Towards AI-guided process metrics with the rapid prediction of red wine mouthfeel

Abstract

Mouthfeel is a key component of red wine sensory perceptions and impacts consumer preference for certain wine styles. Sensory evaluation of wine mouthfeel with experts or trained panellists is necessary for production decisions and research, but is time-intensive and costly. This study aimed to develop machine learning models to predict wine mouthfeel attributes using absorbance-transmittance and excitation-emission matrix (A-TEEM) spectroscopy as a rapid analysis technique. A-TEEM combines fluorescence and UV-Vis measurements, generating thousands of spectral variables per sample. These capture signals from chromophores and fluorophores (primarily phenolics that contribute to wine mouthfeel) and the modulating components that influence fluorescence through matrix effects. An A-TEEM dataset is thus of high dimensionality and capable of capturing wine spectral fingerprints incorporating mouthfeel-related compounds. In combination with the wine’s corresponding sensory data, the application of machine learning modelling makes the prediction of wine mouthfeel attributes feasible. Twelve commercial red wines differing in variety were evaluated using descriptive analysis by nine trained panellists. Eleven out of twelve in-mouth descriptors showed significant differences among samples: acidity, bitterness, pucker, dry, grippy, coating, adhesive, graininess, heat, fullness, and sweetness. A-TEEM spectra collected for the wines were pre-processed and modelled against the sensory scores using machine learning regression algorithms, with strong predictive performance obtained from cross-validation. For example, using extreme gradient boosting, the model for fullness achieved an R2 of 0.986 and root mean square error (RMSE) of 0.71 (for a score range of 0-100), coating afforded an R2 of 0.888 and RMSE of 2.76, and graininess gave an R2 of 0.925 and RMSE of 1.57. These results demonstrate the capacity of combining A-TEEM spectroscopy and machine learning techniques to predict wine mouthfeel attributes. With only minutes required for spectral data acquisition, the developed approach appears viable for the rapid prediction of wine mouthfeel sub-qualities. Ultimately, A-TEEM spectroscopy with machine learning prediction of sensory traits could potentially be used in real time to guide the winemaking process and achieve targeted mouthfeel outcomes that meet consumer expectations.