Artificial intelligence smelling machine as decision-making tools for food

Authenticating the origins of raw materials is essential in the food industry, not only to verify supply authenticity but also to assess their quality over time, ensuring a consistent final product. However, factors such as climate change can lead to variations in plant characteristics and, consequently, in raw materials. In this context, molecular analysis techniques offer a fast and objective tool for both authentication and quality assessment. Despite their advantages, recent analytical tools for quantifying food quality markers generate highly complex and large datasets that are not easily interpretable. Combining them with Artificial Intelligence (AI) models can address this issue by extracting meaningful insights from these datasets, facilitating decision-making and ensuring product consistency despite variations in raw material availability. Moreover, these tools play a key role in developing new blends by employing more accessible or cost-effective supplies. AI models can evaluate the quality of these blends, identifying similarities and differences compared to established industrial origins, to maintain the consistent quality of finished products[1] [2][3]. For this study, 48 samples of cocoa liquor from four different origins (Ecuador, Ivory Coast, Ghana, and Nigeria) were analysed using Headspace-Solid Phase Microextraction-Gas Chromatography- Mass Spectrometry (HS-SPME-GC-MS) with a targeted method that includes 73 volatile compounds. AI model validation was performed by splitting the dataset into a training set (70% of the data) and a test set (30% of the data), ensuring that the proportional distribution of samples across the different origins was preserved. The models were evaluated using two validation techniques: k-fold crossvalidation on the training set and independent validation on the test set. Precision, recall, and accuracy were then calculated to assess the model's performance. The raw data were pre-processed using autoscaling and first analysed using Principal Component Analysis (PCA) to assess clustering and identify outliers. Subsequently, the data were examined using supervised methods including PCA-Linear Discriminant Analysis (PCA-LDA), Partial Least Squares-Discriminant Analysis (PLSDA), Random Forest (RF), K-NN, Logistic Regression (LR), Support Vector Machines (SVM), and Multi-Layer Perceptron (MLP). Logistic Regression was the best classification model, achieving an accuracy of 0.87, with only CCN-51 being misclassified, and it achieved an average weighted precision of 0.90 and an average weighted recall of 0.87. This study demonstrates the effectiveness of combining HS-SPME-GC-MS with AI for cocoa liquor quality assessment, as Logistic Regression provided accurate and reliable classification results, offering a promising alternative to traditional sensory evaluation methods.