Determination of free amino nitrogen in beer mash with an inline NIR transflectance probe and data evaluation by machine learning algorithms

Free amino nitrogen (FAN) concentrations in beer mash can be determined with machine learning algorithms from near-infrared (NIR) spectra. NIR spectroscopy is an alternative to a classical chemical analysis and allows for the application of inline process quality control. This study investigates the capabilities of different machine learning techniques such as Ordinary Least Squares (OLS) regression, Decision Tree Regressor (DTR), Bayesian Ridge Regression (BRR), Ridge Regression (RR), K-nearest neighbours (KNN) regression as well as Support Vector Regression (SVR) to predict the FAN content in beer mash from NIR spectra. Various pre-processing strategies such as principal component analysis (PCA) and data standardization were used to process NIR data that were used to train the machine learning algorithms. Algorithm training was conducted with NIR data obtained from 16 beer mashes with varying FAN concentrations. The trained models were then validated with 4 beer mashes that were not used for model training. Machine learning algorithms based on linear regression showed the highest prediction accuracy on unpre-processed data. BRR reached a root mean square error of calibration (RMSEC) of 2.58 mg/L (R2 = 0.96) and a prediction accuracy (RMSEP) of 2.81 mg/L (R2 = 0.96). The FAN concentration range of the investigated samples was between approx. 180 and 220 mg/L. Machine learning based NIR spectra analysis is an alternative to classical chemical FAN level determination methods and can also be used as inline sensor system.