Noise2Noise Denoising Pipeline for High-Throughput Raman Spectroscopy

A novel, efficient denoising pipeline for high-throughput Raman spectroscopy has been created, utilizing a one-dimensional convolutional autoencoder that employs a Noise2Noise technique. This method removes the necessity for external spectral libraries or high signal-to-noise reference spectra. By training on a limited subset of repeated short-exposure acquisitions, the model effectively reconstructs Raman spectra while minimizing stochastic noise. When tested on a diverse mineral sample, the technique demonstrated a strong alignment with reference data, achieving integration times as brief as 5 ms per spectrum and maintaining chemically coherent maps. Quantitative evaluations included RMSE, SNR, and SSIM, along with task-specific metrics derived from unsupervised K-means classification.