Use a case that must be symmetric to measure the noise floor of a CFD comparison metric

Context: comparing injector layouts with short unsteady VOF runs. One metric was the share of one fluid on each side of a channel (inner half against outer half), from time-mean fluxes over a 0.1 ms window.

Pitfall: a mirror-symmetric layout must give 0.50 by symmetry. It did on the coarse grid, but gave 0.65 on a finer grid with the same averaging window. The liquid sheet flaps, and 0.1 ms holds only a few flaps, so the time mean had not converged.

Practice that worked: - Before trusting differences between layouts, run at least one case whose answer is known by symmetry, and read its deviation (here about +-0.15) as the noise floor of that metric. - Report only differences well above it. Here an asymmetric layout gave 0.14 and a strongly asymmetric one 0.02, which is far outside the noise, so the conclusion stood. - Prefer metrics that converge fast (here, flux-weighted mixing efficiency averaged over several planes changed only a few per cent between 0.1 and 0.2 ms windows) for the ranking. Present slow-converging ones as about, with the noise floor stated. - If compute allows, lengthen the window, or average the two halves of a known-symmetric case to remove sampling bias.

Side note on cloud capacity (Google Cloud): the binding limit for a new project is often CPUS_ALL_REGIONS (12) rather than the regional family quotas. With PREEMPTIBLE_CPUS at 0, Spot VMs count against the same CPU quota, so they do not add capacity.