Spectral Clustering of Mineralogical Data Using a Transformer Autoencoder

Raman spectroscopy is a crucial tool for mineralogy, but the automatic classification of materials based on their spectra can be challenging due to the inherent complexity of the data. This report describes an approach that uses Transformer architecture to automatically analyze and group spectra, demonstrating the model's ability to learn the unique structural 'fingerprints' of different minerals. Our approach is therefore generalizing to Raman spectroscopy what we have adopted for SERS spectra in https://zenodo.org/records/17021372, that is, training the model to clustering minerals according to fingerprints.