Why meshing is where most wrong answers come from

Why dividing up the geometry matters more than the physics settings.

Meshing is where most wrong answers come from because it decides, before the solver has computed a single value, exactly what information the physics equations will actually have to work with, and no amount of correctness in the physics settings can repair a mesh that has already lost the information the answer depended on.

A study I ran once sat computing for six days and returned an answer that was wrong in a way nobody in the room could see just by looking at the coloured plot, and the cause, once finally traced back, turned out to be nothing more exotic than the shape of the cells the geometry had been cut into.

Choices made before the physics is switched on

Meshing is the step where a continuous shape is cut into the finite collection of small cells a solver can compute on. Every choice made during that cut, how many cells to use, how they are shaped, where they cluster and where they thin out, happens before any physics is switched on, and every one of those choices survives, invisibly, into the final answer. A physics setting that is slightly wrong tends to announce itself, with a result that looks implausible or a flow that clearly should not be doing what it is doing. A slightly wrong mesh stays quiet, because a badly shaped cell still produces a number, a coloured patch on the final plot, that looks as confident and as smooth as the correct cells beside it.

Turbulence models, material properties and boundary conditions are all visible choices, sitting in dialogue boxes with names attached, easy to review and easy to question. Mesh quality lives inside a file of coordinates most engineers never look at directly, inspected only through summary statistics, and those averages are part of the danger. A report showing that 99 percent of a million-cell mesh is well shaped still leaves ten thousand poor cells unaccounted for, and says nothing about where they sit, and it is routinely a few of them, sitting at a tight fillet or a sharp internal corner, that decide whether the stress or flow reported at the location an engineer cares about can be trusted. A mesh can pass every average quality check a piece of software offers and still be wrong in the one place the study was commissioned to look at.

Experienced practitioners learn to check tight fillets, thin walls and awkward transitions between feature sizes first, before opening any quality report, because those are the features most likely to have forced a mesh generator into a compromise nobody asked for, squeezing cells into slivers or stretching them long and thin to bridge a change in scale.

Wallpaper cut unevenly at the table

Hanging wallpaper in a hurry and cutting each vertical strip to slightly different widths produces a wall that looks fine from a distance and wrong up close, however carefully every strip is then pasted and smoothed into place. The problem was decided at the cutting table, before the brush ever touched a wall, and by the time the mismatch at the seams becomes visible it is built into every strip hung after the first uneven one.

In a mesh, the pasting is the solving: careful, technically sound, and unable to correct a mismatch introduced upstream of it. A cell cut too thin here, too stretched there, or clustered densely in the wrong place carries that flaw into every iteration that touches it, and the solver has no more ability to notice a badly cut cell than a paste brush has to notice a badly cut strip of paper.

The analogy also says which flaws matter. A strip cut unevenly but hung behind a bookshelf, where nobody will stand close enough to notice, does no harm, and a badly shaped cell somewhere the engineering question never looks can likewise be ignored, while an identical flaw at the location the study was built to examine cannot.

Hours on the mesh, minutes on the physics

When a simulation disagrees with a physical test and the discrepancy is finally traced, the cause is often a slightly deflating one. It rarely involves anything as satisfying as a wrong equation or a mistyped boundary value, and turns out instead to be a cluster of stretched, oddly shaped cells sitting quietly in a fillet nobody inspected closely, having decided the answer weeks before anyone thought to ask about them.

Time spent reviewing a mesh before a solve is therefore a large part of the real engineering. A study that spends an hour on meshing and ten minutes on physics settings, reviewed by someone other than the person who built it, is far more likely to produce a trustworthy answer than the reverse. The six days of computing behind that wrong answer went, patiently and thoroughly, into cells that had already decided what the answer would be before the solver started counting.