Massive Atomic Diversity: a compact universal dataset for atomistic machine learning.

Area of Science:

• Computational Materials Science
• Machine Learning for Science

Background:

• Machine learning models for atomic-scale simulations rely on large materials property databases derived from electronic-structure calculations.
• Existing databases often focus on equilibrium structures, limiting model generalizability to arbitrary atomic configurations.

Purpose of the Study:

• To introduce a novel dataset designed to train machine learning models for accurate predictions across diverse atomic structures.
• To develop a dataset that achieves massive atomic diversity (MAD) for enhanced model training.

Main Methods:

• Constructed a dataset by modifying stable structures to achieve "massive atomic diversity" (MAD).
• Employed highly consistent electronic-structure calculation settings for property computations.
• Developed low-dimensional structural latent space descriptors for materials analysis.

Main Results:

• The MAD dataset, with fewer than 100,000 entries, enabled training of universal interatomic potentials rivaling those from much larger datasets.
• Demonstrated the effectiveness of the MAD dataset in training general-purpose machine learning models for atomic simulations.

Conclusions:

• The MAD dataset design philosophy prioritizes structural diversity for improved machine learning model performance.
• The developed latent space descriptors serve as a valuable tool for materials cartography.