From Geochemical Fingerprints to Food Authentication: Integrating Explainable and Cost-Aware Machine Learning for Provenance Analysis

The increasing demand for reliable food authentication highlights the need for scalable and innovative tools to link geochemical fingerprints in food products with their geographic provenance. Food authentication is not only essential for preventing fraud but also offers a unique opportunity to relate agricultural products to their underlying geochemical signatures. Here we present a unified framework that combines stable isotopes (e.g. ⁸⁷Sr/⁸⁶Sr) and trace-element fingerprints measured in food products with explainable and cost-aware machine learning to support provenance verification.

We first develop a cost-aware binary classification model for French sparkling wines, demonstrating how high-precision ⁸⁷Sr/⁸⁶Sr ratios can be partially substituted by low-cost elemental proxies (e.g. Rb) while maintaining strong discriminative power. To address scalability constraints, we extend this approach to a multiclass setting using cost-sensitive logistic regression to classify wines from multiple Portuguese and Chilean regions, explicitly handling class imbalance and feature redundancy. Finally, we introduce TeaPrint, an unsupervised multimodal clustering framework that jointly integrates isotopic, elemental and volatile organic compound data to uncover coherent regional geochemical patterns in international tea samples without requiring prior labels.

Across these case studies, we show that food products carry integrated geochemical signatures that can be exploited for robust provenance authentication across heterogeneous datasets. By bridging forensic geochemistry and explainable machine learning, our approach offers a cost-efficient and scalable pathway towards robust provenance authentication and transparent food supply chains.