A tutorial teaches you where the buttons are. It does not teach you what to choose when your own part is in front of you and nobody is sitting next to you.
And bad results almost never come out of the solver. They come from the places where a default setting was used instead of a decision — and a default setting never knows what you are building the model for.
Below are thirteen decisions in the order in which they have to be made. They read top to bottom: the objective sets the idealisation, the idealisation sets the mesh. Never the other way round. If you find the mesh forcing a change in the idealisation, you have gone too far back and something higher up is wrong.
The boxes marked ⛔ are the mistakes I see most often when reviewing other people's analyses. Every one of them looks perfectly fine on screen.
01 · The objective of the analysis
The first question is not “what model”, but “what decision comes out of it”.
It is the question any more experienced colleague opens with when you show them a model — and not in order to catch you out. Without an answer there is no way to say what in the model is superfluous. Beginners build models too faithfully: exact geometry, every hole, the whole assembly. It looks conscientious and it is the most expensive way to arrive at a number nobody needed.
| If you want | Then do this |
|---|---|
| A quick “pass / fail” | conservative model with bounding values, one load case, coarse simplifications |
| A number for a report, a customer or an authority | documented assumptions, material from a certificate, validation |
| Design and optimisation | parametric model, coarser mesh, many variants |
| Failure investigation | faithful model including imperfections and real boundary conditions |
A conservative model is a legitimate result, not laziness — provided you can justify why it is conservative in exactly the direction you need.
02 · Geometry idealisation
| If this holds | Then do this |
|---|---|
| Thin-walled, thickness below ~1/20 of the dimension | shell on the mid-surface, thickness as a property |
| Contact, local peaks, changing thickness | solid |
| Frames, stiffeners, bars | beam; shell + beam combined with a connection |
| Detail locally, whole part globally | coarse global model + submodel of the detail |
Solid is the default because it comes straight from CAD and nothing has to be done for it. That does not mean it is more accurate. A solid with one element through the thickness of a plate describes bending worse than a shell, which has the bending formulation built into the element definition. Faithful geometry and faithful physics are not the same thing.
03 · What to remove from the CAD
| If this holds | Then do this |
|---|---|
| Small holes, radii, logos outside the area of interest | remove — they save mesh and time |
| A radius at the location of interest | keep it, that is the stress raiser the whole job is about |
| Symmetric geometry + loading + constraints | use symmetry |
Cleaning up geometry is the first thing worth learning properly, and beginners err in both directions. Either they clean nothing — and then the mesh falls apart on thirty fillets at a place nobody cares about. Or they clean everything, including the radius the whole problem was asking about.
There is a single decision criterion: does that detail lie in the region you are assessing?
Symmetry is the cheapest saving there is, but it only holds when the geometry and the loading and the constraints are all symmetric at once. One eccentric force is enough and symmetry is gone.
⛔ Symmetry in buckling and modal analysis. You lose the non-symmetric shapes — and those tend to be the governing ones. The model will obediently compute a higher critical load than the part actually has.
04 · Material model
| If this holds | Then do this |
|---|---|
| Stresses below the proportional limit | linear elastic |
| Metal above yield | elastic-plastic, bi- or multilinear |
| PMMA, glass, cast iron, ceramics | brittle: maximum principal stress or strain |
| Polymer, long duration, temperature | viscoelasticity, creep, temperature dependence |
| Laminate, timber | orthotropy + a composite failure criterion |
“Structural Steel” from the library is a decision too, just one somebody else made for you. For metals in the elastic range it usually does not matter — the elastic modulus varies little between steels. For polymers it always matters: they change with temperature, with loading rate and with time, and the library value holds for one combination of those.
Where the material data came from belongs in the report. It is one line that takes three seconds to write and without which the analysis cannot be repeated a year later.
05 · Boundary conditions
The most common source of errors in the whole list, and at the same time the one place you cannot see in a colourful picture at all. Boundary conditions are what you assume about the surroundings — and an assumption does not check itself.
| If this holds | Then do this |
|---|---|
| Surroundings markedly stiffer than the part | fixed support — cheapest, but the harshest |
| Surroundings of comparable stiffness | elastic support, equivalent stiffness |
| The interface itself governs | model a piece of the neighbouring part |
| Remote mass or force | remote point, point mass |
A useful question to ask of every model: which part of the assembly is missing here and what did I replace it with? The answer says more about the quality of the model than any stress plot.
Note: Saint-Venant's principle says that the detail of how a load is introduced stops mattering at a distance comparable with the size of the area through which you introduce it. If your location of interest is further away, the simplification need not worry you. If it is right at the fixing, it decides everything.
⛔ A fixed support where the surroundings give. You artificially create a stress peak at the edge of the support and then spend hours treating it as a real problem. It is a singularity, not a finding.
