RamanFormerSSL: A self-supervised learning based transformer model for Raman mixture component quantification

In this study, we introduce RamanFormerSSL, a self-supervised learning (SSL) based transformer methodology for the quantification of constituents in Raman mixture spectra. Raman scattering provides unique vibrational signatures of molecules, making it a powerful tool for material identification through their Raman spectra. However, discerning individual components in a mixture spectrum is complex due to the overlapping of spectral features, particularly when these features are similar. Additionally, quantifying these components is even more challenging because individual spectra contribute to the mixture spectrum at varying rates across different wavenumbers. Our approach, RamanFormerSSL, leverages the Transformer model architecture, renowned for its effectiveness in capturing complex patterns. Initially, our model undergoes pre-training through a self-supervised learning approach, specifically using masked-spectrum modeling. This step is followed by fine-tuning the model on experimental mixture data. The results demonstrate that RamanFormerSSL is adept at identifying intricate spectral patterns in Raman data, showing superior performance compared to various other deep learning strategies and traditional machine learning methods. This advancement highlights the potential of Transformer-based models in enhancing the analysis of Raman spectra, particularly in complex mixture scenarios. Note that codes will be released.