Molecule Identification with Rotational Spectroscopy and Probabilistic Deep Learning.

Area of Science:

• Computational Chemistry
• Machine Learning
• Spectroscopy

Background:

• Identifying unknown molecules from experimental data is challenging.
• Rotational spectroscopy provides rich information about molecular structure and composition.
• Deep learning offers powerful tools for analyzing complex datasets.

Purpose of the Study:

• To develop a proof-of-concept framework for identifying molecules of unknown elemental composition and structure.
• To leverage experimental rotational data and probabilistic deep learning for molecular identification.
• To create neural network architectures that predict molecular properties from spectroscopic parameters.

Main Methods:

• Developed four neural network architectures for molecular identification.
• Translated spectroscopic parameters into Coulomb matrix eigenspectra.
• Utilized deep learning networks for stoichiometry, SMILES string generation, and functional group prediction.
• Employed dropout layers for probabilistic predictions and trained models on a dataset of ~83,000 organic molecules.

Main Results:

• The framework successfully infers chemical and structural information from minimal experimental rotational data.
• Probabilistic predictions were generated for molecular composition and functional groups.
• Models trained on categorized datasets (hydrocarbons, oxygen/nitrogen-bearing species) acted as domain experts.
• Demonstrated practical inference on four unknown molecules.

Conclusions:

• The presented framework shows promise for identifying molecules using rotational spectroscopy and deep learning.
• The probabilistic approach enhances the reliability of molecular identification.
• Future work can refine these architectures for broader applications in chemistry and materials science.