Predicting soymilk odors using a multilayer perceptron neural network model.

The presence of undesirable beany odors in soymilk products is a long-lasting issue for the soymilk manufacturers as it largely affects the consumers acceptance. Promising soybean varieties that generate satisfactory soymilk odors (SV-SSO) may offer a potential solution, and the automatic odor prediction from volatile profiles may facilitate in the discovery of SV-SSO. In this study, a database of soymilk key volatile (SKV) contents and the sensory scores for their corresponding soymilk was established. A multilayer perceptron (MLP) neural network trained on these data achieved odor predictions within an averaged relative error of ≤5%. The MLP weight matrices indicated that hexanal, hexanol, E-2-hexenal, 1-octen-3-ol formed the primary mathematical contributor to the pleasant beany odor, whereas all SKVs mathematically determined the unpleasant beany odor and overall satisfaction. The concentration dependent effects of each SKV on the sensory scores of soymilk were discussed. The findings provided mathematical insights into the different roles of soymilk headspace SKVs in altering soymilk odors and offered technical directions for developing better-tasting soymilk products.