Predicting Infrared Spectra with Message Passing Neural Networks.

Area of Science:

• Computational Chemistry
• Machine Learning
• Spectroscopy

Background:

• Infrared (IR) spectroscopy is crucial for chemical characterization.
• Accurate prediction of IR spectra aids in molecular identification and analysis.
• Existing methods may lack the generalizability to capture diverse spectral forms.

Purpose of the Study:

• To introduce Chemprop-IR, a software package for predicting IR spectra using machine learning.
• To provide a general-purpose, pre-trained model for easy IR spectral prediction.
• To offer a framework for training new, customized IR spectral prediction models.

Main Methods:

• Molecules are encoded using a directed message passing neural network.
• Latent molecular representations are learned and optimized for spectral prediction.
• Model training incorporates specialized spectral metrics, normalization, pretraining with quantum chemistry, and ensembling.

Main Results:

• Chemprop-IR achieves high-quality IR spectral predictions.
• The model effectively captures diverse spectral forms across chemical space.
• Complex peak structures in spectra are accurately represented.

Conclusions:

• Chemprop-IR provides a robust and accurate method for IR spectral prediction.
• The software facilitates both direct use of a pre-trained model and custom model training.
• This approach enhances chemical characterization and identification capabilities.