Spectroscopy from Machine Learning by Accurately Representing the Atomic Polar Tensor

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• Vibrational spectroscopy reveals microscopic structure and dynamics but often requires theoretical support for interpretation.
• Ab initio molecular dynamics is effective but computationally expensive for detailed spectral analysis.
• Directly linking IR spectral features to nuclear motion is challenging due to high computational costs.

Purpose of the Study:

• To develop a computationally efficient method for calculating accurate infrared (IR) spectra.
• To enable direct assignment of IR spectral features to specific nuclear motions.
• To overcome the computational bottleneck in analyzing vibrational spectra from molecular dynamics simulations.

Main Methods:

• Utilized the E(3)-equivariant neural network, e3nn, to fit the atomic polar tensor.
• Applied the machine learning model *a posteriori* to existing molecular dynamics simulations.
• Benchmarked the model by calculating the IR spectrum of liquid water.

Main Results:

• The e3nn model accurately reproduced the IR spectrum of liquid water, showing excellent agreement with reference calculations.
• The methodology provides a direct link between spectral features and nuclear motion, previously hindered by computational cost.
• The developed approach is general and transferable to other molecular systems.

Conclusions:

• The presented methodology offers a novel and efficient route to compute accurate IR spectra from molecular dynamics.
• This approach significantly facilitates the microscopic understanding of vibrational spectra.
• The method overcomes previous computational limitations, paving the way for advanced spectral analysis.