This study compares three different methodologies for the quantification of the fat content of ultra-high temperature (UHT) milk using benchtop proton nuclear magnetic resonance (1H NMR) spectroscopy, a flagship of green, accessible, and state-of-the-art technology suitable for modern laboratory environments. The evaluated approaches included traditional calibration curve and machine learning algorithms, with emphasis on partial least squares regression (PLS-R) and artificial neural networks (ANN), to estimate the fat content in skimmed, semi-skimmed and whole milk. Among these, ANN provided the most accurate results for all types of milk, particularly in skimmed milk, with a relative standard deviation (RSD) of 14.9% and an accuracy of −7.3%. The calibration curve showed higher variability, with an RSD of 34.1% and trueness of 25.3% for skimmed milk. PLS-R improved accuracy in relation to the calibration curve approach, reducing RSD to 18.9% and trueness to −17.7%. The developed method has been successfully applied to determine the fat content in 51 samples of UHT milk purchased in different Spanish supermarkets, providing adequate results for each of the three categories considered, including goat's milk, sheep's milk, and milk coffee. Furthermore, the application of machine learning has proven its validity by successfully distinguishing between lactose and lactose-free UHT milk.
Read full abstract