Smart Spectral Matching (SSM) is a new cloud-based, highly flexible software framework designed to integrate the storage and analysis of scientific datasets in a user-friendly way. The SSM framework comprises many services, including extensions to allow custom analysis modules, and experimental metadata to allow more extensive data inference. Not only is this data readily searchable via the SSM user interface, but pre- and postprocessing tools can also be easily incorporated into the SSM workflow.<br/><br/>To better understand and assign spectroscopic features in solid-state uranium minerals, we leveraged the SSM framework to perform machine learning (ML) analyses of structural and optical vibrational spectral data. Data from our Compendium of Uranium Raman and Infrared Experimental Spectra (CURIES) including raw structural and spectroscopic data is stored along with any additional metadata in SSM. Examples of metadata compiled for samples in CURIES include secondary chemistry, structure type, coordination of uranium sites in the material, data source, and data collection parameters. Using CURIES as training data, an ML tool has been developed to identify new structure–property relationships and apply these results to the identification of unknown mineral species. First, principal component analyses were performed on the data in CURIES to identify regions of the spectra where specific chemical properties contribute the most to variation across the full dataset. From these results, the SSM ML interface can be used to select for the desired properties, spectral features, and targeted spectral regions, and ML models can be trained on the existing data in CURIES. Highly selective models are saved and stored and can be run on unknown samples to assign spectral features to underlying structural properties. A collection of these ML models can then be used to make inferences about the fundamental nature of the underlying mineral structure and coordination characteristics from optical vibrational spectra.

Smart spectral matching: A new framework for machine learning structure–property relationships in uranium minerals

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.