Machine Learning Density Functionals from the Random-Phase Approximation.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Kohn-Sham density functional theory (DFT) is standard for first-principles calculations.
• More accurate methods like random-phase approximation (RPA) are computationally expensive.
• Machine learning offers a path to reduce computational cost for accurate theories.

Purpose of the Study:

• To develop a machine learning model (ML-RPA) that approximates RPA calculations.
• To create a nonlocal density functional extending the standard gradient approximation.
• To assess ML-RPA accuracy for material properties and chemical systems.

Main Methods:

• Mapping RPA to a pure Kohn-Sham density functional using machine learning.
• Employing nonlocal density descriptors and RPA optimized effective potentials for training.
• Training a single ML-RPA functional on diamond, its surfaces, and liquid water.

Main Results:

• ML-RPA achieves accuracy comparable to state-of-the-art van der Waals functionals for diamond surface formation energies.
• ML-RPA performance for liquid water does not yet surpass the standard gradient approximation.
• The study demonstrates ML's potential to extend RPA applicability to larger systems and timescales.

Conclusions:

• Machine learning can effectively approximate advanced theoretical methods like RPA.
• ML-RPA shows promise for accurate calculations in materials science, particularly for solids.
• Further development is needed for ML-RPA to accurately model complex systems like liquids.