Al-O Combustion Dataset for Machine Learning Potentials

1. Introduction

This dataset contains high-quality structural data for the Aluminum-Oxygen (Al-O) system. It was generated to support the development of Deep Learning Potentials (DP) capable of simulating aluminum combustion, surface oxidation, and phase transitions. The dataset includes a wide range of configurations from pure Al, pure O, and various Al-O stoichiometric ratios (90,000 in total).

2. Dataset Generation & Labeling

The structures were processed and labeled using a multi-method approach as seen in the provided workflow:

• Reference Methods: DeepMD-kit (DP), ReaxFF (Reactive Force Field), and EAM (Embedded Atom Method).
• Sampling: Active learning (DP-GEN style) and AIMD trajectories.
• Labels: Total potential energies and per-atom forces.

3. Data Format

The data is provided in the DeepMD-kit (NPY) format, which is optimized for training Deep Potential models. It is organized into set.* directories containing:

• box.npy: Simulation cell dimensions.
• coord.npy: Atomic coordinates.
• energy.npy: Total potential energy of the system.
• force.npy: Atomic forces.
• type.raw: Atomic species mapping (Al and O).

4. How to Use

Prerequisites

• Python 3.x
• dpdata

Loading the Data

You can load the dataset using the dpdata package:

from dpdata import MultiSystems

# Load the training/test systems
ms = MultiSystems().load_systems_from_file(file_name='Al-o-combustion-test', fmt='deepmd/npy')
print(ms)