Enhancing wine authentication: leveraging 12,000+ international mineral wine profiles and artificial intelligence for accurate origin and variety prediction

For the wine industry, ensuring quality and authenticity hinges on the precise determination of wine origin. In our study, we developed a fast semi-quantitative method to analyse 41 chemical elements in wine, employing inductively coupled plasma mass spectrometry (ICP-MS). This methodology characterises what we term the mineral wine profile (MWP). In contrast to an organic molecular profile, the mineral composition of a wine remains constant from the moment it is bottled. Mineral elements play a crucial role in the terroir of wine: they pass primarily from soil to grape and are then influenced by various vinification techniques. Indeed, it is widely recognised that the original soil characteristics are altered by a multitude of winemaking procedures, presenting a considerable challenge when endeavouring to extract origin-related information in a typical scenario. Our study demonstrates that statistical analyses and artificial intelligence (AI) could be a tool for accurately deciphering origin information within the MWP, provided sufficient mineral elements are measured and a comprehensive database of wine samples is employed to establish effective learning. In this study, a dataset comprising 12,966 MWPs was created in just over a year. The first analysis revealed correlations between the elements in wine, especially between rare earth elements, between macronutrients and between micronutrients. A machine learning method was then developed to assess wine origin and principal grape variety. Six models were tested by comparing the area under the receiver operating characteristic curve (AUC), with eXtreme Gradient Boosting as the chosen model. Mean accuracies of 92 % for country classification, 91 % for the French wine region, and 85 % for the main grape variety were obtained, and mean AUC scores of 0.964 for country classification, 0.961 for the French wine region and 0.914 for the main grape variety. This study represents the first comprehensive investigation at this scale on wine samples, and underscores the importance of using a comprehensive MWP dataset for AI applications when verifying wine origin. The authentication of a wine with over 99 % specificity could be routinely achievable through this approach.