Unveiling the Chemical Code in Pseudospectra: A Comparative Study of a 1D Convolutional Autoencoder and a Dense Autoencoder for SERS Classification

Raman spectroscopy, particularly its surface-enhanced (SERS) variant, is a powerful tool for chemical analysis, yet the inherent noise and complexity of spectral data often challenge tradi-tional classification methods. This work presents a novel approach to unsupervised chemical clas-sification of SERS spectra using a 1D Convolutional Autoencoder (Conv1D). The performance of the Conv1D model is directly compared with a standard Dense Autoencoder. The results demonstrate that the Conv1D model successfully extracted intrinsic chemical information from the spectra, generating highly coherent and chemically meaningful clusters. We show that mole-cules were grouped based on shared vibrational fingerprints of their core functional groups, such as C-S bonds in sulfur-containing compounds and indole rings in tryptophan derivatives. In stark contrast, the Dense Autoencoder failed to identify these underlying chemical relationships, pro-ducing random and incoherent clusters. This highlights the superior ability of the Conv1D model to learn hierarchical features from spectral data. Ultimately, this study proves that a Conv1D au-toencoder can serve as more than just a denoising tool. It represents a new paradigm for chemical classification, transforming raw SERS data into a structured library of pseudo-spectra that can be used for future identification and analysis. This approach offers a powerful alternative to traditional methods, enabling the robust classification of complex chemical mixtures.