RePAP

Overview

RePAP provides a atom structure generation model using recurrent probabilistic axis projections.

Prerequisites

• Python 3.5+
• NumPy 1.10+: pip3 install numpy
• SciPy 0.17+: pip3 install scipy
• PyTorch: refer to PyTorch website
• ASE: refer to ASE website

Contents

RePAP_notebook.py contain the RePAP model.

Model overview:

Create densities

The probabilistic axis projection framework consists of representing each atom (in each snapshot of the molecular training trajectory) as a discretized Gaussian density cloud that is projected onto each of the three axes x, y, and z.

With 30 atoms in total, this corresponds to 90 axis projections (i.e. 90 atom coordinates). The 30 atoms in a snapshot are ordered w.r.t. mean distance to origin over the entire trajectory, from smallest to largest distance. The 90 atom coordinates are then ordered as x0, y0, z0, x1, y1, z1, ..., x29, y29, z29. This is the sequential order fed to the recurrent network.

With a 50-dimensional grid along each axis, the density of the i’th atom coordinate at the k’th element of the grid vector is

Usage

(applies to all create_densites_xxx.py files)

In order to take an ASE atom object and make the gaussian expansion run the following in the terminal

python3 create_densities_xxx.py "/location_of_ASE_traj_file.traj"

Optional arguments:

• The number of grid points n_grid_points, default=50
• Kernel parameter for density smoothing sigma, default=0.1
• Number of training examples n_train, default=1
• Types of atoms n_type, default=1

Model performance (radial_distribution_function)

As a measure of the quality of the generated samples, we consider the radial distribution function (RDF) which measures the average number density of atoms at a distance r.

We define the RDF reconstruction error as

and is used to guide model development.

Compare radial distribution functions

The compare_structure_rdf.py script can be used to compare the obtained radial distibution functions from our generation. You run the following in the terminal

python3 compare_structure_rdf.py "/location_of_generated_structure_converted_to_numpy.npy" "/location_of_original_ASE_traj_file.traj"

Optional arguments:

• Maximum distance up to which RDF is calcuated in Ångstrøm rng, default=10
• Kernel parameter for density smoothing sigma, default=0.1
• Resolution of the RDF bins, default=100

Here is a comparison of the radial distribution function of one of our generated (silicon) structures and the original (silicon) structure

Google Colab

Main setup

Testing setup