A Machine-Learned "Chemical Intuition" to Overcome Spectroscopic Data Scarcity.

Area of Science:

• Computational chemistry
• Machine learning applications
• Spectroscopy

Background:

• Predicting infrared ion spectroscopy (IRIS) spectra for molecular ions is challenging due to limited experimental data.
• Existing machine learning models are not well-suited for ion spectra prediction.

Purpose of the Study:

• To develop a machine learning model for accurate prediction of IRIS spectra for molecular ions.
• To leverage transfer learning from neutral molecule models to enhance ion spectra prediction.

Main Methods:

• Utilized the Graphormer-IR model, pre-trained on neutral molecules, as a foundation.
• Employed transfer learning with a dataset of 10,336 computed and 312 experimental IRIS spectra.
• Incorporated graph encodings for molecular charge states and additional ion spectra fine-tuning.

Main Results:

• The Graphormer-IRIS model achieved 21% higher accuracy compared to standard DFT quantum chemical models.
• Successfully captured spectral red-shifts caused by phenomena like sodiation.
• Dimensionality reduction revealed 'chemical intuition' regarding functional groups, electron density, and charge sites.

Conclusions:

• The developed Graphormer-IRIS model offers a significant advancement in predicting IRIS spectra for molecular ions.
• This approach enables rapid spectral predictions, facilitating the structural elucidation of unknown small molecules in biological samples.