Introduction

This Python package facilitates the construction of NEP using active learning techniques. It employs the same strategies as MTP [J. Chem. Phys. 159, 084112 (2023)] and ACE [Phys. Rev. Materials 7, 043801 (2023)].

The package automates active learning by submitting and monitoring jobs. The workflow is as follows:

1. Calculating Energy, Force, and Virial via Ab Initio Calculations

If this is the first iteration and no structures are available, you can generate simple structures. For instance, you can perturb a crystal structure to create an initial training set.

If a train.xyz file (containing energy and force data) already exists before the first iteration, this step will be skipped.

Note: The initial training set should not be too small for active set selection. For a NEP with a 30×30 neural network, approximately 1000 local environments (e.g., 10 structures with 100 atoms each) are required. If your NEP includes N elements, you will need N × 1000 local environments for each element type.

2. Selecting an Active Set

The MaxVol algorithm is used to select an active set, which consists of reference environments for calculating the extrapolation level.

3. Training the NEP

Train the NEP using the structures from Step 1. If a nep.txt file is provided, this step will be skipped in the first iteration.

4. Performing Molecular Dynamics (MD) to Generate New Structures

Run GPUMD to actively select new structures based on an extrapolation level cutoff. If no structures are generated, the loop terminates. You can run multiple MD simulations under different conditions to accelerate exploration.

5. Selecting Structures to Add to the Training Set

While the structures from Step 4 have high extrapolation levels, they may lack diversity. Therefore, another MaxVol selection is performed to identify the most representative structures.

---

Installation

Dependencies

First, install the required dependencies, including PyNEP and CuPy. Refer to this page for detailed instructions.

Installing the Package

1. Clone the code by running:

   git clone --recursive https://github.com/psn417/nep_maker.git

2. For Bash users:

   • Open ~/.bashrc in a text editor.
   • Add the following line at the end of the file:

     export PYTHONPATH=$PYTHONPATH:/path/to/my_packages
   • Save the file and reload the shell configuration:

     source ~/.bashrc

3. Place the nep_maker package into the my_packages directory to make it accessible to Python.

4. Verify the installation by opening Python and typing import nep_maker. If the import is successful, the package is installed correctly.

---

Input Files

Prepare the following input files:

1. init_structures.xyz: Structures used to train the NEP in the first iteration.
2. (Optional) train.xyz and nep.txt: If you already have an NEP, you can perform active learning based on it.
3. run.in: Specifies how to train the NEP.
4. Job Scripts: Define how to run NEP, GPUMD, and other tasks. Examples are provided for LSF; modify them according to your job system.
5. vasp.yaml: Specifies VASP parameters and pseudopotentials. Include the location of pseudopotentials. Refer to the ASE documentation for more details.
6. run_active.py: The main script for the active learning process. Only one CPU core is needed for it.
7. submit_command.sh: Specifies how to submit your jobs. For LSF, it is bsub < job.sh. For slurm, it is sbatch job.sh.

An example folder, example_Na, is provided with all necessary files.