Atom-density representations for machine learning

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning (ML) applications in chemistry and materials science are rapidly expanding.
• A key challenge is developing complete yet concise representations of atomic systems for ML models.
• Current methods for converting atomic structures into ML inputs are diverse and numerous.

Purpose of the Study:

• To introduce an abstract, basis set-independent definition of chemical environments.
• To unify and generalize existing representations for atomic systems in ML.
• To provide a framework for systematically tuning ML representations for materials discovery.

Main Methods:

• Defined chemical environments using a smoothed atomic density.
• Employed bra-ket notation for basis set independence.
• Computed correlations via inner products of feature kets, with explicit representations in real and Fourier space.

Main Results:

• Developed a formalism equivalent to smooth overlap of atomic positions power spectrum and n-body correlations.
• Demonstrated connections between the abstract definition and popular existing representations.
• Introduced operators for systematic tuning of structure-composition-property correlations.

Conclusions:

• The proposed formalism unifies recent developments in ML representations for materials science.
• It offers a pathway toward more effective and computationally efficient ML schemes.
• Enables systematic tuning of representations for improved prediction of material properties.