Session: Analytical Division Poster:

FTIR is an ubiquitous non-destructive analytical method that provides information about chemical functional groups present in samples. Facile, streamlined analysis of FTIR spectra via a machine learning (ML) model would enable high-throughput applications of the method. The ML model will reduce the time required to evaluate spectra and determine present functional groups, while also eliminating subjectivity of interpretation. FTIR spectra from the NIST Chemistry WebBook are obtained via a web scraping method. Using Python scripts, each spectrum is converted from the downloaded file into a csv file. During the conversion, any spectrum not in units of absorbance or wavenumbers is removed. Spectra are normalized with respect to the most intense peak and saved as images. Spectra are sorted using the unique SMARTS key to determine the functional groups and image files are copied to corresponding directories. Many compounds have multiple functional groups and to address this, spectra are copied to each functional group directory that a compound contains. A convolutional neural network (CNN) is utilized to create a ML model for seventeen organic functional groups. The convolutions reduce data dimensionality over many nodes without reducing accuracy or specificity. Backpropagation is employed to improve model accuracy. Lastly, filters that are specific to features are used to direct spectral information throughout nodes. The implementation of filters reduces the reliance of the algorithm on one node and channels spectral information more efficiently. The trained functional group models are used to predict the functional groups present in unknown spectra to evaluate accuracy. All models are run simultaneously for an unknown spectrum. Fifteen of the seventeen models accurately predict functional groups of unknown spectra. The ML approach expedites analysis (15 seconds for a spectrum), without compromising accuracy or efficiency. The results aid in the application of ML to FTIR analysis because the model is generalizable and transferable without any computational or storage requirements. The described work is the first implementation of CNNs as a tool for spectral analysis and provides a platform in that future models can be pursued for specific FTIR applications or other spectral methods.

Abigail Enders

Graduate Research Assistant, The Ohio State University

Appearances