06 · Analysis type by timescale
Let be the period of the first mode and the time over which the load rises. The ratio of the two decides the analysis type, not what the problem looks like.
| If this holds | Then do this |
|---|---|
| static | |
| quasi-static with a dynamic amplification factor | |
| , mild non-linearities | implicit transient |
| Impact, milliseconds, failure, large contacts | explicit dynamics |
| Excitation by frequency or by a spectrum | harmonic analysis / random vibration PSD |
You do not have to guess the period. A modal analysis is cheap — it runs in a fraction of the time of a static analysis and the first natural frequency drops straight out of it. It is worth running before you decide what problem you are actually solving.
For quasi-statics the load is multiplied by a dynamic amplification factor
It is a conservative estimate, not an exact value: the real response to a linearly rising load is lower, and at multiples of the period it falls to unity. The upper bound of 2 is the classical result for a suddenly applied load — twice the static deflection.
07 · When to switch non-linearity on
| If this holds | Then do this |
|---|---|
| Deflection above ~½ the thickness of a thin shell | large deflections — membrane stiffening changes the result by an order of magnitude |
| Bodies touch and can separate | contact |
| Yield strength exceeded | plasticity |
| None of the above holds | linear — but verify it and write down why |
Non-linearity is not a higher grade of analysis, it is a different problem. It is switched on because the assumption of a linear problem has fallen, not because we want a more accurate result.
For large deflections there is a cheap check: run the same problem with them and without them and compare. If the result does not move, you were right and you have something to show for it. If it does move, you have just avoided an error. It costs one extra run.
08 · When stability governs, not strength
| If this holds | Then do this |
|---|---|
| A slender strut in compression | buckling — a hand Euler estimate first, linear buckling for the mode shapes and factors |
| A thin-walled shell or plate in compression or shear | local buckling — linear buckling is only an upper bound; the real value comes from a non-linear run with an imperfection |
| An open section in bending (U, C, I) | lateral-torsional buckling — needs It, Iw and the position of the shear centre |
| A shallow arch or dome under pressure | snap-through — a limit point, which a linear analysis will never find |
| A deflection of the order of the wall thickness | large deflections from section 07, and check whether buckling is hiding behind them |
The critical buckling load depends on stiffness and shape, not on the strength of the material — the ultimate strength does not appear in Euler's formula at all. That is why a structure can be designed to pass the stress check with room to spare and still buckle.
A linear buckling run is cheap and worth doing for the same reason as a modal analysis: it tells you whether the problem is in play at all. Do not trust it as a result, though. For shells it overestimates by a factor of several — an axially compressed cylinder typically carries 20 to 30% of what it predicts, and the difference is governed by manufacturing imperfections.
⛔ Assessing a slender or thin-walled structure on stress alone. Green on von Mises says nothing about stability.
I cover the theory in Buckling and instability and a specific case in Non-linear FEA not converging. The quantities needed for lateral-torsional buckling — It, Iw and the shear centre — come from the section properties calculator; CAD does not give them.
09 · Choice of contact
| If this holds | Then do this |
|---|---|
| Bonded, welded, will not move | bonded — cheapest, linear |
| Touching without separation | no separation |
| Separation yes, friction unknown | frictionless — conservative for sliding |
| Real sliding, μ known | frictional |
| μ uncertain | compute the bounding values μ = 0 and μ = max |
Automatic contact detection in commercial software sets bonded everywhere faces are close to each other. It is the cheapest option and in many models the correct one as well. In an assembly where something is meant to separate or slide, however, it welds the whole assembly together — the model comes out unnaturally stiff and the stresses move elsewhere.
Going through the contact list after automatic detection is five minutes of work and belongs to the routine.
Frictional contact is non-linear and expensive. When you do not know the coefficient of friction, use the bounding values (μ = 0 and an upper estimate) — they say more than one made-up number in the middle.
10 · Mesh
| If this holds | Then do this |
|---|---|
| Always, no exceptions | coarse mesh first, then a convergence study — on stress, not on displacement |
| High stress gradient | local refinement, not across the board |
| Bending, stress peaks | quadratic elements, ≥ 3 elements through the thickness |
| Contact, explicit | linear elements, finer |
The first run is not about accuracy. It is about finding out whether the model holds together at all — whether a constraint is missing, whether the reactions add up, whether anything is floating in space. You want an answer to that in two minutes, not in two hours. A fine mesh is only worth running once the model runs and gives sensible numbers.
Convergence is tracked on the stress at a specific location, not on displacement and not on the global maximum. Displacement converges quickly and almost always, but that is no proof of stress convergence.
⛔ A sharp corner is a singularity — it never converges. The finer the mesh, the higher the stress, indefinitely. Put a radius on it, or assess away from it. It is worked through in detail in the article FEA basics.
