I am trying to simulate solid/liquid interfaces of water with different solids like Zn or graphite. I have trained the force field for water, and now want to continue training the solid phases.
For graphite, this works well, I have trained the bulk and the surface with 96 C atoms just following the procedure described on the vasp wiki.
The ML_ABN file has this header (the whole file is too big to upload):
Code: Select all
1.0 Version
**************************************************
The number of configurations
--------------------------------------------------
1313
**************************************************
The maximum number of atom type
--------------------------------------------------
3
**************************************************
The atom types in the data file
--------------------------------------------------
H O C
**************************************************
The maximum number of atoms per system
--------------------------------------------------
192
**************************************************
The maximum number of atoms per atom type
--------------------------------------------------
128
**************************************************
Reference atomic energy (eV)
--------------------------------------------------
0.000000000000000E+000 0.000000000000000E+000 0.000000000000000E+000
**************************************************
Atomic mass
--------------------------------------------------
8.00000000000000 16.0000000000000 12.0110000000000
**************************************************
The numbers of basis sets per atom type
--------------------------------------------------
4713 2258 196
**************************************************
Basis set for H
--------------------------------------------------
1 1
1 2
1 3Now, with Zn I have problems. I cannot get the continuation training to run with a Zn, despite using exactly the same ML_ABN as in the graphite case with (almost) the same settings. I adjusted ISMEAR, but ENCUT, PREC and so on should be identical.
When I remove the ML_AB with the stored H and O positions, the training proceeds, when I add the file I always get stuck at the initialization:
Code: Select all
Loading vasp6/6.4.3-intel-2021.9.0-ayhsd2e
Loading requirement: intel-oneapi-mpi/2021.9.0-intel-2021.9.0-35bd42h
pkgconf/1.8.0-intel-2021.9.0-hgzow6c hdf5/1.12.2-intel-2021.9.0-tiadmsp
intel-oneapi-tbb/2021.9.0-intel-2021.9.0-xyehvoh
intel-oneapi-mkl/2023.1.0-intel-2021.9.0-7aranik
running 128 mpi-ranks, on 1 nodes
distrk: each k-point on 128 cores, 1 groups
distr: one band on 64 cores, 2 groups
vasp.6.4.3 19Mar24 (build Jul 17 2024 16:12:06) complex
POSCAR found type information on POSCAR Zn
POSCAR found : 1 types and 64 ions
Reading from existing POTCAR
scaLAPACK will be used
-----------------------------------------------------------------------------
| |
| W W AA RRRRR N N II N N GGGG !!! |
| W W A A R R NN N II NN N G G !!! |
| W W A A R R N N N II N N N G !!! |
| W WW W AAAAAA RRRRR N N N II N N N G GGG ! |
| WW WW A A R R N NN II N NN G G |
| W W A A R R N N II N N GGGG !!! |
| |
| For optimal performance we recommend to set |
| NCORE = 2 up to number-of-cores-per-socket |
| NCORE specifies how many cores store one orbital (NPAR=cpu/NCORE). |
| This setting can greatly improve the performance of VASP for DFT. |
| The default, NCORE=1 might be grossly inefficient on modern |
| multi-core architectures or massively parallel machines. Do your |
| own testing! More info at https://www.vasp.at/wiki/index.php/NCORE |
| Unfortunately you need to use the default for GW and RPA |
| calculations (for HF NCORE is supported but not extensively tested |
| yet). |
| |
-----------------------------------------------------------------------------
Reading from existing POTCAR
-----------------------------------------------------------------------------
| |
| ----> ADVICE to this user running VASP <---- |
| |
| You enforced a specific xc type in the INCAR file but a different |
| type was found in the POTCAR file. |
| I HOPE YOU KNOW WHAT YOU ARE DOING! |
| |
-----------------------------------------------------------------------------
LDA part: xc-table for (Slater+PW92), standard interpolation
POSCAR found type information on POSCAR Zn
POSCAR found : 1 types and 64 ions
Machine learning selected
Setting communicators for machine learning
Initializing machine learningThe memory used by vasp for Zn without the continuation training is 135087 kBytes, and I tested this on nodes with up to 2048 GB of memory. Meanwhile the graphite runs fine on a node with 384 GB memory, according to OUTCAR the calculation uses 256992 kBytes. So the issue seems to be the memory of the already stored basis set, but at the same time it is no problem in the graphite case.
Here are the input and output files for Zn without the ML_AB file present:
Please note that the high cut-off energy of 700 eV is due to the preceded training of water, and lowering this to 400 eV did not help at all. When I do continuation training with Zn, the calculation gets stuck every time.
This is the header of the ML_AB file with I use for Zn:
Code: Select all
1.0 Version
**************************************************
The number of configurations
--------------------------------------------------
1123
**************************************************
The maximum number of atom type
--------------------------------------------------
2
**************************************************
The atom types in the data file
--------------------------------------------------
H O
**************************************************
The maximum number of atoms per system
--------------------------------------------------
192
**************************************************
The maximum number of atoms per atom type
--------------------------------------------------
128
**************************************************
Reference atomic energy (eV)
--------------------------------------------------
0.0000000000000000 0.0000000000000000
**************************************************
Atomic mass
--------------------------------------------------
8.0000000000000000 16.000000000000000
**************************************************
The numbers of basis sets per atom type
--------------------------------------------------
5043 2309
**************************************************
Basis set for H
--------------------------------------------------
1 1
1 2
1 3
1 4
1 5
1 6
218 70
28 128
743 56The basis set is slightly smaller after the C training because of sparsification, but initally the ML_AB file was identical for both cases.
Why does the continuation training work for graphite when it does not work for Zn? The memory issue should be identical in both cases. And how do I proceed with Zn to get a working force field?
