Comparing machine learning methods on Raman spectra from eight different spectrometers

In biotechnology, Raman Spectroscopy is becoming increasingly popular as a process analytical technology (PAT) for measuring substrates, metabolites, and product-related concentrations. By recording the vibrational modes of molecular bonds, it provides information non-invasively in a high-dimensional spectrum. Machine learning models are used to transform these spectral data into meaningful concentrations of species. Typically, one assumes a linear relationship between intensity and concentrations and learns these relationships using a partial least squares (PLS) model. However, in biological cultivations with a very large number of components, nonlinear models such as convolutional neural networks (CNN) offer significant advantages. In this work, we show that training one CNN on spectra from eight different spectrometers significantly outperforms PLS models. Specifically, we created samples with known concentrations of glucose, sodium acetate and magnesium sulfate and measured more than 2200 spectra of these samples with eight different spectrometers. We trained one CNN on the spectra from all eight datasets simultaneously. This shows great potential for laboratories with data from more than one spectrometer as they do not need to spend extra effort in calibrating individual PLS models, but they can use a joint CNN, which even improves the overall accuracy. In addition, we compare the eight different spectrometers against each other. The results suggest that three spectrometers are better suited for quantifying glucose, sodium acetate, and magnesium sulfate given the models.