Orbital-free bond breaking via machine learning.

Area of Science:

• Computational chemistry
• Quantum mechanics
• Machine learning applications

Background:

• Accurate kinetic energy density functionals are crucial for electronic structure calculations.
• Approximating these functionals, especially for complex systems, remains a challenge.

Purpose of the Study:

• To investigate machine learning's capability in approximating the non-interacting kinetic energy density functional (TF) for diatomic molecules.
• To assess the accuracy and scalability of a machine learning approach based on nonlinear interpolation.

Main Methods:

• Development of a one-dimensional model to train machine learning algorithms.
• Utilizing Kohn-Sham reference calculations to generate training data.
• Employing a projection method for generating self-consistent densities.

Main Results:

• The machine learning model accurately predicts the dissociation of diatomic molecules.
• The model's accuracy systematically improves with an increased amount of reference data.
• Accurate self-consistent densities were generated, avoiding regions with sparse data.
• Interpolation errors were found to be smaller than typical errors in standard exchange-correlation functionals.

Conclusions:

• Machine learning offers a promising route to accurate kinetic energy density functionals.
• The developed method demonstrates potential for improving density functional theory calculations.
• This approach can be systematically enhanced with more computational data.