Proceedings of the 3rd IAGC International Conference

RESOCONTI /14 Proceedings of the 3rd IAGC International Conference Water-Rock Interaction - 18 & Applied Isotope Geochemistry - 15 Cagliari, Italy, 16-21 June 2025 edited by Franco Frau, Richard B. Wanty, Elisabetta Dore, Dario Fancello UNICApress/ateneo The third IAGC International Conference, held in Cagliari (Sardinia, Italy) in June 16-21, 2025, was widely attended by researchers from around the world. This resulted in the present volume of proceedings containing as many as 308 abstracts, both short and extended, distributed among 22 sessions in addition to the plenary sessions. The subjects covered in the abstracts span numerous topics in geochemistry, from low to high temperature, from frozen to hydrothermal systems, and natural and manmade environments. The conference was hosted by the Department of Chemical and Geological Sciences at the University of Cagliari (UNICA). UNICApress/ateneo RESCONTI 14 Proceedings of the 3rd IAGC International Conference Water-Rock Interaction - 18 & Applied Isotope Geochemistry - 15 Cagliari, Italy, 16-21 June 2025 edited by Franco FRAU, Richard B. WANTY, Elisabetta DORE, Dario FANCELLO Cagliari UNICApress 2025 IAGC-3 Organizing Committee Giovanni B. De Giudici – Secretary General Stefania Da Pelo, Franco Frau, Elisabetta Dore, Stefano Naitza, Dario Fancello, Daniela Medas, Patrizia Onnis, Francesca Podda, Riccardo Biddau, Elodia Musu Silvio Ferrero, Laura Pioli, Richard B. Wanty, Bruna B. Carvalho, Maurizio Testa, Roberto Dessì, Elisa Sacchi, Laura Sanna, Pier Andrea Marras Contacts: Frau F. ([email protected]) Department of Chemical and Geological Sciences, University of Cagliari, Cagliari, Italy Wanty R.B. ([email protected]) Colorado School of Mines, Golden, CO 80401, USA Dore E. ([email protected]) Department of Chemical and Geological Sciences, University of Cagliari, Cagliari, Italy Fancello D. ([email protected]) Department of Chemical and Geological Sciences, University of Cagliari, Cagliari, Italy Sezione Ateneo RESOCONTI /14 ISSN 2974-6671 Proceedings of the 3rd IAGC International Conference Water-Rock Interaction - 18 & Applied Isotope Geochemistry - 15 Cagliari, Italy, 16-21 June 2025 edited by Franco Frau, Richard B. Wanty, Elisabetta Dore, Dario Fancello Cover image: Devil’s saddle with flamingo by Dario Fancello and Emanuele Pucci © Authors and UNICApress 2025 CC-BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/) Cagliari, UNICApress, 2025 (http://unicapress.unica.it) e-ISBN 978-88-3312-187-1 DOI https://doi.org/10.13125/unicapress.978-88-3312-187-1 Proceedings of the 3rd IAGC International Conference Water Rock Interaction-18 & Applied Isotope Geochemistry-15 243 Toward greater transparency: Stable isotopes and explainable Machine Learning solutions for food traceability Ogrinc, N.1,*, Drole, J.2, and Eftimov, T.2 1: Department of Environmental Sciences, Jožef Stefan Institute, Ljubljana, Slovenia 2: Department of Computer Systems, Jožef Stefan Institute, Ljubljana, Slovenia *Corresponding author: [email protected] ABSTRACT The growing demand for authentic and high-quality food has driven the need for advanced traceability systems to verify geographical origin and prevent fraud. This study integrates stable isotope and elemental analysis with machine learning (ML) to enhance food authentication. By employing ML techniques for data analysis, feature selection, and classification, we improve accuracy and provide explainable insights into food provenance. A case study on saffron demonstrates the method’s effectiveness, achieving up to 90% accuracy in origin classification. This approach offers a scalable and adaptable solution for various food products, strengthening food safety, transparency, and regulatory compliance. INTRODUCTION The connection between food and territory has gradually been lost over time due to advances in production, transportation, and exposure to globalization (Luykx and van Ruth, 2008). In response, consumer demand for authentic, high-quality, and sustainably produced food is rising. Traditional cultivation and processing methods are valued for their authenticity, safety, environmental benefits, and quality, often commanding premium prices. Among the exploitable techniques, stable isotope and elemental fingerprinting are leading the way in establishing authenticity and geographical origin of food products (Danezis et al., 2016). The basis of the stable