How can grain boundaries be sampled using AIMD？

[qingyu_wang]

Hello,
I am currently training a machine-learning interatomic potential, and the present training errors are:
grep '^ERR' ML_LOGFILE | awk '{print $1,$2,$3,$4,$5}'
ERR 0 4.75870264E−03 1.87127251E−01 2.09885476E+00
I find that the energy error, 1.87127251E−01 eV/atom, is relatively large. I suspect that this is mainly due to an insufficient number of AIMD sampling structures. My current sampling strategy is to introduce Ag atoms into SiC supercells, but this approach is likely inadequate for capturing grain-boundary environments. May I ask what methods are available to further improve the sampling of grain boundaries, for example through surface sampling or other relevant configurations?

[fabien_tran1]

Hi,

Could you please provide the input files as mentioned in the posting guidelines (https://www.vasp.at/forum/viewtopic.php?t=17928)?

[qingyu_wang]

Hello,
https://1drv.ms/f/c/b8ec5e7d661b03f9/Ig ... M?e=VsNYKp
This directory contains the INCAR file used for AIMD training, the POSCAR files with various doping concentrations, the corresponding ML_AB datasets, and the final ML_FF files generated after the refit step, as well as the ML_LOGFILE.

[jonathan_lahnsteiner2]

Dear Qingyu Wang,

I was checking your output files. In the ML_LOGFILE I recognized ML_RCUT1 and ML_RCUT2 are set to rather small values. Additionally ML_RCUT1 is smaller than ML_RCUT2. Usually ML_RCUT2 should be chosen smaller than ML_RCUT1. I would recommend to use the default ML_RCUT1 and ML_RCUT2 and try the refit step again. This should strongly improve your fitting errors.

All the best Jonathan

[qingyu_wang]

Dear Mr. Lahnsteiner,
Following your suggestion, I carried out the refitting procedure; however, the results did not exhibit significant improvement. I would like to inquire whether it might be possible for you to further train additional structures to enrich my potential and help reduce the remaining errors.
grep '^ERR' ML_LOGFILE | awk '{print $1,$2,$3,$4,$5}'
ERR 0 4.54506989E-03 1.84518361E-01 2.02963293E+00
Sincerely,

[jonathan_lahnsteiner2]

Dear Qingyu Wang,

Upload the new ML_LOGFILE otherwise I can not further analyze your issue.

All the best Jonathan

[qingyu_wang]

Dear Mr. Lahnsteiner,
Thank you for your reply! And below is my ML_LOGFILE.
Sincerely.

[jonathan_lahnsteiner2]

Dear Qingyu Wang,

I was checking your calculations again. I was looking at your ML_AB files you supplied. You mention that you are doing the training in stages for different concentrations and then concatenate the ML_AB files. When checking your ML_AB-refit. It is very suspicious that you have the same number of local reference configurations for every atomic species. This seems that you created the selection of local reference configurations by hand or with some algorithm not implemented in vasp. Therefore my recommendation would be to run a ML_MODE=select. This will select a new set of local reference configurations from the supplied ML_AB file. Additionally I would set ML_MB = 3000 during the select calculation. Then again do a refit with the newly obtained ML_AB file and keep the default cutoff radii for 2 and 3 body descriptors.
I hope this is of help, otherwise contact us again.

All the Best Jonathan