Electronic-Structure Properties from Atom-Centered Predictions of the Electron Density.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Electron density is a key molecular property targeted by machine learning models.
• Traditional methods face challenges with non-orthogonal atomic basis sets for electron density prediction.
• Developing transferable and linear-scaling models is crucial for computational efficiency.

Purpose of the Study:

• To devise a gradient-based machine learning approach for accurate electron density prediction.
• To overcome limitations of atom-centered models for complex molecular systems.
• To achieve accurate and computationally efficient predictions using optimized feature spaces.

Main Methods:

• Representing electron density using a multicentered atomic basis.
• Employing a gradient-based approach to minimize a regression loss function.
• Utilizing a highly sparse feature space for optimized learning.

Main Results:

• The method achieves highly accurate electron density predictions on complex datasets.
• Testing on liquid water demonstrates high accuracy (0.1 meV/atom error) for total energy components.
• The approach requires a small training set for accurate ground-state total energies on the QM9 dataset.

Conclusions:

• The developed framework enables accurate and efficient machine learning predictions of electron density.
• The method overcomes basis set non-orthogonality challenges.
• It provides a pathway to accurate total energy calculations from predicted electron densities.