DeePMD-kit

  • Level of openness 3 ★★★
  • Document quality 2 ★★☆

Python/C++ based software package that employs deep learning techniques for construction of interatomic potentials. It implements the Deep Potential, which defines atomic environment descriptors with respect to a local reference frame. The output of many first-principles and molecular dynamics applications can be used as training data, and the trained potentials can be used for molecular dynamics calculations using LAMMPS and path integral molecular dynamics calculations using i-PI.

To Detail

Theano

  • Level of openness 3 ★★★
  • Document quality 3 ★★★

An open-source numerical library for machine learning. Various functions related to deep learning are implemented. This package directly treats equations as such, and have useful routines such as matrix operation and auto partial derivative. Users can convert their codes into C language, and can compile it. High speed operation by GPGPU parallel calculation is supported. A number of tutorials are available.

To Detail

QUIP

  • Level of openness 3 ★★★
  • Document quality 2 ★★☆

A collection of software tools for molecular dynamics calculations. Various interatomic potentials and tight binding models are implemented, and numerous external applications can be invoked. It also supports training and evaluation of GAP (Gaussian Approximation Potential), which is a form of machine learning potential.

To Detail

Torch

  • Level of openness 3 ★★★
  • Document quality 3 ★★★

An open-source numerical library for machine learning. Various functions related to deep learning based on neural networks are implemented. Users can implement complex network with flexible description, and can try various state-of-the-art methods. This package is used in a number of companies in the world. This package is written by the script language, lua.

To Detail

NequIP

  • Level of openness 3 ★★★
  • Document quality 2 ★★☆

Open source software for building and using machine learning potentials based on E(3)-equivariant graph neural networks, which can be trained on output files of simulation codes that can be read by ASE. Molecular dynamics calculations with LAMMPS can be performed using the trained potentials.

To Detail

PyTorch

  • Level of openness 3 ★★★
  • Document quality 3 ★★★

An interface package to use Torch (the open-source numerical library for machine learning) from Python. Users can easily implement deep learning based on neural networks, and can use various state-of-the-art methods. This package supports GPGPU parallel computation, and realises high-speed operation. A front-end interface for C++ is also prepared.

To Detail

Allegro

  • Level of openness 3 ★★★
  • Document quality 2 ★★☆

Open source software for constructing the Allegro potential model based on E(3)-equivariant graph neural networks and using the potential model for molecular dynamics simulations. The code depends on NequIP and can be run in a similar manner. Allegro scales better than NequIP since it doesn’t rely on message passing and the architecture is strictly local with respect to atom-wise environments.

To Detail

Keras

  • Level of openness 3 ★★★
  • Document quality 3 ★★★

An open-source numerical library for machine learning. Using other machine learning numerical libraries (TensorFlow, CNTK, Theano, etc.), users can construct neural networks by relatively short codes. Since a number of methods in machine learning and deep learning are implemented, users can try state-of-the-art methods easily. This package is written by Python.

To Detail

pacemaker

  • Level of openness 3 ★★★
  • Document quality 2 ★★☆

Software tool for constructing interatomic potentials based on nonlinear atomic cluster expansion. It requires the user to either prepare a fitting dataset based on pandas and ASE, or it can automatically extract data from VASP calculation results. The obtained potentials can be used for molecular dynamics simulations using LAMMPS, and it also provides the capability to calculate extrapolation grades for on-the-fly active learning.

To Detail

NetKet

  • Level of openness 3 ★★★
  • Document quality 2 ★★☆

NetKet is an open-source project delivering cutting-edge methods for the study of many-body quantum systems with artificial neural networks and machine learning techniques. Users can perform machine learning algorithms to find the ground-state of many-body Hamiltonians such as supervised learning of a given state and optimization of neural network states by using the variational Monte Carlo method.

To Detail