jaxip 0.5.3

JAX-based Interatomic Potential

JAX-based Interatomic Potential

Description

JAXIP is an optimized Python library on basis of JAX that helps in the development of emerging machine learning interatomic potentials for use in computational physics and chemistry. These potentials are necessary for conducting large-scale molecular dynamics simulations of complex materials at the atomic level with ab initio accuracy.

JAXIP is designed to develop potentials for use in molecular dynamics simulations, rather than a package for performing the simulations themselves.

Documentation: https://jaxip.readthedocs.io.

Main features

• The design of JAXIP is simple and flexible, which makes it easy to incorporate atomic descriptors and potentials

• It uses autograd to make defining new descriptors straightforward

• JAXIP is written purely in Python and optimized with just-in-time (JIT) compilation.

• It also supports GPU-accelerated computing, which can significantly speed up preprocessing and model training.

Note

This package is under heavy development and the current focus is on the implementation of high-dimensional neural network potential (HDNNP) proposed by Behler et al. (2007).

Examples

Defining an atomic environment descriptor

The following example shows how to create an array of atomic-centered symmetry functions (ACSF) for a specific element. This descriptor can be applied to a given structure to produce the descriptor values that are required to build machine learning potentials.

from jaxip.datasets import RunnerStructureDataset
from jaxip.descriptors import ACSF
from jaxip.descriptors.acsf import CutoffFunction, G2, G3

# Read atomic structure dataset (e.g. water molecules)
structures = RunnerStructureDataset('input.data')
structure = structures[0]

# Define ACSF descriptor for hydrogen element
descriptor = ACSF(element='H')

# Add radial and angular symmetry functions
cfn = CutoffFunction(r_cutoff=12.0, cutoff_type='tanh')
descriptor.add( G2(cfn, eta=0.5, r_shift=0.0), 'H')
descriptor.add( G3(cfn, eta=0.001, zeta=2.0, lambda0=1.0, r_shift=12.0), 'H', 'O')

# Compute descriptor values
descriptor(structure)

# Compute gradient
descriptor.grad(structure, atom_index=0)

Training a machine learning potential

This example illustrates how to quickly create a high-dimensional neural network potential (HDNNP) instance from an in input setting files and train it on input structures. The trained potential can then be used to evaluate the energy and force components for new structures.

from jaxip.datasets import RunnerStructureDataset
from jaxip.potentials import NeuralNetworkPotential

# Read atomic data
structures = RunnerStructureDataset("input.data")
structure = structures[0]

# Instantiate potential from input settings file
nnp = NeuralNetworkPotential.create_from("input.nn")

# Fit descriptor scaler and model weights
nnp.fit_scaler(structures)
nnp.fit_model(structures)
nnp.save()

# Or loading from files
#nnp.load()

# Total energy
nnp(structure)

# Force components
nnp.compute_force(structure)

License

This project is licensed under the GNU General Public License (GPL) version 3 - see the LICENSE file for details.