A charge-density machine-learning workflow for computing the infrared spectrum of molecules

TL;DR

This paper introduces a machine-learning workflow that predicts charge density and electronic properties of molecules, enabling efficient computation of infrared spectra and electronic observables from a single model, streamlining molecular simulations.

Contribution

The work presents a novel machine-learning approach using Jacobi-Legendre cluster expansion to predict charge density and electronic properties simultaneously, reducing the need for multiple models in molecular simulations.

Findings

Successfully applied to uracil molecule in gas phase

Accurately predicts infrared spectrum and electronic observables

Integrates with ab-initio molecular dynamics for enhanced efficiency

Abstract

We present a machine-learning workflow for the calculation of the infrared spectrum of molecules, and more generally of other temperature-dependent electronic observables. The main idea is to use the Jacobi-Legendre cluster expansion to predict the real-space charge density of a converged density-functional-theory calculation. This gives us access to both energy and forces, and to electronic observables such as the dipole moment or the electronic gap. Thus, the same model can simultaneously drive a molecular dynamics simulation and evaluate electronic quantities along the trajectory, namely it has access to the same information of ab-initio molecular dynamics. A similar approach within the framework of machine-learning force fields would require the training of multiple models, one for the molecular dynamics and others for predicting the electronic quantities. The scheme is implemented here within the numerical framework of the PySCF code and applied to the infrared spectrum of the uracil molecule in the gas phase.