Stable and Accurate Orbital-Free Density Functional Theory Powered by Machine Learning.

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Materials Science

Background:

• The Hohenberg-Kohn theorems establish a theoretical foundation for density functional theory (DFT).
• Accurate approximations to the exact energy functional remain a significant challenge in DFT.
• Existing functionals often lack the precision required for diverse chemical applications.

Purpose of the Study:

• To develop an empirically derived density functional using machine learning.
• To achieve chemical accuracy in energy calculations and meaningful electron densities for molecules.
• To bridge the gap between theoretical DFT and practical computational chemistry.

Main Methods:

• Utilized rotationally equivariant atomistic machine learning.
• Trained a model on the QM9 dataset of organic molecules.
• Augmented training data with electron densities from perturbed potentials.

Main Results:

• Developed the STRUCTURES25 density functional.
• Achieved energies with chemical accuracy relative to Kohn-Sham calculations.
• Obtained convergent and meaningful electron densities for organic molecules.

Conclusions:

• Machine learning offers a viable path to learning accurate density functionals.
• This work demonstrates practical progress towards the Hohenberg-Kohn vision.
• Enables more efficient and accurate electronic structure calculations for large molecular systems.