Get a better fitting and lower CV score by excluding str.

Dear Prof. Axel van de Walle and other ATAT users,

I am a new learner of ATAT, and I am focusing on the system of hcpTi with the addition of interstitial oxygen in octahedral sites of hcpTi which owns 2 oct. sites totally in one unit cell. After I used the command ‘maps -d’ to establish the cluster expansion, I totally got 141 structures, and the ‘maps.log’ and ‘gs.out’ contains:
Among structures of known energy, true and predicted ground states agree
No other ground states of 32 atoms/unit cell or less exist.
Concentration range used for ground state checking: [0,1].
Crossvalidation score: 0.0119978’

0.000000 -0.000000 0.000294 0
0.166667 -0.148776 -0.146545 39
0.333333 -0.248734 -0.248522 41
0.500000 -0.315864 -0.314769 221
0.666667 -0.220127 -0.220467 47
1.000000 0.000000 -0.010312 1

However, there are several questions I am confused:

  1. In reality, only 33 at% O can dissolve into the hcpTi, that means only one oct. site can be occupied by O in hcpTi. The structure 221 above has the O concentration of 33 at%. Thus, the ground states of 47, and 1 (two oct. sites are both occupied by O) above are inconsistent with the reality. My supervisor advised me to delete some unreal structures like structure 47 and 1, so that we can get a better fitting (a lower CV score) and a real cluster expansion.
    My question is how to look for the useless structures for us we are able to delete. I am not sure whether we can delete the structures far away from the convexhull, which have formation energies very different from those structures in convex hull. And whether do I need to delete all the structures which have concentration of O > 33 at%. Besides, if we do the two above, whether it is also a good and reasonable way to lower CV score. Our final purpose is to find the O solubility in hcpTi at finite temperature by doing MC simulations later, does this means I can ingore these ordered ground states in gs.out? Because I focus on the phase transition, however these ground states still have hcp crystal structure.

  2. If I do not delete any structures, whether I can do this by setting a lower weight for those structures I need to exclude, namely setting various weights in ‘weight.in’ file. Does the ‘maps’ work for this and read the ‘weight.in’ file automatically? Do not I need to specify any tags after typing ‘maps’? I tried to set lower weight for those structures and then use ‘corrdump’ and ‘cv’ command to do this, but there is no output file containing the updated CV score. Do I need to calculate the CV score manually according to the ‘fit.out’?

  3. When I used the calculation equation of CV score on the ATAT website to verify the value of CV score in the ‘maps.log’, it shows two different value: 0.00696 by manually calculating, 0.011998 in maps.log. Do the Ei and E(i) represent the Evasp and Efit in the fit.out, respectively (i.e. the formation energy)? If I use total energy in ‘the energy’ file in each structure folder instead to calculated CV score, I also got a value different from that in ‘maps.log’.

I am very much appreciated it if someone could give me some advice. If any other files or information need, please let me know.

Many thanks to everyone here.

Kindest regards,
Ning Zhang