https://doi.org/10.15255/KUI.2023.011
Published: Kem. Ind. 72 (11-12) (2023) 639–650
Paper reference number: KUI-11/2023
Paper type: Original scientific paper
Download paper:  PDF

Application of Neural Networks for Estimating the Concentration of Active Ingredients Solution using In-situ ATR-FTIR Spectroscopy

T. Herceg, Ž. Ujević Andrijić, M. Gavran, J. Sacher, I. Vrban and N. Bolf

Abstract

Process analytical technology (PAT) is increasingly applied in the crystallization process for continuous monitoring of some of the key process parameters and product quality features. Very important process variables for cooling crystallization are the temperature and concentration of the mother liquor. Continuous measurement of concentration is made possible by advanced in situ spectroscopic instruments. Attenuated total reflectance Fourier transform infrared spectroscopy (ATR-FTIR), used in this work belongs to such methods. The calibration model which estimates the concentration of the solution in real time can be developed using machine-learning methods. The aim of this work was to develop and analyse partial least squares regression (PLSR) and neural network models for modelling the dependence of the concentration of the active ingredients, xylometazoline hydrochloride in n-butanol, on temperature and spectral data obtained by measurements with an ATR-FTIR spectrometer. In this work, pre-processing of the collected data was performed with MSC technique (multiplicative scatter correction), Min-Max and Z-score normalization; the number of neurons in the first and second hidden layers, the number of hidden layers, the type of learning algorithm applied (ADAM, NADAM, RMSprop), and the influence of the type of transfer function (ReLU, sigmoid, tanh) on the quality of the developed neural networks were analysed. Considering values of coefficient of determination and mean square error, developed models gave very good results on all four datasets. The neural network model gave coefficients of determination in the range of values from 0.9979 to 0.9989, and the mean square error from 0.0020 to 0.0011. With the PLSR model, coefficients of determination from 0.9990 to 0.9995, and mean square errors from 0.0009 to 0.0005, were obtained. Obtained results showed that the pre-processing of the data and the addition of a second hidden layer of the neural network in this study did not have a major impact on the final results. This type of monitoring and control of the process would lead to more efficient production with a lower probability of error, enabling the pharmaceutical industry to bring products to market faster.

This work is licensed under a Creative Commons Attribution 4.0 International License

Keywords

crystallisation, process analytical technology, ATR-FTIR spectrometry, machine learning, xylometazoline hydrochloride