Machine learning-based prediction of sensory quality in tea blends using a semi-trained panel assessment

Abstract

Blending various types of teas with complementary biochemical profiles offers a promising path for enhancing the sensory quality of tea blends. Traditionally, blending is done by professional tea tasters, which makes it dependent on human sensitivity, time-consuming and difficult to scale. Thus, the prediction of tea's sensory quality and identification of its biochemical drivers by reducing the cost of experimentation and expediting the blending process could have tremendous techno-economic value. In this study, we curated a meta-dataset from our own experiments involving blends of four major varieties of Assam tea. Their key biochemical composition was analyzed, and sensory score was estimated using a lexicon-based descriptive technique with a semi-trained consumer-focused panel. Combining it with multiple machine learning models, we showed that while an increase in the (+)-C, TP, TR, and pH values led to a lower sensory score, higher protein/TP, TSS, organic acid, CAF, CAF/TP and TF/TR values resulted in an increased sensory score. Further, we adopted a game theory-based model agnostic interpretation technique called SHAP (Shapley Additive exPlanations) to identify features contributing to higher sensory scores and their relative significance. The key quality indicators were found to be the (+)-C, protein/TP, protein, TP, TSS, TR, citric acid, malic acid, and ascorbic acid contents. By integrating a consumer-centric evaluation with interpretable machine learning, we demonstrated how meta dataset, cutting-edge machine learning techniques, and model interpretability methods could be seamlessly integrated to reduce the number of experiments, minimize dependency on expert intuition, and enable automated quality assessment for developing superior tea blends.