Reasonable NPT constraints in ML

[paulfons]

I would like to melt quench some poor glass formers (phase change materials) in the computer and look at the electronic structure (among other qualities). I have successfully done this using NVT ensembles with MD and have found some interesting results. In a test, I melt-quenched MoS2 and then recrystallized it successfully using all ab-initio calculations. The amount of time and size limitations of these calculations are strong constraints. To get around this, I plan to use the ML feature of Vasp to "learn" the potential to allow me to vary both the size of the cluster as well as to carry out further investigations of the effects of quench rate. To train the system, I have been carrying out training using 10,000 steps for various 400K, 1000K, 1500K temperatures. For the training, I used a hexagonal cell with 64 atoms. I note that when I exceeded the melting point for higher temperatures, the cell shape became highly distorted and I realized, it was necessary to constrain the system. Initially I was using ICONST just to monitor for the system, but for temperatures approaching the melting point, I fixed all of the angles of the cell and two of the bond-lengths and only left the "c" direction as free to vary. Is this a reasonable way forward or should I use fewer constraints, e.g. for example just fixing the a and b axes while letting the c axis vary along with the cell angles. Do you have a recommendation for this (or other keywords in my INCAR file). The results (up to melting seem very reasonable) so I am looking forward to trying the same idea on other systems.

LR 1 7
LR 2 7
LR 3 7
LA 2 3 7
LA 1 3 7
LA 1 2 7
LV 7
LR 1 0
LR 2 0
LA 1 2 0
LA 2 3 0
LA 1 3 0

[paulfons]

I thought I would add an additional fact that I forgot to include in the original. For each temperature from 400,1000, 1500, 1800, 2200, 2400K I included a temperature ramp of about 400 degrees and 10,000. I included 8 kpoints to improve accuracy as well.

[alexey.tal]

Hi,

We have some tips in the best-practice guide for machine-learning calculations here.
If you have an orthorhombic system you can use LATTICE_CONSTRAINTS = F F T to allow the change in one of the directions.

I fixed all of the angles of the cell and two of the bond-lengths and only left the "c" direction as free to vary

I am not sure about the bond-lengths constraints. I don't see this in your ICONST, but I'm not even sure it is required. I would say that fixing the angels and the first two vectors should be enough. Maybe you can explain why you think that constraining angles and the first two vector is not be enough and you need to constrain the bonds too?