I've checked your ML_AB file. You made a mistake when putting the files together:
After the 145th structure (line 30301) the separator line to the next structure is missing:
instead of this:
--------------------------------------------------
XY YZ ZX
--------------------------------------------------
-6.412172937574392E-002 0.116719608405456 -0.113494542001003
Configuration num. 146
==================================================
System name
--------------------------------------------------
C O H
it should look like this:
--------------------------------------------------
XY YZ ZX
--------------------------------------------------
-6.412172937574392E-002 0.116719608405456 -0.113494542001003
**************************************************
Configuration num. 146
==================================================
System name
--------------------------------------------------
C O H
Also some remarks to your calculation:
1) The first 105 structures are containing only two atoms of the "perfect" solid. While I understand to motivation to also add these kind of structures I consider it obsolete in practice. I've run reselection of new local reference configurations on these structures and the sparsification algorithm throws out all atoms as candidates for the local reference configurations, besides the first one. This means that all other atoms give totally redundant information. I guess these structures were sampled from an MD, but and MD on such a small cell will only lead to tiny vibrations around the atom positions.
2) The two-atom structures have a smaller lattice constant than the cut-off radii employed. Hence each atom will have unphysical interactions to it's replica in periodic boundary conditions. It is generally never good to train with a larger cut-off radius than your lattice constants. So this is another reason to throw out the two-atom training structures.
3) You need to reselect the local reference configurations (run with ML_MODE="SELECT") before refitting the force field. Please see the following wiki pages for that:
wiki/index.php/ML_MODE
wiki/index.php/Best_practices_for_machi ... rce_fields
Mind: Be careful by setting ML_MCONF_NEW to larger values since you have 180 atoms for the max number atoms of a given type. 180* ML_MCONF_NEW * 4 atom types can give a large overhead in memory if ML_MCONF_NEW is chosen too large. Maybe set it your case ML_MCONF_NEW to 4 or even 2, but this will result in much longer calculation times.