Unlocking the “black box” of food prediction models: Applications and perspectives of explainable artificial intelligence in agri-food chemical analysis and quality control

With the deep integration of global agri-food supply chains, artificial intelligence and machine learning have been widely applied in fields such as food traceability, adulteration detection, and quality grading. However, high-precision models like CNN and XGBoost are often considered “black boxes” due to opaque decision-making, undermining their credibility and robustness. Explainable Artificial Intelligence (XAI) offers a pathway to model transparency. This paper systematically reviews XAI progress in food chemistry over 2021–2026, covering inherently interpretable models and post-hoc methods (e.g., SHAP, LIME, Grad-CAM). Focusing on three major scenarios—geographical traceability and adulteration detection, flavor chemistry and sensory evaluation, food safety risk assessment—it analyzes XAI applications in quantifying feature contributions, identifying key spectral bands, and identifying candidate biomarkers for subsequent targeted validation. Application paradigms for different data types (spectra, images, chemical composition, sensor data) are summarized. In view of the challenges in current research, such as methodological biases under high-dimensional collinearity and the lack of an integration mechanism between interpretation results and prior chemical knowledge, this paper proposes key directions for future breakthroughs, with the expectation of promoting food science research toward a more intelligent, transparent and trustworthy new paradigm.