Fermentation process tracking through enhanced spectral calibration modeling

Abstract

The FDA process analytical technology (PAT) initiative will materialize in a significant increase in the number of installations of spectroscopic instrumentation. However, to attain the greatest benefit from the data generated, there is a need for calibration procedures that extract the maximum information content. For example, in fermentation processes, the interpretation of the resulting spectra is challenging as a consequence of the large number of wavelengths recorded, the underlying correlation structure that is evident between the wavelengths and the impact of the measurement environment. Approaches to the development of calibration models have been based on the application of partial least squares (PLS) either to the full spectral signature or to a subset of wavelengths. This paper presents a new approach to calibration modeling that combines a wavelength selection procedure, spectral window selection (SWS), where windows of wavelengths are automatically selected which are subsequently used as the basis of the calibration model. However, due to the non-uniqueness of the windows selected when the algorithm is executed repeatedly, multiple models are constructed and these are then combined using stacking thereby increasing the robustness of the final calibration model. The methodology is applied to data generated during the monitoring of broth concentrations in an industrial fermentation process from on-line near-infrared (NIR) and mid-infrared (MIR) spectrometers. It is shown that the proposed calibration modeling procedure outperforms traditional calibration procedures, as well as enabling the identification of the critical regions of the spectra with regard to the fermentation process.

Application of near‐infrared spectroscopy for monitoring and control of cell culture and fermentation

“…The optimal wavelength intervals to build a model may not be the regions where the functional groups of a substance show the highest absorbance, because the signal of groups of different analytes may overlap. 50 Despite the ability of PLS to resolve that information, still a better solution may be to use a region where the signal is not necessarily the highest, but instead more specific. 50 As stated earlier, several computerized methods to select the optimal wavelength regions have been developed.…”

“…The number of LVs of a PLS model is usually optimized (Figure 4) by performing a cross-validation and minimizing the corresponding standard error of cross-validation (SECV) or the root mean squared error of cross-validation (RMSECV). 7,15,21 Different pretreatments and wavelength interval selection methods are often tested 7,21,50 to find a chemometric model as robust as possible (i.e, with the minimum SEP or RMSEP).…”

“…10,14,21,30,50 Optimal selection of wavelength intervals may be achieved by use of criteria based on spectroscopy knowledge and statistics describing the performance of the calibration model. 57 Numerous algorithms for automatic selection of optimal wavelength intervals have been developed.…”

Bagging for Gaussian process regression

“…It has been observed that this simple averaging rule can significantly improve the model accuracy and robustness in various applications [4,6,22,23].…”

“…the models that are sensitive (in terms of model parameters and predictive performance) to small changes in the training data. In practice, bagging has been applied to a large number of model structures, including regression trees [3], regression with variable selection [4] and neural networks [5,6].…”