Generalized convolutional many-body distribution functional representations.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Modern machine learning (ML) models demand extensive data and computational resources, leading to high carbon footprints.
• Lightweight methods offer faster training and reduced environmental impact.

Purpose of the Study:

• Introduce generalized convolutional many-body distribution functionals (cMBDF) as efficient atomic representations.
• Enhance accuracy and reduce computational cost in machine learning for chemical and materials systems, especially in low-data scenarios.

Main Methods:

• Developed cMBDF by generalizing the MBDF framework using translationally and rotationally invariant functionals.
• Encoded local chemical environments compactly using smooth atomic densities weighted by interaction potentials.
• Utilized fast Fourier transforms for efficient evaluation and storage of functional values on predefined grids.

Main Results:

• Achieved highly compute and data-efficient atomic representations (cMBDF) that excel in low-data regimes.
• Demonstrated that cMBDF vectors are compact and constant in size, independent of system size or composition.
• Showed cMBDF is more accurate than popular representations for learning quantum properties across organic and inorganic datasets (QM7b, QM9, VQM24).

Conclusions:

• cMBDF provides a significant reduction in model training and testing times (e.g., 23h to 8min), lowering the carbon footprint.
• The compact and efficient nature of cMBDF makes it suitable for a wide range of chemical and materials applications.
• This approach advances the development of sustainable and efficient machine learning models in scientific discovery.