isotope approach lies in the transfer of the isotopic signals (isotopic fingerprint) of bio-elements (H, C, N, O, S) present in local feed and water to animal and plant tissue, processes that are generally well understood. The verification of regional origin can be even more effective when stable isotopes are combined with elemental composition since a plant's elemental profile is related to the soil composition of the location where the plant grows; consequently, bioavailable nutrients can provide direct information about an agricultural products' geographical origin (Drivelos and Georgiou, 2012). However, stable isotopes and elemental fingerprinting alone are insufficient. A comprehensive reference dataset of authenticated products is essential but challenging to establish due to high time and cost demands. This dataset must include a wide range of samples representing diverse geographical, seasonal, dietary, and production Proceedings of the 3rd IAGC International Conference Water Rock Interaction-18 & Applied Isotope Geochemistry-15 244 conditions (Kelly et al., 2005). Authenticity is assessed by comparing values from commercial samples against dataset-derived limits, using a statistical model to determine the best fit. The presentation will describe a comprehensive authenticity and traceability system that integrates stable isotope analysis, elemental composition, databases, statistical evaluations, and emerging AI-driven technologies, particularly machine learning (ML) applications. Special emphasis will be placed on how ML enhances data processing, pattern recognition, and fraud detection in food authentication. In the final section, a real-world implementation of these technologies in a traceability system will be presented, demonstrating their practical impact. METHODS The analysis of stable isotope ratios and elemental composition was conducted using precise and well-established techniques. The two main analytical methods used were Isotope Ratio Mass Spectrometry (IRMS) and Inductively Coupled Plasma Mass Spectrometry (ICP-MS) for elemental profiling. DATABASE A comprehensive database and database management system (DBMS) called IsoFoodTrack (http://isofoodtrack.ijs.si) was developed (Terro et al., 2025). This system provides extensive data on the stable isotopes of light elements and the elemental composition of authentic samples from various food commodities such as oils, milk and dairy products, meat, spices, truffles, seafood and vegetables. Furthermore, IsoFoodTrack is designed to be interoperable, allowing connection with other databases or centralized repositories and represents a significant advancement over traditional food databases by prioritizing both accessibility and standardization. Additionally, the database integrates standardized metadata protocols and harmonized data entry formats, which streamline cross-study comparisons and enhance reproducibility. MACHINE LEARNING ENHANCED DATA PROCESSING Although classical statistical approaches have historically been used in food traceability research, machine learning (ML) methods are now increasingly employed and gaining popularity. In most cases, these studies use isotopic measurements of food samples to predict or classify their geographic origin, employing classification models such as Random Forest, XGBoost, or Neural Networks and comparing their performance metrics. Despite an abundance of published literature and growing pressure from reviewers to include these analyses, many studies simply apply ML models without adhering to a rigorous experimental design or proper benchmarking, highlighting the need for standardized methods in ML-based food traceability research. Proceedings of the 3rd IAGC International Conference Water Rock Interaction-18 & Applied Isotope Geochemistry-15 245 To address the aforementioned challenges and provide deeper insights into food traceability, we propose a novel ML-based framework consisting of the following steps: (1) selecting representative training data, (2) performing feature selection to identify measurements that improve model performance without overfitting, (3) training the model, (4) evaluating the model using diverse train–test splits to ensure robust results, and (5) providing sample-specific explanations through feature importance. Selecting representative training data may eliminate the need to use all available samples. Instead, we repeatedly select training and test