Predicting Sweetness Intensity and Uncovering Quantitative Interactions of Mixed Sweeteners: A Machine Learning Approach

Sweeteners are commonly blended to exploit synergistic effects, enabling the desired sweetness to be attained while reducing total usage. However, establishing a quantitative relationship between mixed sweeteners' concentration and sweetness intensity remains a key challenge. This study developed a sensory evaluation-machine learning approach to construct prediction models for binary/ternary mixtures of five sweeteners (sucrose, glucose, fructose, mannitol, and sorbitol). After feature selection of molecular descriptors and comparison of seven machine learning regression models, the Multilayer Perceptron achieved superior performance for the binary mixtures (R² = 0.9828), while the Support Vector Regression exhibited optimal performance for the ternary mixtures (R² = 0.9825). Concentration-sweetness intensity curves of mixed sweeteners at specific concentrations were generated using these two optimal prediction models. Results showed that at low concentrations, ternary blends of one sugar and two polyols (mannitol and sorbitol) exhibited stronger synergism than binary mixtures in the same concentration range. Specifically, blending the composite system of 1% mannitol and 2% sorbitol with 1% sucrose, 1% glucose, and 1% fructose separately increased the sweetness intensity by 39.6%, 42.8%, and 37.4%, respectively. This work confirms that machine learning can establish a quantitative relationship between multi-component sweeteners' concentration and sweetness intensity, reveal their complex interactions, and provide a novel approach for intelligent sensory assessment and formulation design.