Fruit distillates, like "țuică", "palincă", and "horincă", are traditional alcoholic beverages in Central and Eastern Europe. Their composition is complex, influenced by fruit variety, geographical origin, production techniques, and aging. Ensuring quality and preventing fraud, such as mislabeling of origin, requires reliable and efficient analytical methods. Vibrational spectroscopy techniques, particularly Raman spectroscopy, offer a promising solution due to their speed, cost-effectiveness, and suitability for high-water content samples. Raman spectroscopy generates large datasets, necessitating advanced data processing like Machine Learning algorithms for effective analysis. This study aimed to evaluate the potential of combining Raman spectroscopy and Machine Learning for classifying fruit distillates based on fruit variety, geographical origin, and trademark.

Previous research has used vibrational spectroscopy (IR and Raman) for quantitative determination of ethanol and methanol in fruit distillates. Raman spectroscopy is particularly advantageous for high-water content products. The combination of Raman spectroscopy and Machine Learning has been successfully applied in various fields, including food analysis, bacteria identification, and diagnostics. This study builds upon this existing research by applying these techniques to the specific challenge of differentiating fruit distillates.

Thirty fruit distillate samples from eight Romanian producers (two large processing companies and three small manufactures) were analyzed. Samples included various fruit types (apples, apricots, cherries, grapes, pears, plums, quince, sour cherries) from four Transylvanian regions. Alcoholic strength was determined using GC-FID. Raman spectra were recorded using a JASCO NRS-3300 spectrometer with a 785 nm laser. Data processing involved baseline subtraction and [0,1] normalization in OriginPro 2017. Machine Learning analysis was performed using MATLAB R2018b's Classification Learner app, employing five predictive modeling approaches: decision trees, discriminant analysis, support vector machines (SVM), nearest neighbor classifiers (KNN), and ensemble classifiers. Different training and testing sets were used for each classification criterion (fruit variety, producer, geographical origin).

Analysis of Raman spectra revealed two regions (200-600 cm⁻¹ and 1200-1400 cm⁻¹) with subtle variations potentially related to production methods, geographical origin, or fruit variety. Attempts to classify distillates solely by fruit variety yielded low accuracy (27.3%). However, classification based on producer (processing company vs. manufacture) achieved high accuracy (95.5%), indicating that producer-specific processing and storage conditions significantly influence the Raman fingerprint. This high accuracy (95.5%) was obtained using an Ensemble (subspace KNN) model, and this model's predictive ability was successfully validated on a separate testing set. Classification based on alcoholic strength showed poor correlation, suggesting that the producer fingerprint is independent of ethanol concentration. Classifying fruit variety within each producer yielded improved accuracy, with 100% accuracy achieved for processing companies, indicating that consistent processing methods enhance discrimination. Geographical classification achieved 90.9% accuracy within the Transylvania region, correctly identifying seven out of eight samples, even when including samples from producers not included in the initial training set.

The study's findings highlight the dominance of the producer's processing and storage methods over fruit variety in shaping the Raman fingerprint of distillates. While direct fruit variety classification is challenging due to these overriding influences, classification within individual producers is feasible and highly accurate. The high accuracy of the geographical classification, even across different producers, demonstrates the potential of this technique for authentication purposes, especially within geographically close regions. These results demonstrate the potential of Raman spectroscopy coupled with Machine Learning for rapid and cost-effective verification of fruit distillate trademarks.

This study successfully demonstrated the use of Raman spectroscopy and Machine Learning for differentiating fruit distillates. The producer fingerprint significantly influences the Raman spectra, showcasing the impact of processing and storage conditions. Fruit variety classification is more successful within individual producers. Geographical origin can also be effectively determined, demonstrating this method’s potential for authenticity verification. Future research could explore broader geographical regions and investigate the specific minor components contributing to the distinctive Raman fingerprints.

The study's sample size, while sufficient for a pilot study, could be expanded for broader generalization. The geographical analysis was limited to the Transylvania region. Future studies should explore the technique's applicability to other regions and broader ranges of fruit varieties and production methods.