Cross-modal fusion of chemical language and physicochemical features enables accurate and interpretable multi-label...

Odor prediction from molecular structure is essential for monitoring emissions from municipal solid waste landfills but remains challenging due to the chemical diversity and long-tailed distribution of odor descriptors. Therefore, we introduce ChemPaDELBoost, a cross-modal framework that integrates ChemBERTa SMILES embeddings with PaDEL physicochemical descriptors through bi-directional attention and gradient-boosted classification. Using a curated dataset of 5296 odorants and 112 descriptors, the model achieves micro-F1 = 0.500 ± 0.012, micro-AUC = 0.914 ± 0.004, and macro balanced accuracy = 0.691 ± 0.013 under 5-fold cross-validation. Ablation studies reveal that PaDEL descriptors markedly outperform ChemBERTa embeddings alone (micro-PR-AUC 0.455 ± 0.032), while cross-modal fusion yields further gains (micro-PR-AUC 0.471 ± 0.006). A dual-threshold applicability domain improves micro-F1-0.546 at 0.3/0.3 thresholds. Interpretability analyses show that the fused representation forms coherent t-SNE manifolds and reconstructs perceptually meaningful "odor islands" aligned with established odor families. These results demonstrate that cross-modal fusion significantly enhances both predictive accuracy and semantic organization, offering a scalable foundation for AI-assisted odor monitoring and environmental management.