NRL-TB paramererization#
The DeePTB package support TB parameters with NRL-TB format, and provide a useful tools to transcript the NRL-TB format to our DeePTB loadable and readable json checkpoint. Moreover, utilizing the efficient differentiable optimization, we can further finetune the transcripted parameters in the DeePTB training framework, to achieve ab initial accuracy without training from scratch.
NRL-TB Introduction:#
The NRL Tight-binding method provides an efficient method for calculating properties of materials. The advantage of the NRL-TB method over classical potential simulations is that it explicitly incorporates the real electronic structure and bonding of the material, obtained by an interpolation from a database of first-principles results.
For full theory of NRL-TB method, we refer to the official website http://esd.cos.gmu.edu/tb/.
Transcript NRL-TB parameters to DeePTB#
Here we show a few lines of a templete NRL-TB parameter file, with bulk silicon system using s, p orbital set. This file is located at DeePTB/examples/NRL-TB/silicon/Si-sp.par:
NN00001 (New style Overlap Parameters)
Silicon (Si) -- sp parametrization
1 (One atom type in this file)
12.5 0.5 (RCUT and SCREENL for 1-1 interactions)
4 (Orbitals for atom 1)
28.086 (Atomic Weight of Atom 1)
2.0 2.0 0.0 (formal spd valence occupancy for atom 1)
.110356625153E+01 0 1 lambda (equation 7)
-.532334619024E-01 0 2 a_s (equation 9)
-.907642743186E+00 0 3 b_s (equation 9)
-.883084913674E+01 0 4 c_s (equation 9)
.565661321469E+02 0 5 d_s (equation 9)
.357859715265E+00 0 6 a_p (equation 9)
.303647693101E+00 0 7 b_p (equation 9)
.709222903560E+01 0 8 c_p (equation 9)
-.774785508399E+02 0 9 d_p (equation 9)
.100000000000E+02 1 10 a_t2g (equation 9)
.000000000000E+00 1 11 b_t2g (equation 9)
.000000000000E+00 1 12 c_t2g (equation 9)
.000000000000E+00 1 13 d_t2g (equation 9)
.100000000000E+02 1 14 a_eg (equation 9)
.000000000000E+00 1 15 b_eg (equation 9)
.000000000000E+00 1 16 c_eg (equation 9)
.000000000000E+00 1 17 d_eg (equation 9)
.219560813651E+03 0 18 e_{ss sigma} (equation 11) (Hamiltonian)
-.162132459618E+02 0 19 f_{ss sigma} (equation 11) (Hamiltonian)
-.155048968097E+02 0 20 fbar_{ss sigma} (equation 11) (Hamiltonian)
.126439940008E+01 0 78 g_{ss sigma} (equation 11) (Hamiltonian)
.101276876206E+02 0 21 e_{sp sigma} (equation 11) (Hamiltonian)
You can download more parameters from internet and use them in our code. To transcript the file, we can use the command:
usage: dptb n2j [-h] [-v {DEBUG,3,INFO,2,WARNING,1,ERROR,0}] [-l LOG_PATH]
[-nrl NRL_FILE] [-o OUTDIR]
INPUT
positional arguments:
INPUT the input parameter file in json or yaml format
optional arguments:
-h, --help show this help message and exit
-v {DEBUG,3,INFO,2,WARNING,1,ERROR,0}, --log-level {DEBUG,3,INFO,2,WARNING,1,ERROR,0}
set verbosity level by string or number, 0=ERROR,
1=WARNING, 2=INFO and 3=DEBUG (default: INFO)
-l LOG_PATH, --log-path LOG_PATH
set log file to log messages to disk, if not
specified, the logs will only be output to console
(default: None)
-nrl NRL_FILE, --nrl_file NRL_FILE
The NRL file name (default: None)
-o OUTDIR, --outdir OUTDIR
The output files to save the transfered model and
updated input. (default: ./)
For example:
dptb n2j DeePTB/examples/NRL-TB/silicon/input_n2j.json -nrl DeePTB/examples/NRL-TB/silicon/Si_sp.par -o nrl_out
For doing this, we need to have a config file contains the common_options as the input_n2j.json:
{
"common_options": {
"basis": {
"Si": [
"3s",
"3p",
"d*"
]
},
"device": "cpu",
"dtype": "float32",
"overlap": true,
"seed": 3982377700
}
}
Then, you can find the transcripted DeePTB json model in ./nrl_out/nrl_ckpt.json:
{
"version": 2,
"unit": "eV",
"common_options": {
"basis": {
"Si": [
"3s",
"3p"
]
},
"device": "cpu",
"dtype": "float32",
"overlap": true
},
"model_options": {
"nnsk": {
"onsite": {
"method": "NRL",
"rc": 6.6147151362875,
"w": 0.2645886054515,
"lda": 1.517042837140912
},
"hopping": {
"method": "NRL1",
"rc": 6.6147151362875,
"w": 0.2645886054515
}
}
},
"model_params": {
"onsite": {
"Si-3s-0": [
-0.7242781465186467,
-12.349108629101169,
-120.1498233699409,
769.6214351454472
],
"Si-3p-0": [
4.868929466977601,
4.131337329837268,
96.49469181636128,
-1054.1493863419682
]
},
"hopping": {
"Si-Si-3s-3s-0": [
2987.2770523703775,
-416.85931392897385,
-753.3335086381961,
1.738135839617497
],
"Si-Si-3s-3p-0": [
137.79420981142889,
-113.22319395024688,
11.013502291392534,
1.2683722943630147
],
"Si-Si-3p-3p-0": [
-312.3734908328387,
44.24279542724431,
68.95104019693315,
1.4177954447585202
],
"Si-Si-3p-3p-1": [
139.6685524393218,
120.11742815429497,
-107.67596847852917,
1.5277405887687854
]
},
"overlap": {
"Si-Si-3s-3s-0": [
9.746427246194099,
2.3569360238929495,
-0.5502870702732275,
1.5233364156496074
],
"Si-Si-3s-3p-0": [
16.768761923322074,
-57.99684419121014,
34.971872964934846,
1.7054914546860056
],
"Si-Si-3p-3p-0": [
21.260342995500338,
-4.178618273015186,
-7.1474883718956805,
1.5638674454024022
],
"Si-Si-3p-3p-1": [
-1308.0386246487092,
1414.6889330892875,
-93.24315909334297,
2.1616536435817872
]
}
}
}
and a recommend model options for continuous training using this checkpoint:
{
"common_options": {
"basis": {
"Si": [
"3s",
"3p"
]
},
"device": "cpu",
"dtype": "float32",
"overlap": true,
"seed": 3982377700
},
"model_options": {
"nnsk": {
"onsite": {
"method": "NRL",
"rc": 6.6147151362875,
"w": 0.2645886054515,
"lda": 1.517042837140912
},
"hopping": {
"method": "NRL1",
"rc": 6.6147151362875,
"w": 0.2645886054515
},
"freeze": false,
"std": 0.01,
"push": false
}
}
}
You can use the transcripted model as any other DeePTB model.