Skip to content

Active Learning Workflow

Tutorial for workflow_activate_learning_dev.sh

This tutorial explains the steps and considerations involved in workflow_activate_learning_dev.sh script. The workflow is used to generate, select, and analyze molecular structures, followed by simulations and data collection.

Overview

This script is designed to handle molecular dynamics simulations, sampling, and SCF (self-consistent field) calculations for NEP. Below is a step-by-step guide explaining how the script works and what each part does.

1. SLURM Directives

#!/bin/bash -l
#SBATCH -p intel-sc3,intel-sc3-32c
#SBATCH -q huge
#SBATCH -N 1
#SBATCH -J workflow
#SBATCH -o workflow.log
#SBATCH --ntasks-per-node=1

The SLURM directives are used to define how the job will be submitted to the cluster:

  • #SBATCH -p defines the partition, in this case, it's intel-sc3 and intel-sc3-32c.
  • #SBATCH -q specifies the queue, in this case, huge.
  • #SBATCH -N allocates 1 node for the job.
  • #SBATCH -J names the job workflow.
  • #SBATCH -o defines the output log file as workflow.log.
  • --ntasks-per-node=1 specifies that only one task should run per node.

NOTE: If your machine does not have the SLURM environment, you can also run the script directly from the command line. For example:

nohup bash workflow_activate_learning_dev.sh &>workflow.log &

2. Basic Setup

cd $SLURM_SUBMIT_DIR

Ensure the working directory is correct, and that all necessary files are present for the job.

source ${GPUMDkit_path}/Scripts/workflow/submit_template.sh  # Load the submit template
python_pynep=/storage/zhuyizhouLab/yanzhihan/soft/conda/envs/gpumd/bin/python  # Python executable
  • GPUMDkit_path is the environment variable that stores the path for the GPUMDkit .
  • python_pynep points to the Python environment needed for pynep-related scripts.

3. Variable Definitions

work_dir=${PWD}  # Set the working directory
prefix_name=LiF_iter01  # Prefix for calculations
min_dist=1.4  # Minimum atom distance
box_limit=13  # Simulation box limit
max_fp_num=50  # Maximum number of single point calculations
sample_method=pynep  # Sampling method (options: uniform, random, pynep)
pynep_sample_dist=0.01  # Sampling distance for pynep

You can customize:

  • prefix_name to reflect the name of your current work.
  • min_dist, box_limit, and max_fp_num based on your own system.
  • sample_method (choose between uniform, random, or pynep).
  • pynep_sample_dist is the sampling distance for pynep.

4. Check Required Files

if [ -f nep.txt ] && [ -f nep.in ] && [ -f train.xyz ] && [ -f run.in ] && [ -f INCAR ] && [ -f POTCAR ] ; then
    # Check for the required files before proceeding.
else
    echo "Please put nep.in nep.txt train.xyz run.in INCAR POTCAR [KPOINTS] and the sample_struct.xyz in the current directory."
    exit 1
fi

Make sure the necessary files (nep.txt, nep.in, train.xyz, run.in, INCAR, POTCAR, KPOINTS) are available in the working directory. If any of these are missing, the script will terminate. train.xyz and nep.txt are used for pynep sampling, and run.in is the simulation parameters of MD in the current iteration.

5. File Organization

mkdir 00.modev common
mv ${work_dir}/{nep.txt,nep.in,*.xyz,run.in,INCAR,KPOINTS,POTCAR} ./common
cp ${work_dir}/common/$sample_xyz_file ${work_dir}/00.modev

The script organizes the working files into two folders:

  • 00.modev: For molecular dynamics simulation.
  • common: For shared resources such as nep.txt, run.in, and the structure files, etc.

6. Molecular Dynamics Simulation Submission

submit_gpumd_array modev ${sample_struct_num}
sbatch submit.slurm

After preparing the input files, the script submits an array of molecular dynamics (MD) tasks using submit_gpumd_array.

7. Monitoring Task Completion

while true; do
    logs=$(find "${work_dir}/00.modev/" -type f -name log -path "*/sample_*/log")
    finished_tasks_md=$(grep "Finished running GPUMD." $logs | wc -l)
    error_tasks_md=$(grep "Error" $logs | wc -l)

    if [ "$error_tasks_md" -ne 0 ]; then
        echo "Error: MD simulation encountered an error."
        exit 1
    fi
    if [ $finished_tasks_md -eq $sample_struct_num ]; then
        break
    fi
    sleep 30
done

The script continuously checks whether all MD tasks have finished by searching for a Finished running GPUMD. message in the logs. If an error is encountered, the script terminates.

8. Analysis and Filtering

mkdir ${work_dir}/01.select
...

In the 01.select folder, all structures in the trajectory file dump.xyz during the MD process will be analyzed and filtered to avoid the generation of non-physical structures as much as possible. Specifically, the get_min_dist.py, filter_structures_by_distance.py and filter_exyz_by_box.py scripts will be employed to check whether there are situations where the distance between atoms is too close or the simulated box exceeds the limit value, and such structures will be filtered out.

9. Sampling Methods

case $sample_method in
    "uniform")
    "random")
    "pynep")

The script supports three sampling methods: uniform, random, and pynep. It selects structures according to the chosen method and checks if the number exceeds max_fp_num, ensuring the final structure count is within limits.

10. SCF Calculations

submit_vasp_array scf ${selected_struct_num} ${prefix_name}

After sampling, SCF calculations are submitted using submit_vasp_array. The process is similar to the MD submission, but for SCF calculations.

11. Prediction Step

submit_nep_prediction

The final step involves submitting the NEP prediction task to check the accuracy of the NEP model.

Thank you for using GPUMDkit! If you have any questions or need further assistance, feel free to open an issue on our GitHub repository or contact Zihan YAN (yanzihan@westlake.edu.cn).