A novel method of using neural networks to predict wine composition.

The present study proposes a method for diagnosing the main components of wine based on their infrared absorption spectra using machine learning techniques. The developed approach involves the application of Partial Least Squares regression, perceptrons, and convolutional neural networks, trained on IR absorption spectra of 1734 solutions simulating white and red wines. These models are used to analyze the IR spectra of real wines in order to determine the concentrations of ethanol, sugars, acids, glycerol, and sulfur dioxide. The method proposed in this study ensures the measurement of the specified components with accuracy either comparable to commercial counterparts (for ethanol and sugars) or only slightly inferior to them. Comparative analysis of the results obtained using the three methods demonstrated that convolutional neural networks exhibit significantly higher quality in solving the inverse problem for ethanol and sugars compared to alternative models. For determining the other parameters, all models provide similar levels of accuracy in estimating their concentrations. The proposed approach enables a substantial reduction in the cost of developing software for instruments designed for rapid chemical composition analysis of wines based on IR spectroscopy, which could serve as an alternative to existing commercial solutions.