Phase-Resolved VOC Fate During Banana Cold Storage Enables Stage-Specific Marker Rules Guided by Quantum and Physicochemical Descriptors

Banana cold-chain storage is accompanied by pronounced, non-linear headspace volatile organic compound (VOC) dynamics, which complicates robust stage classification and marker selection for non-destructive monitoring based on headspace signals.

Cavendish bananas were stored at 15 ± 0.5 °C and 85 ± 3% relative humidity. Headspace VOCs were monitored daily for 20 days across five sealed containers. Based on coordinated VOC evolution, the storage timeline was partitioned into three VOC-defined phases. Quantum-chemical descriptors derived from density functional theory calculations, together with physicochemical properties and temporal features, were integrated into an interpretable machine learning framework across 22 VOCs. Eleven regression algorithms were evaluated.

The optimized Gradient Boosting model showed strong performance on a held-out container test set (R² = 0.936,RMSE = 0.247, MAE = 0.092) and stable leave-one-container-out grouped validation on the training containers (CVR² = 0.958 ± 0.009). Y-randomization confirmed that the model performance was not driven by random descriptor-response associations. SHAP analysis identified LUMO, total energy, Bertz complexity index, storage time, and TPSA as influential contributors to model-predicted VOC dynamics. Quantum-chemical descriptors accounted for 53.6% of the global feature importance in the present dataset.

This work provides a stage-resolved, descriptor-assisted strategy for VOC marker prioritization and validation design for non-destructive headspace monitoring under controlled cold-storage conditions. The findings also highlight the need for external validation across cultivars and storage environments.