Highly accurate real-space electron densities with neural networks

Area of Science:

• Quantum Chemistry
• Computational Chemistry
• Materials Science

Background:

• Variational ab initio methods provide direct access to wave functions in quantum chemistry.
• Extracting observables beyond energy from wave functions is often computationally challenging.
• Electron density is a key observable for understanding molecular properties.

Purpose of the Study:

• To develop a novel method for obtaining accurate electron densities from many-electron wave functions.
• To utilize neural networks for representing and learning electron density properties.
• To overcome the practical difficulties in extracting electron densities from wave functions.

Main Methods:

• Employed variational quantum Monte Carlo with deep-learning Ansätze to generate accurate wave functions.
• Developed a neural network representation for electron densities, incorporating known asymptotic properties.
• Trained the neural network using score matching and noise-contrastive estimation from wave functions.

Main Results:

• Achieved highly accurate electron densities, free from basis set errors.
• Demonstrated the accuracy of the method by calculating various density-based properties.
• Successfully extracted properties such as dipole moments, nuclear forces, and contact densities.

Conclusions:

• The novel neural network approach provides accurate electron densities from accurate wave functions.
• This method offers a computationally practical solution for extracting electron density and related properties.
• Advances the capability to compute and utilize electron densities in quantum chemistry.