conditioning is about the ratio between the largest and smallest stiffnesses. One accidental member can set that ratio for the whole model, and the geometry that caused it is invisible at any sensible zoom.

Soil springs and steel in the same matrix

advanced

A building model includes very soft soil springs beneath the foundations and stiff steel members above, in a single analysis.

Symptoms

  • The condition number is large but the model solves
  • Displacements at the top of the building are consistent between runs
  • Foundation spring forces vary by several per cent between otherwise identical runs
Work it out first, then open the diagnosis
1. First check
Compare the largest and smallest stiffness terms in the model. A range beyond about 10⁸ is where results at the soft end become unreliable.
2. Likely cause
A very wide range of stiffnesses in one matrix, degrading precision at the soft end.
3. Confirmatory test
Re-run with the model renumbered, or with a slightly different but equivalent spring representation. Quantities that move are the ones that have lost precision.
4. Correction
Consider a two-stage analysis — superstructure on a stiffness-matched support, then foundation with the resulting reactions — or scale the units so the range narrows. Report the foundation forces to a precision the arithmetic supports.
5. Independent verification
Foundation forces are stable between runs, and the reported precision is consistent with the sensitivity study.
6. The lesson
Precision is spent by the range of stiffnesses in the model, and it is spent at the soft end first. A result that changes when nothing physical changed is telling you how much of it to believe.

An equilibrium residual nobody read

practitioner

A large model runs successfully and produces a full set of results. The solver reports an equilibrium residual of 2 kN against applied loads totalling 4 000 kN.

Symptoms

  • No warnings are shown to the user
  • The residual appears in the log file only
  • Reactions sum to slightly less than the applied load
Work it out first, then open the diagnosis
1. First check
Read the residual on every run, before reading any result. It is the one number that tests the solution rather than the model.
2. Likely cause
Something absorbing load without reporting it — most often a constraint, an unconnected freedom being quietly restrained, or an iterative solver stopped short of convergence.
3. Confirmatory test
Compare the sum of reactions with the sum of applied loads. The difference will match the residual and points at where the load went.
4. Correction
Find what is absorbing the load. Check constraints, check for restrained-but-unreported freedoms, and if an iterative solver is in use, tighten its convergence and re-run.
5. Independent verification
The residual falls to rounding level, and the reactions sum to the applied load exactly.
6. The lesson
The equilibrium residual is the cheapest and most powerful check a model offers, and it is usually buried in a log nobody opens. A number you can read in kilonewtons is never rounding.

Reading the results (7)

The output that only reported the centre

advanced

A building with its stability elements placed off the centreline, analysed in software that reports one displacement per floor, at the centre of mass.

Symptoms

  • Reported inter-storey drifts are within limits
  • Torsional modes appear early in the modal results
  • No per-node displacement output was examined
Work it out first, then open the diagnosis
1. First check
Ask what the number you are reading is a displacement OF. If the software reports one value per floor, find out which point it belongs to before using it.
2. Likely cause
A limitation of the output, unrecognised, in a structure where it mattered.
3. Confirmatory test
Extract displacements at the plan corners. Compare with the centre value; in a torsion-dominated response they will be substantially larger.
4. Correction
Use software or an output route that gives displacements at the points that govern, and design for the worst corner rather than the centre.
5. Independent verification
Corner drifts are within limits, not just the centre value. The torsional modes are identified and their participation is reported.
6. The lesson
The most dangerous limitation of a tool is the one you do not know it has. Before relying on any output, establish what it is a measurement of.

The nonlinear analysis that converged beautifully

practitioner

A nonlinear analysis that converged in three iterations per step, with a tight residual tolerance, on a model whose supports were wrong.

Symptoms

  • Convergence is fast and clean at every step
  • The residual falls below tolerance easily
  • The deflected shape is not what the engineer expected
Work it out first, then open the diagnosis
1. First check
Take the disagreement between expectation and result seriously and resolve it, in writing. One of the two is wrong and you need to know which.
2. Likely cause
Convergence mistaken for correctness.
3. Confirmatory test
Sketch the deflected shape you expect from the load path, then compare. Where they differ, work out what the model would have to be for its shape to be right.
4. Correction
Fix whatever the comparison exposes — here, the supports — and re-run.
5. Independent verification
The deflected shape matches the expectation, or the expectation has been revised for a reason you can write down.
6. The lesson
Convergence, a clean run and no warnings are all statements about the arithmetic. None of them is a statement about the structure.

The averaged contour that looked perfect

practitioner

A shell model whose stress contours are displayed with nodal averaging on, as the software defaults to.

