Transformer-Based Models for Predicting Molecular Structures from Infrared Spectra Using Patch-Based Self-Attention.

Area of Science:

• Chemistry
• Spectroscopy
• Artificial Intelligence

Background:

• Infrared (IR) spectroscopy is vital for determining molecular structures.
• Analyzing experimental IR spectra is challenging due to specialized knowledge requirements and spectral variability.

Purpose of the Study:

• To develop an effective and simple transformer-based model for analyzing IR spectra.
• To improve the accuracy and robustness of molecular structure determination from IR spectral data.

Main Methods:

• A transformer-based model with a patch-based self-attention spectrum embedding layer was proposed.
• A data augmentation approach introducing selective vertical noise at absorption peaks was employed.

Main Results:

• The model achieved state-of-the-art performance on simulated datasets.
• A top-1 accuracy of 55% was attained on real experimental spectra, surpassing previous methods by ~10%.
• The model effectively analyzed complex fingerprint regions, extracting critical structural information.

Conclusions:

• The proposed transformer model offers a significant advancement in IR spectral analysis.
• The data augmentation strategy enhances the model's ability to handle spectral variability.
• This approach provides a more effective tool for chemists in molecular structure elucidation.