Machine learning-assisted aroma profile prediction in Jiang-flavor baijiu

Abstract

The complex flavor of Jiang-flavor Baijiu (JFB) arises from the interaction of hundreds of compounds at both physicochemical and sensory levels, making accurate perception challenging. Modern machine learning techniques offer precise and scientific approaches for predicting sensory attributes. This study applied flavoromics and sensory profiling to 27 representative JFB samples from main regions in China, integrating five machine learning algorithms to establish a novel strategy for predicting global aroma characteristics. The results indicate that the neural network (NN) model outperformed others, effectively capturing the intricate interactions among flavor compounds. Model dissection identified 18 chemical parameters potentially influencing the overall aroma profile. The importance of these factors was further validated through spiking and omission tests, which notably enhanced the sensory experience of commercial liquor. This study demonstrates the potential of machine learning in JFB flavor research and offers valuable insights into the mechanisms underlying its flavor formation.