Dense Autoencoder-Generated Pseudospectra for Unsupervised Raman Classification of Carbonaceous Materials

This study presents a novel application of a Dense Autoencoder for the unsupervised classification of carbonaceous materials based on their Raman spectra. While convolutional autoencoders are commonly employed for spectral denoising, we demonstrate that a dense network, when applied to a dataset with a well-defined and low-dimensional spectral signature, acts as a powerful ''generalist'' classifier. The method focuses on the most significant spectral features - the G and D bands - disregarding extraneous noise and minor fluctuations. The autoencoder successfully clustered approximately 150 spectra into four distinct groups, each represented by a unique pseudospectrum, which is a clean, ideal spectral signature generated by the model itself. The resulting classification aligns significantly with the established categories of carbonaceous materials (e.g., highly graphitized, mildly graphitized, disordered, and amorphous) identified in previous manual studies. This work validates the Dense Autoencoder as a robust tool for unsupervised learning, capable of autonomously extracting meaningful chemical information from spectral data and creating a library of reference pseudospectra for future analysis. This approach bypasses the need for complex, manual spectral deconvolution and demonstrates a new pathway for automated material characterization.