The advantages of pottery jars in the aging process of Baijiu are evident, but the impact of their material composition and pore structure on the flavor of Baijiu has not been widely studied. This study systematically analyzed the effects of six types of pottery jars on metal ions and flavor substances during the storage process of Maotai-flavored Baijiu. It was found that changes in the content of Fe and Zn metals, as well as pore parameters in the jars, significantly affected the content of AL, Mg, K, Na, and Ca ions in Baijiu. Based on three feature ranking methods and three machine learning models, a feature selection method related to flavor substances was established, identifying the key features (i.e., key metal ions) for each flavor group. The final key features of each flavor group can accurately predict the corresponding flavor substance content (Rp² > 0.87). The comprehensive analysis results indicate that the increase in the content of Fe, as well as the increases in P-max and P-min in the pottery jar, collectively promoted the formation of three flavor groups represented by ethyl valerate (G2), ethyl lactate (G7), and ethyl linoleate (G10), with an increase of 3% to 5%. In contrast, the increase in Zn inhibited the formation of the flavor group represented by 2,3-butanediol (G3), with a decrease of 14%. These results further clarify the impact of pottery jar formulations on the changes in flavor substances and provide a more effective method for analyzing the influence mechanism of jars on Baijiu.

In this study, the optimization algorithm was combined with the visible light hyperspectral imaging technology to distinguish the multi-grain broken grain varieties and determine their mixing ratio. The Otsu method combined with the watershed algorithm was used to extract the ROI spectral reflectance of the sample particles, DBSCAN-MD was used to remove outliers, IVSO-CARS was used to extract the characteristic wavelength, establish four identification models (PNN, GRNN, BPNN, RBFNN) based on full wavelength and characteristic wavelength. Then input three sets of mixed samples from the outside to verify the effect of the BPNN model trained by 7 different kernel functions. The results show that the BPNN model trained with the One Step Secant function has the best effect, and the average recognition accuracy of the three groups of samples is 99%. Therefore, this method can realize the rapid identification and research of multigrain and multigrain, and determine the mixing ratio.