Learning Electron Densities in the Condensed Phase.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate prediction of electron densities is crucial for understanding material properties.
• Existing methods for periodic systems can be computationally intensive.
• Machine learning offers a potential avenue for accelerating these calculations.

Purpose of the Study:

• To introduce a novel local machine-learning method for predicting electron densities in periodic systems.
• To demonstrate the accuracy and efficiency of this method across various material types.
• To showcase the scalability of the approach for large systems.

Main Methods:

• Development of a local machine-learning framework using an atom-centered auxiliary basis.
• Application of symmetry-adapted Gaussian process regression models.
• Adjustment of models for nonorthogonal basis sets to handle diverse periodic systems.

Main Results:

• Accurate prediction of electron densities for metals, semiconductors, and molecular crystals.
• Efficient calculation of electronic properties with errors on the order of tens of meV/atom compared to ab initio methods.
• Demonstrated scalability: models trained on small systems accurately predicted properties of much larger systems (up to 512 molecules) with no increase in error.

Conclusions:

• The proposed machine-learning method provides an accurate and efficient way to predict electron densities and electronic properties of periodic systems.
• The approach shows remarkable scalability, outperforming direct machine-learning models for larger systems.
• The SALTED framework achieves high accuracy (below 4% error) on heterogeneous datasets.