This paper presents a novel solution to the problem of IR spectrum interpretation by applying a 1-dimensional convolutional neural network to classify molecular spectra as either containing a functional group or not containing a functional group. 16 mod- els were trained using a single general model architecture, and 11 were highly effective with accuracy, precision, recall, F1, and AUC scores, all greater than 90%. Phenome- nal classification performance was achieved on aldehydes and ketones, where previous attempts have struggled. Furthermore, a tool was developed to generate saliency maps for each model, allowing for the analysis and study of how the model interprets the IR spectra to make classifications. Our results have applications in fields such as the pharmaceutical industry and drug development, as well as the rapid assay screening of environmental pollutants while being generalizable to be a competitive option for any high-volume IR spectroscopy screening process.
Supplementary Materials for Elucidating Functional Group Presence by Analyzing IR Spectra with 1-Dimensional Convolutional Neural Networks
Supplementary materials. Contains information on how to install the needed software and where to find more data gathered from the project.