Mid-infrared spectroscopy and machine learning as a complementary tool for sensory quality assessment of roasted cocoa-based products

Monitoring sensory quality in cocoa-based products is time-consuming and requires expert panelists. Integrating Mid-infrared (MIR) spectroscopy and chemometric models is a promising tool for real-time quality inspection. This study evaluated machine learning (ML) models based on the latent relationship between spectral and sensory information to predict the overall quality of roasted cocoa. Fifty-four roasted cocoa samples were analyzed using ATR-FTIR in the 4000-650 cm(-1) range and sensory evaluated by four trained panelists. Spectral data were preprocessed using Multiplicative Scatter Correction (MSC) and combined with sensory data. Subsequently, the block-scale Principal Component Analysis (PCA) was performed. Secondly, a PCA was calibrated only on the spectral data to obtain uncorrelated regressors as input to the supervised ML techniques. Supported Vector Machine Regression Model (SVMR) and the Random Forest Regression Model (RFR) were used to predict the overall quality of roasted cocoa samples. The training (75 %) and validation (25 %) of the ML techniques were performed 1000 times, and the hyperparameters optimization of each method was assessed via multifactor Analysis of Variance (ANOVA). According to the tasting panel results, the cocoa beans from different growing areas, initially appeared to have similar sensory characteristics. However, using PCA, a distinction was identified in the northern beans. The SVMR and RFR models demonstrated an outstanding ability to describe the overall quality of roasted cocoa samples. Further, the statistical results revealed the potential of MIR coupled with SVMR as a reliable and robust tool for the rapid (CT < 0.02 s) and accurate prediction (MRE < 2 %, R-2 > 99.9 %) of the overall quality of roasted cocoa-based products. This work demonstrates that it is possible to implement artificial intelligence tools to support decisions in cocoa evaluation, ensuring compliance with quality standards and allowing segmentation according to origin and characteristics.