Authentication of grape varieties in wines using 1H-NRM spectra and convolutional neural networks: first results using a database containing more than 3000 observations

Abstract

Authentication in the wine industry aims to prevent fraud and strengthen confidence in trade. A previous study proposed 1H NMR as a solution for authenticating several characteristics such as grape variety, origin, and vintage (Godelman et al, 2013). However, a database of 600 observations could be considered significant for a publication, but not robust enough for an application. The same approach was adopted, on a much larger scale. Wines and their metadata were provided by oenological laboratories. Samples were prepared as in Leleu et al, 2025. NMR spectra were acquired on a Bruker 400 MHz spectrometer, after water suppression (zgpr) or after water and ethanol suppression (noesy). An observation was a set of 10 to 40 metadata plus a zgpr and noesy NMR spectrum. More than 4,000 observations were collected in 2024 and 2025. The calibration dataset was a subset of 2,683 observations. Predictions were focused on grape variety, three of them were selected for model construction using one-dimensional convolutional neural networks (1D-CNN). Models were then applied to a series of test datasets, completely independent of the calibration dataset, used only once and representing hundreds of observations. Results showed that predictions for wines of the same origin and vintage as the model were fairly good, around 80 %. Performance dropped to 50 % when the models were applied to wines from the new vintage (2025) and to 35 % when they were applied to wines from the new vintage and another French region, Aquitaine. In such conditions, out of the calibration scope, models were strongly challenged, but despite this, they retained a certain prediction ability. These results were therefore considered very encouraging. New observations to be collected in 2026 and 2027 at different locations will improve the models robustness. An application could be available when the number of observations reach an estimated 10,000 or more. Acknowledgments PRRI EWAP-NMR project (Occitanie region) funded the 400 MHz NMR device and support staff, thereby providing an analytical taskforce. ANR WAP-NMR project provided the standard operating procedure (SOP) for acquiring NMR spectra and metadata tables. ANR AIOLY laboratory provided CNN assistance and Python notebooks. FFLŒI association provided test samples. Calculations were performed on the CoLab.IA INRAE web application.

References

Godelman,R.; Fang F.; Humpfer E.; Schutz, B.; Bansbach, M.; Schafer, H.; Spraul,M. (2013). Targeted and non targeted wine analysis by 1H NMR spectroscopy combined with multivariate statistical analysis. Differentiation of important parameters: grape variety, geographical origin, year of vintage. J. Agric. Food Chem., 61,5610-5619

Leleu, G.; Butelle, R.; Jacob, D.; Kurkiewicz, LA.; Boulet, JC.; Deborde, C.; Dubernet, M.; Gaillard, L.; Galvan, A.; Gaudin, K.; Gosse, A.; Herderich, M.; Moing, A.; Rosset, S.; Watson, F.; Da Costa, G.; Richard, T. (2025). Molecules, accepted.