Predicting the One-Particle Density Matrix with Machine Learning

S Hazra,U Patil,S Sanvito

DOI: https://doi.org/10.1021/acs.jctc.4c00042

2024-06-01

Journal of Chemical Theory and Computation

Abstract:Two of the most widely used electronic-structure theory methods, namely, Hartree-Fock and Kohn-Sham density functional theory, require the iterative solution of a set of Schrödinger-like equations. The speed of convergence of such a process depends on the complexity of the system under investigation, the self-consistent-field algorithm employed, and the initial guess for the density matrix. An initial density matrix close to the ground-state matrix will effectively allow one to cut out many of...

chemistry, physical,physics, atomic, molecular & chemical

What problem does this paper attempt to address?

This paper mainly discusses how to use machine learning, especially neural networks, to predict the single-particle density matrix in Kohn-Sham density functional theory (DFT). DFT plays an important role in the calculation of electronic structure in molecules and solids, but its iterative solution process for solving Schrödinger-like equations requires multiple steps to achieve convergence. The proposed method in the paper is to construct a neural network that only depends on atomic position information to provide a better initial density matrix estimation than existing methods, thereby accelerating the convergence process and enabling the evaluation of interatomic forces for accelerated ab initio molecular dynamics simulations.

In the study, the authors compared different strategies for generating initial density matrices and directly predicted the density matrix through neural networks, instead of converting from charge density to avoid errors introduced by different basis set projections. They used H2O, S2O, and [Fe(HO)6]2+ as test molecules, and the training and testing data were obtained from all-electron DFT calculations using the PySCF software package with the BLYP functional.

The results show that the density matrix generated by the neural network as the starting point of the SCF cycle can significantly reduce the number of iterations. For H2O and S2O molecules, convergence can be achieved in just one iteration, and for the metal complex [Fe(HO)6]2+, only two iterations are needed. Furthermore, these neural network-predicted density matrices can accurately estimate energy and forces, and can be used for non-self-consistent ab initio molecular dynamics simulations, showing good performance compared to fully converged results.

In conclusion, the paper addresses the issue of accelerating convergence in DFT calculations and proposes an effective and accurate method for predicting density matrices using machine learning, which will have a positive impact on the efficiency of electronic structure calculations.