Predicting the One-Particle Density Matrix with Machine Learning.

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Hartree-Fock and Kohn-Sham density functional theory (DFT) are crucial for electronic-structure calculations.
• These methods involve iterative solutions of Schrödinger-like equations, with convergence speed depending on system complexity, algorithms, and initial guesses.
• A good initial guess for the density matrix significantly reduces computational steps.

Purpose of the Study:

• To develop a novel method for predicting the density matrix in Kohn-Sham DFT.
• To improve the efficiency of electronic-structure calculations and molecular dynamics simulations.
• To provide a superior initial guess for the density matrix using only atomic positions.

Main Methods:

• Construction of a neural network that takes atomic positions as input.
• The neural network predicts the density matrix for Kohn-Sham DFT.
• Evaluation of the predicted density matrix for interatomic force calculations.

Main Results:

• The neural network provides an initial density matrix guess significantly better than existing methods.
• The predicted density matrix quality is sufficient for accurate interatomic force evaluation.
• Accelerated ab initio molecular dynamics simulations are enabled with minimal self-consistent steps.

Conclusions:

• Neural network-based density matrix prediction offers a substantial improvement in computational efficiency for electronic-structure theory.
• This approach facilitates faster and more accurate molecular dynamics simulations.
• The method holds promise for advancing computational materials science and quantum chemistry research.