Prediction of cocoa liquors sensory profiles using e-tongue and e-nose data by the implementation of Machine Learning to enhance the cocoa production for various associations across multiple municipalities in Colombia

Cocoa production in Colombia holds substantial significance, as the cultivation and processing of this crop have had a considerable social impact, serving as a mechanism for peace transition and the eradication of illicit crop cultivation. However, cocoa liquor encounters a challenge in its production chain, as 90% of the producers are smallholders. To assess quality, human sensory panels are utilized as the primary resource. However, due to insufficient knowledge, inadequate training, or economic inaccessibility, many producers are unable to accurately assess and maintain the quality of their products. This research proposes a hybrid machine learning framework that integrates electronic tongue and nose data to predict sensory profiles comparable to those generated by human evaluators, thereby addressing the challenges associated with human sensory evaluation of cocoa liquor in various municipalities in Colombia. The framework integrates six unsupervised learning methods, including k-means, kernel and classical Principal Component Analysis (kPCA, PCA), Gaussian Mixture Model (GMM), among others, alongside five data augmentation techniques, such as Gaussian Mixture Density Augmentation (GMDA), Generative Adversarial Networks (GANs), and Sliding Window Augmentation (SWA), to address limitations in dataset quantity and variability. Combinations of DA and UL were examined and subsequently paired with three supervised learning models: XGBoost, Support Vector Machines (SVM), and Multilayer Perceptron (MLP), implemented within a two-stage hybrid classification framework for predicting attribute presence and intensity. The findings indicated that the integration of VBGMM (UL), GMDA (DA), and XGBoost (SL) consistently yielded the highest accuracy and interpretability, achieving a test accuracy of 0.86. The model exhibited robust generalization for both fundamental sensory attributes (e.g., co-coa, bitterness) and intricate secondary descriptors (e.g., floral, nutty), attaining superior performance despite data imbalance conditions. This method provides an economical and scalable instrument for evaluating sensory quality, equipping smallholder cocoa associations with rapid, understandable feedback for decisions concerning fermentation, drying, and roasting procedures. The model enhances traceability, standardizes quality, and differentiates the market, supporting Colombia's fine-flavor cocoa sector through innovation and inclusive technology.

Maestría

Machine Learning

Sensory Analysis

Product optimization