Why a finer mesh is not always a better mesh

Where refinement helps, and where it only costs time.

A finer mesh is not always a better mesh because refinement only fixes errors caused by cells being too coarse to capture how a quantity changes across space, and a large share of simulation error comes from other sources entirely that adding more, smaller cells does nothing to touch.

The one error refinement can reach

Refining a mesh means replacing each cell with several smaller ones, giving the solver more points at which to evaluate how a quantity like pressure or temperature is changing. Where that quantity varies sharply across a small region, a coarse mesh can badly misrepresent it, drawing a straight line across a curve, and refinement there closes a real gap between the model and the physics. Keep refining and that particular error keeps shrinking, at the cost of more cells to update on every iteration.

That is the only error refinement touches. A wrong turbulence model, a mistaken boundary condition, or a material property borrowed from the wrong specification sheet has nothing to do with how many cells are present, and once the mesh is fine enough to represent the true variation in the field, the remaining gap between simulation and reality sits somewhere refinement cannot reach.

Eight times the cells for a quarter of the error

Refinement's benefit flattens while its cost climbs. With the second-order schemes most solvers use by default, halving the cell size cuts the error from coarseness to roughly a quarter of what it was, but in three dimensions the same halving multiplies the cell count about eightfold, each extra cell updated on every iteration. The first pass on a coarse mesh removes three quarters of that error, while the fourth or fifth pass, costing eight times as much as the one before, is chipping away at an error already too small to matter. A study that could have finished overnight on a sensibly sized mesh can stretch into several days on one refined past the point of usefulness, waiting on a number that had already stopped moving.

The practical test is how much a finer mesh shifts the answer, since it almost always shifts it a little. Once doubling the cell count again moves a result by too little to change the decision being made, the mesh has done its job. Deciding a design is safe or unsafe rarely depends on the fourth significant figure of a stress value, and chasing that figure solves a problem of curiosity. An engineer waiting on an over-refined study also has that much less of the week left to explore other variants of the design, often the better use of the same time.

Mowing a lawn until it is scalped

Mowing a lawn shorter fixes an uneven cut only up to a point. A shaggy, unevenly grown lawn looks tidier as the blade height comes down, but a lawn with bald patches, weeds or dead grass is not improved by cutting it shorter still. Below a certain height the mower starts scalping, exposing bare soil and stressing the healthy grass, so the same action that helped earlier now makes the lawn worse.

Mesh refinement follows the same curve. Early refinement is the even trim of a shaggy lawn, and refinement past the point where the physics is already resolved stops helping and starts costing, in wasted computing time at best. At worst, a mesh refined aggressively in one place produces cells so small relative to their neighbours that the solver struggles to keep the iteration stable, trading a resolution problem for a convergence problem that did not exist before the extra cutting began.

Putting the cells where the lawn is shaggy

Refinement still earns its place in the regions where a coarse mesh cannot represent what is happening: sharp corners, thin gaps, areas of fast-changing flow. Where finer stops meaning better depends on where the physics changes quickly, which is why uniform refinement everywhere is so often the wrong choice. The efficient mesh puts extra cells where the lawn is shaggy and leaves the even parts alone, usually finishing faster and giving a more trustworthy answer because it spends the computing budget where the physics demands it.