Bio-diesel quality characterization towards bio-jet fuel production using infrared spectroscopy and an interpretable...

Efficient utilization of renewable biomass resources requires advanced modeling and monitoring frameworks for resource optimization and sustainable management systems. Biodiesel-based bio-jet fuel is a green, sustainable, and carbon-neutral alternative for aviation, but its complex and unstable composition presents significant challenges in industrial production-particularly in the rapid and accurate characterization of key quality parameters, which remains unresolved. This study proposes an interpretability-oriented machine learning framework that leverages infrared spectral descriptors to predict and rationalize biodiesel quality attributes. Seven machine learning models were evaluated, with the ensemble model outperforming others by achieving a mean error of 6.83%, RMSE of 1.43, and R² of 0.95. The descriptor-driven approach significantly improved both predictive accuracy and model robustness under ±3σ perturbation. Shapley value-based comparative analysis across biodiesel indicators revealed both shared and indicator-specific infrared drivers, with olefinic = C-H stretching (3007 cm^-1), and CH₃ symmetric bending (1363 cm^-1) consistently contributing across multiple targets, while ester C=O stretching (1741 cm^-1), ester C-O stretching (1233 cm^-1), and O-H stretching (3318 cm^-1) exhibiting indicator-dependent importance. Furthermore, the study elucidates the mechanistic relationships between spectral peak positions, functional group types, and biodiesel properties. A user-friendly interface was developed to enable real-time fuel quality monitoring and data-driven process adjustment for online industrial deployment. Overall, this framework combines spectral analysis with AI-based assessment to enhance bioenergy utilization and supports scalable deployment of biomass-derived aviation fuels within sustainable resource management systems.