Interpretable machine learning for the grade prediction of strong flavor yuanjiu (crude Baijiu) based on HS-GC-IMS

The complexity of the Baijiu matrix makes it a key challenge to map specific flavor compounds to Baijiu grading. Headspace-gas chromatography-ion mobility spectrometry (HS-GC-IMS) and machine learning (ML) techniques were combined to explore the grade classification of strong flavor yuanjiu (crude Baijiu, SFY), and a preliminary prediction model for SFY grades was constructed. A total of 59 volatile components were identified using HS-GC-IMS. Six ML models were used to predict the grade of SFY samples, of which the classification accuracy of the neural network was 88.2%. Through interpretable analyses, seven compounds, such as pentanal, 2-methylbutanal, were screened as potential grade markers. The monomers and dimers, as part of the inherent characteristics of HS-GC-IMS, should both be preserved when carrying out statistical analyses. This interdisciplinary study further breaks through the bottleneck of complex flavor prediction and advances the scientific grading of SFY samples. • Using HS-GC-IMS, the volatiles of different strong flavor yuanjiu were analyzed. • The neural network model achieves 88.2% accuracy in the grade prediction. • Feature importance, SHAP, and ICE plots were used to interpret neural network models. • Seven compounds were identified as grade discriminators. • The effects of dimers on machine learning accuracy were explored.