Machine Learning and Deep Learning methods for predictive modelling from Raman spectra in bioprocessing
This addresses the need for faster, non-invasive online sensing of chemical species in bioprocessing, though it appears incremental as it builds on existing regression approaches with improved methods.
The paper tackled the problem of predicting chemical concentrations from Raman spectra in bioprocessing, where conventional sensors are limited, by developing pretreatment methods and new regression models using machine learning and deep learning, which outperformed conventional models in prediction error and robustness.
In chemical processing and bioprocessing, conventional online sensors are limited to measure only basic process variables like pressure and temperature, pH, dissolved O and CO$_2$ and viable cell density (VCD). The concentration of other chemical species is more difficult to measure, as it usually requires an at-line or off-line approach. Such approaches are invasive and slow compared to on-line sensing. It is known that different molecules can be distinguished by their interaction with monochromatic light, producing different profiles for the resulting Raman spectrum, depending on the concentration. Given the availability of reference measurements for the target variable, regression methods can be used to model the relationship between the profile of the Raman spectra and the concentration of the analyte. This work focused on pretreatment methods of Raman spectra for the facilitation of the regression task using Machine Learning and Deep Learning methods, as well as the development of new regression models based on these methods. In the majority of cases, this allowed to outperform conventional Raman models in terms of prediction error and prediction robustness.