Hello,
We are trying to perform a cluster expansion manually, based on clusters in a published article. We are generating the clusters manually in clusters.out, using clusterexpand then lsfit to look at the predicted and true energies:
clusterexpand -e -cv energy
lsfit -x=allcorr.out -y=allenergy.out -ty -p -e > fit_manual.out
This procedure works very well (low CV score, good fit between predicted and true energies) for all 1-body and 2-body clusters. However, inclusion of the 3-body terms is causing some problems, with a reasonably low CV score but poor fit of the energies. The pattern of the poor fit is very characteristic, where the points form a type of "grid pattern" with many of the structures of similar true energy splitting out into groups of rather different predicted energies and many structures with different true energies, showing very similar predicted energies. I’ve attached an image of this to hopefully explain this.
We have tried to identify any "problem" clusters but this pattern appears for certain groups of 3-body terms but not for other groupings (apologies, again this is hard to explain) so we don’t think it is to do with one specific cluster but somehow a combination of them.
Any insight anyone has about this would be greatly appreciated.
Anna