Mesh independence, and knowing when to stop

The check that tells you refinement has stopped changing the answer.

Mesh independence is the check that tells an engineer refinement has stopped changing the answer, run by comparing results from progressively finer meshes of the same geometry until the number of interest settles down and further refinement no longer moves it by anything that matters.

Tasting the soup after each pinch

Seasoning a pot of soup by taste works on the same principle as a mesh independence check. The first few pinches of salt clearly change the flavour on the spoon. Continue and each additional pinch makes a smaller difference, until eventually another pinch stops changing the taste by anything a tongue can detect. At that point the soup has reached the resolution of taste itself, and any further precision in seasoning could not be perceived even if it existed.

A mesh study reaches the same kind of point. Each refinement is a pinch, tasted by comparing the new result against the last one, and once another pinch stops changing the number of interest, further refinement is seasoning a soup nobody can taste any difference in, with a much longer wait between attempts. The tasting cannot tell a cook whether salt was the right seasoning to reach for, though. A soup can be demonstrably well salted and still be missing pepper, or missing time on the stove, and a mesh study that has converged can still be answering a question posed with the wrong boundary condition or the wrong physics model.

Why one fine mesh proves nothing on its own

A finer mesh generally gets closer to the true physics, with diminishing returns, so running the same study on several mesh densities shows where those diminishing returns set in. A coarse and a medium mesh often disagree widely, a medium and a fine mesh by less, and a fine and a still finer one by less again, until the gap between successive refinements drops below the precision the engineering decision needs. A peak stress that reads 160 on a coarse mesh, 185 on a medium one, 193 on a fine one and 195 on the finest has moved by about 1 percent in its last step, against 15 percent in its first, and that shrinking gap is the signal that the mesh has become independent of further refinement for this particular question.

Building the finest mesh time allows and trusting it is a different exercise, because a single mesh, however fine, offers no evidence about whether it has settled or is still drifting, and only a sequence of meshes compared against each other can show that. The gap matters most when a single result looks plausible, since a plausible number from an unconverged mesh prompts nobody to go back and check it.

The result does not always settle smoothly either. Some quantities overshoot and undershoot the eventual value from one mesh to the next before settling, which is why the check needs at least three points on the curve. Two points alone cannot distinguish steady convergence from the early swing of an oscillation that has yet to damp out.

Stopping at a percent or two

A common working rule is to keep refining until the result of interest changes by only a percent or two between one mesh and the next finer one, and to treat that as evidence the mesh is no longer the dominant source of error. Chasing a smaller change is rarely worth the computing time, since the remaining uncertainty in most engineering questions, from material variation, load uncertainty or manufacturing tolerance, is usually larger than anything further refinement could remove. Matching the mesh's precision to the precision the rest of the engineering supports keeps effort where it can change a decision, and away from a decimal place nobody downstream will use.

Teams unwilling to run a full sequence can extrapolate from two or three results, using how quickly the answer has been changing between them, to estimate what an infinitely fine mesh would report. That is useful for gauging how far a current result still has to move, but it remains an estimate built from limited data, and running the finer mesh is the only way to check whether the extrapolation was right.

The overnight run that saves a week

A mesh independence study is one of the few checks in the process that produces direct evidence, because it compares the simulation against itself across a range of resolutions. Skipping it is a common shortcut under deadline pressure and one of the more expensive ones, since a result from an unconverged mesh carries an unquantified error that stays hidden until something downstream, a test, a failure, a second opinion, forces the question back open. Two or three meshes run overnight while other work continues cost less than discovering, weeks after a decision was made on a single unconverged number, that a finer mesh would have told a different story.