pantea 0.11.0

Key features

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

• It uses automatic differentiation to make defining new descriptors straightforward.

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

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

I. Descriptor (ACSF)

Atom-centered Symmetry Function (ACSF) descriptor captures information about the distribution of neighboring atoms around a central atom by considering both radial (two-body) and angular (three-body) symmetry functions. The values obtained from these calculations represent a fingerprint of the local atomic environment and can be used in various machine learning potentials.

Script below demonstrates the process of defining multiple symmetry functions for an element, which can be utilized to evaluate the descriptor values for any structure.

from pantea.datasets import Dataset
from pantea.descriptors import ACSF
from pantea.descriptors.acsf import CutoffFunction, NeighborElements, G2, G3

# Read atomic structure dataset (e.g. water molecules)
structures = Dataset.from_runner("input.data")
structure = structures[0]
print(structure)
# >> Structure(natoms=12, elements=('H', 'O'), dtype=float64)

# Define an ACSF descriptor for hydrogen atoms
# It includes two radial (G2) and angular (G3) symmetry functions
cfn = CutoffFunction.from_type("tanh", r_cutoff=12.0)
g2 = G2(cfn, eta=0.5, r_shift=0.0)
g3 = G3(cfn, eta=0.001, zeta=2.0, lambda0=1.0, r_shift=12.0)

descriptor = ACSF(
        central_element='H',
        radial_symmetry_functions=(
                (g2, NeighborElements('H')),
        ),
        angular_symmetry_functions=(
                (g3, NeighborElements('H', 'O')),
        ),
)

print(descriptor)
# >> ACSF(central_element='H', num_symmetry_functions=2)

values = descriptor(structure)
print("Descriptor values:\n", values)
# >> Descriptor values:
# [[0.01952943 1.13103234]
#  [0.01952756 1.04312263]
# ...
#  [0.00228752 0.41445455]]

gradient = descriptor.grad(structure)
print("Descriptor gradient:\n", gradient)
# >> Descriptor gradient:
# [[[ 4.64523585e-02 -5.03786078e-02 -6.14621389e-02]
#   [-1.04818547e-01 -1.84170755e-02  4.76021411e-02]]
#  [[-9.67003098e-03 -5.45498827e-02  6.32422634e-03]
#   [-1.59613454e-01 -5.94085256e-02  1.72978932e-01]]
# ...
#  [[-1.36223042e-03 -8.02832759e-03 -6.08306094e-05]
#   [ 1.29199076e-02 -9.58762344e-03 -9.12714216e-02]]]

II. Potential (NNP)

This example illustrates how to quickly create a high-dimensional neural network potential (HDNNP) instance from an input setting file.

from pantea.datasets import Dataset
from pantea.potentials import NeuralNetworkPotential

# Dataset: reading structures from RuNNer input data file
structures = Dataset.from_runner("input.data")
structure = structures[0]

# Potential: creating a NNP from the RuNNer potential file
nnp = NeuralNetworkPotential.from_runner("input.nn")
nnp.load()  # this will require loading scaler and model parameter files.

total_energy = nnp(structure)
print(total_energy)

forces = nnp.compute_forces(structure)
print(forces)

Download example input files from here.