Optimal radial basis for density-based atomic representations.

Area of Science:

• Computational materials science
• Machine learning for chemistry
• Atomic-scale modeling

Background:

• Machine learning models for atomic-scale properties require symmetric representations of atomic coordinates.
• Current representations often rely on heuristic basis set optimization.
• These methods focus on optimizing for specific regression targets.

Purpose of the Study:

• To develop an unsupervised approach for identifying optimal basis sets for atomic representations.
• To determine a basis that compactly encodes dataset-relevant structural information.
• To enhance the accuracy and efficiency of machine learning models in materials science.

Main Methods:

• Unsupervised learning to identify optimal basis functions.
• Approximation of optimal basis with splines for efficient computation.
• Evaluation of representations for molecular and condensed-phase systems.

Main Results:

• A unique, optimal basis can be constructed for any training dataset and number of basis functions.
• The proposed method is computationally efficient, with no additional cost over primitive bases.
• The resulting representations are accurate, especially for high-body order correlations.

Conclusions:

• This unsupervised approach provides a data-driven method for selecting optimal atomic representations.
• The spline-approximated optimal basis offers a computationally efficient and accurate alternative for materials machine learning.
• The method is applicable to diverse machine learning models across different phases of matter.