Toward Complete Molecular Structure Prediction from Infrared Spectroscopy Using Deep Learning

Infrared (IR) spectroscopy is a broadly used tool to solve the molecular structures of unknown compounds. Though the theory of generating IR spectra from molecules is well established, the inverse problem of solving molecular structures from given spectra is still challenging. Because complex organic compounds often produce ambiguous spectra with overlapping peaks, even human researchers may struggle to interpret them accurately. Prior attempts to automate the interpretation of spectra using machine learning have been limited by the availability of high-quality data and primarily focused on predicting the presence of individual functional groups rather than predicting the structure of entire molecules. In this work, we propose a dual-loss deep learning architecture, inspired by image-captioning models, to predict the complete molecular structure as a Simplified Molecular Input Line Entry System (SMILES) string from only IR spectral absorbances. To train the models, we leveraged the high-performance computing resources of Lawrence Berkeley National Laboratory's Lawrencium cluster to generate a rich data set of over 17,000 unique IR spectra through quantum mechanical density functional theory calculations. Our best-performing model predicts the complete structure correctly on a single attempt for 16.26% of an unseen test set of 1710 spectra, while regenerating individual functional groups with up to 88% accuracy. We show that the dual loss function of our model architecture enables it to learn important chemical properties from both SMILES strings and spectral data. This research suggests the potential of using deep learning to enhance the capabilities of IR spectroscopy in analytical chemistry, medicine, and other fields.