Domain adaptation and self-training for improved cross-batch classification of spectroscopic data: a comparative analysis of feature normalization techniques

Surface-Enhanced Raman Spectroscopy (SERS) enables sensitive, label-free chemical identification, but machine-learning models trained on a single acquisition session frequently underperform when applied to spectra from a different batch, owing to substrate, instrument, and session variability. We study this cross-session domain shift problem on the publicly available Rhodamine 6G (R6G) SERS dataset released by Park et al. [1], which contains two batches (Batch-1 and Batch-2; 1500 spectra each) collected under nominally similar conditions but with clear differences in intensity scale, baseline behavior, and concentration range. We propose a domain-adaptive feature-learning pipeline that combines a transformer-based denoising autoencoder with Correlation Alignment in the latent space (T-DAE+CORAL) to reduce inter-batch discrepancies without requiring target labels. In the more challenging transfer direction (B1→B2, where the no-adaptation baseline is lowest), a logistic regression classifier on T-DAE+CORAL latent features improves balanced accuracy from 0.844 to 0.884 ± 0.019, with specificity rising from 0.688 to 0.939. In the reverse direction (B2→B1), adding 100 labeled target spectra (~ 6.7% of the target batch) in a semi-supervised setting raises balanced accuracy from 0.887 to 0.955 ± 0.044. We frame this work as a methodological proof-of-concept for cross-batch alignment in SERS classification using a single benchmark analyte; broader validation across additional analytes, biomolecular targets, and multi-instrument datasets is identified as a priority direction for future work.