Revolutionizing Spectroscopic Analysis Using Sequence-to-Sequence Models I: From Infrared Spectra to Molecular Structures

Infrared (IR) spectroscopy reveals molecular and material features via their characteristic vibrational frequencies in an efficient and sensitive style and has thus become one of the most popular analytical tools in broad areas involving chemical discovery. These fields include material synthesis, drug design, pharmacokinetics, safety screening, pollutant sensing, and observational astronomy. However, in situ molecular or material identification from spectral signals remains a resource-intensive challenge and requires professional training due to its complexity in tracking effects to causes. Motivated by the recent success of sequence-to-sequence (Seq2Seq) models from deep learning, we developed a direct, accurate, effortless and physics-informed protocol to realize such a in-situ spectrum-to-structure translation, and provided the proof-of-concept of our models using IR spectra. We expressed both the input IR spectrum and the output molecular structure as alphanumerical sequences, treated them as two sentences describing the same molecule from two different languages, and translated them into each other using Seq2Seq models from recurrent neural networks (RNNs) and Transformers. Trained and validated using a curated data set of 198,091 organic molecules from the QM9 and PC9 databases, our Seq2Seq models achieved state-of-the-art accuracy of up to 0.611, 0.850, 0.804, and > 0.972 in generating target molecular identities, chemical formulas, structural frameworks, and functional groups from only IR spectra. Our study sets the stage for a revolutionary way to analyze molecular or material spectra by replacing human labor with rapid and accurate deep learning approaches.