11 · Solving and solver convergence
| If this holds | Then do this |
|---|---|
| It will not converge | smaller substeps, line search, reduce contact stiffness |
| Rigid body motion | find the cause in the constraints; weak springs only afterwards |
| Very large model | iterative solver; ill-conditioned → direct |
This is a different convergence from the one in section 10 and the two are often confused. The solver converges when it finds equilibrium in a non-linear problem — that only tells you the run finished. The mesh converges when the value you are tracking stops changing as you refine. An analysis that completed smoothly on a coarse mesh is still a bad result.
⛔ “Fixing” a model so that it runs, without understanding why it did not. Non-convergence is information about the model, not an obstacle between you and the result. Weak springs insert springs that have no business being there, and the problem runs — only now you are computing something other than what you set up.
12 · Assessment criterion
| If this holds | Then do this |
|---|---|
| Ductile metal | von Mises against the yield strength |
| Brittle material | maximum principal stress or strain |
| Repeated loading | amplitudes, S-N, Haigh / Goodman |
| Compressive instability | linear buckling for an estimate, non-linear with an imperfection for the value |
Von Mises stress is the default output and most people take it as a universal stress. It is not — it is derived from a distortion-energy hypothesis that holds for ductile materials. On cast iron, glass or ceramics it assesses the wrong quantity and systematically overlooks tensile failure.
⛔ Von Mises on a brittle material. And contact pressure as a criterion — it is only an input to stress, not a conclusion.
13 · Verification
Without this section everything above is just a picture with colours in it.
| When | What |
|---|---|
| Before the analysis | a hand estimate of the order of magnitude; if the two disagree, one of them is wrong — find out which |
| After the analysis | sum of reactions = applied load |
| Mesh | stress convergence at the location of interest, not globally |
| Uncertain inputs | sensitivity study — a range of values, not one number |
| Critical part | a test |
A hand estimate before the analysis is the habit that does the most work for a beginner. It does not have to be accurate — the order of magnitude is enough. A bending formula on paper exposes an error in the units, in the loading and in the boundary conditions before a day is spent on it.
The sum of reactions is a ten-second check that reveals a surprising amount: when it does not match the applied load, something is holding the model somewhere other than where you think.
How to use this map
Do not memorise the tables. Remember instead that at every one of those thirteen points you decided something, even if you clicked on nothing — a default is a choice too. The difference between an analysis that can be defended and a colourful picture lies in whether you can state that decision and justify it.
Three habits worth adopting straight away:
- Modal analysis before a static run. Cheap, fast, and it tells you whether a static analysis makes any sense at all (section 06).
- A hand estimate before the run. An order-of-magnitude check on paper (section 13).
- A list of simplifications written as you go. Not at the end, when you already cannot remember half of them. It is also the only way to repeat the analysis a year later.
And when two rows in the same table fit at the same time — run both variants and compare. It is the cheapest form of learning there is in this field, and the difference between them will teach you more than any manual.
If you are working on a specific problem and cannot decide between two routes, write to me — ales@makewise.sk. An answer to a question like “is a static analysis enough here, or does it have to be dynamic?” tends to take a few lines and a few minutes of work.
v1 · September 2026. The map is a summary of practice, not a standard — for certification analyses the relevant code and the agreed methodology always govern.
Frequently asked questions
- When should I use shell elements and when solids?
- A shell on the mid-surface wherever the thickness is an order of magnitude smaller than the other dimensions, roughly below one twentieth. Solids where there is contact, a local stress peak, or a changing thickness. A solid generated straight from CAD with one element through the thickness is worse in bending than a shell, not better.
- How do I tell whether a static analysis is enough?
- Compare the rise time of the load with the period of the first mode. You get that from a modal analysis, which is cheap and can be run before the analysis proper. If the rise time is at least ten times longer than the period, a static analysis is enough.
- What is the difference between solver convergence and mesh convergence?
- They are two independent things. The solver converges when it finds equilibrium in a non-linear problem — that only tells you the run finished. The mesh converges when the value you are tracking stops changing as you refine. A run that completed on a coarse mesh is still a bad result.
- When does strength have to be checked for stability as well?
- Whenever the structure is slender or thin-walled and something in it is in compression — a slender strut in compression, a plate or shell in compression or shear, an open section in bending, a shallow arch. The critical buckling load depends on stiffness and shape, not on the strength of the material, so a stress check will not catch it: the part passes on von Mises and buckles anyway.
- Why is bonded contact a dangerous default?
- It welds the whole assembly together. Nothing can separate or slide, the model comes out markedly stiffer than reality and the stresses move elsewhere. Automatic contact detection sets it everywhere faces are close to each other — including places where the parts in reality only touch.