Active Learning of NEP

Features

• Selecting active set with MaxVol algorithem by CPU or GPU.
• Calcualting extrapolation grade with PyNEP.
• Calcualting extrapolation grade during MD simulations with GPUMD.

Installation

1. (Optional) Create a conda environment

conda create -n nep_active python=3.10
conda activate nep_active

If you choose to install it in your current env, jump to the next step.

2. Install PyNEP

Install the latest PyNEP. You can check its own repository for details.

pip install git+https://github.com/bigd4/PyNEP.git

3. (Optional) Install CuPy

When selecting the active set, you may use cupy or numpy. cupy uses your GPU and is much faster when performing MaxVol. Since you are using GPUMD, I assume you have a GPU. You can check its website for installation details.

pip install cupy-cuda12x

Usage

1. Training an NEP potential

You need to have a NEP.

2. Selecting an Active Set

An active set invsersion (.asi file) is needed when calculating the extrapolation grade. The active set can also be considered as the environments with the maximum diversity. You can use select_active.py to get an active set inversion (.asi file) by MaxVol and corresponding structures (.xyz file).

3. Selecting structures with large gamma

If you want to select some structures to add to the training set, you can calculate their extrapolation grade (gamma) and judge if their are outside the training set. This can be performed by the compute_extrapolatione command in GPUMD or by select_gamma.py. You may modify the gamma cutoff to control how far they are from the training set. The default value in select_gamma.py is 1.

However, the selected structures can be dupelicated, so you need to perform the next step.

4. Extending your training set

If you want to select some structures to add to the training set, your can put them together and perform a MaxVol selection. This is in select_extend.py.