The statistical error of structural analysis depends on the number of environments found in the structure (not just the sample size). We should find a way to indicate this and warn users if the number of environments for a given element is too small.
I notice in particular in the structural analysis that while g(r) and CN(r) may have reasonable converged (and there is little concerning "wiggling" of their plots), the distribution of coordination numbers for a given r may converge significantly more slowly between samples (particularly if g(r) is substantially > 0 at its first minimum).
Example: 10k atoms with ~50 environments per sample. Seeing 15% deviation between some coordination numbers in histograms from sample to sample
The statistical error of structural analysis depends on the number of environments found in the structure (not just the sample size). We should find a way to indicate this and warn users if the number of environments for a given element is too small.
I notice in particular in the structural analysis that while g(r) and CN(r) may have reasonable converged (and there is little concerning "wiggling" of their plots), the distribution of coordination numbers for a given r may converge significantly more slowly between samples (particularly if g(r) is substantially > 0 at its first minimum).
Example: 10k atoms with ~50 environments per sample. Seeing 15% deviation between some coordination numbers in histograms from sample to sample