Unsupervised Domain-Adversarial Adaptation of Neural Networks for Analysis of Real Wines Using IR Absorption Spectroscopy
Louisa S. Utegenova¹, Sergey A. Burikov¹, Artem A. Guskov¹,², Kirill A. Laptinskiy1,2, Tatiana A. Dolenko¹,², and Sergey A. Dolenko²; ¹ Faculty of Physics, Lomonosov Moscow State University, Moscow, Russia, ² Skobeltsyn Institute of Nuclear Physics, Lomonosov Moscow State University, Moscow, Russia
Abstract
The determination of concentrations of the main components of wines by IR absorption spectroscopy using neural networks is complicated by the fact that models trained on a representative database of model solutions perform poorly when applied to spectra of real wines, since the two datasets differ substantially in their feature distributions.
We address this problem by applying unsupervised domain adversarial adaptation, in which a neural network is trained on a labeled source dataset of model wine solutions while the target dataset — IR absorption spectra of 150 real wine samples — remains completely unlabeled. The domain -adversarial neural network architecture, comprising a feature extractor, a regressor, and a discriminator connected through a gradient reversal layer, was trained so that the discriminator could not distinguish source from target features, forcing the extractor to learn domain-invariant representations relevant for concentration prediction.
The application of the proposed approach to the analysis of real wine samples demonstrated satisfactory accuracy in determining the concentrations of the studied components, confirming that unsupervised domain adversarial adaptation allows combining knowledge from spectra of model solutions and real, unlabeled wine spectra without requiring reference concentration measurements for the target samples. The results indicate that this approach is a promising tool for solving the inverse problem of IR spectroscopy in practical, real-world conditions.
The study has been conducted at the expense of the Russian Science Foundation grant No. 24-11-00266,
https://rscf.ru/en/project/24-11-00266/. The work of L. Utegenova was supported by the Foundation for the Advancement of Theoretical Physics and Mathematics "BASIS" (Agreement No. 25-2-1-136-1).
Speaker
Utegenova Louisa Salavatovna
Lomonosov Moscow State University
Russia
Discussion
Ask question
