Machine learning electronic structure methods based on the one-electron reduced density matrix.

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Density Functional Theory (DFT) establishes relationships between external potentials and electron density.
• The one-electron reduced density matrix is a key component in electronic structure calculations.
• Existing electronic structure methods can be computationally expensive.

Purpose of the Study:

• To develop machine learning models that act as surrogates for traditional electronic structure methods.
• To utilize the one-electron reduced density matrix as the central learning quantity.
• To enable efficient computation of molecular properties and dynamics.

Main Methods:

• Trained machine learning models on the one-electron reduced density matrix.
• Generated surrogate models for DFT, Hartree-Fock, and full configuration interaction.
• Applied models to systems from small molecules (water) to larger compounds (benzene, propanol).

Main Results:

• Surrogate models accurately predicted molecular observables, energies, and atomic forces.
• Generated band gaps, Kohn-Sham orbitals, and infrared spectra.
• Enabled ab-initio molecular dynamics simulations without self-consistent field theory.

Conclusions:

• Machine learning surrogates based on the one-electron reduced density matrix offer a computationally efficient alternative to traditional methods.
• These surrogates can reproduce the capabilities of standard electronic structure theories.
• The QMLearn Python package provides an accessible implementation.