SpectraFormer: an Attention-Based Raman Unmixing Tool for Accessing the Graphene Buffer-Layer Signature on SiC

TL;DR

SpectraFormer tackles the challenge of strong, variable SiC substrate backgrounds in Raman spectra of graphene on SiC by learning a reference-free reconstruction of the substrate signal from partially masked spectra using a transformer with self-attention. Subtracting the reconstructed SiC contribution reveals ZLG-related vibrational features that align with ab initio predictions, enabling direct mode assignments for the buffer layer. The ab initio analysis decomposes the ZLG vibrations into B, L, and G components and identifies D-band subfeatures, linking spectral signatures to specific interfacial motifs. This approach enables robust, real-time Raman analysis and can be integrated into closed-loop AI-assisted growth optimization, with potential applicability to other substrate-dominated spectroscopies.

Abstract

Raman spectroscopy is a key tool for graphene characterization, yet its application to graphene grown on silicon carbide (SiC) is strongly limited by the intense and variable second-order Raman response of the substrate. This limitation is critical for buffer layer graphene, a semiconducting interfacial phase, whose vibrational signatures are overlapped with the SiC background and challenging to be reliably accessed using conventional reference-based subtraction, due to strong spatial and experimental variability of the substrate signal. Here we present SpectraFormer, a transformer-based deep learning model that reconstructs the SiC Raman substrate contribution directly from post-growth partially masked spectroscopic data without relying on explicit reference measurements. By learning global correlations across the entire Raman shift range, the model captures the statistical structure of the SiC background and enables accurate reconstruction of its contribution in mixed spectra. Subtraction of the reconstructed substrate signal reveals weak vibrational features associated with ZLG that are inaccessible through conventional analysis methods. The extracted spectra are validated by ab initio vibrational calculations, allowing assignment of the resolved features to specific modes and confirming their physical consistency. By leveraging a state-of-the-art attention-based deep learning architecture, this approach establishes a robust, reference-free framework for Raman analysis of graphene on SiC and provides a foundation, compatible with real-time data acquisition, to its integration into automated, closed-loop AI-assisted growth optimization.

Figures (16)

• Figure 1: (a) Ball-and-stick atomic structure model of investigated materials (SiC substrate, ZLG and MLG); (b) Raman spectroscopy of 3 different samples (solid line - mean across the dataset, shaded area - 1 standard deviation (STD) area around mean value): bare SiC substrate, SiC substrate with ZLG, and SiC substrate with both ZLG and free standing MLG. From insets (c,d) it is easy to see the appearance of surface material's signal. Measurement conditions are all same for all represented spectra: 5% laser power, 5 seconds acquisition time, 1 accumulation per spectrum.
• Figure 2: Model's pipelines for training and usage cases. For both approaches the model generates a bare SiC substrate signal reconstruction based on masked input, but it differs in the input type (for training - SiC spectra, for usage - mixed spectra) and whether model parameters being updated or not.
• Figure 3: Model output after the training with different inputs (solid line - mean across the dataset, shaded area - 1 STD area around mean value): (a, c) MLG/ZLG/SiC Raman spectrum is given and (b, d) ZLG/SiC Raman spectrum is given, allowing to reveal targeted features by subtraction of generated SiC spectrum; middle gray shaded region is the region of data available to the model. Intensity values at (c, d) can be interpreted as normalized to the SiC peak at 1514 cm$^{-1}$, while at (a, b) are also shifted by +0.4 a.u. (for details of data preprocessing, see Supplementary).
• Figure 4: (a) Zoomed region of experimental data in Fig.\ref{['fig:model_outputs']}d with fit (solid line - mean across the dataset, shaded area - 1 STD area around mean value), where each component is taken from ab initio calculations; (b,c) Spatial maps of atomic contribution to each of vibrational modes used (with shared gray color bar).
• Figure S1: XPS measurement of ZLG/SiC sample.