LASP SSW-DFT-NN auto-train python-lib

Author

Modified: James.Misaka.Bourbon.Liu

Original: ZPLiu's Group (SDHuang, SCMa, ZPLiu et. al.)

Last Update: 2023-05-06

Version: V1.2.7

Update log

V1.2.7:

1. Update lasp2reann.py and related compoents to transform lasp allstr/for.arc or TrainStr/For.arc to REANN configuratuon formula.
2. Update lasp2extxyz.py and related compoents to transform lasp allstr/for.arc or TrainStr/For.arc to extxyz formula for ASE, Ovito, VMD and NequIP

V1.2.6:

1. Script Traindata_analysis_more.py for give the full infomation of TrainStr.txt and TrainFor.txt
2. Script Traindata_analysis_more_para.py for parallel doing infomation subtraction
3. bug fixed of allstr_new.py and sturcture_new.py

Program Structure

Requirement

Python 3.6+ (better 3.8)

NumPy, matplotlib, Pandas. Scipy, multiprocessing et. al.

Miniconda or Anaconda to construct Python-3.8 env is recommended

Original LASP_PythonLib use Python-2.7, which is TOTALLY OUT-Of-DATE

Function

1. SSW-NN-autotrain Machine Learning Potentials of LASP.
2. python (and shell/fortran) scripts which can be used in LASP Calculation. e.g:
   1. vasp2lasptrain.py transfer VASP-label result to TrainStr.txt and TrainFor.txt. (independently)
   2. shiftformat.py transfer between arc-file and Traindata-file. (independently)
   3. traindata_analysis.py to give infomation of TrainStr.txt and TrainFor.txt.
   4. splitarc_auto.py to split muti-struc-arcfile to each input.arc(lasp.str) or DFT_label_inputfile.
   5. pos_arc_shift.py transfer between POSCAR and struc-arc-file.
   6. shellscript in little_script dir.
   7. computeQ.py for calculate dafa for PlotQE usage.

Scripts related to files manipulation of LASP jobs

1. related to traindata
   1. shiftformat.py: arc2train or train2arc usage (independent)
   2. vasp2lasptrain.py: vasp-dft result directly to TrainStr.txt and TrainFor.txt (independent)
   3. cut_traindata.py: cut TrainStr/TrainFor by
   4. traindata_analysis.py: print-out basic statistic infomation of TrainStr/TrainFor, which can have overview.
   5. traindata_analysis_more(_para).py: print-out full infomation of TrainStr/TrainFor, which can be used for SI.
2. related to arc_data
   1. findGM.py: find top100 (can be set) global minimum structure from SSW result
   2. splitarc_auto.py: split all-str arc file to one-str arc file (or dft-project dir), for doing LASP or VASP
   3. nodejob.py and nodejob_coor.py: collect all-str from SSW to VASP-DFT
   4. collect_vasp_label.py: collect and screen all-str from SSW to VASP-DFT (not test)
   5. screen_data.py: from auto.py, used to screen all-str from SSW result (not test)
3. related to coordination patterns
   1. dft_setting dir: used for coor_verlet_sample.py
   2. coor_verlet_sample.py: two-steps dynamic verlet sampling based on stucture similarity described by coordination patterns
   3. update_patterns(_parallel).py: generate coordination patterns of all-str (and add to exist database)

Main Function of Scripts: Run SSW-NN-autotrain and auto.py

1. modify console file

1.1 some following are parameters often need to change

StartfromVASP 0   # 0 start from SSW sampling provided with NN pot , 1 start from allstr.arc-0 in VASP dir, which is often used to train the first NN pot
Nbad   40    # structures for VASP every cycle
cpupernode 96   # CPU total core (not suggested running between 2 or more CPUs)
SSWcheckcycle  600   # SSW time clock 600 seconds

%block cpuperjob
SSW  24         # cores per SSW job
VASP 24         # cores per VASP job
NN   0          # should be designated in jobs.sh
%endblock cpuperjob

1.2 provide the binary program

%block prog
SSW  /home10/bin/lasp-1.0-release/lasp
VASP  /home10/bin/lasp-1.0-release/lasp
VASPgamma  /home10/bin/lasp-1.0-release/lasp.gamma
NN  /home10/bin/lasp-1.0-release/lasp
%endblock prog

1.3 provide the element name in console file

%block base
O   0.0
H   0.0
%endblock base

2. modify jobs.sh

1. make sure the name of NN pot is correct, e.g. sed -i 's/H2O/PtOH.pot/g' jobs.sh
2. modify the number of cycles, default is 100, in: for i in {1..100}
3. modify the cpu/cores required for your computing cluster (modify it in jobs.sh)

3. create arc file for SSW sampling: allstr-ini.arc

in SSW/sourcedir/allstr-ini.arc

you may get allstr-ini.arc from the examples of structure which you need to add and train in your pot

4. check NN directory

In rootdir/NN you should prepare:

   lasp.in   # lasp_NNtrain input file
   H2O.pot          # not required if start from scratch
   H2O.input        # if start from scratch, use "newrun" for pot
   TrainStr.txt     # if start from scratch, just creat an empty file
   TrainFor.txt     # if start from scratch, just creat an empty file
   adjust_factor  # can ignore

lasp already has a lot of Train*.txt files for different systems please first download TrainStr.txt TrainFor.txt from www.lasphub.com

5. make sure add python exec path in jobs.sh

You'd better have anaconda env in your server, otherwise you can use intel-python:

You can use intel-python by define it in .bashrc:

export PYTHONPATH=/data/apps/intel/intelpython3/bin:$PYTHONPATH

6. qsub jobs.sh (or sbatch jobs_local.slurm)

some tips

1. SSW/sourcedir have input.i for SSW-NN input file, may need to check
2. remember check SSW VASP NN dir before finally running
3. Remember check auto.py(SSW_choosing_mode) and jobs.sh before qsub/sbatch
4. Coordination-Patterns method waiting for using
5. auto.py and SSW-DFT-NN auto still need to be tested