Toward Complete Molecular Structure Prediction from Infrared Spectroscopy Using Deep Learning.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Infrared (IR) spectroscopy is vital for determining molecular structures.
• Interpreting complex spectra with overlapping peaks is challenging, even for experts.
• Previous machine learning efforts focused on functional groups, not complete structures, due to data limitations.

Purpose of the Study:

• To develop a deep learning model for predicting complete molecular structures from IR spectra.
• To address the challenge of spectral ambiguity in complex organic compounds.
• To automate molecular structure elucidation using only spectral data.

Main Methods:

• A dual-loss deep learning architecture, inspired by image-captioning models, was proposed.
• A dataset of over 17,000 IR spectra was generated using quantum mechanical density functional theory calculations.
• The model predicts molecular structures as Simplified Molecular Input Line Entry System (SMILES) strings.

Main Results:

• The best model achieved 16.26% accuracy in predicting complete molecular structures in a single attempt on an unseen test set.
• The model demonstrated up to 88% accuracy in regenerating individual functional groups.
• The dual loss function effectively learned chemical properties from both spectral data and SMILES strings.

Conclusions:

• Deep learning offers a promising approach to enhance IR spectroscopy for molecular structure determination.
• This AI-driven method has potential applications in analytical chemistry, medicine, and other scientific fields.
• The study highlights the capability of AI to overcome limitations in spectral interpretation.