Efficient implementation of atom-density representations.

Area of Science:

• Computational chemistry
• Materials informatics
• Machine learning

Background:

• Atom-centered representations are crucial for atomistic machine learning, enabling property prediction and structure visualization.
• Many effective representations share a formal connection as discretizations of n-body correlation functions of local atom density.
• Optimizing the evaluation of these representations offers significant potential for computational efficiency.

Purpose of the Study:

• To present librascal, a modular implementation for developing and optimizing atomistic representations.
• To demonstrate the optimization of smooth overlap of atomic position (SOAP) features using librascal.
• To analyze the computational scaling and explore data reduction techniques for these representations.

Main Methods:

• Development of the modular librascal implementation for atomistic representations.
• Optimization of local atom density expansions for smooth overlap of atomic position (SOAP) features.
• Application of kernel ridge regression and data reduction techniques in feature space.

Main Results:

• The librascal implementation facilitates refinement and rapid prototyping of rotationally equivariant atomistic representations.
• Optimization of SOAP feature expansion was demonstrated for various radial basis sets.
• Data reduction techniques reduced total computational cost by up to 4x without significant accuracy loss or impact on symmetry properties.

Conclusions:

• The density-based formalism provides a unifying framework for atomistic representations.
• Librascal enables efficient optimization and development of these representations.
• Computational cost reduction is achievable through feature space optimization and data reduction, enhancing the applicability of atomistic machine learning.