I’m interested in using ATAT to perform a CE on H-terminated ZnO surfaces in order to identify the stable ordered and disordered H-configurations at different concentrations of H. For this, I’m planing to use the ground state energies of ~50 configurations, which will be calculated with the FHI-aims code (https://aimsclub.fhi-berlin.mpg.de/).
As I don’t have a wrap that will allow me to proceed with a fully automated process, I’d like to use ATAT as a separate tool that will allow me to (a) generate the clusters based on a "lat.in" file via corrdump -clus -2=pairlength -3=tripletlength -4=… and (b) perform the CE, i.e. generate the convex hull for the know and predicted gs. I understand that for the latter, I need to manually insert the FHI-aims calculated energies in "energy" and change the geometry files into the appropriate "str.out" format within each directory.
My questions are the following:
Will the cross-validation score still be presented in "log.out", although the CE will not be constructed based on maps?
Can I still use mapsrep to generate the four plots demonstrated in the user guide?
wait for the code to spit out a directory. Go into it, run your code manually. write the energy into the energy file.
maps will then update its cluster expansion
repeat
Then, log.out is valid and you can use mapsrep.
If you want to have full control on the clusters and the structures you use,
then you need to create all structure directories with str.out and energy.out in them.
Generate your clusters with corrdump
Count each type of clusters (pair, triplet, etc.) and place these numbers
in a nbclusters.in file
run maps
Thanks for the clarification. Let me share another thought based on my recent DFT energy outputs as a follow up on this thread: FHIaims uses numeric atom-centered orbitals instead of plane waves, which (among others) also reflects on the absolute energy values. Therefore, in contrast to VASP where the energy of a H-terminated ZnO slab is in the order of some eVs, FHIaims predicts it in the order of a few thousands of eVs. The same is also true for the energy/atom. Since the energy file in each directory is responsible for the CV score, I wonder how I can circumvent the issue of these thousands of eVs that will interfere with my CV score and not allow it to reach < 0.025.
Normalizing everything wrt the energy of the the pristine slab, results in differences in the order of tenths of eVs, so the CV score is still nowhere close to the anticipated low value. Would the energy of formation of the H-terminated slabs be a meaningful way to circumvent this issue, i.e. instead of the energy/atom use the energy of formation/atom in the energy file?
Having large energies should not be a problem. maps internally converts energies into formation energies (relative to the end-member energies). These large shifts should cancel before the fit is performed.
If you have a serious fitting problem, look for points that are dramatic outliers - perhaps these ab initio runs have a problem…