Deep Learning for Generating Phase-Conditioned Infrared Spectra.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Spectroscopy

Background:

• Infrared (IR) spectroscopy is crucial for chemical compound identification.
• Current simulation methods often overlook phase dependency, limiting accuracy.
• Accelerating IR spectrum analysis is vital for chemical science.

Purpose of the Study:

• To develop an efficient, phase-aware machine learning method for generating phase-conditioned IR spectra.
• To address the limitations of existing methods that assume gas-phase molecules.
• To enable accurate IR spectrum prediction for real-world complex molecules.

Main Methods:

• A novel phase-aware graph neural network was devised.
• The network was combined with a transformer decoder for spectrum generation.
• The method generates phase-conditioned IR spectra from 2D molecular structures.

Main Results:

• The proposed method successfully generates phase-conditioned IR spectra.
• It is the first known IR spectrum generator for complex molecules considering phase.
• Outperformed state-of-the-art methods on a large benchmark dataset.

Conclusions:

• The developed phase-aware machine learning method significantly advances IR spectrum simulation.
• This approach provides more accurate and realistic IR spectra by considering phase.
• The publicly available implementation facilitates further research and application.