Leveraging deep learning for accurate and automated interpretation of molecular IR and Raman spectra

Abstract

Evaluating molecular and material properties is a fundamental aspect of science with significant practical applications. While these properties can be inferred from spectra, traditional methods require meticulous calibration and rely heavily on the expertise and intuition of researchers. In this study, we evaluated four machine learning models — convolutional neural networks (CNN), multilayer perceptrons (MLP), Transformers, and gated recurrent units (GRU) — to predict key molecular properties, including the number of atoms, HOMO-LUMO gap, dipole moment, aromaticity, and first excitation energy. Using the QM9S dataset that includes IR and Raman spectra of 127468 molecules as a benchmark, we found that combining IR and Raman spectra consistently outperformed individual spectra in prediction accuracy. By incorporating data augmentation and model fusion, we further improved prediction accuracy by 2 %–7 % for all molecular properties. Finally, we proposed the best single or fused deep learning model to predict the number of atoms, HOMO-LUMO gap, dipole moment, aromaticity and first excitation energy with the accuracy of 99 ​%, 96 ​%, 83 ​%, 99 ​% and 97 ​%, respectively. The present work offers an automated and precise interpretation of IR and Raman spectra, enabling accurate predictions of several critical molecular properties.