Data sets accompanying "Neural-network-enabled molecular dynamics study of HfO2 phase transitions"

Data sets accompanying "Neural-network-enabled molecular dynamics study of HfO2 phase transitions"

This repository contains four files in addition to README.md:

- training_data.json.xz: a collection of 1651 configurations used to train the neural-network force field.

- validation_data.json.xz: an additional set of 415 configurations employed to calculate the validation statistics of the force field during training.

- model_params.pkl: saved coefficients of the final committee of neural-network force fields.

- trajectory.traj: trajectory file containing 24000 steps of an NpT molecular-dynamics run for 768 atoms, with a target temperature of 3000 K and a target pressure of 1 bar.

Details about each file are provided below.

Format of training_data.json.xz and validation_data.json.xz

These are xz-compressed JSON files. Each file contains a single list of records representing configurations, each of which is a object (i.e. a collection of key:value pairs) with the following elements:

- system: a string, one of "cubic", "tetragonal", "monoclinic" or "flipped". The first three describe the three phases of HfO2 under study, while "flipped" refers to configurations where the chemical identity of some atoms has been changed at random to introduce more short-distance homocoordination.

- positions: a list of lists, with shape (n_atoms, 3), containing the atomic positions expressed in angstrom.

- types: a list of n_atoms integers describing the chemical identity of each atom. 0 means Hf, 1 means O.

- energy: the DFT-calculated energy of the configuration, in eV.

- forces: a list of lists, with shape (n_atoms, 3), containing the DFT-calculated forces of atoms expressed in eV/angstrom.

Usage of model_params.pkl

To make sure the the committee of neural-network force fields used for this work can be reconstructed and applied, the specific version of the NeuralIL force field used in this work has been archived with D.O.I. 10.5281/zenodo.10786377.

Details of trajectory.traj

This is a standard ASE trajectory file, last tested with ASE version 3.22.1. The ase.io.trajectory module can be used to read all information about each frame. The time step used for generating the trajectory was 0.5 fs, but only one out of every 10 steps was saved. The characteristic times of the thermostat and barostat were set to 50 fs and 500 fs, respectively.

The flexible-cell extension for JAX-MD used to create the trajectory is distributed under an Apache 2 license and can be found at Zenodo, archived with D.O.I. 10.5281/zenodo.10787829.