Machine learning for a melt-quench process

[paulfons]

I am looking at melt-quenched chalcogenide amorphous structures and I am hoping to use machine learning to sample more structures (quench rate dependence) than I would otherwise be able to achieve all first principle calculations. In the melt-quench process (think phase-change), the allow is quenched from the liquid to the amorphous state. Thus I would like to use the machine learned potential to investigate two degrees of freedom, namely cooling rate and the effect of system size (more atoms).

The system I am studying is Ge-Hf-Te. For the calculations, I started with the binary Ge-Te in a 1:6 ration. As there does not crystallize in a single phase, I created the cell by using the structure of GeTe substituting Te for Ge to ensure the correct composition. I then adjusted the density to match experiment.

I carried out the machine learning process in several stages using 64 atoms in total with a total of 8 kpoints to improve accuracy. I am using Vasp 6.4.3.

600K-800K in 10,000 steps
refit
800K-1000K in 10,000 steps
1000K-1400K in 10,000 steps
1400K-1800K in 10,000 steps

This resulted in the following number of reference sites

# LCONF ###############################################################
# LCONF nstep el nlrc_old nlrc_new el nlrc_old nlrc_new
# LCONF 2 3 4 5 6 7 8
# LCONF ###############################################################
LCONF 567 Ge 3332 3361 Te 5340 5383

I then substituted about 10% Hf onto the Te sites and continued the machine learning process.

400K in 10,000 steps
refit
400-600K in 10,000 steps
refit
1000K-1200K in 10,000 steps

This resulted in the following number of reference sites

# LCONF ###############################################################
# LCONF nstep el nlrc_old nlrc_new el nlrc_old nlrc_new
# LCONF 2 3 4 5 6 7 8
# LCONF ###############################################################
LCONF 3 Ge 4406 4415 Te 8329 8377 Hf 1716 1722

I am curious to get some feedback on how reasonable this process is and what I should change, if anything.

[alex]

Hello paulfons,

I wouldn't consider myself an expert in ML, however I have two points you may scratch your head about:

a)

I then adjusted the density to match experiment.

This gives me some headache. I'd prefer a density, which fits to your computational setup.

b) do you really need 10k training steps for 200K steps in T? Considering your variety in stochiometry you might be better off with far less steps (1000 or less). Maybe do even do the whole range in one go?!

Best regards,

alex

[paulfons]

Thank you for your reply. The density issue is not that important as I was using a NPT ensemble for the training. After reading the comment in the Vasp wiki "liquids often require 2000-4000 local reference configurations, while 500-1000 reference configurations might be sufficient for simple periodic volume systems." I thought a large number of configurations were necessary to describe the liquid phase potential -- although I can see that I probable went too far. I will try the training again, but if the above statement about needing 2000-4000 reference configurations is necessary, I wouldn't think you suggestion of 1000 steps wouldn't provide a sufficient number of reference configurations. Thanks for any further advice.

[alex]

OK, I see your point, I messed that one.

Good luck!

alex

[ferenc_karsai]

Also try to not have to large discrepencies between species. To our experience in most calculations the species with the lowest number of local reference configurations determines more or less the accuracy. That means for example in H2O if you have 1000 Oxygens the accuracy will not change much if you have 2000 or 10000 Hydrogens.
In your case it means try to cap the maximum number of local reference configurations around 3000-4000 because you have anyway only 1722 Hf local reference configurations. For that simply set that number for ML_MB in the INCAR file.

[paulfons]

Dear Ferenc,

Thank you for your note. I wanted to follow up on your comment about trying to avoid too large a discrepancy between species. In my case, I am working on the ternary system Ge-Hf-Te. The composition of the system is 10% Hf, 15% Ge and 76% Te . As the concentration of Te is much larger, this apparently leads to a larger number of Te reference atoms, followed by Ge, and the Hf.

As I wish to investigate different cooling rates for the melt quench process, I started a training session with 5000 steps (of which about 1300 have completed) and it has lead to the following number of configurations at present. Note that I have stopped this process and have restarted it with the ML_MB keyword as you recommended.

# LCONF ######################################################################################
# LCONF This line shows the number of local configurations
# LCONF which were sampled from ab initio reference calculations.
# LCONF
# LCONF nstep ...... MD time step or input structure counter
# LCONF el ......... Element symbol
# LCONF nlrc_old ... Previous number of local reference configurations for this element
# LCONF nlrc_new ... Current number of local reference configurations for this element
# LCONF ######################################################################################
# LCONF nstep el nlrc_old nlrc_new el nlrc_old nlrc_new el nlrc_old nlrc_new
# LCONF 2 3 4 5 6 7 8 9 10 11
# LCONF ######################################################################################

LCONF 1252 Ge 3721 3766 Te 6992 7083 Hf 1264 1294

For reference, I have included the input files for the current run for reference. Note I wish to study metastable melt-quench structures going from the melting point (>1000 C) down to approximately room temperature. To accomplish the training I am ramping the system temperature from 400-1500K. Any suggestions as to how to further improve the INCAR settings, would be grateful to receive them.

Best wishes,
Paul Fons