The crystallization process is a very important process in the pharmaceutical industry. The complexity of the crystallization process and the dependence of this process on many factors, as well as the strict requirements for high quality and purity of active pharmaceutical ingredients, have led to the need to develop advanced instruments and technologies for continuous monitoring and process control. Process analytical technology (PAT) is increasingly being applied in the crystallization process to continuously monitor some of the key product quality characteristics. The PAT technology involves continuous monitoring of the concentration with advanced analytical instruments such as optical probes and the development of a calibration model that relates the measured spectral data to the concentration of the pharmaceutically active substance. The calibration model can be developed using machine learning methods. The aim of this work was to develop and investigate neural network models for modeling the dependence of the concentration of the active pharmaceutical ingredient, xylometazoline hydrochloride in n-buthanol on temperature and spectral data obtained by ATR-FTIR spectrometer measurements. The effects of the preprocessing of the spectral data, the number of neurons in the first and second hidden layers, the number of hidden layers, and the type of optimizer and transfer function used on the quality of the evolved neural networks were analyzed.