This software package implements the Symmetry-based Combinatorial Crystal Optimization Program (SCCOP) that predicts crystal structure of specific composition.
SCCOP combines graph neural network and DFT calculation to accelerate the search of crystal structure. The following paper describes the details of the SCCOP framework:
Please cite the following works if you want to use SCCOP.
@article{SCCOP,
title = {Minimizing the training dataset for efficient crystal structure prediction by graph neural network},
author = {Li, Chuan-Nan and Liang, Han-Pu and Xu, Si-Yuan and Zhang, Xie and Wei, Su-Huai and Lin, Zijing},
journal = {submitted to Nat. Comput. Sci.}
}
Package requirements:
Hardware requirements:
In SCCOP, we use shell command ssh and scp to transfer files between nodes, you should make sure the ssh between nodes without password, here we provide one way as follows:
cd ~/.ssh
ssh-keygen -t rsa -b 4096
ssh-copy-id -i ~/.ssh/id_rsa.pub user@nodeXXX[Server]
# PBS or SLURM
Job_System = 'PBS'
Job_Queue = 'CPU'
# Obtain node information automatically
Host_Node = 'HOST_NODE'
CPU_Nodes = 'CPU_NODES'
GPU_Nodes = 'GPU_NODES'
Num_GPUs = 'GPU_NUM'
[Environment]
SCCOP_Env = ''
Sources = ''
Modules = ''
Envs = ''
[Absolute path]
# VASP settings
VASP_scf = '/opt/intel/impi/4.0.3.008/intel64/bin/mpirun -np 4 vasp'
VASP_opt = '/opt/intel/impi/4.0.3.008/intel64/bin/mpirun -np 48 vasp'
# LAMMPS settings
LAMMPS_scf = 'lmp_intel_cpu_intelmpi -in input.inp'
LAMMPS_opt = 'mpirun -np 96 lmp_intel_cpu_intelmpi -in input.inp'Note: the SCCOP should under the /local directory of GPU node, e.g., /local/sccop which includes sccop/src, sccop/data and sccop/libs. For researchers who want to change the submission of VASP jobs, see the code in src/core/sub_vasp.py.
To run SCCOP for desired composition, you need to define a customized initial search file, i.e., the src/core/global_var.py should be:
#Base
Dimension = 3
Composition = 'C1'
Num_Atom = [5, 10]
Space_Group = [[1, 230]]
#2-Dimension settings
Vacuum_Space = 15
Thickness = 1
Z_Layers = 3
#3-Dimension settings
Pressure = 0
#Model
Use_Pretrain_Model = False
Update_ML_Model = True
#Recycling
Num_Recycle = 1
Num_ML_Iter = [1]
Energy_Convergence = -1
Use_Succeed = True
#Sampling
Use_ML_Clustering = True
Min_Dis_Constraint = True
Init_Strus_per_Node = 20
Num_Sample_Limit = 50000
Sampling_Time_Limit = 120
Rand_Latt_Ratio = 0.8
#SpaceGroup-based
General_Search = True
Latt_per_Node = 40
SG_per_Latt = 5
#Cluster-based
Cluster_Search = False
Clus_per_Node = 10
Cluster_Num_Ratio = 0.8
Cluster_Weight = [1]
#Tempelate-based
Template_Search = False
Disturb_Seeds = False
Temp_per_Node = 20
Num_Fixed_Temp = 36
#Searching
SA_Path_per_Node = 40
Restart_Times = 10
Exploration_Ratio = .2
Exploitation_Num = 10
SA_Steps = 75
SA_Decay = .97
SA_Path_Ratio = 0.2
#Sample select
SA_Strus_per_Node = 10
SA_Energy_Ratio = 0.5
#Energy calculate
Energy_Method = 'VASP'
Num_Opt_Low_per_Node = 2
Num_Opt_High_per_Node = 1
Refine_Stru = False
Scf_Time_Limit = 72000
Opt_Time_Limit = 72000
#VASP
Use_VASP_Scf = True
Use_VASP_Opt = True
VASP_Opt_Symm = True
#LAMMPS
Use_LAMMPS_Scf = True
Use_LAMMPS_Opt = TrueSCCOP supports the PBS and SLURM job system, we offer the job script here.
For PBS job system
#!/bin/bash
#PBS -N sccop_pbs
#PBS -q queue
#PBS -l nodes=6:ppn=48
#PBS -l walltime=999:00:00
work=$PBS_O_WORKDIR
cd $work
rm -r output
mkdir output
#get info about host and computation nodes
host=`hostname`
nodes=`cat $PBS_NODEFILE | tr ' ' '\n' | sort | uniq`
echo $host > data/host.dat
echo $nodes > data/nodes.dat
#copy neccessary files to each node
for node in $nodes
do
ssh -T $node "
cd /tmp;
rm -rf sccop;
mkdir sccop;
cd sccop;
scp -r ${work}/data .;
scp -r ${work}/libs .;
scp -r ${work}/src .
"
done
#launch sccop on host node
ssh -T $host "
cd /tmp/sccop;
python src/core/job_system.py;
python src/main.py;
scp data/log.sccop ${work}/.
scp data/poscar/optim_strus/POSCAR* ${work}/output/.
scp data/vasp_out/optim_strus/Energy.dat ${work}/output/.
scp -r data/poscar ${work}/output/.
"
#delete temporary files
rm data/host.dat
rm data/nodes.datFor SLURM job system
#!/bin/bash
#SBATCH -J sccop
#SBATCH -o sccop_%j.out
#SBATCH -e sccop_%j.err
#SBATCH -p partition
#SBATCH -N 3
#SBATCH -n 288
work=$SLURM_SUBMIT_DIR
cd $work
rm -r output
mkdir output
#get info about host and computation nodes
nodes=$(scontrol show hostname $SLURM_JOB_NODELIST)
host=$(echo $nodes | awk '{print $1}')
echo $host > data/host.dat
echo $nodes > data/nodes.dat
#copy neccessary files to each node
for node in $nodes
do
ssh -T $node "
cd /tmp;
rm -rf sccop;
mkdir sccop;
cd sccop;
scp -r ${work}/data .;
scp -r ${work}/libs .;
scp -r ${work}/src .
"
done
#launch sccop on host node
ssh -T $host "
cd /tmp/sccop
python src/core/job_system.py
python src/main.py
scp data/log.sccop ${work}/output/.
scp data/poscar/optim_strus/POSCAR* ${work}/output/.
scp data/vasp_out/optim_strus/Energy.dat ${work}/output/.
scp -r data/poscar ${work}/output/.
"
#delete temporary files
rm data/host.dat
rm data/nodes.datIf you install packages in prerequisites, and finish the server and path configuration and initial search file, then you can cd /path/sccop to submit sccop job by:
# PBS job system
qsub jobs.pbs
# SLURM job system
sbatch job.slurmAfter searching, you will get three important files.
data/log.sccop: stores the searching process of SCCOP.data/poscars/optim_strus: stores the POSCAR of searched structures.
We have applied SCCOP to systematic search 82 compositions of B-C-N system, and you can download the newly discovered 28 stable low energy configurations from B-C-N_POSCAR.
This software was developed by SCCOP group.
SCCOP is released under the MIT License.