Identifying Key Odorant Compounds and Assessing Odor Additivity in Complex Mixtures via Machine Learning: A Multi-Regional Case Study in China

With the continuous development of industrial parks in China, odor pollution resulting from the extensive use of solvents has become increasingly prominent, making its accurate analysis and effective prevention and control critically important. To identify key odorant compounds in complex mixtures, this study collected 426 air samples from industrial parks in Guangdong and Zhejiang, China and built a database of odor concentrations and chemical compositions. Machine learning (AutoGluon) combined with SHAP interpretability analysis was employed to build an odor concentration prediction model. The optimal strategy used the Odor Activity Value (OAV) as input for an ExtraTrees model, achieving a pooled out-of-fold R² of 0.52 under nested cross-validation. Combining SHAP (Shapley additive explanations) interpretability analysis of this optimal model with OAV screening identified 15 key odorant compounds, including methanethiol, tetrachloroethane, acrolein, among others. Further investigation assessed the additivity of odor perception: the summed odor activity value (ΣOAV) failed to reproduce the measured odor concentration, and 81% of the valid samples showed supra-additive behavior. This deviation was closely associated with the analytical coverage of the dominant odorants: sulfur-rich samples were close to additive, whereas sulfur-poor samples deviated by up to two orders of magnitude. The additivity index therefore reflects the completeness of the monitoring list and provides no evidence for or against pairwise sensory interactions. This study validates a data-driven approach for deciphering complex odor formation, providing a powerful methodology for key odorant screening and a scientific foundation for targeted odor pollution control.