Raman spectra rarely go directly from the spectrometer to a model. Before extracting quantitative information, they usually need preprocessing, including cropping the useful spectral region, removing spikes, subtracting the baseline and normalizing the data. This tutorial walks through that workflow using real Raman spectra from the publication "Process monitoring of zeolite crystallization: Bridging in situ Raman spectroscopy, ex situ XRD, and predictive machine learning models."
Interzeolite transformation occurs during zeolite synthesis when one crystalline phase transforms into another, so tracking the contribution of each phase over time is valuable. Raman spectroscopy can track this process, but the spectra must be deconvolved. Non-Negative Least Squares (NNLS) is well suited to this task: the measured spectrum is described as a non-negative combination of reference spectra from the individual phases. The tutorial focuses on Raman data processing and NNLS, starting from raw spectra and recovering the relative evolution of faujasite (FAU) and gismondine (GIS) during crystallization.
The complete dataset is openly available in Nexus/HDF5 format on Zenodo, with the tutorial provided as a Jupyter notebook. The Nexus format is convenient because spectra and experimental metadata, including crystallization time, remain stored together. Spectra acquired at 0 and 16 hours were excluded because the zeolite Raman features were not yet sufficiently informative, so the analysis begins at 18 hours.
The first preprocessing step crops the spectra to the 250–1250 cm⁻¹ region. Because the spectra were not recorded on exactly the same Raman-shift axis, each is interpolated onto a common axis before being combined into a single data matrix. Cropping and interpolation come before spike removal and baseline subtraction, so those algorithms operate only on the relevant, harmonized signal.
Source: Towards Data Science · Summarized by HeadlinesBriefing