Accelerating Molecular Vibrational Spectra Simulations with a Physically Informed Deep Learning Model.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Spectroscopy

Background:

• Machine learning (ML) surrogate models accelerate physical and chemical simulations.
• Accurate ML models for predicting molecular vibrational spectra are scarce.

Purpose of the Study:

• To develop a highly efficient multitask ML surrogate model for accurate prediction of molecular vibrational spectra.
• To calculate infrared (IR) and Raman spectra using ML-enhanced molecular dynamics.

Main Methods:

• Development of the Vibrational Spectra Neural Network (VSpecNN) model.
• On-the-fly calculation of dipole moments and polarizabilities via ML-enhanced molecular dynamics.
• Application of VSpecNN to pyrazine, a polyatomic chromophore.

Main Results:

• VSpecNN-predicted energies are within chemical accuracy (1 kcal/mol).
• VSpecNN-predicted forces show half the error of a popular high-performance ML model.
• Predicted IR and Raman frequencies differ by <5.87 cm^-1 from ab initio references.
• IR spectra intensities and Raman spectra depolarization ratios are well reproduced.

Conclusions:

• VSpecNN provides accurate predictions of molecular vibrational spectra.
• The model demonstrates the utility of highly accurate neural network potentials for spectral prediction.
• This work advances the application of ML in computational spectroscopy.