Symptoms

  • Contours are smooth and attractive
  • Turning averaging off reveals large jumps between adjacent elements
  • The mesh looks acceptable at a glance
Work it out first, then open the diagnosis
1. First check
Turn averaging off once, deliberately, on every model, and look at the jumps.
2. Likely cause
A display default concealing the mesh's own accuracy indicator.
3. Confirmatory test
Refine the mesh and watch the unaveraged jumps shrink. If they do not, the mesh is not the problem — look for a singularity or a distortion instead.
4. Correction
Refine where the jumps are large, then present averaged contours in the report.
5. Independent verification
Unaveraged jumps are small relative to the stress. The refined and unrefined answers agree in the quantities used for design.
6. The lesson
Software defaults are chosen for presentation, not for checking. Know which of your displays is smoothing something, and turn it off before you believe anything.

The stress under the point load

practitioner

A shell model of a slab has a heavy point load applied at a single node. The peak stress beneath it is used to size reinforcement.

Symptoms

  • The stress at that node is far above anything nearby
  • Refining the mesh increases it further, without settling
  • The stress a short distance away is reasonable and stable
Work it out first, then open the diagnosis
1. First check
Refine once and compare. A converging quantity settles; a singular one grows by a roughly constant ratio each time.
2. Likely cause
A concentrated load applied at a single node, producing a mathematical singularity.
3. Confirmatory test
Plot stress against distance from the node. A singular field falls steeply and then flattens into the usable region.
4. Correction
Spread the load over its real bearing area, and read the stress outside the disturbed zone. Both are what the real detail does.
5. Independent verification
The peak stress settles under refinement, and it agrees with a hand calculation for the load spread over its bearing area.
6. The lesson
The largest number in a result is often the least trustworthy. A singularity looks exactly like a stress concentration, and only refinement behaviour tells them apart.

Two stresses at the same node

advanced

A shell model reports both averaged and unaveraged element stresses. The two differ substantially in one region and hardly at all in another.

Symptoms

  • Averaged contours look smooth everywhere
  • Unaveraged contours show visible jumps between elements in one region
  • The design is being taken from the averaged values
Work it out first, then open the diagnosis
1. First check
Plot the unaveraged stresses. Where elements disagree, the mesh is not resolving the field.
2. Likely cause
A mesh too coarse to resolve the stress field in one region, concealed by averaging.
3. Confirmatory test
Refine the region and re-plot unaveraged. The jumps should shrink; if they do not, the cause is a singularity rather than resolution.
4. Correction
Refine where the elements disagree, until the averaged and unaveraged values are close. Then either may be used.
5. Independent verification
Averaged and unaveraged stresses agree to within a stated tolerance across the region of interest.
6. The lesson
Averaging is a display choice that removes a diagnostic. The disagreement between neighbouring elements is the model telling you how much to trust it, and smoothing it away discards the message.

An envelope used as a load case

practitioner

Design forces are taken from an envelope across all combinations. A connection is then designed for the maximum axial force and the maximum moment read from that envelope.

Symptoms

  • The connection is unusually heavy
  • No single combination produces both the maximum axial force and the maximum moment
  • The two maxima come from combinations with different wind directions
Work it out first, then open the diagnosis
1. First check
Ask which combination each envelope value came from. If the answer differs between quantities, they are not concurrent.
2. Likely cause
Non-concurrent envelope extremes combined into a single design case.
3. Confirmatory test
Extract the concurrent force sets — the axial force and moment from the same combination, for every combination — and find the one that is genuinely worst for the connection.
4. Correction
Design from concurrent force sets. Most software will report them; it is rarely the default view.
5. Independent verification
The governing case is a real combination, and its forces are consistent with each other and with the frame's equilibrium under that combination.
6. The lesson
An envelope is a summary, not a state of the structure. Every number in it is real and no two of them necessarily happened at the same moment.

A hogging moment that was reinforced as sagging

foundation

A continuous beam's bending moment diagram is read from a package that plots moment on the tension face, and transferred to a package that plots the opposite way, without either convention being stated.

Symptoms

  • Reinforcement is detailed with the main steel in the wrong face over the support
  • The magnitudes are all correct
  • The pattern of the diagram is correct but inverted
Work it out first, then open the diagnosis
1. First check
Look at a span you know the answer to. A simply supported beam under gravity sags, and whichever way that plots is the convention in use.
2. Likely cause
Opposite plotting conventions between two packages, with neither stated.
3. Confirmatory test
Compare the deflected shape with the moment diagram. Sagging deflection must go with sagging moment, whatever sign the plot gives it.
4. Correction
Establish the convention from a case you know, annotate every diagram with the face the steel goes in, and stop transferring signs between packages.
5. Independent verification
Reinforcement is in the tension face at every location, checked against the deflected shape rather than against a sign.
6. The lesson
A sign is meaningless without its convention, and the convention is the thing that does not travel. Checking against a case whose answer you know is faster than looking up either package's documentation.
← Computational Engineering