hsulab / gdpy

Generating Deep Potential with Python

NOTE: This package has been renamed as gdpx from GDPy as we found, unfortunately, the name of gdpy has been taken for a long time. The x instead of y can be seen as an extended and improved version of the original one. The development of gdpx will still be undergoing in this GDPy repository. Also, the documentation is changed to https://gdpx.readthedocs.io.

NOTE: gdpx is under active development and has not been released. The APIs are frequently changed and we cannot ensure any backward compatibility.

Install

gdpx is a pure python package. Since it does not include any codes that actually perform calculations and training, for example, VASP and DEEPMD, you should install them by yourselves.

Latest Version

$ python -m pip install git+https://github.com/hsulab/GDPy.git

Stable Release

$ conda install gdpx -c conda-forge

Table of Contents

• Overview
• Features
• Architecture
• Modules
• Authors
• License

Overview

Documentation: https://gdpx.readthedocs.io (Changed from gdpyx)

GDPy stands for Generating Deep Potential with Python (GDPy/GDP¥), including a set of tools and Python modules to automate the structure exploration and the training for machine learning interatomic potentials (MLIPs).

It mainly focuses on the applications in heterogeneous catalysis. The target systems are metal oxides, supported clusters, and solid-liquid interfaces.

Features

• A unified interface to various MLIPs.
• A graph-and-node session to construct user-defined workflows.
• Versatile exploration algorithms to construct a general dataset.
• Automation workflows for dataset construction and MLIP training.

Architecture

Modules

Potential

We do not implement any MLIP but offers a unified interface to access. Certain MLIP could not be utilised before corresponding required packages are installed correctly.The calculations are performed by ase calculators using either python built-in codes (PyTorch, TensorFlow) or File-IO based external codes (e.g. lammps).

Supported MLIPs:

MLIPs	Representation	Regressor	Implemented Backend
eann	(Rescursive) Embedded Atom Descriptor	NN/PyTorch	ASE/Python, ASE/LAMMPS
deepmd	Deep Potential Descriptors	NN/Tensorflow	ASE/Python, ASE/LAMMPS
lasp	Atom-Centered Symmetry Functions	NN/LASP	ASE/LASP
nequip	E(3)-Equivalent Message Passing	NN/PyTorch	ASE/Python, ASE/LAMMPS

NOTE: We use a modified eann package to train and utilise.

NOTE: Allegro is supported as well through the nequip manager.

Other Potentials: Some potentials besides MLIPs are supported. Force fields or semi-empirical potentials are used for pre-sampling to build an initial dataset. Ab-initio methods are used to label structures with target properties (e.g. total energy, forces, and stresses).

Name.	Description	Backend	Notes
reax	Reactive Force Field	LAMMPS
xtb	Tight Binding	xtb	Under development
VASP	Plane-Wave Density Functional Theory	VASP
CP2K	Density Functional Theory	CP2K

Expedition

We take advantage of codes in well-established packages (ASE and LAMMPS) to perform basic minimisation and dynamics. Meanwhile, we have implemented several complicated alogirthms in GDPy itself.

Name	Current Algorithm	Backend
Molecular Dynamics (md)	Brute-Force/Biased Dynamics	ASE, LAMMPS
Evolutionary Global Optimisation (evo)	Genetic Algorithm	ASE/GDPy
Basin Hopping	Monte Carlo like Global Optimisation	GDPy
Adsorbate Configuration (ads)	Adsorbate Configuration Graph Search	GDPy
Reaction Event Exploration (rxn)	Artificial Force Induced Reaction (AFIR)	GDPy
Grand Cononical Monte Carlo (gcmc)	Monte Carlo with Variable Composition	GDPy

Workflow

There are two kinds of workflows according to the way they couple the expedition and the training. Offline workflow as the major category separates the expedition and the training, which collects structures from several expeditions and then trains the MLIP with the collective dataset. This process is highly parallelised and is usually aimed at a general dataset. Online workflow, a really popular one, adopts an on-the-fly strategy to build a dataset during the expedition, where a new MLIP is trained to continue exploration once new candidates are selected (sometimes only one structure every time!). Thus, it is mostly used to train an MLIP for a particular system.

Type	Supported Expedition
Offline	md, evo, ads, rxn
Online	md

Authors

Jiayan Xu ([email protected])

under the supervision of Prof. P. Hu at Queen's University Belfast.

License

GDPy project is under the GPL-3.0 license.

gdpy's Issues

Request: Driver for ABACUS calculator

ABACUS is an open-source DFT package using both PW and LCAO as basis, which is developing by research group of Lixin He from USTC, Xinguo Ren from IPCAS, Mohan Chen from PKU, and also AI for Science Institute. This electronic structure software support LDA, GGA(PBE), mGGA(SCAN) and Hybrid Functional (HSE, PBE0) calculation, and by using LCAO basis, the calculation efficiency have large improvement in large system which can be more than hundreds and thousands of atoms.

From my personal usage, ABACUS have much efficiency improvement by VASP. Additionally, the input files of ABACUS is much easier than CP2K, and in most of magnetic system, ABACUS showed better SCF performance than CP2K, which is good to use in heterogenous catalysis calculation. I'm also developing ASE workflow for using ABACUS.

English doc : https://abacus.deepmodeling.com/en/latest/
Chinese doc: https://mcresearch.gitee.io/abacus-user-guide/

Now, ABACUS have interface with ASE https://gitlab.com/1041176461/ase-abacus, and also dpdata. It is not hard for us to implement ABACUS as a driver in GDPy code.