Bypassing the Kohn-Sham equations with machine learning.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Density functional theory (DFT) using the Kohn-Sham scheme is crucial for electronic structure calculations.
• Current DFT methods require solving complex Kohn-Sham equations, limiting system size and simulation time.
• Machine learning (ML) offers a potential alternative to learn energy functionals, bypassing direct equation solving.

Purpose of the Study:

• To develop a machine learning approach for electronic structure calculations that bypasses the need to solve Kohn-Sham equations.
• To overcome limitations in previous ML attempts by directly learning density-potential and energy-density maps.
• To demonstrate the feasibility of ML-driven density functionals in molecular dynamics simulations.

Main Methods:

• Directly learning density-potential and energy-density maps using machine learning models.
• Developing and applying a machine-learned density functional.
• Performing the first molecular dynamics simulation using a machine-learned density functional on the malonaldehyde molecule.

Main Results:

• Successfully performed molecular dynamics simulations with a machine-learned density functional.
• Captured the intramolecular proton transfer process in malonaldehyde accurately.
• Demonstrated that ML models can learn density maps, enabling accurate density functional construction for molecular systems.

Conclusions:

• Machine learning can effectively learn density models, leading to accurate density functionals.
• This approach significantly reduces computational cost, enabling larger system sizes and longer simulation timescales.
• The developed method opens new avenues for computational chemistry and materials science research.