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

Area of Science:

• Computational chemistry
• Machine learning applications
• Spectroscopy

Background:

• Accurate calculation of molecular properties like infrared spectra is crucial in chemistry.
• Traditional methods often require separate models for dynamics and electronic property prediction.
• Machine learning offers potential for integrated and efficient calculations.

Purpose of the Study:

• To develop a unified machine-learning workflow for calculating infrared spectra and other temperature-dependent electronic observables.
• To enable simultaneous molecular dynamics simulations and electronic property evaluations using a single model.
• To demonstrate the workflow's application to the infrared spectrum of uracil.

Main Methods:

• Utilizing a Jacobi-Legendre cluster expansion to predict real-space charge density from density-functional-theory calculations.
• Developing a machine-learning model that provides access to energy, forces, and electronic observables (dipole moment, electronic gap).
• Implementing the workflow within the PySCF computational chemistry code.

Main Results:

• The developed workflow can simultaneously drive molecular dynamics and evaluate electronic quantities, mimicking ab initio molecular dynamics.
• This approach avoids the need for multiple specialized machine-learning models.
• The method was successfully applied to calculate the infrared spectrum of uracil in the gas phase.

Conclusions:

• The presented machine-learning workflow offers an efficient and integrated approach for calculating molecular infrared spectra and electronic properties.
• This method provides a powerful alternative to traditional computational chemistry techniques, especially for temperature-dependent observables.
• The successful application to uracil demonstrates the workflow's potential for broader use in computational molecular science.