subsets using unsupervised learning or graph-based methods to ensure an approximately uniform distribution of samples in the feature space. Next, we apply feature selection techniques (e.g., correlation analysis) to identify the measurements that yield improved classification models without overfitting. The model is then trained on each train–test split using the selected features. To enhance explainability, we perform SHAP (SHapley Additive exPlanations) analysis on each sample in the test set, thereby revealing the interactions and relative importance of features that contribute to origin classification for each sample separately. This approach elucidates the features most critical for accurate predictions and highlights patterns leading to misclassifications. By aggregating explanations across all test splits, we identify robust insights that are unlikely to occur by chance. Ultimately, these results generate a “food origin footprint”, indicating the most influential features and their interactions for each origin. PRACTICAL APPLICATION The example of practical application includes the saffron data sourced from the main saffron-producing countries: Iran, Italy, Spain, Morocco and Greece. Specifically, we trained robust classifiers capable of detecting the geographical origin of the saffron samples. We tested different scenarios using only isotopic variables, trace elements and a combination of both. By repeating the learning process five times to assess the robustness of the model across all scenarios, we achieved average accuracy of around 90% across the splits. Additionally, through explainable post-hoc analysis (SHAP), we identified which variables are important for each country to obtain reliable origin differentiation. The model's accuracy was 80% when only stable isotope analysis was included in the evaluation. The most reliable parameters for Morocco and Italy are presented in Figure 1. For both countries, δ18O values contribute the most to their separation, followed by δ13C in Italy and δ34S in Morocco. This analysis has been done for all the countries of origin, but only for the correctly predicted samples, since they give insight into which attributes contribute to the correct prediction. Proceedings of the 3rd IAGC International Conference Water Rock Interaction-18 & Applied Isotope Geochemistry-15 246 Figure 1: Decision plot analysis for Morocco and Italy showing how specific isotopic ratios are critical in distinguishing saffron origin. When both isotopic and elemental data are considered, the separation improves slightly, reaching an accuracy of 89.8%, with the most discriminative parameters being the elemental variables. Although this approach has been used for saffron as an example, it is readily adaptable to other raw materials. CONCLUSIONS This study demonstrates the effectiveness of integrating stable isotope and elemental analysis with machine learning (ML) for food traceability. The method is adaptable to various food commodities, offering a scalable solution for strengthening food quality control, authenticity verification, and fraud prevention in the global food supply chain. ACKNOWLEDGEMENTS We acknowledge the financial assistance of the Horizon Europe projects FishEUTrust (Grant agreement no. 101060712) and PROMEDLIFE project PRIMA programme supported by the European Union (Grant agreement No. 2132). REFERENCES Danezis, G.P., Tsagkaris, A.S., Camin, F., Brusic, V., and Georgiou, C.A., 2016. Food authentication: techniques, trends & emerging approaches. Trends in Analytical Chemistry, v.85, pp. 123–132. Drivelos, S.A., and Georgiou, C.A., 2012. Multi-element and multi-isotope-ratio analysis to determine the geographical origin of foods in the European Union. Trends in Analytical Chemistry, v.40, pp. 38– 51. Kelly, S., Heaton, K., and Hoogewerff, J., 2005. Tracing the geographical origin of food: the application of multi-element and multi-isotope analysis. Trends in Food Science and Technology, v.16, pp. 555– 567. Luykx, D.M.A.M., and van Ruth, S.M., 2008. An overview of analytical methods for determining the geographical origin of food products. Food Chemistry, v.107(2), pp. 897-911. Terro, C., Modic, R., Ogrinc, M., Simčič, A., Drole, J., Eftimov, T., Koroušić Seljak, B., and Ogrinc, N., 2025. IsoFoodTrack: a comprehensive database and management system based on stable isotope ratio analysis for combating food fraud. Frontiers in Nutrition, v.12, 1516521.